一种TCN的改进模型及其在短期光伏功率区间预测的应用

为了提高光伏功率预测的精度，提出了一种基于时序卷积网络(temporal convolutional network, TCN)的新型短期光伏功率区间预测模型。首先，采用深度残差收缩网络(deep residual shrinkage network, DRSN)的软阈值和注意力机制来改进TCN的残差模块以增强其对有用特征提取能力，并削弱冗余特征的不利影响；然后，利用樽海鞘群算法(slap swarm algorithm, SSA)对TCN的卷积层的卷积核大小和TCN层数等超参数进行自动寻优，以克服原TCN感受野不足的问题；接着，采用核密度估计(kernel density estimation, KDE)方法对所建改进TCN短期光伏功率预测模型的点预测结果进行误差分析，获得模型预测输出的区间。最后，通过对比仿真实验得到的结果表明，提出的SSA-DRSN-TCN模型的RMSE平均值为0.27,优于LSTM、GRU、CNN-LSTM和TCN等模型；而且，KDE方法能够在80%、90%和95%的置信度下准确描述光伏功率波动区间，验证了所提模型在提高光伏功率预测性能上的有效性。

Improved TCN model and its application in short-term photovoltaic power interval prediction

In order to improve the accuracy of PV power prediction, this paper proposed a new short-term PV power interval prediction model based on TCN. Firstly, the model used the soft threshold and attention mechanism of DRSN to modify the residual module of TCN so as to enhance its capability to extract useful features, and reduce the impact of redundant features. Then, it adopted the SSA to search the optimal hyper-parameters automatically, such as the convolutional kernel size and the number of TCN layers in the convolutional layer of the TCN, to overcome the drawback of the original TCN with insufficient receptive fields. Next, this paper applied the KDE to analyze the error of the point prediction results of the proposed improved TCN short-term PV power forecasting model to obtain its output interval. Finally, comparative simulation experiments show that the RMSE of the proposed SSA-DRSN-TCN model can reach 0.27, which is better than those of LSTM, GRU, CNN-LSTM and TCN, and the KDE method can accurately describe the PV power volatility intervals at 80%, 90% and 95% confidence levels, the proposed model verifies the effectiveness in improving the performance of PV power prediction.