Decomposable Transformer with Inter-series Dependencies and Intra-Series Temporal Modeling for Multi-Horizon Photovoltaic Power Forecasting

Accurately predicting photovoltaic (PV) power is essential for improving grid stability, enabling grid operators to more effectively manage power resources, and reducing the operational cost of the grid system. However, PV power generation is subject to intermittency, volatility, and uncertainty due to weather conditions like irradiance, rainfall, humidity, and temperature, which cause relatively large errors in traditional prediction methods. Consequently, accurate forecasting of PV power is gaining increasing attention from academia and industry. In this work, inspired by the architectures of PatchTST and iTransformer, we propose our model, named TSformer, based on trend-seasonal decomposition. TSformer is a Transformer encoder-only model, which models trend and seasonal parts of time series separately based on their distinct variation characteristics. TSformer effectively learns the inter-series dependencies and the intra-series temporal dynamics. An experimental analysis conducted on two solar sites from the State Grid Corporation of China shows that TSformer offers higher prediction accuracy and enhanced robustness compared to other baseline models, indicating the effectiveness of TSformer in photovoltaic power forecasting.