Leveraging Spatio-Temporal Autocorrelation to Improve the Forecasting of the Energy Consumption in Smart Grids

Smart grids are networks that distribute electricity by relying on advanced communication technologies, sensor measurements, and predictive methods, to quickly adapt the network behavior to different possible scenarios. In this context, the adoption of machine learning approaches to forecast the customer energy consumption is essential to optimize network planning operations, avoid unnecessary energy production, and minimize power shortages. However, classical forecasting methods are not able to take into account spatial and temporal auto-correlation phenomena, naturally introduced by the spatial proximity of consumers, and by the seasonality of the energy consumption trends.In this paper, we investigate the adoption of several solutions to take into account spatio-temporal autocorrelation phenomena. Specifically, we investigate the contribution provided by the explicit representation of temporal i-nformation related to historical measurements using multiple strategies, as well as that of simultaneously predicting multiple future consumption measurements in a multi-step predictive setting. Finally, we investigate the effectiveness of injecting descriptive features to make the learning methods aware of the spatial closeness among the consumers.The experimental evaluation performed on a real-world electrical network demonstrated the positive contribution of making the models aware of spatio-temporal autocorrelation phenomena, and proved the overall superiority of models based on the multi-step predictive setting.