Enhancing Solar Power Generation Forecasting using CLDNN with Lag and Rolling Window Features

This study explores advanced methods for forecasting solar power generation. We propose a hybrid deep learning model, the Convolutional Long Short-Term Memory Deep Neural Network (CLDNN), which combines the spatial capabilities of Convolutional Neural Networks (CNNs) with the temporal processing strengths of Long Short-Term Memory (LSTM) networks. Our research compares the performance of CNN, LSTM, and CLDNN models with and without feature engineering techniques, namely lag features and rolling window calculations. Additionally, we split the test set data into four seasons: spring, Summer, Autumn, and Winter. This seasonal approach captures variability in solar generation and provides insights into each model’s performance under different conditions. We evaluate the models using metrics like Root Mean Square Error (RMSE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>). Our results show that the CLDNN model outperforms CNN and LSTM in accuracy, particularly when enhanced with feature engineering. The CLDNN with feature engineering achieved the lowest error rates (RMSE: 0.0187) and the highest R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value (0.9915). The findings highlight the potential of combining hybrid architectures with advanced preprocessing techniques to improve solar power forecasting, contributing to more efficient energy management in VPPs and the stability of renewable energy grids.