On the parametric and nonparametric prediction methods for electricity load forecasting (original) (raw)

Accurately forecasting of electricity load is of great importance in deregulated electricity markets. Market participants can reap significant financial benefits by improving their electricity load forecasts. Electricity load exhibits a complex time series structure with nonlinear relationships between the variables. Hence, new models with higher capabilities to capture such nonlinear relationships need to be developed and tested. In this thesis, we present a parametric and a nonparametric method for short-term load forecasting, and we compare the performance of these models for different lead times ranging from one hour to one week. These methods include a modified version of Holt Winters Double Seasonal Exponential Smoothing (m-HWT) model and a Nonlinear Autoregressive with Exogenous Inputs (NARX) neural network model. Using hourly load data between 1/1/2008 to 1/1/2013 from the Dutch electricity grid, an extensive empirical study is carried out for five different provinces. Due to the promising results, in the second part of the study, NARX is applied to long-term load forecasting in one Dutch province. Our results indicate that NARX clearly outperforms m-HWT in one-hour ahead forecasting. Additionally, our modification to HWT leads to significant improvement in model accuracy. Despite its simplicity, m-HWT outperformed NARX for 6, 12, 24, and 48-hour ahead forecasts. However NARX performs better in 1 week ahead forecasting. In addition, NARX performs clearly superior to m-HWT in terms of maximum error and on special days. The results also indicate that with a well-trained closed loop NARX neural network model, electricity load can be forecasted successfully one and a half years ahead for hourly intervals. NARX can successfully capture nonlinear effects of special days and temperature. NARX has brought a performance improvement of 30% in terms of mean absolute percent error (MAPE) compared to the existing methodology. In the long-term forecasting part of the study, a closed-loop network is trained for one pilot province Brabant, which is the largest grid that Enexis owns. In this phase, temperature and sunlight are added to the input set. Architecture search is presented in detail, which was limited to two hidden layers for long-term load forecasting. Results indicate an improvement of approximately 28% compared to current methodology in terms of MAPE. MaxAPE levels almost halved with NARX forecasting compared to time shifting and regression models. Additionally for both lead times we observed that NARX performs very well in load forecasting on special days. Special days are known as most difficult periods of the year to forecast due to divergence from regular pattern on those days. As expected NARX performed good at capturing complex relationships on these days and outperformed the conventional models.