基于时间卷积网络与图注意力网络的分行业日售电量预测方法

Daily Electricity Sales Forecasting by Industries Based on Time Convolution Network and Graph Attention Network

为控制电力成本及提高电力部门绩效考核能力,需要高效准确地进行日售电量预测。深度学习卷积神经网络常被用于电力数据预测,但由于其输入数据信息量有限,现有模型预测存在上限,致使其存在难以捕捉长时特征等问题。为高效准确地预测日售电量,提出了一种基于时间卷积网络与图注意力网络相结合的分行业日售电量预测方法,搭建了高维度分行业日售电量预测模型。该方法可同时输入多个行业的日售电量,提取反映单个行业时序特征的高维变量,将多个行业的高维变量进行拼接学习,得到各行业之间的影响因素。通过多个行业日售电量的集成增加输入数据的信息量,从而实现各行业的日售电量预测。以中国东南某城市的21个行业日售电量为实际算例,上述方法的平均误差为4.03%。与时间卷积网络、门控循环单元网络、Facebook的Prophet模型、M4冠军模型指数平滑递归神经网络等进行对比,实验表明,所提出的分行业日售电量预测模型具有更高的预测精度。

In order to control the power cost and improve the performance appraisal ability of the power department,it is necessary to make efficient and accurate forecast for the daily electricity sales.The deep learning Convolutional Neural Network(CNN) is often used in the power data prediction.However,the existing model has an upper limit for prediction,which makes it difficult to capture the long-term characteristics due to the limited information of its input data.For predicting the daily electricity sales efficiently and accurately,a prediction for the daily electricity sales of different industries based on the combination of the Time Convolution Network(TCN) and the Graph Attention Network(GAT) is proposed,and a high-dimensional prediction model for different industries is established.This method inputs the daily electricity sales of multiple industries at the same time,extracts the high-dimensional variables reflecting the timing characteristics of a single industry,and combines the high-dimensional variables of multiple industries to learn the influencing factors among the industries.The information of input data is increased through the integration of the daily electricity sales for multiple industries,so as to realize the forecast of the daily electricity sales for various industries.Taking the daily electricity sales of 21 industries in a city in Southeast China as an example,the average error of this method is 4.03%.Compared with the time convolution network(TCN),the Gated Recurrent Unit(GRU),the Facebook’s Prophet model and the M4 champion ESRNN model,the results show that the forecast model proposed in this paper has higher prediction accuracy.