INDIAN JOURNAL OF SCIENCE AND TECHNOLOGY

RESEARCH ARTICLE

An Efficient Short-Term Solar Power

Forecasting by Hybrid WOA-Based LSTM
Model in Integrated Energy System
OPEN ACCESS Amit Kumar Mittal1 ∗ , Kirti Mathur2
Received: 18-08-2023 1 Assistant Professor, Computer Engineering, IET- DAVV, Indore, Madhya Pradesh, India
Accepted: 05-01-2024 2 Associate Professor, IIPS-DAVV, Indore, Madhya Pradesh, India
Published: 23-01-2024

Abstract
Citation: Mittal AK, Mathur K (2024) Objectives: Due to the irregular nature of sun irradiation and other
An Efficient Short-Term Solar Power meteorological conditions, solar power generation is constantly loaded with
Forecasting by Hybrid WOA-Based
LSTM Model in Integrated Energy
risks. When solar radiation data isn’t captured and sky imaging equipment
System. Indian Journal of Science isn’t available, improving forecasting becomes a more difficult endeavor.
and Technology 17(5): 397-408. http So our objective to improve the forecasting accuracy for next year solar
s://[Link]/10.17485/IJST/v17i5.2020
power generation data. Methods: Our research used a real numerical
∗
Corresponding author. solar power dataset of Australia and Germany and a standard approach
amittal@[Link] for preprocessing. The feature selection in this research uses the Whale
Funding: None Optimization Algorithm (WOA). A Long Short-Term Memory (LSTM) method
is utilized to determine the accuracy of solar power forecasts. The HHO
Competing Interests: None
(Harris Hawks Optimization) technique is also used to improve solar power
Copyright: © 2024 Mittal & Mathur.
forecasting accuracy. The performances were analyzed and the proposed
This is an open access article
distributed under the terms of the method is employed in the python platform. Findings: The findings show that
Creative Commons Attribution the suggested technique considerably increases the accuracy of short-term
License, which permits unrestricted
use, distribution, and reproduction
solar power forecasts for proposed method is 3.07 in comparison of LSTM
in any medium, provided the and SVM at different data types and 15 min and 60 min interval. Novelty: The
original author and source are key novelties of this research is hybrid strategy for improving the precision
credited.
of solar power forecasting for short periods of time. Including the Whale
Published By Indian Society for Optimization Algorithm (WOA), Long Short-Term Memory (LSTM), and Harris
Education and Environment (iSee)
Hawks Optimization (HHO).
ISSN
Keywords: Power generation; Solar power forecasting; Whale optimization
Print: 0974-6846
Electronic: 0974-5645 algorithm; Long ShortTerm Memory; Harris hawk’s optimization

1 Introduction
Nowadays, solar power generation has increased significantly compared to other
types of renewable energy. Natural variations, on the other hand, make solar energy
production unstable (1) . When it comes to bringing this unexpected renewable energy
into the system, solar electricity forecasting is a major challenge (2) . The more volume
of integration, the more volatile the system gets. If this unstable energy is not efficiently
managed, problems with power supply reserve and frequency management will arise.
To address these difficulties, very accurate forecasting far in advance will allow us to

