Multi-Source Data-Driven Day-Ahead Load Forecasting for a Provincial Power Grid: A Case Study in Hunan, China

Accurate day-ahead load forecasting at the provincial scale is a basic input to unit commitment, reserve scheduling and market settlement, but the forecast has to capture periodic load behaviour, spatially distributed weather sensitivity and historically conditioned operating regimes together. This paper develops a multi-source forecasting model that couples a TimeXer backbone with a full-grid encoder for five-channel archived ECMWF IFS HRES fields over the service territory, an availability-controlled similar-pattern retrieval whose output is mediated by a learned correction gate, and a validation-selected daily-cycle level correction. The model is evaluated on one year of 15-min Hunan provincial load under a chronological train–validation–test split with 61 daily test origins and five random seeds. It attains a 726.2 MW MAE, 2.04% MAPE, R2=0.9699 and 207.8 MW mean quantile loss, ranking first on a six-metric mean-rank criterion, with lower mean MAE and MQL values than the hyperparameter-tuned and level-corrected DLinear and iTransformer comparators. A staged ablation shows that the gridded weather field supplies the largest MAE reduction (308.6 MW), while the similar-pattern prior produces the largest mean-rank movement (3.17 to 2.25); the eight learned weather tokens receive 90.5% of the averaged cross-attention weight. Over the 22 origins spanned by five documented large-scale cooling events, the model retains the best mean rank among the twelve compared models that produce a predictive distribution. These results indicate that multi-source information fusion supports accurate and computationally feasible day-ahead forecasting for the Hunan provincial system.