Journal of Physics:
Conference Series
PAPER • OPEN ACCESS You may also like
- Third-order resonance correction using
Research on Ultra-short-term combination new Trim-S system in J-PARC MR
Y Tan, Y Morita, M Yoshii et al.
forecasting algorithm of power load based on - Assessment of creep rupture properties of
martensitic heat resistant steel
machine learning Yuan Lu, Ke Yun, Ming Yang et al.
- Advances in muometric navigation system
To cite this article: Jinggeng Gao et al 2024 J. Phys.: Conf. Ser. 2846 012046 and its use in underwater localization
Mailun Chen, Guo Wei, Chunfeng Gao et
al.
View the article online for updates and enhancements.
This content was downloaded from IP address [Link] on 28/12/2025 at 12:27
2024 2nd International Conference on Power, Grid and Energy Storage IOP Publishing
Journal of Physics: Conference Series 2846 (2024) 012046 doi:10.1088/1742-6596/2846/1/012046
Research on Ultra-short-term combination forecasting
algorithm of power load based on machine learning
Jinggeng Gao1,2, Kun Wang1, *, Xiaohua Kang2, Huan Li1, Qi Hou1, Gaixia Kang1
and Jiaxu Zhou2
1
State Grid Gansu Electric Power Company Marketing Service Center, Lanzhou,
730030, China
2
State Grid Gansu Electric Power Research Institute, Lanzhou, 730030, China
*
Corresponding author’s e-mail: gs_jlzx@[Link]
Abstract: Power load forecasting is of great significance to the power grid marketing
department. To obtain accurate load forecasting results, a minute-by-minute forecasting
method for electricity load based on multi-stage is proposed (TPE-WXL) by combining the
non-linear and time-series attributes. Firstly, the historical series of specific areas in the city are
pre-processed. Then, in order to obtain accurately predicted results, XGBoost and LightGBM
are applied to extract attributes from the series to build a hybrid model. Moreover, TPE is
introduced to enhance the hyperparameters of the hybrid model and the attributes of the load
series. Finally, the power load dataset of the region is used as an example to conduct
experimental analysis. Experimental results revealed that the method can accurately forecast
the trend of power load, that is, R2=0.981, RMSE=2.643.
1. Introduction
As the core application of power grid marketing, power load forecasting has been widely concerned
[1-3]. Accurate minute-by-minute power demand estimation is beneficial to realize stable dispatching
and pre-management of the power grid. To address this challenge, machine learning is used for
electricity demand forecasting. This method does not need accurate physical models and prior
knowledge of experts and has broad prospects in the area of electricity load forecasting. This method
reduces the model’s intricacy and helps the model to take the autocorrelation effect of the predicted
objects [4-6]. Jiang et al. [7] introduced the characteristics of load variation and combined the
characteristics of power load and date. On this basis, they completed the feature extraction through
TCN and the estimation of minute-by-minute power demand by using GRU. The results show that
analysis of the attributes of power load is helpful in enhancing the precision of model prediction.
Zhang et al. [8] introduced an attention model to LSTM to further highlight the input attributes that
play a key role in load prediction. The power load prediction model of LSTM-Attention is constructed.
The effective information from a vast number of power system data collection can be mined better to
enhance the precision of electricity load forecasting. The above research shows that power load
forecasting mainly includes data input and models. Emphasis is placed on the derivation of artificial
features and the construction of a prediction model. The influence of power load sequence on the
model is not deeply considered in the attributes. Although the power load trend can be predicted
accurately, the spatial angle and generalization ability are limited due to the observation data of a
Content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
2024 2nd International Conference on Power, Grid and Energy Storage IOP Publishing
Journal of Physics: Conference Series 2846 (2024) 012046 doi:10.1088/1742-6596/2846/1/012046
single model. In addition, contrasted with machine learning, deep learning is limited in its
interpretability [9]. This impedes trust in the model in power marketing decisions.
