基于时段细分训练时序网络组的超短期负荷预测  

Ultra-short-term Load Prediction Based on Time-interval-differential Training of Time Series Neural Network Groups

作　　者：张晓虎 蒋葓藨 ZHANG Xiaohu;JIANG Hongbiao(College of Electrical and Information Engineering,Hunan University of Technology,Zhuzhou 412000,China)

机构地区：[1]湖南工业大学电气与信息工程学院,湖南株洲412000

出　　处：《电工技术》2023年第21期11-14,19,共5页Electric Engineering

基　　金：国家重点研发计划项目(编号2022YFE0105200)。

摘　　要：为优化能源系统的能源供应调度方案,针对时序类神经网络难以提取与利用影响负荷变化的不固定成分信息的问题,结合相关影响特征在不同时段下重要性不同的特点,提出一种基于时段细分训练时序网络组的预测方法,通过与经典的预测模型长短期记忆网络(LSTM)、可控门循环网络(GRU)、双向长短值记忆网络(BiLSTM)、双向可控门循环网络(BiGRU)的多次对比可知,该方法能有效降低相应模型负荷预测时的RMSE、MSE、MAPE与MAE值。Aiming at optimizing power supply scheduling and addressing the problem of time series neural networks′incompetence in extracting and utilizing unfixed component information that affects load variation,and considering the time-dependent feature importance of relevant influence characteristics,a prediction method based on time-interval-differential training of time series neural network groups is proposed.Comparative analysis with typical methods such as long short-term memory(LSTM),gate recurrent unit(GRU)neural network,bi-directional long-short term memory network(BiLSTM)and bi-directional gate recurrent unit(BiGRU)neural network confirmed the effectiveness of the proposed method in reducing values of RMSE,MSE,MAPE and MAE in load prediction.

关 键 词：超短期负荷预测 LSTM GRU 神经网络 

分 类 号：TM411[电气工程—电器]