A time–frequency collaborative network with endogenous–exogenous variable decoupling for enhanced photovoltaic power forecasting

The growing global energy demand and rising computational loads in the artificial intelligence era have intensified the need for reliable and flexible power systems. Solar photovoltaic (PV) generation, as a key renewable energy source, presents unique forecasting challenges due to its complex temporal dynamics and sensitivity to external environmental factors. However, most existing methods encode endogenous and exogenous variables in a unified manner, neglecting their intrinsic differences, and rarely integrate time-domain and frequency-domain features within a cohesive framework. To address these limitations, this paper proposes a novel time-frequency collaborative forecasting model that employs separate encoding pathways for endogenous and exogenous variables to better capture their distinct characteristics. The model segments endogenous time series for localized temporal modeling and embeds exogenous variables as independent tokens, enhancing robustness to missing or incomplete inputs. Concurrently, a discrete Fourier transform is applied to the endogenous sequence to extract complementary spectral features. A gating-enhanced Kolmogorov-Arnold network then dynamically fuses time- and frequency-domain representations and maps the integrated features to the target forecasting horizon. Extensive experiments on real-world PV datasets demonstrate that the proposed model achieves superior forecasting accuracy compared to state-of-the-art baselines. Code is available at this repository: https://github.com/laowu-code?tab=repositories.