Online incremental regression for electricity price prediction

Modeling methods aiming at predicting electricity price accurately, should be capable of handling a continuous stream of data while keeping responsive to the potential structural changes. To this end, traditional machine learning based approaches are widely applied such as Multi-linear Regression, Artificial Neural Network (ANN), Time Series Models like Auto Regressive Moving Average Models (ARMA), Gaussian Process (GP), random forests and Genetic Algorithm (GA), all of which can fall into two categories: the parametric and non parametric model. While practical challenges in forecasting streaming data come along with the structural variation of the testing samples making the training samples not necessarily representative enough towards the new arriving samples. In such an online forecasting context, an incremental supervised learning based algorithm is better suited in contrast to the batch-mode one. Given the fact that it can adapt to the new coming streaming data by accommodating the possible variations of new samples, as well as allows for the removal of old data if necessary. An incremental learning algorithm is presented in this paper, i.e. the online support vector regression model, which enjoys the merits of less memory capacity and less computational overload compared with the batch methods. Promising results are demonstrated by evaluating with other typical regression methods for the electricity price forecasting task on a publicly available benchmark data set.