Comparative study of learning machine predictors for half-hour and day-ahead electricity price forecast in deregulated markets

The electricity markets all over the world have received several changes since the arrival of smart meters, real time pricing and deregulation waves. In such conditions, electricity price is a key index that reflects competition between generators, percentage of renewable generation and load behavior. An accurate forecast of the future price allows both producers and customers to plan their strategies, increase their profits and avoid contingencies. For this purpose, this paper proposes two predictors that perform half-hour-ahead and day-ahead price forecast, respectively. Both predictors are based on three learning machines, namely artificial neural network, support vector regression and random forest, in addition to persistence model which is used as reference. The built models are tested through real historical time series from New South Wales in Australia, which include half-hourly price and load. The profiles are analyzed with the aim of defining correlations and choosing adequate inputs. Then, the performance of each machine is assessed using the mean absolute percentage error, in order to determine the most suitable prediction machine according to the case. Furthermore, the best forecast situations are plotted to evaluate the ability of predicting sudden and sharp price peaks.