基于人工智能多模式集成的光伏电站总辐射预报方法研究  

RESEARCH ON ARTIFICIAL INTELLIGENCE-BASED MULTI-MODEL ENSEMBLE FORECAST OF GLOBAL HORIZONTAL IRRADIANCE IN PV STATIONS

作　　者：袁彬 于廷照 申彦波[1,2] 莫景越 邓华[3] Yuan Bin;Yu Tingzhao;Shen Yanbo;Mo Jingyue;Deng Hua(CMA Public Meteorological Service Centre,Beijing 100081,China;CMA Wind and Solar Energy Centre,Beijing 100081,China;Guangzhou Institute of Tropical and Marine Meteorology,China Meteorological Administration,Guangzhou 510641,China)

机构地区：[1]中国气象局公共气象服务中心,北京100081 [2]中国气象局风能太阳能中心,北京100081 [3]中国气象局广州热带海洋气象研究所,广州510641

出　　处：《太阳能学报》2025年第4期291-300,共10页Acta Energiae Solaris Sinica

基　　金：中国气象局创新发展专项(CXFZ2024J068);中国气象局公共气象服务中心创新基金(K2023002);新疆“天池英才”引进计划(2023)。

摘　　要：基于2022年CMA-WSP、CMA-MESO、CMA-GD、WRF-SOLAR 4个数值模式预报以及广东省阳江市4个光伏电站实况观测数据,采用LightGBM集成模型,开展逐月总辐射辐照度(GHI)多模式集成预报试验。结果表明:多模式集成可有效降低GHI预报的平均绝对误差(MAE)和均方根误差(RMSE),与每月的最优数值模式预报相比,MAE减少2.47%~32.71%、RMSE减少5.46%~32.29%;多模式集成在不同GHI区间效果差异明显,400 W/m^(2)以下区间内,多模式集成效果最好,全年12个月中有10个月集成有效,MAE减少6.25%~44.44%、RMSE减少14.62%~43.07%,400~700 W/m^(2)区间内多模式集成效果次之,全年12个月中有6个月集成有效,MAE减少0.76%~34.59%、RMSE减少4.14%~31.11%,大于700 W/m^(2)区间内受限于样本量,多模式集成无效果;在晴天、少云、多云、阴天4种典型天气条件下,多模式集成预报与实况观测趋势最为接近,且能体现出因云量变化造成的GHI波动。Based on the forecast of 4 numerical models:CMA-WSP,CMA-MESO,CMA-GD and WRF-SOLAR,as well as the observation data of 2022 collecting from 4 PV stations in Yangjiang city,Guangdong Province,the monthly multi-model ensemble forecasting experiments of global horizontal irradiance(GHI)were conducted by using LightGBM ensemble model.The results show that multi-model ensemble,when compared with the monthly optimal numerical model,can effectively reduce MAE and RMSE of GHI forecast by a range of 2.47%-32.71%and 5.46%-32.29%,respectively.There are considerable differences in the results of multi-model ensemble at different GHI value intervals.When GHI is below 400 W/m^(2),the multi-model ensemble achieves the best performance,with 6.25%-44.44%decrease of MAE and 14.62%-43.07%decrease of RMSE,for 10 months in the entire year.The effect of the ensemble in the GHI interval between 400 W/m^(2) and 700 W/m^(2) ranks second,and the decreasing ranges of MAE and RMSE,for 6 months in the entire year,are 0.76%-34.59%and 4.14%-31.11%respectively.The multi-model ensemble has no effect,due to insufficient sample,when GHI is greater than 700 W/m^(2).Under 4 typical weather conditions,i.e clear,partly cloudy,cloudy and overcast,the multi-model ensemble forecast is the closest to the real observation trend,and it also can illustrate the radiation fluctuation caused by variation of cloud cover.

关 键 词：太阳辐射 预报 人工智能 多模式集成 光伏电站 

分 类 号：P456.1[天文地球—大气科学及气象学]