基于深度递归神经网络的电力系统短期负荷预测模型  

Short-term Load Forecasting Model of Power System Based on Deep Recurrent Neural Network

作　　者：于惠鸣 张智晟[1] 龚文杰 段晓燕 YU Huiming;ZHANG Zhisheng;GONG Wenjie;DUAN Xiaoyan(College of Automation and Electrical Engineering,Qingdao University,Qingdao 266071,China;State Grid Qingdao Power Supply Company,Qingdao 266002,China)

机构地区：[1]青岛大学自动化与电气工程学院,青岛266071 [2]国网青岛供电公司,青岛266002

出　　处：《电力系统及其自动化学报》2019年第1期112-116,共5页Proceedings of the CSU-EPSA

基　　金：国家自然科学基金资助项目(51477078);智能电网教育部重点实验室开放研究基金资助项目(2018)

摘　　要：针对电力负荷非线性动态特性导致的负荷预测困难、预测精度低等问题,本文构建了深度递归神经网络短期负荷预测模型。在深度神经网络多隐层结构的基础上,深度递归神经网络增设了关联层,并以改进粒子群算法作为网络的优化学习算法,对模型权值空间进行深度优化。对某地区电网实际负荷进行预测仿真,结果表明与BP网络、深度神经网络相比,深度递归神经网络的平均绝对误差的周平均值分别降低1.61%和0.56%,验证了深度递归神经网络能够融合前馈与反馈连接,提高网络泛化能力,有效提高负荷预测精度。In view of the difficulty in load forecasting and the low prediction accuracy caused by the nonlinear dynamic characteristics of power load , a short-term load forecasting model based on deep recurrent neural network is established in this paper. Based on the deep neural network ' s multi-hidden-layer structure , a connection layer is added to the deep recurrent neural network , and an improved particle swarm algorithm is adopted as the optimization learning algorithm for the network to optimize the model ' s weight space. The actual load of a regional power grid is forecasted through simu-lations , showing that the weekly average values of mean absolute errors of deep recurrent neural network are reduced by 1.61 % and 0.56 % compared with those of BP network and deep neural network , which verifies that the deep recurrent neural network can combine the feedforward and feedback connections , improve the network ' s generalization capability , and effectively improve the accuracy of load forecasting.

关 键 词：深度神经网络 深度递归神经网络 改进粒子群优化算法 短期负荷预测 电力系统 

分 类 号：TM715[电气工程—电力系统及自动化]