基于LSTM模型的光伏辐照度短期预报订正技术OA

To improve the accuracy of short-term photovoltaic irradiance forecasts from numerical models,the study utilized the hourly short-term radiation forecast products from the China Meteorological Administration Wind and Solar Numerical Forecasting System(CMA-WSP 2.0)for the period 2023-2025,along with ground observational irradiance data from photovoltaic power stations and meteorological observation stations in Ningxia.A sequence correction model integrating a long short-term memory(LSTM)network with an Attention Mechanism(hereafter referred to as attention-enhanced LSTM sequence correction,ALSC)was constructed,in which the LSTM captured the long-term dependence of radiation time series,and the attention mechanism assigned adaptive weights to meteorological factors.The uncorrected CMA-WSP 2.0 forecast products served as a control.The effectiveness of the model correction was evaluated using metrics including the coefficient of determination(R²),root mean square error(RMSE),and mean absolute error(MAE).The results indicate that,using 2025 data as the test set,R² of the corrected radiation forecast significantly improved,while RMSE and MAE decreased by 10.57%and 7.57%,respectively.Evaluation of March data under different sky conditions showed that the average relative errors under overcast and cloudy conditions were reduced from 28.87%and 25.47%to 3.06%and 0.33%,respectively,with skill scores(indicating the reduction in errors of the corrected forecasts relative to the original forecast values)of 89.4%and 98.7%,indicating a significant enhancement in forecasting accuracy.Under weather conditions with cloud cover and rapid radiation fluctuations,the model's responsiveness to abrupt radiation changes was also enhanced.This study provides a theoretical basis and practical guidance for power generation forecasting at photovoltaic power stations and optimization of power grid scheduling.