基于FNN-LSTM-Attention的短期电力负荷预测研究  

Research on Short-term Power Load Forecasting Based on FNN-LSTM-Attention

作　　者：薛文斌 穆晨宇 杜建城 穆羡瑛 田永明 邹德凡 XUE Wenbin;MU Chenyu;DU Jiancheng;MU Xianying;TIAN Yongming;ZOU Defan(Urumqi Power Supply Company of State Grid Xinjiang Electric Power Co.,Ltd.,Urumqi 830001,China;School of Electrical and Electronic Engineering,North China Electric Power University,Beijing 102206,China)

机构地区：[1]国网新疆电力有限公司乌鲁木齐供电公司,新疆乌鲁木齐830001 [2]华北电力大学,电气与电子工程学院,北京102206

出　　处：《微型电脑应用》2024年第12期89-92,共4页Microcomputer Applications

基　　金：国网新疆电力有限公司科技项目(5230WJ220005)。

摘　　要：为了充分挖掘数据潜在规律,解决电力负荷复杂性、非线性等预测难点,提出一种基于FNN-LSTM-Attention的混合预测模型。通过前馈神经网络(FNN)在时间维度上提取数据特征,得到不同特征,利用长短期记忆(LSTM)提取日期、温度等因素对负荷的影响,通过Self-Attention层进一步挖掘数据特征,输出预测值。以新疆某地区实际负荷数据为实例,对不同模型的预测误差进行分析与对比,结果显示,所提出的混合预测模型的预测误差较小,证明了所提模型的有效性。In order to fully explore the potential patterns of data and overcome the forecasting difficulties such as complexity and nonlinearity of power load,this paper proposes a hybrid forecasting model based on FNN-LSTM-Attention.Data features are extracted in the time dimension through feedforward neural network(FNN),different features are obtained,and long short-term memory(LSTM)is used to extract the impact of factors such as date and temperature on load.The Self-Attention layer is used to further explore data features and output predicted values.Taken actual load data from a certain region in Xinjiang as an example,the forecasting errors of different models are analyzed and compared.The results show that the proposed hybrid forecasting model has smaller forecasting errors,proving the effectiveness of the model.

关 键 词：深度学习 电力负荷预测 长短期记忆网络 自注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]