Day-ahead electricity price prediction applying hybrid models of LSTM-based deep

learning methods and feature selection algorithms under consideration of market coupling

Wei Lia,∗, Denis Beckera

a NTNU Business School, Norwegian University of Science and Technology, 7491 Trondheim, Norway

Abstract
The availability of accurate day-ahead electricity price forecasts is pivotal for electricity market participants. In the context of trade
liberalisation and market harmonisation in the European markets, accurate price forecasting becomes difficult for electricity market
participants to obtain because electricity forecasting requires the consideration of features from ever-growing coupling markets.
This study provides a method of exploring the influence of market coupling on electricity price prediction. We apply state-of-
the-art long short-term memory (LSTM) deep neural networks combined with feature selection algorithms for electricity price
arXiv:2101.05249v2 [[Link]] 18 Jul 2021

prediction under the consideration of market coupling. LSTM models have a good performance in handling nonlinear and complex
problems and processing time series data. In our empirical study of the Nordic market, the proposed models obtain considerably
accurate results. The results show that feature selection is essential to achieving accurate prediction, and features from integrated
markets have an impact on prediction. The feature importance analysis implies that the German market has a salient role in the
price generation of Nord Pool.
Keywords: Deep learning, Electricity price forecasting (EPF), Electricity market coupling, Feature selection, Long short-term
memory (LSTM), The Nord Pool system price

1. Introduction the following five categories [5]: multi-agent [7, 8], fundamen-
tal [9, 5], reduced-form [10, 11], statistical [12, 13, 14], and
Over the last two decades, worldwide energy markets have computational intelligence (CI) models [15, 16, 17]. Compared
experienced a transition towards deregulation and harmonisa- with the other four traditional models, CI models are regarded
tion [1]. Under trade liberalisation, the traditional vertically as state-of-the-art techniques, and their superior performance
integrated power utilities are replaced with decentralised busi- contributes to their prevalence in EPF in recent years. In partic-
ness entities whose targets are to maximise their profits. Con- ular, deep neural networks (DNNs) have gradually become the
sequently, a growing number of market participants are ex- most avant-garde CI approach in other disciplines [18, 19, 20]
posed to intense competition, and their need for suitable de- and entered the scientific research field related to EPF.
cision support models to increase margins and reduce risk has DNNs are often categorised into three main classes: feed-
significantly increased [2]. Thus, the availability of accurate forward neural networks (FNNs), recurrent neural networks
day-ahead electricity price forecasts is vital for market partic- (RNNs), and convolutional neural networks (CNNs). Different
ipants to adjust production plans and to perform effective bid- types of DNNs are used to solve different problems. For time
ding strategies to make an economic profit. However, due to series prediction, RNNs have achieved superior performance by
the productive structure and characteristics of electricity prices, building extra mappings to hold relevant information from past
highly accurate forecasting is quite challenging [3, 4]. With inputs. The long short-term memory (LSTM) and gated recur-
the increasing integration of electricity markets, making accu- rent units (GRUs) are important variants of this kind of network,
rate forecasts becomes even more difficult in the complex and which overcome the vanishing gradient problem of RNNs [21].
integrated system. This is because the forecasting of electric- Compared with GRUs, LSTM is more accurate on the dataset
ity prices needs to consider a large number of factors from an using long sequences. Due to the superiority of LSTM in time
ever-growing number of interconnected, neighbouring power series forecasting, researchers have gradually paid attention to
markets. These factors include electricity prices, production, its application in EPF [22, 23, 24]. However, as with other
consumption, and other important features that influence cross- DNNs, when LSTM models are applied to high-dimensional
border electricity markets. data, a critical issue occurs, known as the curse of dimension-
Numerous research efforts have contributed to the exploita- ality [25]. This means that, with a large amount of features,1
tion and development of advanced technologies for day-ahead the performance of LSTM will degrade because of overfitting
electricity price forecasting (EPF), aimed at highly accurate [26]. Thus, LSTM cannot be employed directly for electricity
forecasting results [5, 6]. A considerable amount of literature price prediction with a large number of features as input under
has been devoted to EPF models, which can be classified into consideration of market coupling. While some researchers have

∗ Correspondingauthor 1 In machine learning, features are individual independent variables as input

Email address: [Link]@[Link] (Wei Li) in a model.

Accepted for publication in Energy

 Nomenclature

i Input/update gate’s activation vector

β Coefficients of independent variables
o Output gate’s activation vector
ŷ Predicted output
r1 , r2Random number from the interval [0, 1]
λ Tuning parameter
W Weight matrix
µ Expected electricity exchange capacity
w1 Weight vector between the input and the hidden
ω Constant that determines the significance of →
−v
layer
⊗ Element-wise product
→
−g w2 Weight vector between the hidden layer and the out-
Global current best solution
→
−p put
Particle’s past best solution
→
−v X, Y Variable
Velocity of the particle
x Input variable
Φ Kernel function
y Output variable
φ Activation function
Abbreviations
ρ PC coefficient
CNNs Convolutional neural networks
σX Standard deviation of X
ConvLSTM Convolutional LSTM
σY Standard deviation of Y
DM Diebold-Mariano
σFD Cross-border flow deviation
DNNs Deep neural networks
w Parameter vector
ELM Extreme learning machine
Tanh Hyperbolic tangent function
EPF Electricity price forecasting
ε Error variable
FNNs Forward neural networks
ξ, ξi∗ Slack variable
GA Genetic algorithm
ζ Sigmoid function
GRUs Gated recurrent units
b Bias vector
LSTM Long short-term memory
C Regularisation constant
MAE Mean absolute error
c Cell state vector
MAPE Mean absolute percentage error
c1 Constant that determines the significance of →
−p
MSE Mean squared error
c2 Constant that determines the significance of −g
→
NARMAX Moving Average with eXogenous Input
cov Covariance
PC Pearson’s Correlation
d F1,F2 Loss differential
PSO Particle swarm optimisation
e Forecast error
RMSE Root mean squared error
f Forget gate’s activation vector
RNNs Recurrent neural networks
g(∗) Loss function
SHAP SHapley Additive exPlanations
h Hidden state vector
SMAPE Symmetric mean absolute percentage error

attempted to involve explanatory variables from integrated mar- search for an optimal subset [34]. The current research on fea-
kets to make a prediction of electricity price by neural networks ture selection algorithms can be categorised as filter, wrapper,
[27, 28, 29], no existing research has investigated the state-of- and embedded methods [33]. In particular, the filter methods
the-art LSTM-based deep neural networks for this purpose. Be- use a proxy measure to estimate a feature subset before train-
sides, some research starts to pay attention to the influence of ing a prediction model. Pearson’s Correlation (PC) is a typical
the market integration on Nord Pool [30, 31, 32]. However, indirect assessing measure for the regression problem [35]. In
efficient ways to utilise the ever-growing information from the contrast, the wrapper methods evaluate selected feature subsets
electricity market integration for the Nordic EPF have yet to be by employing a predictive model directly. Each subset is used
explored. to train a new forecasting model, and the optimisation method
Typically, feature selection is an efficient way to avoid the is used to search for the best performing model in the process of
curse of dimensionality. It is the process of selecting a subset feature selection. The embedded methods can implement an au-
of relevant attributes in the dataset when developing a predic- tomatic feature selection in the process of estimating the param-
tive model. It can reduce the computation time, improve model eters of predictive models. This means this catch-all group of
prediction performance, and help to get a better understanding techniques performs the process of feature selection during the
of the dataset [33]. The ideal feature selection is to search the training of the model. The particle swarm optimisation com-
space of all variable subsets with an algorithm, which is imprac- bined with the extreme learning machine method (PSO-ELM)
tical except for quite small sized feature spaces. However, as and genetic algorithm combined with the extreme learning ma-
the space of variables subset grows exponentially with the num- chine method (GA-ELM) are two typical wrapper-based meth-
ber of variables, heuristic search methods are commonly used to ods. They have been widely used for various feature selection