Covering minute-by-minute power load prediction, ensemble learning has the characteristics of fast
computing speed and good interpretability [10]. Therefore, looking at the minute-by-minute forecast
of the power load of the power marketing department, this study proposes a minute-by-minute forecast
method of electricity load based on TPE-WXL.
2. Proposed Algorithm: TPE-WXL
In this study, the time step and the weights of XGBoost and LightGBM are taken as the optimization
variables of TPE based on the data on users' historical usage characteristics. XGBoost and LightGBM
are applied to fully extract the characteristics of the load data. The optimization objective is to
maximize , and the influence of different time step information and weight calculation are
considered. Finally, the ultra-short-term power load can be accurately predicted. The field data set
collected by an electric power marketing department is used for simulation to confirm the
effectiveness of the method.
The flowchart of the minute-by-minute power load prediction method based on TPE-WXL is
shown in Figure 1. The process consists of three segments: constructing the dataset, beginner learner
feature mining, and Bayesian feature optimization.
Figure 1. Flow chart of marketing power load forecasting model.
3. Experimental analysis
3.1. Performance metrics
In this paper, we utilize R2, RMSE, and MAPE as metrics for performance evaluation, with their
respective calculations provided subsequently.
m
2 i <1
( zi' , zi )2
R =1- (1)
m
i <1
( zi , z i ) 2
1 m '
RMSE < ( zi , z i )
m i <1
(2)
2
2024 2nd International Conference on Power, Grid and Energy Storage IOP Publishing
Journal of Physics: Conference Series 2846 (2024) 012046 doi:10.1088/1742-6596/2846/1/012046
1 m zi' , zi
MAPE < |
m i <1 zi
| ≥100% (3)
The experiments in this study were conducted in the Anaconda. The running hardware environment
of the program is shown in Table 1.
Table 1. Program running hardware environment.
Hardware Configuration Type
Central Processor Chip i7-12700H
Graphics Processing Unit GTX-3070ti 8GB
Random Access Memory 16 GB
3.2. Experimental result
To research the efficiency of the TPE-WXL algorithm, this study obtained the real power load data of
a region, which included daily records from January to September 2022 (sampling period of 15 min,
unit MW). The obtained data is divided into training sets and test sets, with the first seven months as
the training set and the last month as the test set.
Table 2. Results of evaluation indicators.
Evaluation Index
Model
R2 RMSE MAPE
model1 0.981 2.643 0.159
model2 0.968 3.949 0.163
model3 0.972 2.802 0.166
model4 0.961 3.891 0.165
model5 0.963 3.693 0.164
Table 2 depicts the evaluation index results for a variety of models. The model1 to model5,
respectively, in Table 2: TPE - WXL LightGBM, XGBoost, CNN, LSTM. It can be seen from this
table that the R2 of TPE-WXL is 0.981. Among the algorithms used in this paper, XGBoost has the
best performance, 0.972. The R2 of TPE-WXL is 0.72% higher than that of XGBoost, and the RMSE is
5.81% lower than that of XGBoost. The outcomes demonstrate that the algorithm not only preserves
the accuracy of the predictions but also markedly boosts the dependability of electricity load
forecasting.
Table 3. Value range of some parameters.
Parameter Name Parameter Range Tuning Results
eta (0.05,4.05) 0.35
Learning_rate (0.05,4.05) 0.4
booster (gbtree, dart) gbtree
num_boost_round (50,200) 120
Time_step (10,1000) 640
max_depth (2,30) 2
subsample (0.1,1) 1.0
rate_drop (0.1,0.1) 0.2
w1 0.601
w1+ w2=1
w2 0.399
Bayesian Optimization is an algorithm used to optimize black-box functions, especially when the
input space is high-dimensional, the function evaluation is expensive, or there are multiple locally
optimal solutions. Bayesian optimization is based on a probabilistic model that captures the unknown
properties of the function and guides the optimization by constantly updating it. Contrasted with mesh
search, Bayesian optimization has a simpler process and faster operation of model parameters, which
effectively improves the speed of hyperparameter tuning. An optimization technique is used to
3
2024 2nd International Conference on Power, Grid and Energy Storage IOP Publishing
Journal of Physics: Conference Series 2846 (2024) 012046 doi:10.1088/1742-6596/2846/1/012046
optimize the hyperparameters of the WXL algorithm to obtain the best time step, weight vector, and
hyperparameter value. Table 3 shows some hyperparameter adjustment ranges and results.
