基于单步预测LSTM的短期负荷预测模型

Short-Term Load Forecasting Model Based on LSTM of Single-Step Forecasting

针对目前智能电网下负荷数据存在特征相关性低与序列非平稳性的特点,为达到提高短期负荷预测精度的目的,提出一种单步负荷预测的双层LSTM模型的方法,除了考虑的地理区域的天气数据外,还使用负荷的时间序列。首先采用最大信息系数(MIC)对多源异构特征进行提取,随后使用随机森林和递归特征消除(RFE)用于特征选择,将对异常值敏感、鲁棒性好的Robust标准化方法应用于所选特征的预处理中,最后由单步预测的双层LSTM模型预测结果。文章所搭建的负荷预测模型具有数据强相关性特征、预测精度高的优点。依据数据实验结果,取得了1.38%的MAPE值。并且与其它预测模型相比,基于单步预测LSTM的短期负荷预测方法的预测精度提高效果明显。

In order to improve the accuracy of short-term load forecasting, a two-layer LSTM model for one-step load forecasting is proposed, which not only considers the weather data of geographical regions, but also uses the time series of loads. In addition to the weather data of the geographic area considered, the time series of the load can also be used. First, the maximum information coefficient(MIC) was used to extract multi-source heterogeneous features. Then the random forest and recursive feature elimination(RFE) were used for the feature selection. Robust standardization method, which is sensitive to outliers and has good robust, was applied to the preprocessing of selected features. Finally, the prediction result of the double-layer LSTM model was predicted by the single step forecasting. The load forecasting model built in the paper has the advantages of strong data correlation and high forecasting accuracy. According to the experimental results of data, the MAPE value of 1.38% was achieved. And compared with other prediction models, the prediction accuracy of the proposed method is significantly improved.