2
problems [36, 37, 38, 39, 40, 41]. Guyon et al. [42] proposed 2. Data description
another popular wrapper approach, known as recursive feature
elimination combined with support vector machine for regres- The Nordic system price is the central reference price in the
sion (RFE-SVR). The Lasso regression method is one of the Nordic electricity market. It is used as a settlement price for
most popular embedded feature selection methods proposed by the derivatives market. Each hourly system price is calculated
Tibshirani [43]. by Nord Pool based on all bids and offers posted in Nordic
bidding zones, which is referred to as a market-clearing price,
1.1. Contributions without taking into account any congestion restrictions. The
daily system price represents the arithmetic average of the 24
To the best of our knowledge, no existing study considers
hourly prices. This paper discusses and evaluates several hy-
how to apply LSTM models in an integrated market EPF and
brid LSTM-based approaches for the prediction of the Nordic
detect the impact of the features from cross-border markets on
hourly and daily system prices.
EPF. To fill this scientific gap, we propose three hybrid archi-
Previous empirical research on the prediction of electricity
tectures of LSTM-based deep learning predictive models com-
prices has considered information from both price and sup-
bined with advanced feature selection algorithms: the two-step
ply/demand sides. To find out what matters when predicting
hybrid architecture, the autoencoder hybrid architecture, and
the day-ahead Nordic system price in coupling markets, we
the two-stage hybrid architecture. Different feature selection
also included the electricity exchange between Nord Pool and
methods have different selection mechanisms, which will lead
its integrated countries. The Russian electricity market is ex-
to different sets of selected features. To explore the influence
cluded because it differs significantly from European models.
of different feature selections on LSTM-based EPF, we em-
To consider the influence of the correspondence between elec-
ployed five feature selection algorithms, PC, PSO-ELM, GA-
tricity flow and capacity, we introduced a new daily feature,
ELM, RFE-SVR, and the Lasso regression method, in the case
namely theq cross-border flow deviation. It can be calculated as
study of Nord Pool and its neighbouring, interconnected coun-
σFD = i=1 (Xi − µi ) /N, where Xi is the hourly electricity
PN 2
tries. The main contributions of this study are as follows:
flow, µi is the hourly expected exchange capacity, and N stands
1. We compare and analyse the forecasting performance of for 24 hours.
the proposed models in the case study of the Nord Pool In summary, we consider eight categories of input features:
system price forecasting, considering six integrated mar- day-ahead price, production, production prognosis, consump-
kets (sixty-two features). The results indicate that the tion, consumption prognosis, currency exchange rate, cross-
cross-border markets influence the Nordic electricity price border electricity flow, and flow deviation. The first five are
formation. As the rapid market coupling development in the basic features from local markets for predicting electricity
Europe, we show that it is necessary to consider cross- price. Some historical/predictive information, such as weather
border information for EPF in future studies. and human social activities, does not directly impact electricity
price but influences the supply/demand for electricity, incorpo-
2. We introduce three architectures of hybrid LSTM-based
rated in those five fundamental variables. The last three are the
deep neural networks for EPF and conclude that different
features spawned from cross-border trade.
feature selection algorithms yield divergent subsets of fea-
tures, which, in turn, affect the prediction accuracy of the
2.1. Data
proposed LSTM models. In addition, the results show that
hybrid models are an efficient way to deal with the ever- We collected data from the Nord Pool,2 Thomson Reuters
growing information and obtain accurate prediction results Eikon,3 and Entsoe.4 The available time series ranges from
in cross-border markets. 01/01/2015 to 31/12/2019. Nord Pool provides cross-border
transmissions with Germany (DE), the Netherlands (NL),
3. We employ a game theoretical approach (SHapley Ad- Lithuania (LT), Estonia (EE), Poland (PL), and Russia (RU).
ditive exPlanations) to explore the relevance of various The map in Figure 1 shows both the Nord Pool markets as well
cross-border features in EPF. The analysis of Shapley val- as transmissions (black dashed lines) between the Nord Pool
ues increases the transparency of the prediction and pro- and its coupling bidding areas. There are five bidding zones
vides advice for policy makers and market participants. in Norway (NO1, NO2, NO3, NO4, and NO5), four in Swe-
den (SE1, SE2, SE3, and SE4), and two in Denmark (DK1 and
The remainder of this paper is organised as follows. Sec- DK2), and one in Finland (FI). Since the transmissions between
tion 2 describes the dataset used in this research. In Section DK1 and NL started at 01/09/2019, the data series is not suffi-
3, we present the methodology. Section 4 describes the model cient for the application of deep learning models. Besides, the
training and introduces evaluation criteria applied in the em- electricity exchange between SE4 and LT started at 09/12/2015.
pirical study. Section 5 reports the forecasting results of the Therefore, the entire available dataset employed in this study
implemented models. Finally, Section 6 concludes the paper
and proposes future research developments.
2 Nord Pool: [Link]
3 Thomson Reuters Eikon: [Link]
4 Entsoe: [Link]

3
 Continued from previous column
Feature Description (Units) Data Source

F22 PL production (MWh) Entsoe

F23 DE production (MWh) Entsoe
F24 NL production (MWh) Entsoe
F25 Nordic production prognosis (MWh) Nord Pool
F26 EE production prognosis (MWh) Nord Pool
F27 LT production prognosis (MWh) Nord Pool
F28 PL production prognosis (MWh) Entsoe
F29 DE production prognosis (MWh) Entsoe
F30 NL production prognosis (MWh) Entsoe
F31 Nordic consumption (MWh) Nord Pool
F32 EE consumption (MWh) Nord Pool
F33 LT consumption (MWh) Nord Pool
Figure 1 Overview of the Nord Pool market coupling. F34 PL consumption (MWh) Entsoe
F35 DE consumption (MWh) Entsoe
F36 NL consumption (MWh) Entsoe
ranges from 09/12/2015 to 31/12/2019. The features included F37 Nordic consumption prognosis (MWh) Nord Pool
in the dataset are shown in Table 1. The hourly data is converted F38 EE consumption prognosis (MWh) Nord Pool
into the daily data by the arithmetic average (e.g., price) or the F39 LT consumption prognosis (MWh) Nord Pool
aggregate (e.g., flow). F40 PL consumption prognosis (MWh) Entsoe
F41 DE consumption prognosis (MWh) Entsoe
2.2. Cross-border electricity transmission F42 NL consumption prognosis (MWh) Entsoe
Figure 2 shows the electricity exports from Germany, the F43 EUR/NOK Nord Pool
Netherlands, Lithuania, Poland, and Russia in 2019.5 The ex- F44 EUR/SEK Nord Pool
ports to the Nord Pool comprised 16.03% of the whole exports F45 EUR/DKK Nord Pool
F46 EUR/PLN Thomson Reuters Eikon
from these coupling countries. In Figure 3, we can see that the
F47 NO2 ↔ NL flow (MWh) Nord Pool
electricity exports of the Nord Pool comprised 4.82% of its total
F48 DK1 ↔ DE flow (MWh) Nord Pool
production in 2019. The EU aims to achieve 15% interconnec-
F49 DK2 ↔ DE flow (MWh) Nord Pool
tion capacity in 2030 for each EU country [44].
F50 SE4 ↔ DE flow (MWh) Nord Pool
Table 1 The features included in the dataset. F51 SE4 ↔ PL flow (MWh) Nord Pool
F52 SE4 ↔ LT flow (MWh) Nord Pool
Feature Description (Units) Data Source F53 FI ↔ EE flow (MWh) Nord Pool
F1 System Day-ahead price 1-Lag (EUR/MWh)Nord Pool F54 FI ↔ Russia flow (MWh) Nord Pool
F2 SE1 Day-ahead price (EUR/MWh) Nord Pool F55 NO2 ↔ NL flow deviation Calculation
F3 SE2 Day-ahead price (EUR/MWh) Nord Pool F56 DK1 ↔ DE flow deviation Calculation
F4 SE3 Day-ahead price (EUR/MWh) Nord Pool F57 DK2 ↔ DE flow deviation Calculation
F5 SE4 Day-ahead price (EUR/MWh) Nord Pool F58 SE4 ↔ DE flow deviation Calculation
F6 FI Day-ahead price (EUR/MWh) Nord Pool F59 SE4 ↔ PL flow deviation Calculation
F7 DK1 Day-ahead price (EUR/MWh) Nord Pool F60 SE4 ↔ LT flow deviation Calculation
F8 DK2 Day-ahead price (EUR/MWh) Nord Pool F61 FI ↔ EE flow deviation Calculation
F9 NO1 Day-ahead price (EUR/MWh) Nord Pool F62 FI ↔ Russia flow deviation Calculation
F10 NO2 Day-ahead price (EUR/MWh) Nord Pool
F11 NO3 Day-ahead price (EUR/MWh) Nord Pool
F12 NO4 Day-ahead price (EUR/MWh) Nord Pool
F13 NO5 Day-ahead price (EUR/MWh) Nord Pool 3. Methodology
F14 EE Day-ahead price (EUR/MWh) Nord Pool
F15 LT Day-ahead price (EUR/MWh) Nord Pool 3.1. LSTM
F16 PL Day-ahead price (PLN/MWh) Thomson Reuters Eikon The LSTM architecture was initially introduced by [45] and
F17 DE Day-ahead price (EUR/MWh) Thomson Reuters Eikon has since been enhanced by other researchers to achieve better
F18 NL Day-ahead price (EUR/MWh) Thomson Reuters Eikon performance [46, 47, 48]. An LSTM network is a special kind
F19 Nordic production (MWh) Nord Pool of recurrent neural network that is capable of learning long-
F20 EE production (MWh) Nord Pool term dependencies. Unlike simple RNNs, an LSTM network
F21 LT production (MWh) Nord Pool
has built-in mechanisms that control how information is mem-
Continued on next column orised or abandoned throughout time. The architecture of the
LSTM network is shown in Figure 4 and is defined by the fol-
5 Fraunhofer ISE provides the electricity exchange data of Germany/Europe: lowing system of equations [49]:
[Link]

4
 Figure 2 The electricity cross-border transmission from the coupling countries to Nord Pool.

Hidden ht

Cell Cell

ct−1 × + ct

ft Tanh
it
× ot
ct ×

Hidden ζ ζ Tanh ζ Hidden

ht−1 ht
Figure 3 The percentage of the Nord Pool production for exporting.

Input xt
ft = ζ(W x f xt + Wh f ht−1 + Wc f ct−1 + b f ) (1)
it = ζ(W xi xt + Whi ht−1 + Wci ct−1 + bi ) (2)
Figure 4 LSTM cell.
ot = ζ(W xo xt + Who ht−1 + Wco ct−1 + bo ) (3)
ct = ft ⊗ ct−1 + it ⊗ Tanh(W xc xt + Whc ht−1 + bc ) (4)
of the predictive models and making [Link] second ar-
ht = ot ⊗ (ct ) (5) chitecture can be referred to as an autoencoder model. Here,
the input data will be turned into a compressed representation
where ft , it , ot , ct , and ht indicate the values of the forget gate
rather than specifically showing which features are selected, as
state, input gate state, output gate state, memory cell, and hid-
shown in Figure 6. The third combines the two aforementioned
den state at time t in the sequence, respectively. ζ and Tanh
architectures, and it is referred to as two-stage feature selection.
are the sigmoid function and hyperbolic tangent function, W
In this architecture, the explanatory variables will be selected
and b are the weight matrix and bias vector, and ⊗ denotes the
by some feature selection method in the first stage. The se-
element-wise product.
lected features will then become the input for the autoencoder
models in the second stage. Figure 7 shows this architecture.
3.2. Architectures of hybrid models
Typically, there are three hybrid architectures for EPF when 3.3. Feature selection methods
working with LSTM. The first architecture consists of two 3.3.1. PC
steps, as shown in Figure 5. The first step includes data process- The PC coefficient is a statistic used to measure the linear
ing and feature selection, and the second step contains training relationship between two data samples. Given two variables
5
 Input Data Output
processing Feature selection Model Prediction

Step 1 Step 2

Figure 5 The flowchart of a two-step hybrid model. The green nodes stand for the first step, and the red nodes stand for the second step.

