Forecasting the aggregated charging load of electric vehicles at different charging sites

Abstract

The anticipated large-scale market ramp-up of electric vehicles in the future represents both an opportunity and a major challenge for the power grid. On the one hand, electric vehicles may support the power grid by offering their flexibility on the Control Reserve market. On the other hand, simultaneous charging could lead to severe grid bottlenecks in the low voltage network. In either case, the prediction of the aggregated charging load is of great importance. Within this work a new multi-variate multi-step forecasting approach based on Artificial Neural Networks, namely the Long Short-Term Memory, is introduced. Using historical charging data, a prediction of the aggregated charging load in 15-min resolution, clustered according to the charging at home, the charging at work, the charging at public car parks and the charging at shopping centers, is conducted. For each charging site two forecast horizons, a 1-hour and a 1-day charging load forecast, are analyzed. The results indicate that no reliable forecast of the charging load can be accomplished with a forecast horizon of one day and that the prediction accuracy can be significantly increased if the period is shortened to one hour.