Combination Approach of LSTM and CNN in Solar Energy Production Prediction

Introduction: The growing demand for cleaner energy alternatives has led to a significant increase in solar photovoltaic (PV) installations. However, the integration of solar energy into the grid remains challenging due to the inherent variability of solar power, influenced by factors such as irradiance, temperature, and wind speed. Accurate forecasting of PV power production is essential to improve grid stability and optimize energy dispatch. Objectives: This study aims to develop and evaluate a hybrid deep learning model that combines Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) to accurately forecast solar energy production. The specific objectives include comparing the forecasting performance of the CNN-LSTM model, preprocessing the dataset, and assessing the model's predictive accuracy using key performance metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-Squared (R²). Methods: This research uses time-series data from 2019 to 2022, including PV production data and environmental factors such as irradiance, temperature, wind speed, and humidity. The data is preprocessed using suitable techniques, and the model architecture incorporates CNN for spatial feature extraction and LSTM for capturing temporal dependencies. The model is trained using a sliding window algorithm and evaluated using performance metrics like MAE, RMSE, and R². Results: The CNN-LSTM hybrid model demonstrated exceptional performance in predicting solar energy production. Over various time forecasting periods, including 1 month, 3 months, and 1 year, the model achieved near-perfect accuracy. Specifically, it recorded 99.99% accuracy for 1-month predictions, 100% accuracy for 3-month predictions, and maintained 99.99% accuracy over a 1-year period. The model showed an excellent R² score of 1 and achieved minimal Mean Absolute Error (MAE) values: 191.92 Wh for 1 month, 2.31 Wh for 3 months, and 3,191.72 Wh for 1 year. These results demonstrate the model's robust ability to accurately capture both short-term and long-term fluctuations in solar power production, outperforming traditional forecasting models in terms of accuracy and reliability. Conclusions: The combination of LSTM and CNN in the hybrid model has proven highly effective for predicting solar energy production. The approach successfully combines the spatial feature extraction strength of CNN and the temporal dependency capturing capability of LSTM to provide accurate forecasts. This model outperforms traditional methods, offering superior prediction accuracy for both short-term and long-term solar energy production. Further research could focus on refining the model through hyperparameter tuning, integrating additional environmental factors, and applying it across different geographical regions to enhance its generalizability.