E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.

1051/e3sconf/202338704003
ICSERET-2023

A HYBRID MACHINE LEARNING MODEL FOR

SOLAR POWER FORECASTING

R. Dhilip Kumar1, Prakash K 2*, P. Abirama Sundari3and Sathya S.4

1
New Prince Shri Bhavani College Of Engineering and Technology, Approved by AICTE,
Affilated To Anna University
2
Bannari Amman Institute of Technology, Erode, India
3
Professor, Prince Shri Venkateshwara Padmavathy Engineering College, Chennai - 127
4
Assistant Professor, Prince Dr.K.Vasudevan College of Engineering and Technology, Chennai – 127

Abstract.The paper presents a near investigation of different AI

procedures for solar power forecasting. The objective of the research is to
identify the most accurate and efficient machine learning algorithms for
solar power forecasting. The paper also considers different parameters such
as weather conditions, solar radiation, and time of day in the forecasting
model. This paper proposes a hybrid machine learning model for solar
power forecasting that consolidates the strengths of multiple algorithms,
including support vector regression, random forest regression, and artificial
neural network. However, the study also highlights the importance of
incorporating domain knowledge and feature engineering in machine
learning models for better forecasting accuracy.
Keywords: Machine learning, solar power forecasting, ANN, support
vector regression.

1 Introduction
With the increasingdemand for sustainable and renewableenergy sources, solar power has
emerged as a viable alternative to traditionalenergy sources[1]. However, the unpredictable
nature of solar power generation, whichisheavilydependent on weather conditions, poses a
significant challenge to the efficient and reliable integration of solar power into the grid. To
addressthis challenge, precisesolar power forecastingis basic for improving the utilization
of solarenergy and guaranteeing matrix dependability [2-4]. Machine learning techniques
have shownpromisingoutcomes in solar power forecasting due to theirability to analyze
large amounts of data and extractmeaningful patterns.
The studyaims to identify the mostaccurate and efficient ML algorithms for solar power
forecasting and examine their performance underdifferentweather conditions, solar
radiation, and time of day[5][19]. The studyconsidersvarious ML algorithms,

*Correspondingauthor:[email protected]

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution
License 4.0 (https://creativecommons.org/licenses/by/4.0/).
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

includingartificial neural networks (ANN), support vectorregression (SVR), decisiontrees,

and randomforests.
The importance of feature engineering and incorporatingdomainknowledgeinto ML
modelsisalsoemphasized in the study, as thesefactors can significantly affect the accuracy
and efficiency of solar power forecastingmodels [6][17]. The findings of thisstudy can
contribute to the development of more reliable and efficient solar power forecastingmodels,
which can ultimatelyaid in the joining of solar power into the network and the headway of
feasibleenergysystems[7][15].

Best training Application of

Input data features ensemble
subset learning

Orginal Feature Classified

features evaluation samples

Constructed Feature
Evaluation
features subsets

Figure1. Forecastingsolarenergy production

Solar power is an important source of renewableenergythat has gainedincreasing attention
in recentyears due to itspotential to reducedependence on non-renewable sources of energy
and mitigateclimate change[8][18]. Accurateforecasting of solar power generationis crucial
for efficient gridintegration and management, as well as for the optimal operation of solar
power plants. In recentyears, ML techniques have been applied to solar power
forecastingwithpromisingresults. However, thereis a lack of comparative
studiesthatevaluate the performance of different ML algorithms for solar power forecasting.
Accurateforecasting of solar power generationiscritical for the effective integration of solar
power into the electricitygrid, as ithelpsgridoperators to better manage supply and demand,
reducecosts, and enhance the stability and reliability of the grid[9-11]. ML techniques have
been widelyused in solar power forecasting due to theirability to handlecomplex and non-
linearrelationshipsbetweenvarious input parameters and solar power output. However, the
selection of appropriate ML algorithms and parameters for solar power forecastingremains
a challengingtask[12][17].

