A comparative study of machine learning approaches for an accurate predictive modeling of solar energy generation

Abstract

Solar energy prediction poses a challenging task that necessitates robust models and precise data to accurately forecast solar energy yield, especially in grid areas with high photovoltaic shares. Existing methods often rely on statistical or physical models, which have limitations in capturing the complex and non-linear relationships between weather variables and solar power generation. In this paper, we address this issue by comparing and evaluating different learning models, ranging from artificial neural networks (ANNs) and random forest models to long- and short-term memory (LSTM) networks, to predict the PV energy yield based on weather forecast data. A methodology has been developed to evaluate various models using real-world datasets from a large-scale industrial solar project, incorporating historical photovoltaic data, meteorological data, and solar irradiation data. The experimental results showed that the Random Forest Algorithm (RFR) consistently outperforms other algorithms, providing a mean absolute error (MAE) of 0.06 and a root mean square error (RMSE) of 0.15 when applied to historical meteorological datasets. The accuracy of the learning model was improved by combining meteorological data with a solar irradiation dataset to obtain an MAE of 0.03 % and an RMSE of 0.09 %. Validation analysis showed the proposed model is effective in both forecast accuracy and stability. The proposed methodology can provide valuable information to PV system operators, grid managers, and energy planners, facilitating the optimization of solar energy resources. • Address the challenging task of predicting solar energy generation with a focus on accuracy using machine learning models. • Comparative study of machine learning models, including ANN, LSTM, KNN, and RFR. • Evaluation based on real-world datasets from a large-scale industrial solar project in Saudi Arabia. Datasets. • RFR consistently outperforms other algorithms with a mean absolute error (MAE) of 0.06 % and a root mean square error (RMSE) of 0.15 %. • Improved accuracy by combining meteorological data with a solar irradiation dataset (MAE of 0.03 %, RMSE of 0.09 %).

Topics & Concepts