Bounded residual Temporal Convolutional Network for multi-site day-ahead numerical weather prediction global horizontal irradiance correction

Day-ahead numerical weather prediction (NWP) forecasts of global horizontal irradiance (GHI) often exhibit systematic errors. A common correction strategy assumes that NWP forecast error behaviour is sky-type-specific, with errors being relatively homogeneous within a given sky type and sufficiently distinct across different sky types. Previous work has shown that sky-type-based grouping can support targeted correction. However, through a cross-site diagnostic analysis, this study finds that the assumed error structure is not consistently observed across locations. To address this limitation, a bounded-residual Temporal Convolutional Network (BRTCN) is proposed for post-processing day-ahead NWP GHI forecasts. The model learns constrained residual adjustments relative to the original forecast using NWP GHI, near-surface air temperature, relative humidity, and time-of-day context, producing a corrected 24-hour GHI sequence for the target day. Experiments are conducted using four years of hourly data from multiple sites, with forecasts from two NWP models and evaluation against ground observations. All methods are evaluated on identical prediction days, including the proposed model, raw NWP forecasts, statistical and machine-learning baselines, plain deep-learning baselines, and benchmark methods. On average across four sites, BRTCN reduces root mean square error from 78.38 to 74.68 W/m 2 and mean daily root mean square error from 66.28 to 61.31 W/m 2 relative to the clustered benchmark. It achieves the lowest errors at all sites across all comparison methods. These results show that clustering can improve correction performance at selected sites, but its benefit is not consistent across locations. BRTCN provides a robust alternative by maintaining consistent gains without requiring a clustering stage.