A neural network-GARCH-based method for construction of Prediction Intervals

A novel hybrid method for construction of high quality Prediction Intervals (PIs) for electricity prices is proposed in this paper. The proposed method uses moving block bootstrapped neural networks and Generalized AutoRegressive Conditional Heteroscedasticity (GARCH) models for forecasting electricity prices and estimation of their variance. Rather than employing the traditional maximum likelihood estimation method, parameters of the GARCH model are adjusted through minimization of a PI-based cost function. Experiments are conducted using hourly electricity prices of Australian and New York energy markets. Demonstrated results indicate that the proposed method generates high quality PIs with a narrow width and a large coverage probability. It is shown that the narrow variable-width PIs constructed using the proposed method are more informative than the fixed-width PIs constructed using the traditional methods. Also, the proposed method is computationally hundreds of times faster than its traditional rivals.

► Quantification of uncertainties associated with electricity price forecasts.
► A new neural network-GARCH-based method prediction interval construction.
► Comprehensive assessment of prediction interval quality.
► Quality Prediction Intervals for price forecasts.