Input Data Output

processing Autoencoder Model Prediction

Figure 6 The flowchart of an autoencoder hybrid model. The orange nodes stand for the autoencoder process.

Input Data Output

processing Feature selection Autoencoder Model Prediction

Stage 1 Stage 2

Figure 7 The flowchart of a two-stage hybrid model. The green nodes stand for the first stage, and the orange nodes stand for the second stage.

(X, Y), the formula of the PC coefficient ρ is given by the fol- selection. In general, GAs will search for the optimal solution
lowing: from a set of possible solutions, called a population. A solution
is referred to as a chromosome or an individual. These chromo-
cov(X, Y)
ρ(X, Y) = (6) somes evolve over a number of generations by recombination
σX σY (cross-over) and mutation [55]. The detailed introduction of
where cov is the covariance, σX is the standard deviation of X, the methods can be found in Appendix A.1 and A.2.
and σY is the standard deviation of Y.
3.3.3. RFE-SVR
3.3.2. PSO-ELM and GA-ELM RFE-SVR is another wrapper-based feature selection
method. The core idea of this algorithm is to search for the best
PSO-ELM and GA-ELM are wrapper-based hybrid methods.
subset of features by starting with all features and discarding the
ELM is a single hidden layer feedforward neural network. Its
less important features. In particular, the RFE algorithm oper-
fast training [50] contributes to the popularity of its employ-
ates with SVR to perform feature selection and regression si-
ment as a predictive model in wrapper-based feature selection
multaneously. SVR performs well in high dimensionality space
[51, 52, 53]. The output of ELM is calculated as follows:
[56]. The detailed explanation of SVR is in Appendix A.3.
L
X
F L (x) = w2i φ(w1i x j + bi ), j = 1, ..., N (7) 3.3.4. Lasso regression
i=1 The Lasso regression aims to increase the prediction accu-
racy of regression models by adding a penalty λ nj=1 |β j | to
P
where L is the number of hidden units, N is the number of train- the loss function. This means that instead of minimising a
j=1 xi j β j ) , the loss function becomes
ing samples, w2 is the weight vector between the hidden layer loss function, m
P Pn 2
i=1 (yi −
and the output, w1 is the weight vector between the input and
j=1 xi j β j ) + λ
Pm Pn 2 P n
i=1 (yi − j=1 |β j |, where y is the vector of the
the hidden layer, φ(∗) denotes an activation function, b is a bias dependent variable, x denotes independent variables, the β are
vector, and x is the input vector. the corresponding coefficients. The algorithm has the advantage
PSO and GA are different types of optimisation algorithms, that it shrinks some of the less critical coefficients of features to
which provide the subsets of features as the input to the ELM zero. Therefore, it removes less relevant features.
to detect the optimal feature selection. The basic idea of PSO is
that a swarm of particles moves through the search space. The 3.4. Autoencoder Model
movement of each particle is guided by its own known best- An autoencoder is typically a neural network that aims to fil-
position and the entire swarm’s known best position. PSO per- ter and compress the representation of its input, which consists
forms the search for the optimum by iteratively updating the ve- of two components: an encoder and a decoder, shown in Fig-
locities of the particles in the swarm [54]. The GA is a search ure 8. The encoder typically accepts a set of input data and
metaheuristic that was inspired by Darwin’s theory of natural compresses the information into an intermediate vector. The
6
 4.1. Statistical performance measures

Intermediate Vector
4.1.1. Evaluation metrics
In this paper, we employ several indicators to evaluate the

output
Input

Encoder Decoder accuracy of predictions: the mean absolute error (MAE), the
root mean squared error (RMSE), the mean absolute percent-
age error (MAPE), and the symmetric mean absolute percent-
age error (SMAPE) as the model estimator. They are commonly
Figure 8 Structure of the Encoder-Decoder model.
adopted in EPF research [1]. Given a predicted output vector,
yˆk = [yˆ1 , .., yˆN ], and a real output vector, yk = [y1 , .., yN ], the
MAE, RMSE, MAPE, and SMAPE can be calculated as fol-
decoder is typically a predictive model. In our case, a decoder lows:
is an LSTM network, and the encoders are LSTM, CNN, and N
1 X
convolutional layers, as described in the following. MAE = |yk − yˆk | (8)
N k=1
3.4.1. LSTM-LSTM Encoder-Decoder model v
u
N
t
In an LSTM-LSTM Encoder-Decoder model, an LSTM 1 X
model is used as the encoder to process the raw input time se- RMSE = (yk − yˆk )2 (9)
N k=1
ries and to transform it into an intermediate vector. LSTM is
capable of extracting the complex dynamic information within N
100 X yk − yˆk
the temporal input series and filtering useful information from MAPE = (10)
long input series via internal memory. N k=1 yk

N
3.4.2. CNN-LSTM Encoder-Decoder model 100 X |yk − yˆk |
SMAPE = (11)
In a CNN-LSTM Encoder-Decoder model, a CNN is the en- N k=1 (|yk | + |yˆk |)/2
coder to filter the input data. CNNs were originally and success-
fully used to process the image input data in image recognition
4.1.2. Diebold-Mariano test
tasks [57] or the sequence of input data in natural language pro-
cessing problems [58]. The convolutional layers are usually fol- The metrics for assessing the forecasting accuracy mentioned
lowed by a pooling layer, which extracts information from the above cannot guarantee that the observed difference from two
convolved features and produces a lower dimensional output. predictive models is statistically significant. In this context, the
Then, the output values are flattened into a long intermediate Diebold-Mariano (DM) test is typically used for evaluating the
vector representation. performance of two models [60, 61]. Given the actual values of
a time series [yt ; t = 1, ..., T ], two forecasts from two models,
3.4.3. Convolutional LSTM Encoder-Decoder model [ŷ1t ; t = 1, ..., T ] and [ŷ2t ; t = 1, ..., T ], and the associated fore-
cast errors, e1t = ŷ1t − yt and e2t = ŷ2t − yt , the DM test defines
The computational mechanism of the convolutional LSTM
the loss differential between the two forecasts by the following:
(ConvLSTM) is similar to that of CNN-LSTM [59]. Unlike
the CNN-LSTM, where the CNN model generates the input for dtF1,F2 = g(e1t ) − g(e2t ) (12)
the LSTM model, in the ConvLSTM model, the LSTM neu-
ral network processes the extracted information directly from where g(∗) stands for loss function. In a one-sided DM test, the
preceding convolutional layers. hypotheses is the following:

H0 : E(dtF1,F2 ) ≥ 0,
4. Experimental details (13)
H1 : E(dtF1,F2 ) < 0.
In this section, we introduce the concepts and methods em-
A one-sided DM test is used to detect whether F2 is better than
ployed in the process of training and evaluation of the con-
F1. If H0 is rejected, the test suggests that the accuracy of F1
structed models. This section includes five parts. First, sta-
is, statistically, significantly better than F2. The complementary
tistical performance measures are specified for evaluation and
one-sided DM test can be expressed as follows:
comparison. Second, theory and preliminary work for train-
ing models are explained. Third, configuration parameters are H0 : E(dtF1,F2 ) ≤ 0,
set for feature selection algorithms and LSTM-based models. (14)
H1 : E(dtF1,F2 ) > 0.
Fourth and fifth, the benchmark model and feature explanation
method used in the empirical study are introduced. They are If H0 is rejected, the test suggests that the accuracy of F2 is,
necessary preparations for conducting the experiments. statistically, significantly better than F1. In this study, we em-
ploy a one-sided DM test to assess the forecasting performance
of the proposed models. We chose dtF1,F2 = |e1t | − |e2t | as the
loss differential.
7
 from which to begin the iterative training [63]. The random
Training subset Validation Inner loop
initial conditions for an LSTM network can result in different
performances each time a given configuration is trained. Thus,
Training set Test set we employed ten experiments for each model to reduce the im-
pact of the variability on performance evaluation. Models were
Step 1
evaluated after taking the average of the experiments.
Train Test

