A high-precision photovoltaic power forecasting model leveraging low-fidelity data through decoupled informer with multi-moment guidance

Accurate power generation forecasting for distributed photovoltaic (PV) systems is essential for the grid with increased distributed PV penetration. This task depends on high-fidelity historical and forecast weather data, but obtaining such data is challenging. This paper proposes a decoupled Informer with multi-moment guidance (DMGformer), using real-world low-fidelity historical and forecast weather data for day-ahead hourly distributed PV power forecasting. Specifically, the framework employs a decoupled history-forecast (DHF) structure where the encoder exclusively captures long-term historical meteorological and power generation dependencies, while the decoder uses forecast data and historical insights to predict future power. Additionally, the multi-moment guidance (MMG) module is designed to introduce domain knowledge that multiple corresponding moments from historical data can contribute to the power forecasting of a future moment in the short term. To evaluate the feasibility and effectiveness of the model, we construct a real-world dataset of 500 sites, containing hourly power generation and low-fidelity historical and forecast weather data. The results highlight the impressive performance of the proposed DMGformer, achieving a 24.11 % reduction in Mean Absolute Error (MAE) and a 1.46 % improvement in accuracy compared to the suboptimal Informer. Furthermore, the DHF and MMG effectively enhance the performance of LSTM (Long Short-Term Memory), Transformer, and Informer models, validating the generalizability of these two paradigms. The DMGformer exhibits superior efficiency in utilizing low-fidelity meteorological data to achieve precise power generation forecasting, especially for distributed PV plants, which facilitates optimized resource allocation for sustainable energy production.