基于深度脊波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 Ridgelet Neural Network

作　　者：岳远波 撖奥洋 张智晟 YUE Yuanbo;HAN Aoyang;ZHANG Zhisheng(College of Electrical Engineering,Qingdao University,Qingdao 266071,China;State Grid Qingdao Electric Power Company,Qingdao 266002,China)

机构地区：[1]青岛大学电气工程学院,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》2020年第4期57-61,68,共6页Proceedings of the CSU-EPSA

基　　金：国家自然科学基金资助项目(51477078)。

摘　　要：针对电力系统负荷的非线性预测问题,本文构造了一种基于深度脊波神经网络的电力系统短期负荷预测模型。该模型的隐含层采用脊波神经元,神经元的激励函数采用脊波变换函数。对该预测模型采用受限的玻尔兹曼机学习原理进行预训练,最后利用粒子群优化算法对其进行深度优化精调。通过对某地区实际电网负荷系统进行仿真预测,结果表明,与传统的BP神经网络、脊波神经网络和常规深度神经网络模型相对比,深度脊波神经网络预测模型的日平均绝对误差百分比分别降低了1.96%、1.12%和0.3%,日最大绝对误差分别降低了3.91%、2.19%和1.78%,验证了深度脊波神经网络预测模型具有较好的预测准确度和稳定性。To solve the nonlinear forecasting problem of power system load,a short-term load forecasting(STLF)model based on deep ridgelet neural network(DRNN)is constructed,in which the hidden layer adopts ridgelet neurons and the ridgelet transform function is taken as their excitation function. This prediction model is pre-trained according to the restricted Boltzmann machine(RBM)learning principle,and it is further deeply optimized using the particle swarm optimization(PSO)algorithm. An actual grid load system in one certain region is predicted through simulations,and results show that compared with models based on the traditional BP neural network,ridgelet neural network,and the conventional deep neural network(DNN),the daily mean absolute percentage error of the prediction model based on DRNN is reduced by 1.96%,1.12%,and 0.3%,respectively,and its daily maximum absolute error is reduced by3.91%,2.19%,and 1.78%,respectively,showing that the novel model has higher prediction accuracy and better predictive stability.

关 键 词：深度脊波神经网络 短期负荷预测 玻尔兹曼机 粒子群优化算法 电力系统 

分 类 号：TM715[电气工程—电力系统及自动化]