Enhancing Training Dynamics of Long-Short Term Memory-Based Models for Accurate Photovoltaic Power Generation Forecasting

Abstract

Accurate forecasting of solar photovoltaic (PV) power generation is essential for reliable grid integration and efficient energy management. To achieve this, the most existing studies improve forecasting performance of deep learning models by modifying their neural network architectures, however this study focuses instead on optimizing the training process of Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) models. Three learning rate decay strategies-linear decay, exponential decay, and cosine annealing-are employed during training and compared with baseline models trained using a fixed learning rate. The forecasting performance of all models is evaluated on unseen test data using the mean absolute error (MAE), root mean square error (RMSE), coefficient of determination $(\mathbf{R}^{\mathbf{2}})$, and mean absolute percentage error (MAPE). Experimental results demonstrate that cosine annealing consistently yields the best forecasting performance across all evaluation metrics, indicating that an appropriate learning rate adaptation strategy can significantly enhance PV power forecasting without altering the underlying neural network architecture.