基于ISSA-CNN-GRU模型的电动汽车充电负荷预测方法

电动汽车用户充电行为的随机性，给电动汽车充电站充电负荷的短期预测带来极大挑战。针对在多因素影响下电动汽车充电站充电负荷短期预测精度低的问题，提出一种基于改进麻雀搜索算法-卷积神经网络-门控循环神经网络(improved sparrow search algorithm-convolutional neural network-gated recurrent unit neural network, ISSA- CNN-GRU)模型的电动汽车充电站充电负荷短期预测方法。首先，构建包含气温、日期类型、节假日3种充电负荷显著影响因素与历史充电负荷的输入特征矩阵。然后，融合CNN在特征提取、数据降维和GRU神经网络在时间序列预测上的优势，搭建CNN-GRU混合神经网络模型，使用基于混合策略的ISSA算法优化混合神经网络模型的超参数。最后，在优化后的CNN-GRU模型中输入特征矩阵实现充电站充电负荷的短期预测。以美国ANN-DATA公开数据集中充电站的历史负荷数据作为实际算例，与随机森林、CNN、GRU神经网络、CNN-GRU模型以及分别用贝叶斯优化、粒子群优化、标准麻雀优化算法进行超参数调优的CNN-GRU模型相比，实验结果表明所提方法具有更好的预测效果。

Charging load prediction method for electric vehicles based on an ISSA-CNN-GRU model

The randomness of EV user charging behavior poses a great challenge to the short-term prediction of EV charging station charging load. There is a problem of the influence of multiple factors, and so a short-term prediction method of charging load of electric vehicle charging stations based on an improved sparrow search algorithm-convolutional neural network-gated recurrent neural network (ISSA-CNN-GRU) model is proposed. First, an input characteristic matrix containing three significant influencing factors of charging load, that is temperature, date type, and holiday, combined with historical charging load is constructed. Then, the advantages of a CNN in feature extraction, and data dimensionality reduction combined with a GRU neural network in time series prediction are used to build a CNN-GRU hybrid neural network model, and the ISSA algorithm based on hybrid strategy is used to optimize the hyperparameters of the hybrid neural network model. Finally, the input feature matrix in the optimized CNN-GRU model realizes the short-term prediction of charging load of a charging station. Taking the historical load data of charging stations in the public dataset of ANN-DATA in the United States as an actual example, compared with the random forest, CNN, GRU neural network, CNN-GRU model and CNN-GRU model with Bayesian optimization, particle swarm optimization and standard sparrow optimization algorithms respectively, the experimental results show that the proposed method offers a better prediction.