Short-Term Photovoltaic Power Prediction Based on FCformer-BiTCN Method

Accurate prediction of photovoltaic (PV) power and reliability of predictive features are usually compromised due to uncertainties and fluctuations of the PV time series. In this work, a novel short-term PV power prediction method, named the FCformer-BiTCN, is proposed via integrating the enhanced Informer with Fourier convolutional network and bidirectional temporal convolution network (BiTCN). Specifically, raw PV power dataset is preprocessed by the Savitzky-Golay smoothing with wavelet threshold denoising (WTD). Then, the multimodal features of distillation layer and hidden patterns of the preprocessed input sequences are extracted by the proposed FCformer-BiTCN model. Eventually, two experimental datasets of PV power are employed for performance verification. The prediction results, ablation experiment and the Wilcoxon signed-rank test indicate that higher prediction accuracy and robustness are achieved by the proposed model compared with state-of-the-art benchmarks such as Informer and the Informer-TCN, the BiTCN and the Autoformer model.