基于精细化气象预报数据的分布式光伏日发电量预测研究——以南浔地区为例

Daily Power Generation Prediction of Distributed Photovoltaic Systems Using High-Resolution Weather Forecast Data—A Case Study in Nanxun District

摘要: 为提高分布式光伏日发电量预测精度，本文以浙江省南浔地区多个分布式光伏电站为研究对象，收集电站发电数据以及降水、温度、相对湿度、风速、云量、能见度和晴空辐射等逐小时气象数据，构建了基于XGBoost的分布式光伏日发电量预测模型，并分析了不同气象要素的重要性及时变特征，以及模型预测误差与气象要素之间的关系。结果表明，云量和能见度是影响分布式光伏日发电量预测的关键气象因素，其中云量在白天各时段均保持较高的重要性，能见度的重要性主要集中在中午时段，不同气象要素对光伏发电具有明显的时间依赖特征。所建立的XGBoost模型能够有效挖掘多源气象要素与日发电量之间的非线性关系，在验证集上的决定系数达到0.98，平均绝对百分比误差为13%，预测准确率达到87%，预测结果与实测值具有较高的一致性。进一步的误差分析表明，模型预测误差未随云量、能见度、温度及晴空辐射等单一气象要素变化而表现出明显的系统性规律，说明模型能够综合利用多种气象信息实现稳定预测，具有较好的泛化能力和鲁棒性。研究结果可为区域分布式光伏发电预测、电网调度及新能源消纳提供技术参考。

Abstract: To improve the accuracy of daily power generation forecasting for distributed photovoltaic (PV) systems, multiple distributed PV power stations in Nanxun District, Zhejiang Province were selected as the study area. Historical power generation data together with hourly meteorological data, including precipitation, air temperature, relative humidity, wind speed, cloud cover, visibility, and clear-sky solar radiation, were collected to develop a daily PV power generation forecasting model based on XGBoost. The importance and temporal characteristics of different meteorological variables, as well as the relationship between prediction errors and meteorological factors, were further investigated. The results indicate that cloud cover and visibility are the most influential meteorological variables affecting daily PV power generation forecasting. Cloud cover maintains consistently high importance throughout the daytime, whereas the importance of visibility is mainly concentrated around midday, demonstrating a clear temporal dependence of meteorological factors on PV power generation. The proposed XGBoost model effectively captures the nonlinear relationships between multiple meteorological variables and daily power generation, achieving a coefficient of determination (R²) of 0.98, a mean absolute percentage error (MAPE) of 13%, and a prediction accuracy of 87% on the validation dataset. The predicted values show high consistency with the measured observations. Furthermore, error analysis reveals no significant systematic relationship between prediction errors and individual meteorological variables, including cloud cover, visibility, temperature, and clear-sky solar radiation, indicating that the model can comprehensively utilize multiple meteorological inputs to achieve stable predictions with strong generalization ability and robustness. The findings provide a useful technical reference for regional distributed PV power forecasting, power grid dispatching, and renewable energy integration.