基于多模型综合特征选择和LSTM-Attention的短期负荷预测

Short Term Load Forecasting Based on Multi-Model Comprehensive Feature Selection and LSTM-Attention

为提高电力系统短期负荷预测精度和预测效率,提出一种基于多模型综合特征选择和长短期记忆单元(long short time memory, LSTM)-Attention的短期负荷预测方法。首先,利用随机森林算法、自适应集成(adaptive boosting, AdaBoost)算法及梯度提升树(gradient boosting decision tree, GBDT)算法对原始数据进行初步拟合预测,提取3种算法拟合后的结果来获取特征量与负荷大小的相关系数,从而建立综合相关系数。接着,根据综合相关系数的大小,剔除相关系数较小的特征量,将剩余的特征量与历史负荷大小数据结合构成新的数据集。最后,将新的数据集作为LSTM-Attention预测模型的输入,从而得到待预测日的负荷预测曲线。通过分析所提出的预测方法在某地区负荷数据集的预测结果可知,该方法优于其他预测方法。

In order to improve the forecasting accuracy and efficiency of power system short-term load forecasting, this paper proposes a short-term load forecasting method based on multi-model comprehensive feature selection and long short time memory(LSTM)-Attention. Firstly, the random forest algorithm, adaptive boosting(AdaBoost) algorithm and gradient boosting decision tree(GBDT) algorithm are used to preliminarily fit and predict the original data, extract the output results of the three algorithms after fitting, and establish the comprehensive correlation coefficient by using the correlation coefficient between the characteristic quantity and the load size;then, according to the size of the comprehensive correlation coefficient, the feature quantity with smaller correlation coefficient is removed, and the remaining feature quantity is combined with the historical load size data to form a new data set. Finally, the new data set is used as the input of the LSTM-Attention forecasting model, and the load forecasting curve of the day to be predicted is obtained. By analyzing the forecasting results of the proposed forecasting method in the historical load data set of a certain area, it can be seen that the proposed method has higher forecasting accuracy than other methods.