Numerical weather prediction (NWP) global horizontal irradiance (GHI) suffers from inherent biases that are drastically amplified under transitional weather, severely compromising day-ahead photovoltaic (PV) power forecasting accuracy. To address this issue, this study proposes a day-ahead PV forecasting method based on dual-stage NWP GHI correction with transitional weather classification. First, a dual-sliding-window approach identifies GHI turning points and categorizes weather-transition trends. Second, trend-specific attention-bidirectional gated recurrent unit (BiGRU) models perform the first-stage correction of NWP GHI. The residual errors are then decomposed via the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) into multiscale components, each corrected by a second-stage sttention-BiGRU to compensate for frequency-band errors. Finally, the dual-stage-corrected GHI serves as a key input for accurate day-ahead PV power forecasting. Case studies using operational data from Hebei, China show that the proposed method reduces mean absolute error (MAE) by up to 39% and achieves an approximately 10% relative improvement in R² compared with raw NWP data at station 1 for day-ahead forecasting. These improvements are particularly pronounced under transitional weather, demonstrating the method’s effectiveness in enhancing forecast reliability and supporting grid dispatch and renewable energy integration.