[Link] 397
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

offer enough electricity to meet current demand while maintaining frequency stability in the system (3) . A multitude of factors,
including the algorithm utilized, influence the forecasting model’s accuracy (4) .
In most cases, a deep learning algorithm is used to convert weather data into power in solar power forecasting models (5) .
Because solar power sources are found in a range of locations and have varied capacities, small generating sources frequently
do not have weather surveillance systems because it does not cost viable. A major roadblock to solar power being connected
to the grid, as well as a power sector prejudice against PV system deployment (6) . As a result, precise short-term forecasting
and enhancing the accuracy of solar power forecasts are critical for successfully integrating PV systems into the grid (7) . This
gets more challenging if no previous data on solar radiation was acquired and there is no equipment is available for specific sky
imaging (8,9) .
To enhance solar power forecasting accuracy, the HHO (Harris Hawks Optimization) technique is applied (10) . HHO is a
natural-inspired optimization technique based on Harris Hawk behavior modeling. The heart of the algorithm is built on hawks
cooperating to catch their prey. This strategy is used by a group of Harris hawks to attack the target from diverse directions,
catching it off guard. The accuracy of our proposed work is compared to that of other processes to show that it is superior to
those techniques, and the results of the comparison are analyzed.
Liu et al., (4) have presented A simplified LSTM neural network for one-day-ahead solar power forecasting. This work
introduces a simpler LSTM algorithm for forecasting one-day-ahead solar power generation based on the architecture of the
Machine Learning technique. Under varied weather circumstances, the LSTM model’s forecast can successfully capture intra-
hour ramping.
Wang et al., (11) have presented a Photovoltaic power forecasting-based LSTM-Convolutional Network. This research
proposes and uses a hybrid deep learning model (LSTM-Convolutional Network) to anticipate renewable energy. LSTM-
convolutional network technology is used. The proposed hybrid types outperform shows the Convolutional-LSTM Network
and the hybrid prediction model outperforms the solar prediction model.
Yu et al., (12) . Have discussed An LSTM short-term solar irradiance forecasting under complicated weather conditions. An
LSTM-based approach is employed in this work to produce short-term forecasts using a timeline that incorporates global
horizontal irradiance (GHI) one hour and one day ahead of time. The long short-term memory (LSTM) network is the deep
structure of RNN. Low loss and pollution feature significantly ease the energy-environment conflict, making it a vital component
of the future energy system.
Wang et al., (13) . A day-ahead PV power forecasting technique based on the LSTM-RNN model has been provided, as well
as temporal correlation modification in a framework for forecasting daily. Under partial daily basis prediction, a day-ahead PV
power prediction approach was devised, based on the LSTM-RNN model and temporal correlation modification. Under the
suggested PDPP architecture, the forecasting model’s output was further improved for those days with correct daily pattern
predictions if time correlation alteration (TCM) is more exact than the specific LSTM-RNN model.
Li et al., (14) . Deep learning with a hybrid component has been presented for solar forecasting in the short term. The wavelet
packet degradation process (WPD) & LSTM networks are used in this study to create a hybrid deep learning model. The original
solar power series is broken down into sub-series using WPD. Then, for each of these sub-series, four distinct LSTM networks
are built. Traditional solar power forecasting methods can’t provide the level of precision necessary for energy systems. Some
researchers are using deep learning algorithms to enhance the accuracy of projections for solar power or other keys elements
to consider when forecasting
Abdel-Nasser et al., (5) . Exact solar power estimating models depend on deep learning algorithms such as LSTM-RNN- which
have been reported. This study proposes the need for a long short-term memory recurrent neural network (LSTM-RNN) to
correctly predict the output power for PV systems. When LSTM is used instead of other approaches, the predicting error is
reduced even more. The proposed forecasting method could be useful for smart grid planning and control.
Hossain et al., (15) . We’ve talked about using an LSTM neural network to forecast short-term photovoltaic electricity and
creating a synthetic weather forecast. This research proposes a forecasting strategy for photovoltaic (PV) power generation
utilizing the LSTM neural network (NN). To verify the superiority of the LSTM NN with the proposed features, other deep
learning engines such as the RNN, generalized regression neural network, and extreme learning machine are investigated.
Aslam et al., (16) . A two-stage focus on LSTM and Bayesian optimization have been presented for day-ahead solar energy
forecasting. The proposed deep-learning technique for forecast PV power for the next day is based on a 2-stage mechanism of
attention on the LSTM model. Forecasting for the next day, the proposed models were tested to state-of-the-art methods like
LSTM-Attention, CNN-LSTM, and Ensemble models. Some faults in the work will need to be addressed in future research.
Over LSTM, there are two layers of attention in the suggested model, each having decoder-encoder layers. In comparison to
existing models, this technique used many layers and parameters to be learned.

