Reconstruction-Aided Short-Term Power Forecasting for Data-Limited Distributed PV Stations

Short-term forecasting is difficult at distributed photovoltaic (PV) stations that retain hourly energy records but lack power measurements at fine time intervals and site-specific irradiance. We propose a reconstruction-aided method for one-step-ahead forecasting under these conditions. Variational mode decomposition (VMD) separates the target-station energy sequence and the reference-station energy and power sequences into multiscale modes. Hilbert modal features and a multidimensional similarity score identify suitable reference modes, and extremely randomized trees (ExtraTrees) reconstruct the target station's 5-min power profile while preserving each recorded hourly energy total. For the irradiance input, the clear-sky index reduces periodic effects, VMD captures multiscale disturbances at observed grid cells, and ExtraTrees completes global horizontal irradiance (GHI) at the unobserved target cell. Reconstructed power, completed GHI, periodic time features, and aggregated neighboring-station power are then supplied to LightGBM to predict target power at the next time step. Tests use 2021 measurements from distributed PV stations in a UK region and CAMS irradiance on a 7 × 7 grid. The normalized mean absolute error was 0.377%, the normalized root mean square error was 1.168%, and the coefficient of determination was 0.990. The results indicate that the method can recover fine-resolution power and local irradiance from limited target-station information and improve short-term forecasts for data-limited distributed PV stations.