Selecting effective NWP integration approaches for PV power forecasting with deep learning

Accurate forecasting of photovoltaic (PV) power is crucial for reliable energy scheduling and system operation. While deep learning models have demonstrated strong capabilities in this domain, effectively integrating numerical weather prediction (NWP) data into such models remains a challenging problem. In this study, we propose and systematically evaluate five distinct NWP integration strategies — referred to as Method 1 through Method 5 — for enhancing PV forecasting performance. These methods are tested across 14 representative models and four forecasting horizons (4, 24, 72, and 144 steps), covering short-, mid-, and long-term scenarios. Experimental results reveal that the effectiveness of each integration method depends on the model architecture and forecasting horizon. In particular, Method 5 shows strong compatibility with recurrent models such as LSTM in short-term forecasting, while Method 4 performs best with Transformer-based models in long-term settings. Additionally, Method 1 and Method 2 demonstrate consistently reliable performance across various models and tasks. These findings provide practical insights into selecting suitable NWP integration strategies for PV power forecasting applications. • Five strategies are proposed for integrating NWP data into photovoltaic forecasting models. • Unified experiments on 14 models reveal horizon- and architecture-dependent benefits. • Guidelines support robust model selection for reliable PV forecasting.