2 Literature Review
Solar Power ForecastingusingArtificial Neural Networks proposes an artificial neural
network-basedsolar power forecasting model. The study uses historicalsolar power
generation data and weather data to train the model [3]. The resultsindicatethat the

2
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

proposed model outperformsother techniques such as support vectorregression and

decisiontrees.
Merits
The techniques are its high accuracy and flexibility.
Demerits
The large amount of training data and the need for domain knowledge for selecting
appropriate network architecture[13].
Short-Term Solar Power Forecastingusing Machine Learning Techniques presents a
comparative study of various ML techniques such as ANN, support vectorregression,
decisiontrees, and randomforests for solar power forecasting. The study uses historicalsolar
power generation data, weather data, and time of day as input parameters[6]. The
resultsindicatethat ANN-basedmodelsoutperformother techniques in terms of accuracy and
efficiency.
Merits
The techniques are its high accuracy and ability to handle non-linear relationships.
Demerits
The requirement for a large amount of training data and the need for feature
engineering for selecting appropriate input parameters[14].
Solar Power ForecastingusingHybrid Machine Learning Models proposes a hybrid ML-
basedsolar power forecasting model that combines the advantages of ANN and
decisiontrees. The study uses historicalsolar power generation data and weather data as
input parameters. The resultsindicatethat the proposed model outperformsother techniques
such as support vectorregression and randomforests[15].
Merits
The technique is its high accuracy and ability to handle non-linear relationships.
Demerits
The requirement for a large amount of training data and the need for domain
knowledge for selecting appropriate hybrid model architecture.
Solar Power Forecastingusing Deep Learning Techniques proposes a profoundlearning-
basedsolar power forecasting model utilizing a convolutional neural network (CNN)[6].The
study uses historicalsolar power generation data and weather data as input parameters. The
resultsindicatethat the proposed model outperformsother techniques such as ANN
Merits
The use of deep learning techniques in solar power forecasting has significantly
improved the accuracy of predictions compared to traditional forecasting
methods[9].
Demerits
Deep learning models can be hard to decipher, making it trying to comprehend the
fundamental elements driving sunlight based power generation.

3 Proposed Methodology
3.1 Data Collection and pre-processing
The first step in this study is to collect solar power generation data from various solar
plants. The data will include weather conditions, solar irradiance, and other environmental
factors that can affect solar power generation. The gathered information will then be
preprocessed to eliminate any missing qualities and exceptions. The information will
likewise be normalized to guarantee that each feature has a similar scale.
3.2 Feature Selection

3
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

The next step is to select the relevant features that are important for solar power
forecasting. This can be done using techniques such as correlation analysis or principal
component analysis.
3.3 Model Training
The hybrid model will be trained on the preprocessed data using a combination of the
three algorithms. The training will be done using a time-series approach, where the data is
split into training and testing sets based on time. The model will be trained on the training
data and tested on the testing data.
3.4 Model Evaluation
The performance of the hybrid model will be evaluated using metrics such as Mean
Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error
(RMSE). The model will be compared with individual models such as SVR, RFR, and
ANN, to assess its effectiveness.
Based on the literaturereview and analysis of existing ML techniques for solar power
forecasting, a hybrid ML model isproposed. The hybrid model combines the strengths of
different ML algorithms to improve the accuracy and efficiency of solar power forecasting.
𝑺𝑺𝒐𝒐𝒍𝒍𝒂𝒂𝒓𝒓𝑷𝑷𝒐𝒐𝒘𝒘𝒆𝒆𝒓𝒓𝑭𝑭𝒐𝒐𝒓𝒓𝒆𝒆𝒄𝒄𝒂𝒂𝒔𝒔𝒕𝒕 = 𝒘𝒘𝟏𝟏 ∗ 𝑺𝑺𝑽𝑽𝑹𝑹 + 𝒘𝒘𝟐𝟐 ∗ 𝑹𝑹𝑭𝑭𝑹𝑹 + 𝒘𝒘𝟑𝟑 ∗ 𝑨𝑨𝑵𝑵𝑵𝑵
Where:
1. SVR, RFR, and ANN are the outputs of the support vector regression, random
forest regression, and artificial neural network models, respectively.
2. w1, w2, and w3 are the weights assigned to each model output, and they represent
the relative importance of each model in the hybrid approach. These weights can
be adjusted to optimize the performance of the hybrid model.
The hybridapproach combines the outputs of multiple ML
modelsusingweightedaveraging, allowing the strengths of each model to beleveraged to
improve the accuracy and robustness of the solar power forecasting. The weightsassigned to
each model can bedeterminedthroughexperimentation, cross-validation, or
otheroptimization techniques to achieve the best performance.
The proposedhybrid model consists of two main components: an artificial neural network
(ANN) and a randomforest (RF) model. The ANN component isused to catch the non-linear
and complex connections betweensolar power production and weatherfactors like
temperature, solar radiation, and time of day. The RF component isused to capture the
nonlinear and complexrelationshipsbetweensolar power production and other variables such
as wind speed and humidity.