[Link] 398
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

Zhou et al., (17) . Short-term PV power forecasting was demonstrated utilizing a long short-term memory neural network
and an attention mechanism. A time-series hybrid ensemble deep learning framework for predicting short-term PV power
generation is presented in this paper. The temperature is predicted by one LSTM neural network, while the power output is
predicted by the other. Predicting data is flattened & mixed with such a layer that is fully connected to increase forecasting
accuracy.
This study has the following gaps that are identified and to fill this gap-
• To provide the short term solar power forecasting model based on effective feature selection method that is not used in
the literature.
• In our finding Short term PV power that is 15 min and 60 min interval is used which is not used in this study.
• We used a three year solar power datasets to next year power forecasting.
• The HHO (Harris Hawks Optimization) technique is also used to improve solar power forecasting accuracy.

2 Methodology
This study used a standardized approach for preprocessing a real-numerical solar power datasets. Feature selection in this
research uses the Whale Optimization Algorithm. The enhanced LSTM method is used to determine the accuracy of solar
forecasts shown in Figure 1. HHO (Harris Hawks Optimization) technique is also used to improve solar power forecast accuracy.

2.1 Data Collection

The Austria and Germany countries data sets are utilized in the solar power forecasting. Each of these countries is taken across
two time periods, with the first period lasting 15 minutes and the second lasting 60 minutes. Using the two nations mentioned
above, the same time frame.

Fig 1. Solar power forecasting in the short-term using a hybrid WOA-based LSTM model

2.2 Whale Optimization Algorithm

The proposed method’s inspiration is initially explored in this section. Then there’s the mathematical model.

2.2.1 Mathematical model and optimization algorithm

This section introduces a mathematical model for surrounding prey, spiral bubble-net feeding technique, & prey search.
Following that, the WOA algorithm is suggested.

2.2.2 Encircling prey

Humpback whales are capable of detecting prey & encircling it. A WOA technique implies that the current best candidate
solution would be the goal target as well as extremely close to it since the position of an optimum design within the search
space is unknown. Those specific search agents can attempt to enhance overall positions in comparison to other search agents,

[Link] 399
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

as determined by the best solution found. This behavior is represented by the equations below:

−
→= − − ∗ (t) − −
→.→ → (t) (1)
R Q u u

→
− (t + 1) =−
→ ∗ (t) − −
→.−
→ (2)
u u P R

Where,
→
− →
−
t is the current iteration, P and Q coefficient vectors,
u’ is the position vector of the best solution identified,
→
− indicates the position vector,
u
| | is the exact value and is a Multiplication of elements one by one It’s worth noting that u’ must be changed in every cycle
if a better option is discovered. The vectors P and Q are calculated as follows:
P=2→
− .→
− −−
→ (3)
i r i

Q = 2.r (4)
where →
− is reduced linearly from two to zero throughout iterations (in both exploration and exploitation phases) and →
− is a
d r
random vector in (0,1).
The reasoning behind Equation (2) for a 2D situation is shown in Figure 2 (2a). A search agent’s location (u, v) depends
on the location, it’s possible to change the current perfect recorded (u’, v’). Changing the values of the variables −
→ and −
→
P Q
vectors For the current position, there are many areas throughout the country where the best agents can be found. Figure 2
also depicts a search agent’s possible updated position in 3D space in Figure 2 (2b). It should be noted that by defining the
random vector → − , any position in the search space between the key points depicted in Figure 2 can be reached. An agent
r
used for searching might update his position inside the region of both the current best answer and the next best solution as a
consequence of the Equation (2), emulating the behavior of nearby prey. In a search area with n-dimensional space, the search
agents will roam in hypercubes all around the best solution identified.

