摘要： 电力负荷预测对电力系统的部署、规划和运行影响重大，但目前各输入特征对电网负荷情况影响的程度不稳定，且递归神经网络捕获负荷数据的长期记忆能力差，导致预测精度下降。提出一种基于双重注意力机制和GRU网络的预测新模型，利用特征注意力机制自主分析历史信息与输入特征间的关联关系，提取重要特征，并通过时序注意力机制自主选取GRU网络中关键时间点的历史信息，提升较长时间段预测效果的稳定性。在3个公开数据集上的实验结果表明，该模型在预测精度指标上表现良好，对比SVR、KPCA-ELM、DBN、GRU、Attention-GRU、CNN-LSTM、Attention-CNN-GRU模型预测精度分别提高了2.47、1.14、1.93、1.37、1.04、0.74、0.41个百分点。

关键词: 时间序列预测, GRU网络, 特征注意力机制, 时序注意力机制, 短期负荷预测

Abstract: Power load forecasting significantly impacts the deployment, planning, and operation of power systems.However, the impact of input characteristics on power grid load is unstable, and the long-term memory ability of recursive neural networks to capture load data is poor, reducing the forecasting accuracy.A new prediction model based on a dual attention mechanism and Gated Recurrent Unit (GRU) network was established in this study.The feature attention mechanism was used to autonomously analyze the relationship between historical information and input features and extract essential features.Moreover, the historical information of key time points in the GRU network was independently selected based on a temporal attention mechanism to improve the stability of the prediction effect over an extended period.The experimental results for three public data sets show that the prediction accuracy index of the model is satisfactory.Compared with the Support Vector machine Regression(SVR), Kernel Principal Component Analysis-Extreme Learning Machine(KPCA-ELM), Deep Belief Network(DBN), GRU, attention-GRU, Convolutional Neural Network (CNN)-Long Short-Term Memory(CNN-LSTM), and attention-CNN-GRU models, the prediction accuracy improved by 2.47, 1.14, 1.93, 1.37, 1.04, 0.74, and 0.41 percentage points, respectively.

Key words: time series prediction, GRU network, feature attention mechanism, temporal attention mechanism, short-term load forecasting

中图分类号: 

• TP181