Abstract: Numerical weather prediction (NWP) is widely applied in photovoltaic power generation forecasting, and it holds significant importance for the economic dispatch of power grids. Due to the unavoidable inherent errors in numerical methods, the day-ahead solar irradiance — the most direct factor affecting photovoltaic power generation forecasting — leads to a reduction in prediction accuracy. This study proposes a day-ahead solar irradiance correction method based on inferring forecast errors from historical weather forecast data and clustering historically similar days using K-means. On this basis, a COSTLSTM model capable of decomposing trend and seasonal components of input sequences is further proposed for the solar irradiance correction model. Experimental verification was conducted using on-site collected data from a company and numerical meteorological forecast data from a platform. Compared with traditional methods such as LinearRegression, KernelRidge, and GradientBoosting, the method proposed in this study exhibits higher robustness and accuracy.