2.2.3 Exploitation phase (Bubble-net attacking method)

There have been two methodologies created to statistically measure humpback whale bubble-net behavior:
1. Encircling mechanism shrinking: In Equation (3), accomplished by reducing the range of I. It’s important to note that
i reduce the rate of ((p)) volatility. Where, −
→ denoted the random number between [-i, i] with i dropping from 2 to 0 with each
P
repetition. The new location of the search agent can be found somewhere between the beginning place and the current best
agents using random integers for −→ in [-1, 1]. In a 2D space, Figure 2(3a) displays the possible positions that 0≤ P ≤ 1 can
P
accomplish from (u, v) to (u’, v’).
2. Spiral updating position: This approach begins with computing the distance here between the whale (u, v) and the prey
(u’, v’), as illustrated in Figure 2(3b). To emulate the A spiral equation is formed between both the position of the whale and the
place of the prey due to the helical structure movement of humpback whales:
→
− (t + 1) =−
→′
. ebl . cos2π l + −
→′
(t) (5)
u R u

where −
→′
= −
→′
(t) − −
→ (t) and denotes a distance between ith whale as well as the prey (currently the best answer), b is
R u u
indeed a constant used to define the form of a logarithmic spiral, l is a random number in [1,1], and is an element-by-element
multiplication.
While swimming, grey whales swim in a lowering circle around their food in a spiral pattern. We assume there’s a
50/50 chance of either employing the shrinking encircling mechanism or the spiral model to update whale locations during
optimization to simulate this simultaneous behavior. The following is the mathematical model:
{ →
− → −
→
−u ∗ (t) − P . R
→
−u (t + 1) = →
−′ bl
i f P < 0.5
(6)
→
− ′
R .e . cos 2π l + u (t) i f P ≥ 0.5

[Link] 400
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

p indicates random integer between [0,1].

Humpback whales hunt for food at random times, in addition to employing bubble nets. The following is the mathematical
model for the search.

Fig 2. 2 D and 3 D position and Bubble net method

2.2.4 Exploration phase (Search for prey)

A similar strategy that involves changing the B vector can be used to locate prey. In reality, based on their relative locations,
humpback whales seek at random. As a consequence, we use B to move the search agent away from the reference whale by using
random values larger than or equal to 1. Rather than employing the best search agent identified so far, we utilize a randomly
generated search agent that can adjust the position of a search agent throughout the exploration phase. That strategy, combined

[Link] 401
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

with the fact that |B| > 1, The mathematical model looks like this:

−
→= −
→ . −−−→ − −
→ (7)
R Q urand u

→
− (t+) = −−−→ − −
→.−
→ (8)
u urand P R

Where −−−→ denotes randomly generated position vector.

urand
In Figure 3, some of the probable sites surrounding a solution with B>1 are shown.

Fig 3. WOA has an exploration mechanism (’u’ is a search agent chosen at random)

To make searching and attacking easier, a parameter is decreased from II to 0. When |P| > 1, a random search agent is
chosen, however since | −
→ |< 1 is selected, and the good option for updating the search agents’ positions is chosen. WOA may
P
flip between spirals and a round movement depending on the value of p. Lastly when a termination condition is met, the WOA
algorithm is terminated.
Because it includes exploration/exploitation capabilities, WOA may theoretically be called a global optimizer. In addition, the
proposed hypercube approach establishes a search region around the best answer, allowing additional search agents to use the
current better details within that region. The WOA algorithm can quickly transition between exploration and exploitation due
to adaptive changes in the search vector A: by reducing P, certain iterations are given to exploration (|P| 1) and the remainder
to exploitation (|P| 1). It’s odd that WOA only has two internal settings that can be adjusted (P and Q).
Although we might have included mutations and other evolution in the WOA formulations to completely reproduce
humpback whale behavior, we elected to simplify the WOA algorithm by reducing the amount of heuristics and internal
parameters. On the other hand, hybridization utilizing evolutionary search techniques might be the subject of future research.

