Ultra-Short-Term Photovoltaic Power Forecasting Via a Hybrid Physics-Guided and Data-Driven Model With Dynamic NWP Correction

Abstract

The integration of photovoltaic (PV) power into grids is hampered by its intermittency, necessitating highly accurate ultra-short-term forecasting. The quality of numerical weather prediction (NWP) data is a critical bottleneck. Current approaches often treat NWP as static inputs or apply offline corrections, failing to address systematic biases in real-time. Furthermore, the lack of robust physical priors and the misalignment between symmetric loss functions and asymmetric market penalties limit both accuracy and economic utility. To address these challenges, this paper proposes PGD3-Net, a novel hybrid physics-guided and data-driven model with dynamic NWP correction for ultra-short-term PV power forecasting. PGD3-Net synergistically integrates three key innovations: 1) a physics-guided feature enhancement (PGFE) module that encodes solar altitude, azimuth, and precise sunrise-sunset offsets to provide strong astronomical priors; 2) a dynamic NWP correction (D-NWP-C) module that employs multi-scale parallel convolutions and channel attention to adaptively refine future NWP sequences online; and 3) an adaptive asymmetric L1 loss that imposes a heavier penalty on over-estimation errors, aligning model training with grid operation economics. Evaluated on real-world data across four seasons, PGD3-Net significantly outperforms state-of-the-art baseline methods. Compared to the optimal baseline, it reduces the MAE by up to 23.26% and the RMSE by up to 9.72%, while improving the R2 by up to 2.83% across different seasons, demonstrating its effectiveness in leveraging physical knowledge and correcting meteorological uncertainties to deliver reliable and economically aligned forecasts.