When Does Cross-Zonal Learning Help? Graph Neural Network Forecasting of Electricity Demand Across Italy’s Bidding Zones

Zonal electricity demand forecasts underpin market clearing, balancing and demand-side management, yet a market’s bidding zones are not independent: their demand co-moves through shared weather, economic activity and calendar effects. This paper asks when learning jointly across zones improves day-ahead forecasting, using Italy’s seven bidding zones as a case study. A GraphSAGE graph neural network, trained jointly across zones, is benchmarked against per-zone multilayer perceptron, long short-term memory and seasonal-naive models over 2021 to 2024, and against a matched global multilayer perceptron trained jointly across zones without graph aggregation, under a calendar split and three chronological partitions with Diebold–Mariano testing. Joint cross-zonal training lowers average day-ahead MAPE from 6.29% for the matched per-zone model to 5.58%, a gain that is stable across partitions; against the strongest per-zone model, the graph model’s improvement is significant in six of seven zones. The graph and pooled implementations reach that level equally, so the gain is attributable to learning across zones rather than to the graph specifically; at equal accuracy the graph model uses a sixteenth of the parameters, carries lower absolute error, and retains an inductive architecture that can accommodate changes in the zone set without changing the model input dimensionality. The benefit is specific to the day-ahead horizon and is largest under low demand, where a zone’s own recent history is least informative. Performance is insensitive to the specific graph topology in this seven-zone system. A monthly fixed-effects panel indicates that industrial activity co-moves with demand where industry is concentrated, while tourism acts largely through the summer cycle that seasonal terms already capture. The results identify when cross-zonal learning is worthwhile for zonal electricity-market operation.