An Efficient Hybridization of Gaussian Processes and Clustering for Electricity Price Forecasting

Electricity retailers and power generators have an increasing potential to profit from selling and purchasing electricity as wholesale electricity prices are encouraged to be introduced to both industrial and domestic customers. Hence, this paper focuses on developing an efficient method to aid decision-makers in forecasting the hourly price of electricity 4 weeks ahead. The method developed in this paper uses an approach to hybridize Gaussian Process (GP) regression with clustering to improve the predictive capabilities in electricity price forecasting. By first clustering all the input data and introducing a cluster number as a new input variable, the GP is conditioned to aid predictive process through data similarity. This proposed method has been successfully applied to real electricity price data from the United Kingdom, comparing the predictive quality of the novel method (GPc1) to that of an original GP and that of a method which pre-clusters and filters the data for numerous GPs (GPc2). By comparing the predictive price distributions to the observed prices in the month of December 2018, it was found that clustering improves the predicted mean values of a GP while the mean predictive quality of GPc1 and GPc2 are of equal standing. Therefore, the number of outliers at 2 STD's were compared, showing GPc1 to have a predicted distribution with uncertainty that covers more of the true electricity prices than that of GPc2. In conclusion, the novel method provides the decision-maker with greater reliability so that the true electricity prices will be within the confidence limits predicted.