TF-LiteNet: a time–frequency fusion lightweight network for multi-source load and renewable forecasting in virtual power plants

With the rapid integration of renewable and distributed energy resources, Virtual Power Plants (VPPs) have emerged as a key paradigm for the future power system. By aggregating heterogeneous resources for coordinated scheduling and market participation, the operation of VPPs relies heavily on accurate forecasting of load and renewable outputs. However, traditional statistical models fail to capture nonlinear characteristics of multi-source time series, while single-domain approaches in either time or frequency remain limited in accuracy and robustness. To address these challenges, this study focuses on the joint forecasting of load, photovoltaic (PV), and wind power time series for VPPs, where accurate multi-source prediction is essential for reliable scheduling and market participation. We proposes TF-LiteNet , a time–frequency fusion deep learning framework tailored for multi-source forecasting in VPPs. The framework incorporates a parallel time-domain encoder and frequency-domain encoder to extract long-term dependencies, local dynamics, and non-stationary components at different frequency bands. A bilinear interaction module with energy-aware fusion enables deep integration of temporal and spectral features, while a closed-form readout layer enhances efficiency and stability. Experiments on the real-world dataset demonstrate that the proposed method consistently outperforms state-of-the-art models in forecasting load, photovoltaic, and wind generation. The results confirm the framework’s effectiveness across both short- and long-horizon tasks, providing reliable predictive support for VPP scheduling and market participation.