Forecasting Methodologies of the Electrical Load in Urban Distribution Networks: a Case Study in Milan, Italy

The electrification of end-uses is causing a substantial increase in electrical demand in urban distribution systems. In this framework, an accurate forecast of load consumption patterns plays a key role in ensuring an efficient and reliable system operation, proper planning of grid infrastructures, and reduction of operational costs. To this purpose, this work introduces different methods to predict the electric consumption of the medium voltage feeders of a primary substation: two approaches based on machine-learning theory (Random Forest and Generalized Boosting Regressor) and one statistical forecasting model (functional Principal Component Analysis). An extensive analysis of their performance has been conducted considering the urban scenario of Milan as a reference. The methods are trained on a specific database that includes the electric demand of the last three years of a portion of the Milan metropolitan area, as well as exogenous variables such as auto-regressors, meteorological and calendar variables. Among the selected models, the functional Principal Component Analysis provided the most accurate prediction, reducing the index of error compared to the machine-learning models.