Constructing two-stream input matrices in a convolutional neural network for photovoltaic power prediction

Abstract

To enhance the prediction accuracy of photovoltaic (PV) power and provide a more exact decision basis for power system management, an end-to-end reinforced two-stream convolutional neural network (RTS-CNN) is proposed. A temporal data converter is firstly devised to adaptively transform the input series into two different input matrices for improving the representational ability of the input data. One matrix is obtained through the tensor dimension adjustment, while the other uses Fast Fourier Transform (FFT) to acquire periodic information to reconstruct the shape of raw data. Next, to fully extract features from these two input matrices, the RTS-CNN model selects the powerful ConvNeXt block as the foundational building block of the encoder. Subsequently, the PV power prediction is realized using a decoder augmented with the Convolutional Block Attention Module (CBAM). Finally, for better performance and learning speed, the model is trained using the Second-order Clipped Stochastic Optimization (Sophia). Three sets of comparative experiments are conducted on the dataset obtained from the Desert Knowledge Australia Solar Centre (DKASC), and the model performance is further validated using the Diebold-Mariano (DM) test and external validation. The experimental results show that the proposed RTS-CNN model has outstanding performance in predicting PV power under different temporal resolutions and seasonal conditions. In the spring season, with time intervals of 15 min, 30 min, and 1 h, the RTS-CNN model achieves root mean square errors (RMSE) of 0.5113, 0.5984, and 0.7801, respectively.