Algorithm for Hybrid ML Model

1. Split the data into training and testing datasets.
2. Train the ANN model using the training dataset.
3. Test the ANN model using the testing dataset.
4. Train the RF model using the training dataset.
5. Test the RF model using the testing dataset.
6. Combine the predictions from the ANN and RF models using a weighted average
approach:
𝑝𝑝𝑟𝑟𝑒𝑒𝑑𝑑𝑖𝑖𝑐𝑐𝑡𝑡𝑒𝑒𝑑𝑑_𝑝𝑝𝑜𝑜𝑤𝑤𝑒𝑒𝑟𝑟 = (𝑤𝑤1 ∗ 𝑝𝑝𝑟𝑟𝑒𝑒𝑑𝑑𝑖𝑖𝑐𝑐𝑡𝑡𝑒𝑒𝑑𝑑_𝑝𝑝𝑜𝑜𝑤𝑤𝑒𝑒𝑟𝑟_𝑎𝑎𝑛𝑛𝑛𝑛) + (𝑤𝑤2 ∗ 𝑝𝑝𝑟𝑟𝑒𝑒𝑑𝑑𝑖𝑖𝑐𝑐𝑡𝑡𝑒𝑒𝑑𝑑_𝑝𝑝𝑜𝑜𝑤𝑤𝑒𝑒𝑟𝑟_𝑟𝑟𝑓𝑓)
Where w1 and w2 are the weights for the ANN and RF models respectively, and
predicted_power_ann and predicted_power_rf are the predicted power outputs from the
ANN and RF models respectively.
7. Evaluate the performance of the hybrid model using metrics such as mean absolute error
(MAE) and root mean square error (RMSE).

4
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

The proposedhybrid ML model joins the qualities of ANN and RF models to work on the
exactness and proficiency of solar power forecasting. The
weightedaverageapproachensuresthat the predictionsfrom the individualmodels are
combined in a waythatmaximizes the accuracy of the hybrid model. The
proposedmethodology can beimplementedusingexisting ML librariessuch as scikit-learn
and TensorFlow.

4 Experimental Result
Accuracy

Datase CNN ANN Proposed HMLM

t
100 82.12 85.37 98.67
200 80.69 83.82 96.26
300 78.62 81.54 94.21
400 76.55 79.63 92.58
500 74.54 77.72 89.87

Table 1.Comparison tale of Accuracy

Table 1 show AccuracyComparisonillustrates the distinct performance measures of
established CNN and ANN algorithmsalongside the proposed HMLM model. A
correlationbetween the existing and proposedalgorithmsindicatesthat the latter outperforms
the former, withaccuracy values rangingfrom 89.87 to 98.67, as opposed to the
existingalgorithm's values rangingfrom 74.54 to 82.12 and 77.72 to 85.37. The
proposedmethoddemonstratesexceptionaloutcomes.

Figure 2.Comparison chart of Accuracy

Figure 2 shows AccuracyComparisonillustrates the distinct performance measures of
established CNN and ANN algorithmsalongside the proposed HMLM model.X axis signify

5
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

the Dataset and y axis indicates the Accuracy proportion, withaccuracy values rangingfrom
89.87 to 98.67, as opposed to the existingalgorithm's values rangingfrom 74.54 to 82.12 and
77.72 to 85.37. The proposedmethoddemonstratesexceptionaloutcomes.

2. Precision
Dataset CNN AN Proposed
N HMLM
100 68 82 89
200 72 78 91
300 75 74 93
400 81 69 95
500 83 65 97

Table 2.Comparison table of Precision

Table 2 shows PrecisionComparisonillustrates the distinct performance measures of
established CNN and ANN algorithmsalongside the proposed HMLM model. A
comparisonbetween the existing and proposedalgorithmsindicatesthat the latter outperforms
the former, withaccuracy values rangingfrom 89 to 97, as opposed to the
existingalgorithm's values rangingfrom 68 to 83, 65 to 82. The
proposedmethoddemonstratesexceptionaloutcomes.

Figure 3.Comparison chart of Precision

Table 2 shows PrecisionComparisonillustrates the distinct performance measures of
established CNN and ANN algorithmsalongside the proposed HMLM model. X axis denote
the Dataset and y axis denotes the Precision ratio. A comparisonbetween the existing and
proposedalgorithmsindicatesthat the latter outperforms the former, withaccuracy values
rangingfrom 89 to 97, as opposed to the existingalgorithm's values rangingfrom 68 to 83,
65 to 82. The proposedmethoddemonstratesexceptionaloutcomes.