Step 2 4.3. Model configuration parameters

Train Test Outer loop 4.3.1. Parameters of feature selection
Step 3 The feature selection stopping criterion varies by algorithm,
Train Test
which is controlled by the parameters of models. The applied
configuration of PSO was [c1 : 0.5, c2 : 0.3, ω: 0.7], and the
Figure 9 Walk forward nested cross-validation. stop condition is satisfied after 10,000 iterations. For GA, the
crossover possibility and mutation possibility were set to 0.5
and 0.2, respectively. The population size was 100, and the
4.2. Model training
maximum number of generations was 10,000. On the basis
4.2.1. Walk forward nested cross-validation of the predictive ELM, the amount of the selected features by
To avoid over-fitting, it is common to include a validation PSO-ELM and GA-ELM was automatically set to 30. For the
set to evaluate the generalisation ability of the training model. sake of input consistency, the magnitude of the selected fea-
The cross-validation is referred to as a method for tuning the tures of the rest of the models was set to 30 as well. For the PC
hyperparameters and producing robust measurements of model method, we ranked all features attributable to the correlation
performance. In [62], a nested cross-validation procedure was coefficients and selected the first 30 features. In terms of RFE-
introduced, which considerably reduced the bias and provided SVR, we ranked features by importance, discarded the least im-
an almost unbiased estimate of the true error. Because new portant features, and refit the model until 30 features remained.
observations become available over time, in time series mod- The regularisation parameter, λ, in Lasso regression was 0.02.
elling, we implemented a walk forward nested cross-validation
in which the forecast rolls forward in time. More specifically, 4.3.2. Network hyperparameters
we successively considered each day as the test set and assigned Our study aimed to investigate the applications and impacts
all previous data to the training set (Outer loop). The training of different types of feature selection methods in a predictive
set is split into a training subset and a validation set. The val- LSTM architecture. We used a coherent configuration of a
idation set data comes chronologically after the training subset specific LSTM model for comparison and did not perform an
(Inner loop). Walk forward validation involves moving along extensive hyperparameter optimisation to search for the opti-
the time series one time step at a time. The process requires mal configuration. After an inexhaustive grid search, we con-
multiple models to be trained and evaluated, but the additional structed our prediction model from an LSTM model with a sin-
computational cost will provide a more robust estimate of the gle hidden layer of 300 units, followed by a fully connected
expected performance of the predictive model on unseen data. dense layer with 100 neurons that preceded the output layer.
It is shown in Figure 9. The LSTM encoder has a hidden layer with 300 units. In the
CNN-LSTM Encoder-Decoder model, the CNN encoder has
4.2.2. Data division
two convolutional layers, with 96 units to amplify any salient
We divided the whole database into two subsets: a training features, followed by a max-pooling layer. In the ConvLSTM
set and a test set. The training set includes a training subset and Encoder-Decoder model, the encoder is a convolutional layer
a validation subset, as shown in the dashed box in Figure 9. We with 64 units. The input sequence length is 14 days (2 weeks,
initially apportioned the data set into training, validation, and commonly used in EPF). The optimiser is the Adam algorithm,
test sets, with an 80-10-10 split. The magnitude of the test and and the loss function is Mean Squared Error (MSE).
validation set is anchored during the walk-forward test.
4.4. Benchmark model
4.2.3. Data processing
For neural network model training, the input data is usually Among the traditional methods, the statistical models per-
normalised to the intervals [0,1]. This is not only done be- form best for EPF. Thus, we select the Nonlinear AutoRe-
cause the normalised data will require less time to train, but gressive Moving Average with eXogenous Input (NARMAX)
the prediction performance will also increase. In addition, we model as the benchmark (trained with the optimal structure) for
linearly interpolate the missing data and eliminate duplicates our case study. This statistical model is widely used in energy
due to daylight saving. price forecasting to handle multiple nonlinear inputs [32, 64].
The equation is represented as the following:
4.2.4. Ten experiments
y(t) = F ` y(t − 1), . . . , y t − Ny , x(t), . . . , x (t − N x ) ,
h  
Training algorithms for deep learning models have usually (15)
required the initialisation of the weights of neural networks ε(t − 1), . . . ε (t − Nε )] + ε(t)
8
Table 2 time series, given their short-term autoregressive nature. As in-
The proposed models.
troduced in section 3.3.2, PSO-ELM and GA-ELM are widely
Mode Category Model Explanation used in research. However, the optimisation methods, such as
M0 Benchmark NARMAX model PSO and GA, have the problem of trapping in local optima. Al-
M1 Filter method PC-LSTM model though re-setting and experimenting can increase the chance of
M2 Wrapper method PSO-ELM-LSTM model
avoiding traps, when dealing with high dimensional data sets,
M3 Wrapper method GA-ELM-LSTM model
M4 Wrapper method RFE-SVR-LSTM model
the optimisation methods still cannot guarantee they will find
M5 Embedded method LASSO-LSTM model a global optimum solution, and they are not suitable for all
M6 Autoencoder method LSTM-LSTM Encoder-Decoder model cases [72]. The straightforward concept and fast computation of
M7 Autoencoder method CNN-LSTM Encoder-Decoder model ELM contributed to its widespread application in an exhaustive
M8 Autoencoder method CovLSTM Encoder-Decoder model grid search, but it does not consider the sequential relationships
M9 Two-stage method PC-LSTM-LSTM Encoder-Decoder model
in time series data, as with other traditional neural networks.
M10 Two-stage method PSO-ELM-LSTM-LSTM Encoder-Decoder model
M11 Two-stage method GA-ELM-LSTM-LSTM Encoder-Decoder model
These could be the reasons why the two methods eliminate the
M12 Two-stage method RFE-SVR-LSTM-LSTM Encoder-Decoder model lag system price as an input. The other wrapper model M4, se-
M13 Two-stage method LASSO-LSTM-LSTM Encoder-Decoder model lects various types of features, and eliminates less price features
compared to M2 and M3. It is interesting to note that M4 does
not pick up any features from the cross-border flow deviation.
where x(t) is the input and y(t) is the output time-series; ε(t) For the Lasso regression method, M5, we can summarise that it
is the uncertainties and possible noise; Nu , Ny , and Nε are the diversely chooses features such as M4 but puts more emphasis
input, output, and prediction error lags, respectively; and F ` is on electricity transmission.
a nonlinear function.
Table 3 The results of feature selection.
4.5. Feature explanation method
Feature selection model
In this study, we used SHapley Additive exPlanations Feature
M1 M2 M3 M4 M5
(SHAP) values to interpret the impact of certain values of a
given feature on the expected price prediction. SHAP6 is a the- F1 3 7 7 3 3
oretic game method to explain the output of machine learning F2 3 3 3 3 3
models [65, 66, 67]. The Shapley value is used to assess the F3 3 7 7 3 3
feature relevance relative to the expectation of the output [68].
F4 3 3 7 3 7
In particular, a Kernel SHAP is used for explaining an optimal
SVR model obtained by grid-search on the dataset. F5 3 7 7 7 7
F6 3 7 7 7 7
5. Results F7 3 7 7 7 7
F8 3 3 7 7 7
In this section, we report the empirical results obtained by F9 3 7 7 3 7
the application of the introduced models. For similarity of pre-
F10 3 7 7 3 3
sentation, the list of models and their acronyms are shown in
Table 2. F11 3 3 7 3 3
F12 3 7 3 3 3
5.1. Analysis of empirical results F13 3 3 3 3 3
The results of the feature selection are shown in Table 3. F14 3 7 3 7 7
Overall, it can be observed that different selection mechanisms F15 3 7 7 7 7
lead to different selections. From Table 3, we can see that M1
F16 3 7 7 7 7
selects all the day-ahead prices. It is not surprising that the
day-ahead prices from different bidding areas are more rele- F17 3 7 3 7 7
vant to the Nord Pool system price than other feature variables. F18 3 3 7 3 3
However, the over-selection results in information redundan- F19 7 3 7 3 7
cies. Some researchers have recognised that the redundancy F20 7 3 3 3 7
among features decreases the model’s performance [69, 70, 71].
F21 7 7 3 7 7
Compared to M1, the wrapper-based methods, M2 and M3,
eliminate several price variables rather than the other categories F22 7 3 3 3 7
of variables. It is worth noticing the two methods does not F23 7 3 7 3 7
select the lag system price (F1), which is commonly used in F24 3 3 3 7 7
F25 7 7 7 3 3
6 The Python package SHAP is available at [Link]
Continued on next column
slundberg/shap

9
 Continued from previous column
Feature selection model
Feature
M1 M2 M3 M4 M5
F26 7 7 3 3 3
F27 3 3 3 7 3
F28 7 3 3 7 7
F29 7 7 7 3 3
F30 7 3 7 7 3
F31 7 3 7 3 7 Figure 10 The SMAPEs of 10 experiments for M1, M2, M3, M4, M5, M6,
F32 7 3 3 3 3 M7, and M8.

F33 3 3 7 7 7
F34 3 7 3 3 3
F35 7 7 7 3 3
F36 7 3 3 3 3
F37 7 7 7 3 7
F38 3 3 3 3 7
F39 3 3 3 3 3
F40 7 7 3 7 7
F41 3 3 7 3 3
Figure 11 The comparison of SMAPEs between two-step LSTM models and
F42 7 7 7 7 7 two-stage LSTM-LSTM models.
F43 3 7 3 3 7
F44 3 3 3 3 3 els are overwhelmingly better than the benchmark statistical
F45 3 7 3 7 7 model, M0. The superior performance of deep learning models
F46 7 3 7 7 3 to statistical models has been recognised by numerous studies
F47 7 7 7 7 7 [22, 73]. Moreover, the performances of LSTM models in ten
F48 7 7 7 3 3 experiments are depicted in Figure 10, measured in terms of
SMAPE. As seen in Figure 10, of all the models, M4, M5, and
F49 7 7 3 7 3 M6 perform better than the others. The statistical details of the
F50 7 3 3 7 3 model performance in ten experiments are listed in Appendix
F51 7 3 7 7 3 Tables A.7, A.8, A.9, and A.10 by means of MAD, RMSE,
F52 7 3 3 7 7 MAPE, and SMAPE. Based on the analysis of feature selec-
F53 7 7 7 3 3 tion, we conclude that M4 and M5, i.e., the minimum redun-
dancy maximum relevance algorithms, perform better than the
F54 7 7 3 7 3 others. The results are consistent with the observations from
F55 7 7 7 7 7 other literature [74, 75] that an elimination of the redundant
F56 7 7 3 7 7 and less relevant features increases the performance of mod-
F57 7 3 3 7 7 els. In addition, some researchers have attempted to introduce
F58 7 7 3 7 3 the CNN-LSTM model and show its excellent performance in
the energy field [76, 77]. However, we found that M6 per-
F59 3 3 3 7 3 forms better than M7 and M8. This means LSTM-LSTM is
F60 3 3 3 7 7 a better autoencoder structure than CNN-LSTM and ConvL-
F61 7 3 7 7 3 STM for EPF. The results are reasonable because convolutional
F62 7 7 7 7 7 neural networks (CNNs or ConvNets) were originally designed
for image recognition and classification, while recurrent neural
Note: 3 denotes that the feature is selected; networks (LSTM) are for sequence and time series prediction.
7 denotes that the feature is not selected. We show the comparison of the SMAPE of the two-step and
the two-stage models in Figure 11. From this figure, it can be
To evaluate the statistical significance in the difference of seen that M4 has been improved by applying LSTM-LSTM as
predictive accuracy, one-side DM tests, as defined in section predictors. We used a one-sided DM test to detect whether the
4.1.2, were applied, and the results are shown in Table 4. Ta- two-stage LSTM models were statistically better than two-step
ble 5 exhibits the performance comparison of all the models LSTM-LSTM models. The results are shown in Table 6. The
in terms of SMAPE. As expected, the proposed LSTM mod-
10
Table 4
The results of the one-sided DM test.

