TFDN: A reliable hybrid time and frequency domain-based model for photovoltaic power generation time series forecasting

Qiang Hua, HaoRan Chu, Feng Zhang, Yong Zhang, Chunru Dong

Abstract: The rapid expansion of photovoltaic (PV) power generation has highlighted the critical need for accurate PV power forecasting. This paper proposes a novel deep learning-based model for PV power forecasting that effectively integrates time and frequency domain information to achieve more accurate predictions. Within the time-domain module, the input data undergoes an initial decomposition into cycle segments using the Fast Fourier Transform (FFT). These segments are then processed through a mask-based multilayer perceptron and a parameter-efficient inception block, designed to more easily capture both global and local dependencies within the data. In parallel, a distinct module leverages multiscale convolutional kernels of varying sizes to model cross-dimensional dependencies among different variables in the PV data. In the frequency-domain module, a low-pass filter removes high-frequency noise from the frequency components extracted by the FFT. The filtered components are then fed into a complex linear layer to perform linear interpolation, generating the forecasted frequency components. The inverse FFT subsequently converts the frequency components back into the time domain. This process leverages frequency-domain processing to mitigate information loss in the time domain, while also enabling the integration of both time-domain and frequency-domain information for a more comprehensive analysis of the data. Experimental results demonstrate that the proposed model provides accurate PV power forecasts across three sites and outperforms state-of-the-art models.

External IDs:dblp:journals/mlc/HuaCZZD25