Vehicle-to-grid plug-in forecasting for participation in ancillary services markets

TL;DR

This paper develops and compares neural network models to forecast aggregate vehicle-to-grid plug-in levels for electric vehicles, aiding participation in ancillary services markets with improved accuracy.

Contribution

It investigates feature dependencies and introduces a neural network model incorporating historic, calendar, and weather data for accurate V2G plug-in forecasting.

Findings

Neural network considering multiple historic points performs best.

Calendar variables significantly improve forecast accuracy.

Forecasting model achieves high accuracy for day-ahead planning.

Abstract

Electric vehicle (EV) charge points (CPs) can be used by aggregators to provide frequency response (FR) services. Aggregators must have day-ahead half-hourly forecasts of minimum aggregate vehicle-to-grid (V2G) plug-in to produce meaningful bids for the day-ahead ancillary services market. However, there is a lack of understanding on what features should be considered and how complex the forecasting model should be. This paper explores the dependency of aggregate V2G plug-in on historic plug-in levels, calendar variables, and weather conditions. These investigations are used to develop three day-ahead forecasts of minimum aggregate V2G plug-in during 30-minute window. A neural network that considers previous V2G plug-in values the day before, three days before, and seven days before, in addition to day of the week, month, and hour, is found to be the most accurate.