Graph Neural Network Method for Integrating Spatial-Temporal Data in Electricity Demand Forecasting

Forecasting electrical energy demand with machine learning models using only temporal information fails to capture the dynamics and interrelationships between different types of electrical grid customers. This research proposes incorporating spatial information through graph neural networks constructed from a GIS. These graphs are based on customer locations, selecting an objective node and nearby neighbors within a specified radius, with edges weighted by direct connections and inverse geographical distance. Experiments forecast demand at 1, 3, 6, 12, and 24-hour horizons for residential, commercial, and official customers, using graphs with radius values of $\mathbf{2 5 0} \mathrm{m}, 500 \mathrm{~m}$, and 750 m. The MPNNLSTM and GCLSTM models outperform the baseline CNNLSTM model, with MPNNLSTM performing well at 6,12, and 24 -hour horizons, and GCLSTM at 1 and 3-hour horizons. Using the radius to define the graph demonstrates improved forecasting accuracy by including spatial information.