GridCast: Weather-Conditioned Short-Term Load Forecasting for Distribution Networks with Sparse Meter Coverage

Abstract

Accurate short-term electricity demand forecasting is challenging in distribution networks where smart-meter coverage is incomplete and local weather conditions change rapidly. This paper proposes GridCast, a weather-conditioned forecasting model that combines historical electricity consumption with temperature, humidity, solar radiation, calendar information, and regional load observations. Instead of imputing all unavailable meter readings independently, GridCast learns a shared representation across geographically related substations and estimates local demand from both temporal and regional dependencies. The model is evaluated using 18 months of measurements from 137 distribution substations, containing approximately 9.6 million 15-minute load observations. Across 1-hour, 6-hour, and 24-hour forecasting horizons, GridCast achieves average MAPEs of 2.91%, 4.26%, and 5.83%, respectively. For 24-hour forecasting, this represents a 17.4% reduction in forecasting error compared with LightGBM and an 11.2% reduction compared with a standard temporal Transformer. When 30% of local meter observations are removed, the proposed model maintains a MAPE of 6.47%, compared with 8.92% for the strongest baseline. The results demonstrate the value of regional information for demand forecasting under incomplete monitoring conditions.