2.3 LSTM
The LSTM network is a form of RNN that combines representation learning and model training without the requirement for
extra domain knowledge.
LSTM is specifically developed to address the problem of gradient vanishing, which makes it difficult to retain the short-
and long-term correlation between vectors. We also looked at the influence of single parameter optimization on the proposed

[Link] 402
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

approach and found that only learning rate optimization had little impact on the proposed LSTM’s performance. The suggested
LSTM’s overall performance improves when the learning rate, momentum rate, and dropout rate are all optimized together.
Here, the weight of the LSTM is obtained by HHO.

2.4 Harris Hawks Optimization (HHO)

This technique appears to be a metaheuristic optimization method. It replicates Harris Hawks’ co-operative ”surprise pounce”
behavior Exploitation and Exploration levels are present in the HHO strategy, as they are in other metaheuristic algorithms.
HHO is a population-based optimization method that does not use gradients. As a result, when properly formulated, it can be
used to resolve every optimization issue. In the HHO algorithm, exploration have2 stages, and exploitation have 4 stages. There
are two exploration stages and four exploitation steps in the HHO algorithm. This cooperative behavior’s mathematical model
also presents a novel stochastic technique for tackling a number of optimization issues. In the following section, to suggest a
DVR control scheme, the HHO approach is applied.

Fig 4. HHO in various phases

We simulate the proposed HHO’s exploring and exploitative stages in this part, which are based on Harris hawk prey
investigation, surprise pounce, as well as a variety of attack methods. HHO can solve an optimization issue with the appropriate
formulation because it is a population-based, gradient-free optimization strategy. Figure 4 depicts all HHO phases, which have
been described in detail in the subsequent.

2.4.1 Exploration Phase:

Throughout this stage, HHO roams around in random places, looking for prey utilizing one of two methods. To alter the location
of each hawk, an equation is utilized (Equation (9)). The Ki1, Kp1, Kp2, & Ki2 PI controller parameters are the ”prey” in our
scenario, although the ”hawks” are the variety of recommended search agents.
{ }
(yPREY (L) − yM (L)) −C3 (lb +C4 (ub − lb)), k < 0.5
y(L + 1) = (9)
yRAND (L) −C1 |yRAND (L) − 2C2 y(L)|, k ≥ 0.5

Where:
y (L + 1) is the hawks’ next iteration’s location vector,
yPREY (T ) is the prey’s location (Kp1, Ki1, Kp2, and Ki2),
K, c1, c2, c3, c at each cycle, are the random integers inside (0,1)

[Link] 403
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

y (L) Hawks’ current location vector is c1, c2, c3, c4,

yRAND (L) is a hawk chosen at random from the present population,
lb, ub The upper and bottom bounds of the variable, signifying the lowest and highest values, are denoted by Kp1, Ki1, Ki2,
and Ki2.
The hawks arrive at an average location using the Equation (10),

1 n
yM (L) = ∑ yI (L)
n I=1
(10)

where:
yM (L) hawks’ average starting position,
yI (L) each hawk’s location on iteration ‘t’,
n denotes the number of hawks (n=10 here due to the multitude of search engines).

2.4.2 Exploration To Exploitation Transition:

( )
T
e = 2eo 1 − (11)
t

Because the prey tries to run, the change between exploitation and discovery, and the transition from searching to attacking
takes place. The victim expends a great deal of energy striving to escape. Equation (11) simulates the prey’s energy equation:
where:
e denotes the prey’s fleeing energy, t the maximum number of repetitions, andeo the energy’s beginning condition.

2.4.3 Exploitation Phase:

[Link] Soft Besiege:. Before swooping down on the victim, the HHO softly circles it to exhaust it. This action is explained
by Equations (12) and (13).

y (L + 1) = △y (L) − e |kyPREY (L) − y(L)| (12)

△y (L) = yPREY (T ) − y(L) (13)

Where,
In iteration t, △y (T ) represents the difference between the current location of hawks and prey.
K: the strength of the prey when bouncing randomly during the escape is known as k=2 (1-c5). To imitate the natural behavior
of prey, this value fluctuates at random during each cycle.
R: is the escape prey possibility.

