基于深度多步时空神经网络的电动汽车负荷时空动态负荷预测

Spatio-Temporal Dynamic Load Forecasting of Electric vehicle charging demand Based on Deep Multi-step Spatio-Temporal Neural Network

电动汽车充电负荷具有很强的时空随机性,从而加大电网控制的难度以及影响电能质量,提前准确的预测充电负荷是解决此类问题有效方法之一。因此,首先按时间顺序建立充电桩时空动态负荷矩阵,并在时空神经网络的基础上提出一种时空动态负荷预测模型——多步深度时空神经网络。该模型能够根据过去充电负荷规律多步预测未来负荷,模型将ConvLSTM层和3D-ConvNet层然后连接到融合层作为一个单元,通过多个单元的堆叠增强网络学习能力,3D-ConvNet作为模型的网络的输出层。ConvLSTM层可以很好的学习长时规律,3D-ConvNet层可以学习到短时规律,3D-ConvNet作为输出层可以让网络具有多步输出的能力,以消除单步滚动预测带来的更大误差。并与STN模型进行对比,结果证明了所提预测模型的有效性。

The charging load of electric vehicles has strong time-space randomness,which increases the difficulty of power grid control and affects the quality of power.Predicting the charging load accurately in advance is one of the effective methods to solve such proble ms.Therefore,the space-time dynamic load matrix of the charging pile is firstly established in chronological order,and a spatio-temporal dynamic load prediction model,multi-step deep spatio-temporal neural network,is proposed based on the spatio-temporal neural network.The model can predict the future load in multi-steps according to the past charging load law.The model then connects the ConvLSTM layer and the 3D-ConvNet layer to the fusion layer as a unit,and enhances the network learning ability through the stacking of multiple units.3D-ConvNet is used as the model network.Output layer.The ConvLSTM layer can learn long-term rules very well,the 3D-ConvNet layer can learn short-term rules,and the 3D-ConvNet as an output layer can make the network have multi-step output capability to eliminate the larger error caused by single-step rolling prediction.Compared with the STN model,the results demonstrate the effectiveness of the proposed prediction model.