F2
M0 M1 M2 M3 M4 M5 M6 M7 M8
F1

M0 5.83*** 4.99*** 3.90*** 6.65*** 7.09*** 6.60*** 4.61*** 4.27***

M1 -5.83*** -1.15 -1.67* 2.31** 2.98*** 1.59# -0.16 0.88
M2 -4.99*** 1.15 -0.91 2.70*** 3.55*** 3.84*** 0.84 -0.20
M3 -3.90*** 1.67* 0.91 3.19*** 4.01*** 4.31*** 1.33 -0.97
M4 -6.65*** -2.31** -2.70*** -3.19*** 1.12 -0.23 -1.64# 2.57**
M5 -7.09*** -2.98*** -3.55*** -4.01*** -1.12 -0.97 -2.33** 3.19***
M6 -6.60*** -1.59# -3.84*** -4.31*** 0.23 0.97 -1.44 3.54***
M7 -4.61*** 0.16 -0.84 -1.33 1.64# 2.33** 1.44 0.66
M8 -4.27*** -0.88 0.20 0.97 -2.57** -3.19*** -3.54*** -0.66
Note: ***, ** ,* and # denote 1%, 5%, 10%, and 15% significance levels, respectively. The positive sign of the DM value indicates that F2 is
better F1. The negative sign of the DM value indicate that F1 is better F2.

Table 5
The SMAPE of M0, M1, M2, M3, M4, M5, M6, M7, and M8.

Model M0 M1 M2 M3 M4 M5 M6 M7 M8
SMAPE 10.07 6.25 6.58 7.06 5.29 4.89 5.20 6.14 6.53

Table 6
The results of the one-sided DM test when comparing two-step models (F1)
and two-stage models (F2).

F1 M1 M2 M3 M4 M5
Figure 12 The SMAPEs of 10 experiments for predicting H8.
-0.6039 0.2222 2.4556 *** 2.4524 *** -1.7053 ***
F2 M9 M10 M11 M12 M13
Note: ***, ** ,* and # denote 1%, 5%, 10%, and 15% significance levels,
respectively. The positive sign: F2 is better F1. The negative sign: F1 is better
F2.

superior features selection from M4 provides the possibility for

autoencoder models to further process the selected features to
obtain more meaningful information.
Additionally, we detected the forecasting performance for 24
hourly system prices. Figures 12, 13, and 14 show the results Figure 13 The SMAPEs of 10 experiments for predicting H12.
for the three peak hours: H8 (07 - 08), H12 (11 - 12) and H18
(17 - 18), respectively, measured in terms of SMAPE. We ob-
served that the feature selections influence the forecasting ac-
curacy and the models M4 and M5 are still relatively stable,
performing better than other models. Indeed, the RFE-SVR
and Lasso regression feature selection methods are applied to
various areas in energy finance and achieve good performance
for improving forecasting accuracy [78, 79, 80, 81].

5.2. Analysis of feature impact

Figure 15 shows the ranking of features and their impacts Figure 14 The SMAPEs of 10 experiments for predicting H18.
on the predicted price in terms of the selected features of the
RFE-SVR model (M4), which is the model with the best per-
formance. From Figure 15a, we can observe that the features features. In particular, production and consumption and their
from supply/demand sides are more important than the other prognosis in the Nordic and German markets are prioritised

11
 (a) Features ranking (b) Features impact
Figure 15 The feature ranking and feature impact of the selected features of RFE-SVR. (a) Bar chart of the average SHAP value magnitude showing the importance
of the features. (b) A set of beeswarm plots, where each dot corresponds to an individual day-ahead price. The dot’s position on the x-axis shows the impact that
feature has on the model’s prediction for that price. Multiple dots landing at the same x position pile up to show density.

by the model. The significant impact from the German mar- change in predicted price as DE consumption changes. Verti-
ket can be explained by the fact that the German market has cal dispersion at a single value of DE consumption represents
the most electricity cables and the highest electricity exports to the interaction effects with other features. For example, the in-
the Nordic market, as shown in Figures 1 and 2. This indicates teraction effect of DE consumption with the Nordic production
that it is critical to consider features from cross-border markets (F19) is shown in Figure 16c. The dependence plot highlights
with increasing interconnections across Europe for EPF. Be- that the impact of DE consumption differs with different lev-
sides, electricity prices have more impact on EPF than the fea- els of the Nordic production. The results reveal that the Nordic
tures from cross-border electricity trade. From Figure 15b, we price is less sensitive to the German power consumption when
can see that the features have asymmetric predictive influence the Nordic electricity is oversupplied. It indicates, in such a
on electricity price. For instance, the impact of DE consump- case, that the information from the Nordic market rather than
tion (F35) on EPF has a long-tail reaching to the right but not cross-border countries drives the price prediction.
to the left. This indicates that German over-consumption can From Figure 17, we can see that the majority of the
result in high Nordic electricity prices, but scarce consumption EUR/NOK exchange rates (F43) have no contribution to the
cannot significantly lower the price. prediction of the Nordic price (the y-axis value of the dots is
Moreover, we detected the relations between different types zero). In addition, there is no obvious interaction effect of the
of features and the predicted price with their dependence plots. Nordic productions and the exchange rates on the price. Thus,
Figure 16a demonstrates the negative association between DE the predictive importance of exchange rate is extremely limited.
consumption and its conditional expectation of the predicted From Figure 18, we can find that the predicted price is ex-
price. If the DE consumption is high, then its value tends to pected to increase when observing a high DK1→DE cross-
revert to its expectation. Thus, the downward expectation of border electricity flow, indicating a relatively low current
DE consumption will lead to the expected decline of the im- Nordic price. By contrast, a high flow from DE to DK1 implies
port demand from the Nordic market, which further decreases that the Nordic price is relatively high and expected to decline.
the expectation of the Nordic price. Figure 16b represents the In addition, it can be seen via an interaction effect of the DK1

12
 (c) SHAP dependence plot of the interaction effect of DE
(a) SHAP partial dependence plot of DE consumption (b) SHAP dependence plot of DE consumption consumption with the Nordic production (F19)
Figure 16 The SHAP partial dependence (a) and dependence (b, c) plot of DE consumption (F35). E[F35] is the expectation of the DE consumption, and E[f(x)]
is the expectation of the Nordic price. The grey histogram in (a) shows the distribution of the feature in the test dataset. For (a), the x-axis is the normalised DE
consumption. The x-axes in (b) and (c) are the real values of the consumption.

Figure 19 FI ↔ Russia flow (F54) (blue dots) versus FI ↔ Russia flow devia-
tion (F62) (red line).
Figure 17 The SHAP dependence plot of EUR/NOK (F43). The interaction
effect of EUR/NOK with the Nordic production (F19).
The lack of variability results in their inability to provide use-
ful information for forecasting. An example of the flow and
↔ DE flow with the DE production prognosis that the flow has flow deviations between FI and Russia can be seen in Figure
less impact on the predicted price, with high expected produc- 19. From Figure 19, it is evident that the majority of flow devi-
tion in Germany. The high production prognosis from cross- ations are zero. The findings indicate that the non-selection of
border countries will lead to a sharp decline in the expected flow deviations from M4 is essential and reasonable. However,
cross-border transmission. Thus, the impact of the cross-border the capacity utilisation indicates the potential for more electri-
flow on the Nordic price formation on the following day will cal power transmission across the Europe-wide market which
decrease significantly. would increase the overall socio-economic benefits.