[Link] Hard Besiege:. In this situation, the victim is unable to depart due to exhaustion. As a result, hawks have an easier
time catching and pinning their prey. Each hawk takes use of its existing position to improve its circumstances (Equation (14)).

y (T + 1) = yPREY (T ) − e |△y(T )| (14)

[Link] Progressive Rapid Dives In a Soft Besiege. Imagine that hawks can use the following rule to analyze (decide)
their next action.

h = yPREY (L) − e |KyPREY (L) − y(T )| (15)

The HHO technique uses the LF (Lévy Flight) principle can build the differential equation to mimic the prey’s zigzag motion
when attempting to elude. There under the LF principle, hawks should dive for their prey Equation (16). Equations (17) and (18)
LF function.

g = h + s × l f (d) (16)

[Link] 404
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

U× ∝
l f (X) = 0.01 ×
1 (17)
|V | δ

 1
 ( ) δ
 r(1 + δ ) × sin πδ 
 
 2  (18)
α =  
 δ −1 
 ( )  
 1+δ 2 
r ×δ ×2
2

Where, d is denoted as the problem’s dimensionality is denoted a 1×d random vector, if denoted a levy flight function.

[Link] Progressive Rapid Dives In a Hard Besiege. By minimizing the distance between the average location as well
as the prey location, team members’ whereabouts are revealed.
{ }
h IF f (h) < f (y(L))
y(L + 1) = (19)
g IF f (g) < f (y(L))

If h and g were calculated using the additional criteria Equations (20) and (21):

h = yPREY (t) − e |KyPREY (T ) − yM (T )| (20)

g = h + s × l f (d) (21)

Where, yM (T )is obtained from Equation (10).

3 Result and Discussion

The performance metrics of RMSE and MAE are acceptable. Three approaches are compared: the suggested method, the
traditional LSTM, and the SVM.
Analysis of solar power in sunny day V/S time is shown in Figure 5 for the survey of solar power sunny time series 15 mints
data- Austria. The LSTM and SVM in this condition are actual, hypothetical, and conventional. In comparison to other 800 MW
time 10:00:00, conventional LSTM is at its peak value. Comparing actual data to other data, it is in a low range. Over 600MW
is the suggested value.

Fig 5. Solar power Sunny time series (15 mints data- Austria)

[Link] 405
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

Fig 6. Solar power Sunny time series (60 mints data- Austria)

Analysis of solar power in sunny day V/S time is shown in Figure 6 for the survey of solar power sunny time series 60 mints
data for Austria. Comparing actual, proposed, conventional LSTM and SVM under this condition. Proposed is the peak value
when compared to others who use more than 25000 MW at 11:06:40. Actual data are in the low to mid-teens compared to other
data of 20,000 MW.
Analysis of solar power in sunny day V/S time is shown in Figure 7 for the survey of solar power sunny time series 15 mints
data from Germany. In this condition, the proposed, actual, conventional LSTM and SVM are compared. When compared to
others, the conventional LSTM reaches its peak value at 11:06:40. Comparing actual data to other data above 20,000, it is low
range. The suggested amount is 25,000 MW.

Fig 7. Solar power Sunny time series (15 mints data- Germany)

Figure 8 illustrates an analysis of solar power on a sunny day versus time for 60 minutes of data from Germany. The LSTM
and SVM in this condition are actual, hypothetical, and conventional. Compared to other peaks over 30000 MW, the proposed
peak value is 11:06:40. Traditional LSTM and SVM have a limited range when compared to other data above 5000MW time
05:33:20.

