Probabilistic estimation framework on short term electricity load forecasting via parametric and nonparametric approaches

Ergin, Elçin
Accurate electricity load forecasting plays a crucial role for all the electricity market parties. With the deregulation of the markets, electricity load forecasting, especially for short-term, gained a great importance. Electricity load data may show different characteristics according to country, region or customer type as it has a nonlinear relationships with many of variables. Capabilities of only one model may not be enough to capture all these relationships. In this study, parametric, nonparametric and hybrid approaches are developed and employed for short-term electricity load forecasting on hourly load series from Turkish electricity market which is not studied before. These approaches include Double Seasonal Autoregressive Integrated Moving Average (DSARIMA), Nonlinear Autoregressive with Exogenous Inputs Neural Networks (NARX) and ε-Support Vector Regression (ε-SVR) and hybrid models DSARIMA-NARX and DSARIMA-ε-SVR. Additionally, we apply Local Quantile Regression (LQR) method both to combine the results obtained from the other approaches and more importantly construct probabilistic forecasts instead of providing point forecasts. Considering the multiple seasonal cycles existing in load series, examining the load data with grouping is another option to improve forecast accuracy. We also construct 4 different grouping scheme and compared their performances with the whole data results.