Photovoltaic power forecasting with PG-TiDE: Physically guided time-series dense encoder

With a rapid increase on the integration of intermittent renewable energy sources, accurate photovoltaic power forecasting has become increasingly essential for guaranteeing the secure operation of power systems and minimizing the solar energy curtailment. Therefore, this paper proposes a novel photovoltaic power forecasting (PVPF) method. First, the covariates involved in PVPF are categorized into three types: static covariates, known future dynamic covariates, and unknown future dynamic covariates. For the dynamic covariates, joint mutual information is employed to select informative features for the forecasting target. These three types of covariates are then fed into the time-series dense encoder (TiDE) model. TiDE represents a recent advancement in time series forecasting and adopts a specialized MLP-based architecture that balances accuracy and efficiency. Moreover, to address the issue of negative outputs often encountered in PVPF, a physical constraint is incorporated into the TiDE model to guide it toward non-negative predictions. The proposed approach is verified through real-world PV power generation data. Experimental results indicate that: (1) the TiDE model significantly outperforms the baseline models, thus reducing the forecasting errors to a large extent; (2) the physical constraint effectively mitigates the problem of negative outputs; and (3) the inclusion of covariates provides valuable information that improves forecasting accuracy as well.