Interval Forecasting of Hourly Electricity Spot Prices using Rolling Window Based Gaussian Process Regression

Electricity price forecasting is important to the energy companies in planning and decision making. Gaussian process (GP) regression is a powerful tool for probabilistic forecasts of time series data. In this paper, we employ GP regression for prediction interval (PI) based forecasting of electricity spot prices. At each hour of the day, a new parameter set is computed incorporating most recent available electricity price data. We compare performance of several kernels. Likelihood ratio (LR) test statistics are used to measure goodness of the out-of-sample forecasts. Results show that our scheme outperforms other schemes in literature. In one case, LR statistics are slightly better for an existing quantile regression averaging (QRA) based scheme .But QRA scheme employs 12 other forecasting schemes followed by performing regression on the forecasts by those 12 schemes. However, our results significantly better than other individual forecasting schemes such as ARX/SNARX and averaging schemes such as SIMPLE/LAD.