Exploiting electric vehicles for maintaining a stable electricity grid is a promising opportunity, which is recognized by both the academic and industrial world. Although the concept is relatively simple, the operating space containing many different stakeholders, rules and legislation is not. There are various practical issues that complicate large scale implementation, which are not yet addressed in scientific literature. One such issue is that aggregators who engage in energy trading and have the desire to capitalize on flexible power from electric vehicles by active control, need forecasts and real-time information of uncontrolled charging power. The main problem is that under active control of charge sessions, information of uncontrolled charging is lost as energy meters measure the result of such control signals. This thesis proposes an innovative solution for which a patent has been filed, to reconstruct uncontrolled charging profiles from historical data and to forecast future values of the load for a large group of charge stations as input for trading on energy markets.
Three forecasting methods were applied to forecast future values of the uncontrolled load in two different scenarios, e.g. day-ahead and intraday. In the day-ahead scenario, the 24 hours of the next day are predicted, but predictions are to be submitted 12 hours in advance due to the day-ahead market clearing, such that our models need to predict 36 hours into the future. In the intraday scenario, our models need to predict the load for the next 15 minutes.
For forecasting, the following methods were used: a SARIMAX model, Singluar Spectrum Analysis with a Linear Recurrent Formula and a Nonlinear Autoregressive Gaussian Process model. The forecasting accuracy of these models was tested by evaluating the out-of-sample predictions using the mean absolute percentage error and root mean squared error. As a guideline for acceptable accuracy, we require the mean absolute percentage error to be below 10 (percent).
In the day-ahead scenario, for weekdays the best performing model is the SARIMAX(2,0,1)(2,1,0)24$ which returned a MAPE of 4.57% and RMSE of 236 (rounded) for a specific day in March based on three weeks of data. The AR-GP recurrent with order p = 28 followed with an acceptable result with MAPE 6.68% and RMSE 330. The SSA model scored significantly lower than both methods, with a MAPE of 9.47% and RMSE of 413. Cross-validation showed that the current models were unable to adequately predict the load during weekends. The models were retrained by selecting different model orders based on a new training data set, which contained only Saturdays or Sundays and we saw the AR-GP direct models now outperform the SARIMAX model, which was the best model for week days, with MAPE of 4.87% and 6.84% and RMSE 213 and 309 for Saturday and Sunday respectively.
For the intraday scenario, on weekdays, the AR-GP with recurrent forecasting (p = 16) was the best performing model and returned a MAPE of 5.73% and RMSE of 220.
As a general conclusion, using the proposed method we can reconstruct the uncontrolled load for thousands of charge stations and using a combination of different forecast models, our forecasting accuracy goal for all different scenarios is achieved.