AutoPV: An intelligent framework for automated design of photovoltaic forecasting models

Photovoltaic power forecasting (PVPF) is essential for the efficient integration of solar energy into power systems. Despite recent advances in machine learning and deep learning, designing optimal forecasting models for specific PVPF tasks remains challenging due to the varying performance of models across different scenarios and the need for domain expertise in architecture selection. While neural architecture search (NAS) has shown potential in automating model design in fields such as computer vision, its application in time series forecasting (TSF) is limited, largely due to the structural diversity of TSF models. This study proposes a modular approach to TSF model design and constructs a tailored search space specifically for PVPF tasks. Based on this search space, we develop AutoPV, an automated model design framework that constructs optimal forecasting architectures by considering task type and data distribution. Experiments on a real-world PV dataset with 8856 hourly samples and 11 features, covering 12 forecasting tasks across two scenarios (with and without future weather reports) and horizons from half-day to one-month ahead, show that AutoPV achieves the lowest MAE on most tasks and reduces average forecasting error by roughly 5–20% compared with strong deep learning and Transformer-based baselines, while keeping the searched architectures compact and deployment-friendly.