Rethinking the use of deep learning methods for photovoltaic power forecasting

Abstract

Abstract Accurate photovoltaic power forecasting is critical for grid stability, but remains challenged by weather uncertainties and difficulty integrating historical observations with forward-looking forecasts. We re-evaluate existing architectural choices in deep-learning models for time-series forecasting and demonstrate the importance of full encoder-decoder architectures and channel dependence modeling when both weather forecasts and historical data are available. Based on this insight, we propose Cross-Unet, a Transformer-based architecture featuring multi-scale temporal encoding, correlation-aware channel attention, and hierarchical cross-attention decoding to fuse historical generation data with weather forecasts. Evaluated on open-source datasets from four utility-scale plants in northern China and one aggregated plant in central Australia, using three types of forward-looking inputs: numerical weather prediction, satellite-derived irradiance, and AI-based weather model forecasts. Across the majority of evaluated configurations spanning five photovoltaic power stations, five forecasting horizons (4 hours to 7 days), and three forecast sources, Cross-Unet outperforms ten deep learning baselines and traditional operational benchmarks. By integrating advanced forecasting systems, such as modern AI weather models, into an end-to-end forecasting pipeline, Cross-Unet enables operational 15-minute-resolution predictions over 4-hour to 7-day horizons, supporting grid scheduling and energy trading.