5.3. Discussion of practical importance

The obtained results in the empirical study show that the
LSTM-based hybrid models with various features from cross-
border markets have considerably accurate prediction results
for electricity price. An accurate prediction can be highly bene-
ficial for the electricity market participants in practice. A power
market firm that is capable of forecasting the volatile electric-
ity price with a reasonable level of accuracy can reduce trading
risk and maximise profits in the day-ahead market by adjust-
ing its bidding strategy and the schedule for production or con-
sumption. More specifically, a 1% improvement in MAPE of
forecast accuracy (within a 5% to 14% range) leads to about a
Figure 18 The SHAP dependence plot of DK1 ↔ DE flow (F48). The interac-
tion effect of DK1 ↔ DE flow with DE production prognosis (F29).
0.1 - 0.35% cost reduction [82]. On average, a 1% reduction
in the MAPE of short-term price forecasts can result in savings
of $1.5 million per year for a typical medium-sized utility com-
Last but not least, not all of the cross-border electricity flows
pany with 5-GW peak load [83, 29]. Furthermore, electricity
and flow deviations are helpful for forecasting. The reason for
is economically non-storable, and the imbalance between pro-
this is that, in many cases, the flow capacity is fully occupied.
duction and consumption can result in power system instability
13
[84]. Accurate electricity forecasting allows energy firms to ef- allows the autoencoder model to detect more meaningful infor-
ficiently organise production or consumption, and this improves mation for more accurate predictions. (4) The features from the
the stability of the power system. German market (with the most power cables linking to Nord
In view of the findings from the analysis of feature impact, Pool) are more significant for EPF than others. This indicates
some implications are important for policy makers to improve that more interconnections will increase the cross-border influ-
cross-border trading in an integrated European power market: ence on EPF. (5) Compared to other features, the exchange rates
are relatively less important. (6) Flow deviation cannot signif-
1. The external trading capacities from the German market icantly contribute to the price prediction because of its lack of
play a salient role in the generation of Nordic electric- variability. In many cases, the expected flow capacity is fully
ity price, and an increasing influence7 is expected. Thus, occupied. The network congestion implies that more intercon-
all trading capacity between the Nordic and German mar- nections are expected for an efficient Europe-wide electricity
kets allocated to Nord Pool for implicit auction in the day- market.
ahead price formation could lead to a notable contribution For future studies, several extensions of the current study can
to achieve better allocation of cross-border network capac- be developed. Indeed, although the forecasting performance of
ity, such as the Nordic and Baltic bidding areas. the proposed models is considerable, we did not conduct an
2. The German production prognosis has a significant pre- extensive grid search to optimise hyperparameters. It is reason-
dictive impact on the Nordic price. However, the pro- able to believe that the LSTM-based models with more com-
gressive introduction of intermittent renewable energies prehensive architecture will achieve better forecasting perfor-
in Germany8 makes it difficult to yield accurate predic- mance. The results will benefit spot electricity traders and poli-
tions. Thus, it is essential to establish formal obligations cymakers, who make decisions based on accurate price predic-
for cross-border markets to collaborate by sharing useful tions. Moreover, we envision that more testing on other feature
prospective information. It can better serve the effective selection models can obtain more and different feature selection
demand needed of the electricity trade. subsets. They can provide more possibilities for researchers and
industries to understand how different features affect prediction
3. The presence of physical transmission constraints can accuracy. Finally, the study was carried out using the data from
cause a substantial disruption to the market integration and the Nord Pool market, but the generality of the proposed mod-
the network congestion implies the shrinkage of the com- els ensures a possible application to other integrated markets,
mercial capacity. Thus, an optimal network at a European such as EPEX and OMIE.
level should be constructed by a decisive plan.
Acknowledgement
6. Conclusion
This work acknowledges research support by COST Ac-
In this paper, we present three LSTM-based hybrid archi- tion “Fintech and Artificial Intelligence in Finance - Towards a
tectures for the EPF. This study puts emphasis on the influ- transparent financial industry” (FinAI) CA19130, and has been
ence of feature selection methods in the proposed hybrid mod- performed within the +CityxChange9 (Positive City ExChange)
els. In particular, we compare the prediction performance of project under the Smart Cities and Communities topic that has
the two-step feature selection, the autoencoder, and two-stage received funding from the European Union’s Horizon 2020
feature selection models based on the empirical study on the research and innovation programme under Grant Agreement
Nord Pool day-ahead system price. In addition, we employ a No. 824260. Critical comments and advice from Florentina
SHAP method to evaluate the importance and impact of the Paraschiv, Rüdiger Kiesel, and Frode Kjærland are gratefully
features on predicting this price. The main findings are the acknowledged. The computations were performed on resources
following: (1) We conclude that the different feature selection provided by UNINETT Sigma2 - the National Infrastructure for
methods will lead to different feature selections. As input, di- High Performance Computing and Data Storage in Norway.
verse features will have a comparably significant impact on the
performance of LSTM-based predictive models. (2) Compared
to CNN-LSTM and ConvLSTM, LSTM-LSTM is a better au-
toencoder structure for EPF. (3) The two-stage models can im-
prove the forecasting accuracy of two-step models to some ex-
tent. The superior feature selection from the RFE-SVR model

7 The NordLink, the power cable being built between Norway and Germany,

is expected to commence operation in 2021.

8 Renewable power generation covered more than 46% of Germany’s power

consumption in 2020. The forecasting information of intermittent renewable

energies, wind and photovoltaic infeed forecasts, in Germany are essential
supply-side variables for the adjustment positions of intraday trading and are
updated every 15 minutes until the physical delivery of electricity [85]. 9 [Link]

14
Appendix A.
Feature set F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 ... Fn

Appendix A.1. PSO Features

Encoding ...
1 1 0 0 1 0 0 0 1 0 1 Continue
Each particle has knowledge about its current velocity, its until the
own past best solution (→ −p (t)), and the current global best solu-
Features
conditional
terminal
→
−
tion ( g (t)). Based on this information, each particle’s velocity Selected F1 F2 F5 F9
...
Fn

is updated such that it moves closer to the global best and its
past best solution at the same time. The velocity update is per- F1 ELM
Global optimum update.
formed according to the following equation: F2

→
−v (t + 1) = ω→−v (t) + c r (→ − →− Predictive F5
1 1 p (t) − x (t)) Model Individual updates:
..
(A.1) . optimal set of features.
+ c r (→−g (t) − →
−x (t)) F9
2 2
..
.
where c1 and c2 are constants defined beforehand, which de-
termine the significance of → −p (t) and →
−g (t). →
−v (t) is the velocity Fn

→−
of the particle, x (t) is the current particle position, r1 and r2 Input layer Hidden layer Output layer

are random numbers from the interval [0,1], and ω is a constant

(0 ≤ ω ≤ 1). The new position is calculated by summing the Figure A.20 The workflow of the two-step wrapper-based feature selection
previous position and the new velocity as follows: model.

→
−x (t + 1) = →
−x (t) + →
−v (t + 1) (A.2) y
support vector
This iterative process is repeated until a stopping criterion is f itted by SV R
satisfied. data points inside

f (x) + 
Appendix A.2. GA-ELM workflow ξi f (x)

Figure A.20 shows the workflow of PSO-ELM and GA-ELM f (x) − 

models for feature selection. The process flow of GA-ELM can ξi∗
be described as follows:

Step 1: Initialise the population with a set of random individ- x

uals, each individual representing a particular subset
of features. For a specific individual (feature set), the Figure A.21 Fitted SVR.
features are encoded as ”1” or ”0”, as shown in Figure
A.20. ”1” means that the feature is selected, and ”0”
Appendix A.3. SVR
means that it is not selected.
To minimise the forecasting errors, SVR individualises the
Step 2: The selected features are the input for the ELM. The hyperplane by maximising the margin. To solve a nonlinear
prediction results of the ELM are used to evaluate the regression problem, the following linear estimation function is
fitness value of the individuals. The fitness value is considered as follows [86]:
calculated based on the MSE.
f (x) = (w × Φ(x)) + b (A.3)
Step 3: Select the best individual with regard to the fitness
value. If its fitness is higher than the lowest value in where w is the parameter vector, Φ(x) is a kernel function and b
the existing mating pool, it will replace the individual is a bias vector. The function formulation of the SVR model can
with the worst fitness. Furthermore, the global opti- be transformed into the following convex minimisation prob-
mum will be updated accordingly. lem:
n
1 X
Step 4: The child individuals are generated by crossover and min ||w||2 + C (ξi + ξi∗ ) (A.4)
2
mutation. The new generation is composed of a set of i=1
new individuals that are encoded and prepared to be subject to the following constraints:
evaluated. The whole process continues until meeting
the iteration terminal. The best feature subset in the (w × Φ(xi ) + b) − yi ≤ ε + ξi
mating pool is the optimal selection. yi − (w × Φ(xi ) + b) ≤ ε + ξi∗
ξi , ξi∗ ≥ 0; i = 1, 2, ..., n

