Very-short term forecasting of electricity price signals using a Pareto composition of kernel machines in smart power systems

Abstract

In smart power distribution systems price forecasting is an indispensable participant tool for developing purchase strategies. This paper places itself in price directed power systems, where participants respond with their load demand to incoming price signals. To that end, a two hour-ahead forecasting method using a linear composition of kernel machines for electricity prices is introduced. Initially, two kernel machines, i.e., a Gaussian process (GP) and a relevance vector machine (RVM), are utilized for next-two-hour prediction making. Subsequently, a linear predictor composed of the two kernel machines, which are both equipped with the Gaussian kernel, is built that integrates the individual predictions to provide a single one. The linear coefficients are obtained as the solution of multiobjective problem that is sought by a genetic algorithm utilizing Pareto optimality theory. Method testing is performed on a set of historical data obtained from the New England area. The proposed Pareto optimal kernel machine composition outperforms in terms of accuracy the autoregressive moving average (ARMA) predictor as well as the individual kernel machines in the vast majority of the tested cases.