| Photovoltaic power generation has the advantages of clean and sustainable.With the "double carbon" goal proposed,the photovoltaic industry has developed rapidly..Irradiance is the most important factor affecting photovoltaic power generation output.Accurate irradiance forecast is a necessary prerequisite to ensure the stability of photovoltaic on-grid power.However,the irradiance is characterized by mutation,randomness and intermittency under the influence of weather conditions.In this paper,a photovoltaic power station was selected as the research object.Based on WRF-Solar regional climate model,the optimal irradiance simulation physical parameter scheme was determined through sensitivity analysis,and the dynamic downscaling simulation experiment of regional irradiance was completed.At the same time,two artificial intelligence algorithms were used to establish irradiance correction models under different weather types,and the forecast results of WRF-Solar model were revised.The results showd that the forecast errors of WRF-Solar initial results were 16.00%,33.33%and 64.03%in sunny,sunny and rainy weather,respectively,showing poor forecasting effect.Stepwise clustering(SCA)and multilayer sensing(MLP)neural network significantly improved the prediction accuracy of WRF-Solar model.The two models had the best correction effect in sunny days,with relative percentage errors of 11.91%and 11.11%,respectively.The performance of MLP neural network was slightly better than that of SCA model.Under the condition of clear skies and cloudy skies,the correction effect of SCA model was better than that of MLP neural network,and the error after correction was 14.03%and 18.01%,respectively.The output irradiance error of WRF-Solar model was large under cloudy and rainy weather.After correction by the two models,the error was reduced from 192.77 W/m2 to 73.79 W/m2 and 71.51 W/m2 respectively.In addition,relative to the initial results of WRF-Solar,SCA performed better than MLP correction results on peak simulation.From the perspective of intraday variation,most of the deviations between the setting values and the measured values of the two models occurred in the noon period(such as 0 to 61 hours).The correction result of SCA was higher than that of MLP neural network,so its simulation accuracy in the noon period was slightly lower.There are still some challenges to improve the irradiance simulation accuracy under cloudy and rainy weather conditions.Although the revised results of SCA roughly simulate the intra-day fluctuation and abrupt change process,there are still large deviations,which can be improved by adding training samples and optimizing the correction model in the future. |
