Electricity demand forecasting has been a popular topic for quite some time, as electricity cannot be stored in large quantities, whereas the stability of the electric power system requires a constant balance between electricity produced and consumed. With the deregulation of the electricity market, price forecasting has become valuable. This master’s thesis offers an approach to electricity price and demand forecasting based on ARMA-GARCH models (time series) and support vector machines (machine learning). We focus on the day-ahead electricity prices. In both electricity price and demand, the analyzed data are available hourly, and our models provide 24-hour forecasts for the next day.
The crucial part in the elaboration of the forecasts is the model selection and identification. Both price and demand are predicted using only past values of price and demand, without additional external variables. The selection of models follows the Box-Jenkins methodology, a standard tool in time series analysis. We focus on a few (similar) models in each approach, and the optimal parameters are selected using different criteria. For time series models, these are the AIC criterion and the Ljung-Box test, and for the support vector machines, we use cross-validation.
Finally, we predict the future values of electricity price and demand with the previously selected models. We compare the models based on different measures for forecast accuracy (MAPE, RMSE). In both cases, the results show that the predictions of the ARMA-GARCH models are better for the given data. In the case of electricity demand, no model is considerably better than the rest, and in the case of price, the ARMA-GARCH-t model stands out.
|
