Three Hour Ahead PV Power Forecasting with Bidirectional Recurrent Networks: Insights into Monthly Variability

Abstract

Despite extensive research into PV power forecast models, their monthly performance is rarely thoroughly examined, creating gaps in our understanding of their accuracy and applicability across different times of the year. This paper focuses on evaluating the monthly performance of four predictive models designed for a three-hour forecast horizon. The studied models are based on distinct architectures of Recurrent Neural Networks (RNNs): Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), and Bidirectional GRU (Bi-GRU). By analyzing statistical indicators across various months, we uncover seasonal fluctuations, with performance hitting its lowest points in winter (November and December) and peaking in summer (June to August). Notably, Bi-GRU consistently outperforms the other models, displaying lower error rates and higher accuracy across diverse months. As a result, it emerges as the preferred choice for forecasting due to its superior predictive capability. Additionally, we observe variations in daily  performance across seasons, highlighting the complexities of data sequences and underscoring the importance of careful model selection. This research significantly contributes to the advancement of predictive modeling in timeseries analysis, providing valuable insights into model performance and seasonal variations, and equipping practitioners and researchers with enhanced methodologies for improving forecast accuracy.