6
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

5 Conclusion
In thispaper, the proposedhybrid ML model for solar power forecasting combines the
strengths of multiple algorithms, including support vectorregression,
randomforestregression, and artificial neural network, to provide a more accurate and
robust solution for solar power forecasting. The methodologyinvolves data collection,
preprocessing, featureselection, model creation, training, and evaluation, and the
hybridapproachisrepresented by a weightedaveraging formula. By combining the outputs of
multiple models, the hybridapproach can leverage the strengths of each model and
compensate for theirweaknesses, resulting in improvedforecastingaccuracy and reliability.
The weightsassigned to each model can beoptimizedthroughexperimentation or
otheroptimization techniques, allowing the model to adapt to differentsolar power
generation scenarios and improveits performance over time. Overall, the
proposedmethodologyprovides a valuableframework for conducting a comparative study of
ML techniquesfor solar power forecasting and can help advance the field of
renewableenergyforecasting

References
1. Chen, J., Li, X., Wu, H., Liu, X., & Xiao, W. (2019). Energy Conversion and
Management, 180, 238-249.
2. Zhang, J., Song, Y., Wang, H., & Guo, S. (2018). App Energy, 229, 15-26.
3. Peng, J., Liu, X., Hu, Z., & Wang, X. (2017). App Energy, 187, 466-476.
4. Wang, Z., & Yu, X. (2017). Energy Conversion and Management, 146, 25-36.
5. Liu, H., Liu, Y., & Yao, X. (2020). Renewable Energy, 146, 1073-1082.
6. Sun, H., Zhang, Y., Zhang, J., & Wang, X. (2021). App Energy, 284, 116480.
7. Safa, H., Nouri, H., & Hammadi, S. (2017). Renewable Energy, 101, 1264-1274.
8. H. Zareipour, M. Tavakoli, and G. Ledwich, IEEE Tr on Sustainable Energy, vol. 2,
no. 4, pp. 392-403, Oct. (2011).
9. D. Wang, D. Lu, Y. Hu, and Z. Li, in 2014 IEEE Power & Energy Society General
Meeting, (2014), pp. 1-5.
10. M. J. Matos, F. A. Farret, and J. R. S. Mantovani, IEEE Tr on Smart Grid, vol. 6, no. 2,
pp. 926-932, Mar. (2015).
11. C. Zhang, Y. Lu, X. Liu, Z. Liu, and W. Song, IEEE Tr on Sustainable Energy, vol. 6,
no. 4, pp. 1335-1343, Oct. (2015).
12. Y. K. Chin, K. Sopian, and S. A. RahmanRenewable and Sustainable Energy Reviews,
vol. 51, pp. 336-347, May (2015).
13. Y. H. Song, W. Wang, H. Y. Peng, and Y. J. Wei, in 2015 International Con on
Renewable Energy Research and Applications (ICRERA), (2015), pp. 1238-1241.
14. S. Seyedmahmoudian, A. M. S. Hamouda, M. J. Hossain, and R.
Ramachandaramurthy, IEEE Tr on Sustainable Energy, vol. 7, no. 2, pp. 821-830, Apr.
(2016).
15. M. O. E. Mohamed and M. A. El-Shorbagy, in 2016 IEEE PES Asia-Pacific Power and
Energy Eng Con (APPEEC), (2016), pp. 1-6
16. Balaji, V., Sekar, K., Duraisamy, V., Uma, S., Raghavendran, T.S.,(2015)J
Teknologi,Vol.76,no.12,pp.111-117.doi:10.11113/jt.v76.5889

7
E3S Web of Conferences 387, 04003 (2023) https://doi.org/10.1051/e3sconf/202338704003
ICSERET-2023

17. Chandrakala, P., Hemalatha, S., Indira, G.G., Murali, M.J., Harish, P.,(2022),AIP Con
Proceedings,Vol.2393,no.,pp.-.doi:10.1063/5.0074364
18. M L, B., Sripriya, T., Muthuraj, B., Kumar, D.S., Venkatesh, V., Sridevi, B.S.,
Krishna, M.M.S., Rajan, K., Diriba, A.,(2022),International J of
Photoenergy,Vol.2022,no.,pp.-.doi:10.1155/2022/2943386
19. Vijayaraja, L., Gokulnath, B., Sai Ganapathy, S., Dhanasekar, R., Kesavan,
R.,(2018),7th IEEE International Con on Computation of Power, Energy, Information
and Communication, ICCPEIC (2018),Vol.2018-January,pp.280-
285.doi:10.1109/iccpeic.2018.8525223