基于贝叶斯神经网络短期负荷预测模型

The Short-Term Load Forecasting Model Based on Bayesian Neural Network

本文提出了基于贝叶斯神经网络(BNN)短期负荷预测模型。根据气象影响因素和电力负荷的样本数据,针对权向量参数的先验分布分别为正态分布和柯西分布两种情况,应用混合蒙特卡洛(HMC)算法学习了BNN的权向量参数。由HMC算法和Laplace算法学习的贝叶斯神经网络以及BP算法学习的传统神经网络分别对4月(春)、8月(夏)、10月(秋)和1月(冬)每月25天的每个整点时刻的负荷进行了预测。这些神经网络的输入层有11个节点,它们分别与每个整点时刻和的气象因素、上一个整点时刻的气象因素和时间变量相对应,输出层只有一个节点,它与负荷变量对应。试验结果表明HMC算法学习的BNN的预测结果的百分比平均绝对误差(MAPE)和平方根平均误差(RSME)取值远远小于由Laplace算法学习的BNN和BP算法学习的人工神经网络的MAPE和RMSE。而且,HMC算法学习的BNN在测试集和训练集上的预测误差MAPE和RMSE的相差很小。实验结果充分说明HMC算法学习的BNN具有较高的预测精度和较强的泛化能力。

A short term load forecasting model based on Bayesian neural network learned by the Hybrid Monte Carlo（HMC） algorithm is presented in this paper.The weight vector parameter of the Bayesian neural network is considered as multi-dimensional random variables.Using the weather factors and load recorders in training set,HMC algorithm is used to learn the weight vector parameter with respect to normal prior distribution and Cauchy prior distribution respectively.Two Bayesian neural networks learned by Laplace algorithm and HMC algorithm and the artificial neural network learned by the BP algorithm are used to forecast the hourly load of 25 days of April（spring）,August（summer）,October（autumn） and January（winter） respectively.There are eleven nodes in input layer,ten nodes representing the ten weather factor variables of current hour and the previous hour and one hour variable.There is one node in output layer,corresponding to the load on each hour.The experimental result shows that the roots mean squared error（RMSE） and the mean absolute percent errors（MAPE） of the Bayesian neural network learned by hybrid Monte Carlo algorithm both are much smaller than those of the neural networks learned by Laplace algorithm and BP algorithm.Hence,the forecasting model based on BNN learned by the HMC algorithm has higher forecasting precision,and can be used to short-term load forecasting.