4. Conclusion
To enhance the estimation performance of minute-by-minute power consumption, considering the
information of different series, a method of minute-by-minute power load prediction founded on TPE-
WXL is proposed for the pre-management and resource optimization in power marketing. Through
experimental simulation and result analysis, the following conclusions are drawn:
(1) Various machine learning methods selected in this paper can achieve good forecasting results
for short-term power load data with nonlinear characteristics after a reasonable selection of model
parameters. Therefore, machine learning technology is suitable for solving very short-term electric
load forecasting problems.
(2) The hybrid model built on XGBoost and LightGBM is optimized by TPE, which enhances the
predicted results contrasted with other prediction models. Compared with LightGBM, XGBoost, CNN,
and LSTM, the very short-term load forecasting effect of the proposed TPE-WXL algorithm is the best.
Acknowledgments
The research presented here was funded by the Science and Technology Project of State Grid Gansu
Electric Power Company (Project Number 522731220001).
References
[1] Eren, Y and Küçükdemiral, İ. 2024 A comprehensive review on deep learning approaches for
short-term load forecasting. RENEW SUST ENERG REV. 189 114031.
[2] Wang L, Qu Y, Wang S, Liu Y, Yang R, Ren X, and Ma Y. 2024 Optimal allocation of customer
energy storage based on power big data and improved LSTM load forecasting. ENERGY
REP. 11 3902-3913.
[3] Shi Z, Zheng R, Shen R, Yang D, Wang G, Liu, Y and Zhao J. 2024 Building heating load
forecasting based on the theory of transient heat transfer and deep learning. ENERG
BUILDINGS. 313 114290.
[4] Wang X, Wang H, Bhandari, B and Cheng L. 2024 AI-empowered methods for smart energy
consumption: A review of load forecasting, anomaly detection and demand response. INT J
PR ENG MAN-GT. 11(3) 963-993.
[5] Mao Q, Wang L, Long Y, Han L, Wang Z and Chen K. 2024 A blockchain-based framework for
federated learning with privacy preservation in power load forecasting. KNOWLEDGE-
BASED SYSTEMS. 284 111338.
[6] Cao Z, Wang J, Xia, Y. 2024 Combined electricity load-forecasting system based on weighted
fuzzy time series and deep neural networks. ENG APPL ARTIF INTEL. 132 108375.
[7] Jiang B, Yang H, Wang Y, Liu Y, Geng H, Zeng H, and Ding J. 2024 Dynamic temporal
dependency model for multiple steps ahead short-term load forecasting of power
system. IEEE T IND APPL. 11(4) 11-20.
[8] Zhang L, Huang Z, Wang Z, Li X, and Sun F. 2024 An urban charging load forecasting model
based on trip chain model for private passenger electric vehicles cased study in
Beijing. ENERGY. 299 130844.
[9] Lotfipoor A, Patidar S, and Jenkins, D P 2024 Deep neural network with empirical mode
decomposition and Bayesian optimisation for residential load forecasting. EXPERT SYST
APPL. 237 121355.
[10] Yang Y, Xing Q, Wang K, Li C, Wang J, and Huang X. 2024 A novel combined probabilistic
load forecasting system integrating hybrid quantile regression and knee improved multi-
objective optimization strategy. APPL ENERG. 356 122341.