where C is a regularisation constant and ξi and ξi∗ are slack vari-

ables, which are used to handle the situation where no such
15
function f (x) exists to satisfy the constraint |yi −(w× xi +b)| ≤ ε Table A.10
The SMAPE (%) results for M1, M2, M3, M4, M5, M6, M7 and M8.
for all points. They are regarded as the soft margin to allow re-
gression errors, ε, to exist up to ξi and ξi∗ and still satisfy the M1 M2 M3 M4 M5 M6 M7 M8
constraint. Only the points outside the ε-radius contribute to
count 10 10 10 10 10 10 10 10
the final cost. The error parameter, ε, represents the region of
the tube located around the regression function, f (x), as shown mean 7.42 7.58 8.46 6.76 6.85 6.87 9.51 8.38
in Figure A.21. std 0.50 0.67 0.95 0.65 0.72 0.60 0.44 0.90
min 6.49 6.63 6.77 5.76 5.46 5.71 8.49 7.38
Appendix A.4. The statistical details of the model performance 25% 7.22 7.10 7.95 6.33 6.43 6.77 9.38 7.59
50% 7.34 7.48 8.30 6.87 6.91 6.85 9.67 8.22
Table A.7 75% 7.81 8.10 9.05 7.01 7.37 7.27 9.77 9.13
The MAD (%) results for M1, M2, M3, M4, M5, M6, M7 and M8.
max 8.13 8.70 10.09 7.87 7.74 7.75 9.95 9.93
M1 M2 M3 M4 M5 M6 M7 M8
Note: 25%, 50%, and 75% denote 25%, 50%, and 75% percentiles.
count 10 10 10 10 10 10 10 10
mean 2.78 2.99 3.31 2.54 2.61 2.67 3.67 3.28
References
std 0.17 0.29 0.40 0.25 0.30 0.25 0.15 0.38
min 2.42 2.58 2.64 2.16 2.04 2.20 3.30 2.84 [1] Weron R. Modeling and forecasting electricity loads and prices: A statis-
tical approach. Wiley; 2006.
25% 2.72 2.77 3.10 2.34 2.45 2.59 3.65 2.99 [2] Bunn D. Modelling prices in competitive electricity markets. Wiley;
50% 2.77 2.95 3.25 2.59 2.65 2.65 3.69 3.18 2004.
[3] Nogales FJ, Contreras J, Conejo AJ, Espinola R. Forecasting next-
75% 2.92 3.24 3.56 2.63 2.74 2.85 3.73 3.60 day electricity prices by time series models. IEEE Transactions on
max 2.98 3.41 3.93 2.99 3.00 3.09 3.89 3.95 Power Systems 2002;17(2):342–348. [Link]
TPWRS.2002.1007902.
Note: 25%, 50%, and 75% denote 25%, 50%, and 75% percentiles. [4] Bunn DW. Forecasting loads and prices in competitive power mar-
kets. Proceedings of the IEEE 2000;88(2):163–169. [Link]
10.1109/5.823996.
Table A.8 [5] Weron R. Electricity price forecasting: A review of the state-
The RMSE (%) results for M1, M2, M3, M4, M5, M6, M7 and M8. of-the-art with a look into the future. International Journal of
Forecasting 2014;30(4):1030 – 1081. [Link]
M1 M2 M3 M4 M5 M6 M7 M8 [Link].2014.08.008.
[6] Nowotarski J, Weron R. Recent advances in electricity price forecast-
count 10 10 10 10 10 10 10 10 ing: A review of probabilistic forecasting. Renewable and Sustainable
Energy Reviews 2018;81:1548 – 1568. [Link]
mean 3.60 3.79 4.22 3.25 3.46 3.33 4.74 3.99 [Link].2017.05.234.
std 0.24 0.42 0.55 0.33 0.43 0.32 0.29 0.46 [7] Ventosa M, Baillo A, Ramos A, Rivie M. Electricity market model-
ing trends. Energy Policy 2005;33(7):897 – 913. [Link]
min 3.17 3.15 3.35 2.61 2.79 2.85 4.13 3.50 10.1016/[Link].2003.10.013.
25% 3.51 3.45 3.87 3.04 3.27 3.13 4.62 3.60 [8] Kiose D, Voudouris V. The acewem framework: An integrated agent-
based and statistical modelling laboratory for repeated power auctions.
50% 3.64 3.78 4.16 3.30 3.50 3.30 4.90 3.92
Expert Systems with Applications 2015;42(5):2731 – 2748. https://
75% 3.71 4.05 4.55 3.45 3.71 3.47 4.91 4.33 [Link]/10.1016/[Link].2014.11.024.
max 3.96 4.51 5.11 3.73 4.08 3.91 4.97 4.77 [9] Burger M, Schindlmayr G, Graeber B. Managing Energy Risk: An Inte-
grated View on Power and Other Energy Markets. Wiley; 2007.
Note: 25%, 50%, and 75% denote 25%, 50%, and 75% percentiles. [10] Islyaev S, Date P. Electricity futures price models: Calibration and fore-
casting. European Journal of Operational Research 2015;247(1):144 –
154. [Link]
[11] Weron R, Misiorek A. Forecasting spot electricity prices: A comparison
Table A.9 of parametric and semiparametric time series models. International Jour-
The MAPE (%) results for M1, M2, M3, M4, M5, M6, M7 and M8. nal of Forecasting 2008;24(4):744 – 763. [Link]
[Link].2008.08.004.
M1 M2 M3 M4 M5 M6 M7 M8
[12] Conejo AJ, Contreras J, Espı́nola R, Plazas MA. Forecasting electric-
count 10 10 10 10 10 10 10 10 ity prices for a day-ahead pool-based electric energy market. Interna-
tional Journal of Forecasting 2005;21(3):435 – 462. [Link]
mean 7.24 7.83 8.75 6.66 6.79 7.01 9.31 8.73 10.1016/[Link].2004.12.005.
std 0.46 0.73 1.05 0.65 0.78 0.68 0.47 1.06 [13] Misiorek [Link] Trueck S, Weron R. Point and interval forecasting of
min 6.28 6.82 6.89 5.72 5.32 5.67 8.56 7.50 spot electricity prices: Linear vs non-linear time series models. Studies in
Nonlinear Dynamics & Econometrics 2006;10(3). [Link]
25% 7.05 7.22 8.19 6.18 6.35 6.84 9.08 7.90 1558-3708.1362.
50% 7.24 7.71 8.62 6.73 6.85 6.95 9.16 8.44 [14] Gonzalez JP, Roque AMSMS, Pérez EA. Forecasting functional time
series with a new Hilbertian ARMAX model: Application to electricity
75% 7.61 8.49 9.35 6.87 7.15 7.45 9.50 9.59 price forecasting. IEEE Transactions on Power Systems 2018;33(1):545–
max 7.80 8.93 10.37 7.86 7.91 8.09 10.27 10.66 556. [Link]
[15] Catalao J, Mariano S, Mendes V, Ferreira L. Short-term electricity prices
Note: 25%, 50%, and 75% denote 25%, 50%, and 75% percentiles. forecasting in a competitive market: A neural network approach. Electric

16
 Power Systems Research 2007;77(10):1297 – 1304. [Link] The Journal of Machine Learning Research 2003;3(null):1157–1182.
10.1016/[Link].2006.09.022. [36] Chen X, Zeng X, van Alphen D. Multi-class feature selection for tex-
[16] Keles D, Scelle J, Paraschiv F, Fichtner W. Extended forecast methods ture classification. Pattern Recognition Letters 2006;27(14):1685 – 1691.
for day-ahead electricity spot prices applying artificial neural networks. [Link]
Applied Energy 2016;162:218 – 230. [Link] [37] Nguyen HB, Xue B, Liu I, Andreae P, Zhang M. New mechanism
[Link].2015.09.087. for archive maintenance in PSO-based multi-objective feature selection.
[17] Peter S, Raglend I. Sequential wavelet-ANN with embedded ANN-PSO Soft Computing 2016;20:3927 – 3946. [Link]
hybrid electricity price forecasting model for indian energy exchange. s00500-016-2128-8.
Neural Comput & Applic 2017;28:2277–2292. [Link] [38] Shang L, Zhou Z, Liu X. Particle swarm optimization-based feature se-
10.1007/s00521-015-2141-3. lection in sentiment classification. Soft Computing 2016;20:3821 – 3834.
[18] Hinton G, Deng L, Yu D, Dahl GE, rahman Mohamed A, Jaitly N, et al. [Link]
Deep neural networks for acoustic modeling in speech recognition: The [39] Zhou Y, Zhou N, Gong L, Jiang M. Prediction of photovoltaic power out-
shared views of four research groups. IEEE Signal Processing Magazine put based on similar day analysis, genetic algorithm and extreme learn-
2012;29:82 – 97. [Link] ing machine. Energy 2020;204:117894. [Link]
[19] Bahdanau D, Cho K, Bengio Y. Neural machine translation by jointly [Link].2020.117894.
learning to align and translate. arXiv:1409.0473 [[Link]] 2014;Available [40] S. Krishnan G, S. S. A novel GA-ELM model for patient-specific mor-
from: [Link] tality prediction over large-scale lab event data. Applied Soft Computing
[20] Li L, Yuan Z, Gao Y. Maximization of energy absorption for a wave Journal 2019;80:525–533. 10.1016/[Link].2019.04.019.
energy converter using the deep machine learning. Energy 2018;165:340 [41] Luo P, Zhu S, Han L, Chen Q. Short-term photovoltaic generation fore-
– 349. [Link] casting based on similar day selection and extreme learning machine.
[21] Bengio Y, Simard P, Frasconi P. Learning long-term dependencies with IEEE Power and Energy Society General Meeting 2018;2018-January:1–
gradient descent is difficult. IEEE Transactions on Neural Networks 5. 10.1109/PESGM.2017.8273776.
1994;5(2):157–166. [Link] [42] Guyon I, Weston J, Barnhill S, Vapnik V. Gene selection for can-
[22] Lago J, Ridder FD, Schutter BD. Forecasting spot electricity prices: cer classification using support vector machines. Machine Learning
Deep learning approaches and empirical comparison of traditional al- 2002;46(1):389–422. 10.1023/A:1012487302797.
gorithms. Applied Energy 2018;221:386 – 405. [Link] [43] Tibshirani R. Regression shrinkage and selection via the lasso. Journal of
10.1016/[Link].2018.02.069. the Royal Statistical Society: Series B (Methodological) 1996;58(1):267–
[23] Chang Z, Zhang Y, Chen W. Electricity price prediction based on 288. [Link]
hybrid model of adam optimized LSTM neural network and wavelet [44] Greenfish. Shaping our electrical future: Moving towards an inte-
transform. Energy 2019;187:115804. [Link] grated european network. [Link]
[Link].2019.07.134. our-electrical-future-moving-towards-an-integrated-
[24] Kuo PH, Huang CJ. An electricity price forecasting model by hybrid european-network/; 2019. [accessed 13 September 2020].
structured deep neural networks. Sustainability 2018;10(4). https:// [45] Hochreiter S, Schmidhuber J. Long short-term memory. Neural
[Link]/10.3390/su10041280. Computation 1997;74(8):1735 – 1780. [Link]
[25] Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: neco.1997.9.8.1735.
data mining, inference and prediction. 2 ed.; Springer; 2009. [46] Gers FA, Schmidhuber J. Recurrent nets that time and count. In: Pro-
[26] Li J, Cheng K, Wang S, Morstatter F, Trevino RP, Tang J, et al. Fea- ceedings of the IEEE-INNS-ENNS International Joint Conference on
ture selection: A data perspective. ACM Computing Surveys 2017;50(6). Neural Networks. IJCNN 2000. Neural Computing: New Challenges
10.1145/3136625. and Perspectives for the New Millennium; vol. 3. 2000, p. 189–194.
[27] Ziel F, Steinert R, Husmann S. Forecasting day ahead electricity spot [Link]
prices: The impact of the exaa to other european electricity markets. [47] Graves A, Schmidhuber J. Framewise phoneme classification with
Energy Economics 2015;51:430 – 444. [Link] bidirectional LSTM and other neural network architectures. Neu-
[Link].2015.08.005. ral Networks 2005;18(5):602 – 610. [Link]
[28] Panapakidis IP, Dagoumas AS. Day-ahead electricity price fore- [Link].2005.06.042; iJCNN 2005.
casting via the application of artificial neural network based models. [48] Cho K, van Merrienboer B, Gulcehre C, Bahdanau D, Bougares F,
Applied Energy 2016;172:132 – 151. [Link] Schwenk H, et al. Learning phrase representations using RNN encoder-
[Link].2016.03.089. decoder for statistical machine translation. arXiv:1406.1078 [[Link]]
[29] Lago J, De Ridder F, Vrancx P, De Schutter B. Forecasting day-ahead 2014;Available from: [Link]
electricity prices in europe: The importance of considering market in- [49] Graves A. Generating sequences with recurrent neural networks.
tegration. Applied Energy 2018;211:890 – 903. [Link] arXiv:1308.0850 [[Link]] 2013;Available from: [Link]
10.1016/[Link].2017.11.098. abs/1308.0850.
[30] Uribe JM, Mosquera-López S, Guillen M. Characterizing electricity [50] Huang GB, Zhu QY, Siew CK. Extreme learning machine: Theory
market integration in nord pool. Energy 2020;208:118368. https: and applications. Neurocomputing 2006;70(1):489 – 501. https:
//[Link]/10.1016/[Link].2020.118368. //[Link]/10.1016/[Link].2005.12.126.
[31] Marcjasz G, Lago J, Weron R. Neural networks in day- [51] Saraswathi S, Sundaram S, Sundararajan N, Zimmermann M, Nilsen-
ahead electricity price forecasting: Single vs. multiple out- Hamilton M. ICGA-PSO-ELM approach for accurate multiclass can-
puts. arXiv:2008.08006 [[Link]] 2020;Available from: https: cer classification resulting in reduced gene sets in which genes encod-
//[Link]/abs/2008.08006. ing secreted proteins are highly represented. IEEE/ACM Transactions on
[32] Johannesen NJ, Kolhe M, Goodwin M. Deregulated electric energy Computational Biology and Bioinformatics 2011;8(2):452–463. https:
price forecasting in nordpool market using regression techniques. In: //[Link]/10.1109/TCBB.2010.13.
2019 IEEE Sustainable Power and Energy Conference (iSPEC). 2019, p. [52] Chyzhyk D, Savio A, Graña M. Evolutionary ELM wrapper fea-
1932–1938. [Link] ture selection for Alzheimer’s disease CAD on anatomical brain MRI.
8975173. Neurocomputing 2014;128:73 – 80. [Link]
[33] Chandrashekar G, Sahin F. A survey on feature selection methods. Com- [Link].2013.01.065.
puters & Electrical Engineering 2014;40(1):16 – 28. [Link] [53] Ahila R, Sadasivam V, Manimala K. An integrated PSO for parameter de-
10.1016/[Link].2013.11.024; 40th-year commemorative is- termination and feature selection of ELM and its application in classifica-
sue. tion of power system disturbances. Applied Soft Computing 2015;32:23
[34] Sanz H, Valim C, Vegas E, Oller JM, Reverter F. Svm-rfe: selection – 37. [Link]
and visualization of the most relevant features through non-linear kernels. [54] Zhang Y, Wang S, Ji G. A comprehensive survey on particle swarm opti-
BMC Bioinformatics 2018;19(1):432. 10.1186/s12859-018-2451-4. mization algorithm and its applications. Mathematical Problems in Engi-
[35] Guyon I, Elisseeff A. An introduction to variable and feature selection. neering 2015;2015:931256. 10.1155/2015/931256.

