基于时序卷积网络与线性残差连接的光伏发电功率预测

Photovoltaic Power Prediction Based on Temporal Convolutional Network and Linear Residual Connections

摘要: 针对传统神经网络对于多元发电过程数据处理效率低、且容易过拟合的缺陷，提出了一种基于时序卷积网络(TCN)与线性残差连接的光伏功率预测方法。构建时序卷积网络，通过因果卷积与膨胀卷积技巧并行地提取多个时间点间的动态关系，从而在提取非线性时序相关性的同时保持较高的运算效率。引入线性残差连接构建了网络模型输入端与输出端的信息通路，有效地避免了过拟合。以某光伏电站实测数据对所提出方法进行性能验证，所提出的模型的预测结果的均方误差、均方根误差以及平均绝对误差分别为22.63、4.79、2.47，预测性能好于传统方法。

Abstract: Traditional neural networks have low processing efficiency for multi-variable power generation process data and are easy to overfit. In view of this, a photovoltaic power prediction method based on Temporal Convolutional Network (TCN) and linear residual connections is proposed. TCN is constructed to extract the dynamic relationship between multiple time points in parallel through causal convolution and dilation convolution techniques, so as to maintain high computational efficiency while extracting nonlinear temporal correlation. The introduction of linear residual connections constructs an information path between the input and output of the network, effectively avoiding overfitting. The performance of the proposed method is verified by the data collected from a photovoltaic power station. The mean square error, root mean square error, and mean absolute error of the prediction results of the proposed model are 22.63, 4.79, and 2.47, respectively, outperforming traditional methods.