Fig 8. Solar power Sunny time series (60 mints data- Germany)

[Link] 406
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

Fig 9. Solar power time series in future prediction

Figure 9 displays the analysis of solar power V/S date-time for the survey of solar power time series in future prediction. The
starting solar power range for the year 2017 is 700MW, the midyear solar power range is slightly higher than 700MW, and the
year-end solar power range is 800MW. The solar power range for 2018 is 800MW, with an additional 800MW for the middle
and end of the year. The solar power range for the year 2019 starts above 800 MW, increases to above 1200 MW in the middle
of the year, and decreases to above 1000 MW at the end of the year. Analysis of 1200MW of solar power feature predictions for
2020.
Table 1. The comparative analysis of the proposed model and the existing model
Country Data type Processing type Metrics Proposed Conventional SVM
LSTM
Austria 15 mins Training process RMSE (MW) 3.07 15.87 259.95
MAE (MW) 6.62 10.28 172.07
Testing process RMSE (MW) 3.24 15.86 262.63
MAE (MW) 7.03 9.7 168.81
60 mins Training process RMSE (MW) 5.61 19.86 259.87
MAE (MW) 18.64 20.52 191.52
Testing process RMSE (MW) 5.87 17.36 260.12
MAE (MW) 19.63 22.79 187.33
Germany 15 mins Training process RMSE (MW) 13.53 212.66 8212.8
MAE (MW) 120.39 144.82 4892.79
Testing process RMSE (MW) 13.26 211.96 7403.08
MAE (MW) 115.89 141.71 4210.66
60 mins Training process RMSE (MW) 21.18 520.64 8217.02
MAE (MW) 269.52 332.2 4833.61
Testing process RMSE (MW) 21.86 535.22 7356.05
MAE (MW) 279.85 342.43 4109.6

4 Conclusion
In this study, we describe a hybrid technique for enhancing the accuracy of solar power forecasts over short periods. We used
a genuine numerical solar power dataset and a conventional pre-processing method for our study. The Whale Optimization
Algorithm is used to pick features in this study (WOA). The accuracy of solar power estimates is determined using an LSTM
(Long Short-Term Memory) approach. The HHO (Harris Hawks Optimization) method is also employed to increase the
accuracy of solar power forecasts. The findings imply that the suggested method considerably enhances the accuracy of short-
term solar power estimates. Results were examined, and the recommended method was implemented in Python. Solar energy
in sunny day versus time solar energy in sunny time series for 15 minutes – Austria, in this situation, the LSTM and SVM are
at their highest values when compared to other 800 MW times 10:00:00. Proposed, traditional LSTM and SVM for Austrian
solar power sunny time series data for 60 mints. When compared to other users who use more than 25000 MW, proposed is
the peak value at 11:06:40. When compared to other data of 20,000 MW, the actual data are in the low to mid-terms. data from
Austria for 15 minutes comparing solar power on cloudy days to time. To compare to others who use more than 400 MW, the
suggested peak time for conventional LSTM and SVM in this situation is 16:40. The conventional LSTM reaches its maximum
value at 05:33:20 for 60 minutes of Austrian data on solar power in cloudy days.

[Link] 407
Mittal & Mathur / Indian Journal of Science and Technology 2024;17(5):397–408

German data for a 15-min period on solar power in sunny days versus time. At 11:06:40, the conventional LSTM reaches its
maximum value. Solar energy on a cloudy day Traditional LSTM and SVM are at their peak when compared to others above
12000 MW during the survey of solar power over a 60-min period in Germany. 1200MW of solar power feature predictions
for 2020 are based on solar power VS date-time solar power time series. This is a lot less than bench marking errors. Future
research will look at how well the suggested method predicts additional renewable energy sources, like the amount of electricity
generated by wind farms.