17
[55] Whitley D. A genetic algorithm tutorial. Statistics and Computing rfe classification. Renewable and Sustainable Energy Reviews
1994;4(2):65–85. 10.1007/BF00175354. 2017;70:330–341. [Link]
[56] Drucker H, Burges CJC, Kaufman L, Smola A, Vapnik V. Support article/pii/S1364032116309297. [Link]
vector regression machines. In: Mozer MC, Jordan M, Petsche T, [Link].2016.11.155.
editors. Advances in Neural Information Processing Systems; vol. 9. [75] Radovic M, Ghalwash M, Filipovic N, Obradovic Z. Minimum redun-
MIT Press; 1997,[Link] dancy maximum relevance feature selection approach for temporal gene
file/[Link]. expression data. BMC Bioinformatics 2017;18(1):9. [Link]
[57] Szegedy C, Wei Liu, Yangqing Jia, Sermanet P, Reed S, Anguelov D, 10.1186/s12859-016-1423-9. 10.1186/s12859-016-1423-9.
et al. Going deeper with convolutions. In: 2015 IEEE Conference on [76] Kim TY, Cho SB. Predicting the household power consumption using
Computer Vision and Pattern Recognition (CVPR). 2015, p. 1–9. https: cnn-lstm hybrid networks. In: Yin H, Camacho D, Novais P, Tallón-
//[Link]/10.1109/CVPR.2015.7298594. Ballesteros AJ, editors. Intelligent Data Engineering and Automated
[58] Sutskever I, Vinyals O, Le QV. Sequence to sequence learning with Learning – IDEAL 2018. Cham: Springer International Publishing; 2018,
neural networks. In: Ghahramani Z, Welling M, Cortes C, Lawrence p. 481–490.
ND, Weinberger KQ, editors. Advances in Neural Information Process- [77] Kim TY, Cho SB. Predicting residential energy consumption using cnn-
ing Systems 27. Curran Associates, Inc.; 2014, p. 3104–3112. https: lstm neural networks. Energy 2019;182:72–81. [Link]
//[Link]/abs/1409.3215. 10.1016/[Link].2019.05.230.
[59] SHI X, Chen Z, Wang H, Yeung DY, Wong Wk, WOO Wc. Convolu- [78] Sultana T, Khan ZA, Javaid N, Aimal S, Fatima A, Shabbir S. Data an-
tional LSTM network: A machine learning approach for precipitation alytics for load and price forecasting via enhanced support vector regres-
nowcasting. In: Cortes C, Lawrence ND, Lee DD, Sugiyama M, Gar- sion. In: Barolli L, Xhafa F, Khan ZA, Odhabi H, editors. Advances in
nett R, editors. Advances in Neural Information Processing Systems 28. Internet, Data and Web Technologies. Cham: Springer International Pub-
Curran Associates, Inc.; 2015, p. 802–810. [Link] lishing; 2019, p. 259–270.
[Link]. [79] Nawaz M, Javaid N, Mangla FU, Munir M, Ihsan F, Javaid A, et al. An
[60] Diebold FX, Mariano RS. Comparing predictive accuracy. Journal approximate forecasting of electricity load and price of a smart home us-
of Business & Economic Statistics 2002;20(1):134–144. https:// ing nearest neighbor. In: Barolli L, Hussain FK, Ikeda M, editors. Com-
[Link]/10.1198/073500102753410444. plex, Intelligent, and Software Intensive Systems. Cham: Springer Inter-
[61] Harvey D, Leybourne S, Newbold P. Testing the equality of prediction national Publishing; 2020, p. 521–533.
mean squared errors. International Journal of Forecasting 1997;13(2):281 [80] Brusaferri A, Fagiano L, Matteucci M, Vitali A. Day ahead elec-
– 291. [Link] tricity price forecast by NARX model with LASSO based features
[62] Varma S, Simon R. Bias in error estimation when using cross-validation selection. IEEE International Conference on Industrial Informat-
for model selection. BMC Bioinformatics 2006;7:91. https:// ics (INDIN) 2019;2019-July:1051–1056. [Link]
[Link]/10.1186/1471-2105-7-91. INDIN41052.2019.8972263.
[63] Goodfellow I, Bengio Y, Courville A. Deep Learning. The MIT Press; [81] Leerbeck K, Bacher P, Junker R, Goranović G, Corradi O, Ebrahimy R,
2016. et al. Short-term forecasting of CO2 emission intensity in power grids
[64] McHugh C, Coleman S, Kerr D, McGlynn D. Daily energy price fore- by machine learning. Applied Energy 2020;277. [Link]
casting using a polynomial narmax model. In: Lotfi A, Bouchachia H, 10.1016/[Link].2020.115527.
Gegov A, Langensiepen C, McGinnity M, editors. Advances in Compu- [82] Zareipour H, Canizares CA, Bhattacharya K. Economic impact of
tational Intelligence Systems. Cham: Springer International Publishing. electricity market price forecasting errors: A demand-side analysis.
ISBN 978-3-319-97982-3; 2019, p. 71–82. IEEE Transactions on Power Systems 2010;25(1):254–262. https:
[65] Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. //[Link]/10.1109/TPWRS.2009.2030380.
From local explanations to global understanding with explainable ai for [83] Uniejewski B, Nowotarski J, Weron R. Automated variable selection and
trees. Nature Machine Intelligence 2020;2(1):56–67. 10.1038/s42256- shrinkage for day-ahead electricity price forecasting. Energies 2016;9(8).
019-0138-9. [Link]
[66] Janzing D, Minorics L, Blöbaum P. Feature relevance quantification in [84] Kaminski V. Energy markets. Risk Book; 2013.
explainable ai: A causal problem. 2019. arXiv:1910.13413; available [85] Li W, Paraschiv F. Modelling the evolution of wind and solar power
from: [Link] infeed forecasts. Journal of Commodity Markets 2021;:100189https:
[67] Sundararajan M, Najmi A. The many shapley values for model explana- //[Link]/10.1016/[Link].2021.100189.
tion. 2020. arXiv:1908.08474; available from: [Link] [86] Herceg S, Željka Ujević Andrijić, Bolf N. Development of soft sen-
abs/1908.08474. sors for isomerization process based on support vector machine re-
[68] Lundberg S, Lee SI. A unified approach to interpreting model predictions. gression and dynamic polynomial models. Chemical Engineering Re-
2017. arXiv:1705.07874; available from: [Link] search and Design 2019;149:95 – 103. [Link]
1705.07874. [Link].2019.06.034.
[69] Hanchuan Peng, Fuhui Long, Ding C. Feature selection based on mu-
tual information criteria of max-dependency, max-relevance, and min-
redundancy. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence 2005;27(8):1226–1238. 10.1109/TPAMI.2005.159.
[70] Yu L, Liu H. Efficient feature selection via analysis of relevance and
redundancy. J. Mach. Learn. Res. 2004;5:1205–1224.
[71] Langley P. Selection of relevant features in machine learning. In: In Pro-
ceedings of the AAAI Fall symposium on relevance. AAAI Press; 1994,
p. 140–144.
[72] Jamian JJ, Abdullah MN, Mokhlis H, Mustafa MW, Bakar AHA. Global
particle swarm optimization for high dimension numerical functions
analysis. Journal of Applied Mathematics 2014;[Link]
10.1155/2014/329193.
[73] Somu N, Raman M R G, Ramamritham K. A deep learning framework for
building energy consumption forecast. Renewable and Sustainable En-
ergy Reviews 2021;137:110591. [Link]
science/article/pii/S1364032120308753. [Link]
10.1016/[Link].2020.110591.
[74] Shao Z, Yang S, Gao F, Zhou K, Lin P. A new electricity
price prediction strategy using mutual information-based svm-

18