基于Attention-BiLSTM-LSTM神经网络的短期电力负荷预测方法  

Short-term power load forecasting method based on Attention-BiLSTM-LSTM neural network

作　　者：龚飘怡 罗云峰[1] 方哲梅 窦帆 GONG Piaoyi;LUO Yunfeng;FANG Zhemei;DOU Fan(School of Artificial Intelligence and Automation,Huazhong University of Science and Technology,Wuhan Hubei 430074,China)

机构地区：[1]华中科技大学人工智能与自动化学院,武汉430074

出　　处：《计算机应用》2021年第S01期81-86,共6页journal of Computer Applications

基　　金：华中科技大学自主创新基金(2019kfyXJJS17)。

摘　　要：短期电力负荷预测是电力系统中的重要问题之一,准确的预测结果可以提高电力市场的灵活性和资源利用效率,对电力系统高效运行具有重要意义。为了提高预测精度,针对电网负荷数据的时序性特征,提出一种基于Attention-BiLSTM-LSTM神经网络的短期电力负荷预测方法。该方法首先针对电力负荷的影响因素(温度、节假日等)提取特征,并使用双向长短期记忆(BiLSTM)神经网络层进行双向时序的特征学习;将双向时序特征作为长短期记忆(LSTM)神经网络层的输入,用LSTM神经网络建模学习时序数据的内部变化规律;使用attention机制计算LSTM隐层状态的不同权重,以对隐层状态进行选择性地关注;结合注意力权重和LSTM神经网络进行负荷预测,最后使用全连接层输出负荷预测结果。使用EUNIT电力负荷数据集进行实验,采用提前单点预测模式,该方法的平均绝对百分比误差(MAPE)达到1.66%,均方根误差(RMSE)达到814.85。通过与单LSTM网络、基于attention机制的LSTM网络(Attention-LSTM)、前馈神经网络(FFNN)、卷积神经网络联合长短期记忆神经网络(CNN-LSTM)等4种典型的负荷预测模型结果对比,验证了Attention-BiLSTM-LSTM神经网络方法更加准确有效。Short-term power load forecasting is one of the most important issues in power systems.The flexibility of the power market and the efficiency of resource utilization can be improved by accurate prediction,which poses great significance for the efficient operation of the power system.Considering the temporal characteristic of the grid load data and the goal of improving prediction accuracy,an Attention mechanism based Bi-directional Long Short-Term Memory and Long Short-Term Memory(Attention-BiLSTM-LSTM)neural network model for predicting short-term power load was proposed in this paper.Firstly,the features were extracted based on the factors affecting the power load,then the Bi-directional Long Short-Term Memory(BiLSTM)neural network layer was used to learn high-level features.The output of the BiLSTM layer was used as the input of the Long Short-Term Memory(LSTM)neural network layer to learn the internal change rule of temporal data.And different weights of the LSTM hidden layer states were calculated by adopting attention mechanism to focus on the hidden layer state.The attention weight and the LSTM neural network were combined to perform load prediction,and the final load prediction result was generated accordingly.In this paper,the EUNIT power load dataset and the single-point forecast mode in advance were used for experiments,the Mean Absolute Percentage Error(MAPE)of this method was 1.66%,and the Root Mean Square Error(RMSE)was 814.85.Compared with four typical load forecasting models such as single LSTM network,Attention mechanism based Long Short-Term Memory(Attention-LSTM)neural network,FeedForward Neural Network(FFNN),Convolutional Neural Network and Long Short-Term Memory(CNN-LSTM),the proposed Attention-BiLSTM-LSTM neural network method has been proved more accurate and effective.

关 键 词：短期负荷预测 长短期记忆神经网络 注意力机制 循环神经网络 双向长短期记忆神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TM715[自动化与计算机技术—控制科学与工程]