References
1) Liu ZF, Luo SF, Tseng ML, Liu HM, Li L, Mashud AHM. Short-term photovoltaic power prediction on modal reconstruction: A novel hybrid model
approach. Sustainable Energy Technologies and Assessments. 2021;45:101048. Available from: [Link]
2) Jalali SMJ, Khodayar M, Ahmadian S, Shafie-Khah M, Khosravi A, Islam SMS, et al. A New Ensemble Reinforcement Learning Strategy for Solar Irradiance
Forecasting using Deep Optimized Convolutional Neural Network Models. In: 2021 International Conference on Smart Energy Systems and Technologies
(SEST). IEEE. 2021;p. 1–6. Available from: [Link]
3) Wang F, Xuan Z, Zhen Z, Li K, Wang T, Shi M. A day-ahead PV power forecasting method based on LSTM-RNN model and time correlation modification
under partial daily pattern prediction framework. Energy Conversion and Management. 2020;212:112766. Available from: [Link]
enconman.2020.112766.
4) Liu CH, Gu JC, Yang MT. A Simplified LSTM Neural Networks for One Day-Ahead Solar Power Forecasting. IEEE Access. 2021;9:17174–17195. Available
from: [Link]
5) Abdel-Nasser M, Mahmoud K. Accurate photovoltaic power forecasting models using deep LSTM-RNN. Neural Computing and Applications.
2019;31:2727–2740. Available from: [Link]
6) Qu J, Qian Z, Pei Y. Day-ahead hourly photovoltaic power forecasting using attention-based CNN-LSTM neural network embedded with multiple relevant
and target variables prediction pattern. Energy. 2021;232:120996. Available from: [Link]
7) Agga A, Abbou A, Labbadi M, Houm YE. Short-term self consumption PV plant power production forecasts based on hybrid CNN-LSTM, ConvLSTM
models. Renewable Energy. 2021;177:101–112. Available from: [Link]
8) Jung Y, Jung J, Kim B, Han S. Long short-term memory recurrent neural network for modeling temporal patterns in long-term power forecasting for solar
PV facilities: Case study of South Korea. Journal of Cleaner Production. 2020;250:119476. Available from: [Link]
9) Chen B, Lin P, Lai Y, Cheng S, Chen Z, Wu L. Very-Short-Term Power Prediction for PV Power Plants Using a Simple and Effective RCC-LSTM Model
Based on Short Term Multivariate Historical Datasets. Electronics. 2020;9(2):1–19. Available from: [Link]
10) Wang Y, Feng B, Hua QSS, Sun L. Short-Term Solar Power Forecasting: A Combined Long Short-Term Memory and Gaussian Process Regression Method.
Sustainability. 2021;13(7):1–16. Available from: [Link]
11) Wang K, Qi X, Liu H. Photovoltaic power forecasting based LSTM-Convolutional Network. Energy. 2019;189:116225. Available from: [Link]
1016/[Link].2019.116225.
12) Yu Y, Cao J, Zhu J. An LSTM Short-Term Solar Irradiance Forecasting Under Complicated Weather Conditions. IEEE Access. 2019;7:145651–145666.
Available from: [Link]
13) Wang F, Xuan Z, Zhen Z, Li K, Wang T, Shi M. A day-ahead PV power forecasting method based on LSTM-RNN model and time correlation modification
under partial daily pattern prediction framework. Energy Conversion and Management. 2020;212:112766. Available from: [Link]
enconman.2020.112766.
14) Li P, Zhou K, Lu X, Yang S. A hybrid deep learning model for short-term PV power forecasting. Applied Energy. 2020;259:114216. Available from:
[Link]
15) Hossain MS, Mahmood H. Short-Term Photovoltaic Power Forecasting Using an LSTM Neural Network and Synthetic Weather Forecast. IEEE Access.
2020;8:172524–172533. Available from: [Link]
16) Aslam M, Lee SJ, Khang SH, Hong S. Two-Stage Attention Over LSTM With Bayesian Optimization for Day-Ahead Solar Power Forecasting. IEEE Access.
2021;9:107387–107398. Available from: [Link]
17) Zhou H, Zhang Y, Yang L, Liu Q, Yan K, Du Y. Short-Term Photovoltaic Power Forecasting Based on Long Short Term Memory Neural Network and
Attention Mechanism. IEEE Access. 2019;7:78063–78074. Available from: [Link]

[Link] 408