基于最小窥视孔长短期记忆神经网络的电力系统短期负荷预测模型  

Power Load Forecasting Based on Min Peephole Long Short-term Memory

作　　者：蔡鑫祥 撖奥洋 周生奇 魏振 张智晟 CAI Xinxiang;HAN Aoyang;ZHOU Shengqi;WEI Zhen;ZHANG Zhisheng(College of Electrical Engineering,Qingdao University,Qingdao,Shandong 266071,China;Qingdao Power Supply Company of State Grid Shandong Electric Power Company,Qingdao,Shandong 266002,China)

机构地区：[1]青岛大学电气工程学院,山东青岛266071 [2]国网山东省电力公司青岛供电公司,山东青岛266002

出　　处：《广东电力》2021年第3期92-97,共6页Guangdong Electric Power

基　　金：国网山东省电力公司科技项目(2020A-022)。

摘　　要：长短期记忆(long short-term memory,LSTM)模型可以克服循环神经网络(recurrent neural networks,RNN)在短期负荷预测中存在的梯度消失和梯度爆炸等问题,但LSTM模型的门控单元结构复杂,参数较多,模型训练较为困难。为此,提出一种采用改进粒子群优化(particle swarm optimization,PSO)算法优化的最小窥视孔长短期记忆(min peephole long short-term memory,MP-LSTM)模型。与经典LSTM模型相比,MP-LSTM模型舍弃了输入门和输出门,只保留遗忘门,模型包括1个sigmoid网络层和1个tanh网络层,减少了模型参数,优化了模型结构。对RNN模型、经典LSTM模型、MP-LSTM模型进行对比仿真分析,结果表明MP-LSTM模型可有效提高负荷预测精度,缩短收敛时间。The long short-term memory(LSTM)model can overcome the problems of the recurrent neural network(RNN)in short-term load forecast such as gradient disappearance and gradient explosion,but its gating unit is more complicated and the parameters are too many,which makes it difficult to train the model.Therefore,this paper adopts a min peephole long short-term memory(MP-LSTM)model based on the improved particle swarm optimization(PSO)algorithm.This MP-LSTM model abandons the input gate and output gate,and only retains the forget gate.The model includes a sigmoid network layer and a tanh network layer,which reduces model parameters and optimizes the model structure.The paper compares the MP-LSTM model with the RNN model,the classic LSTM model,and the result verifies that the MP-LSTM model can effectively improve the prediction accuracy and shorten convergence time.

关 键 词：短期负荷预测 最小窥视孔长短期记忆神经网络 粒子群优化算法 电力系统 循环神经网络 

分 类 号：TM715.1[电气工程—电力系统及自动化] TP183[自动化与计算机技术—控制理论与控制工程]