Decreasing temporal convolutional method integrated with Fourier transform for precise solar PV and wind power generation prediction

Abstract

Accurately predicting power generation from solar photovoltaic (PV) and wind power systems is paramount for grid scheduling decisions, operational efficiency enhancement, and energy conservation. Many existing deep learning-based methods have more network layers, more complex models, and higher computational costs. They generally have limitations, such as difficulty in accurately extracting features of long sequences and predicting multiple types of energy separately. In response to the challenges above, this paper proposes a novel Decreasing Time-Domain Convolutional Network (DTCN), whose core innovation lies in its seamless integration with Fourier transform methods for the task of power generation prediction for solar photovoltaic and wind power facilities. The method’s key advantage lies in its deep integration of frequency-domain features: by incorporating frequency-domain features captured through the Fourier transform, it reveals the compositional information of the sequence across different frequency components, thereby endowing the model with a global perspective. This characteristic makes it inherently suited for multi-variable time series prediction tasks. Specifically, the method innovatively constructs a ”time-domain - frequency-domain” collaborative processing framework: first, the original time series data is mapped to the frequency domain using the fast Fourier transform (FFT) algorithm to model the frequency characteristics of the data explicitly; then, DTCN is innovatively applied to the frequency domain representation to extract complex dependencies between multiple variables. Notably, DTCN achieves efficient extraction of feature information in long-sequence frequency-domain representations through its core components, including diminishing expansion convolution, causal convolution, multi-step processing mechanisms, and residual connections. This design overcomes the limitations of traditional convolutional networks in modeling dependencies in long sequences. This paper presents experiments conducted on a full-year solar PV and wind power generation dataset in Xinjiang, China. The experimental results demonstrate that the proposed DTCN-FFT method outperforms other methods in accuracy. Taking the results of Xinjiang’s fourth quarter solar PV power generation prediction as an example, compared with Transformer, DLinear, TCN-GRU (Gated Recurrent Unit), TCN-LSTM (Long Short-Term Memory), TCN, and TCN-FFT, the Mean Absolute Error (MAE) of this method is reduced by 94.82%, 93.21%, 69.79%, 71.91%, 46.52%, and 31.85%, respectively.