Corrected Support Vector Regression for intraday point forecasting of prices in the continuous power market

In this paper, we develop a new approach to the very short-term point forecasting of electricity prices in the continuous market. To this end, we introduce a fast, nonparametric method based on the Support Vector Regression with a kernel correction built on an alternative forecast of the dependent variable. It allows for improving forecast accuracy by leveraging information from already strong predictors. Moreover, the kernel parameters are calculated using the distribution of the input data. We test the proposed approach on minutely volume weighted average transaction prices from the German intraday continuous market and compare its forecast accuracy with several benchmarks: classic SVR, LASSO model, Random Forest and the naïve forecast. The analysis is performed for different forecast horizons, deliveries, and lead times. Overall, the proposed cSVR approach yields the highest forecast accuracy among the considered benchmarks and is computationally feasible. The greatest improvement in forecast accuracy is observed for deliveries during the morning and evening peaks.