基于XGBoost算法融合多特征短期光伏发电量预测

针对目前光伏发电过程中由于"弃光"现象导致能源利用率低和经济性差等问题,提出一种基于XGBoost算法融合多种特征的短期光伏发电量预测的方法。文中介绍了XGBoost算法的基本原理,引入正则化惩罚函数和误差函数来构建光伏预测模型的目标函数;分析了光伏发电量和各特征之间的皮尔森相关系数,同时对特征的异常数据进行预处理。在训练过程中为了避免对模型超参数的影响,采用K折交叉验证(K Fold Cross Validation)对数据的训练集、验证集和测试集进行划分。训练完模型参数后把测试集数据放到光伏预测模型中,预测得到未来三天的光伏发电量。对比实验选择SVM和LSTM两种预测方法进行,实验结果表明XGBoost算法在预测光伏发电中具有较高的准确性和实用性。

Multi-Feature Short-term Photovoltaic Generation Forecasting Based on XGBoost Algorithm

For the problems of low energy utilization and poor economy caused by "light abandonment" phenomenon in the current photovoltaic power generation process, a short-term photovoltaic power generation forecasting method based on XGBoost algorithm mixing multiple features is proposed. Firstly, the basic principle of XGBoost algorithm is introduced, and described briefly the advantages compared with other methods. The objective function of photovoltaic prediction model is constructed by introducing regularization penalty function and error function, and the abnormal data is preprocessed. Then Pearson correlation coefficients between photovoltaic power generation and each feature are analyzed.After training the model parameters, the test data are put into the photovoltaic forecasting model, and the photovoltaic power generation in the next three days is forecasted. The other two forecasting methods, SVM and LSTM, are selected for comparison experiments. The absolute values of load forecasting accuracy and load forecasting deviation rate are used as the evaluation indexes of the model. The experimental results show that XGBoost algorithm has high accuracy and practicability in forecasting photovoltaic power generation.