Quantifying uncertainties of neural network-based electricity price forecasts

Abstract

Neural networks (NNs) are one of the most widely used techniques in literature for forecasting electricity prices. However, nonzero forecast errors always occur, no matter what explanatory variables, NN types, or training methods are used in experiments. Persistent forecasting errors warrant the need for techniques to quantify uncertainties associated with forecasts generated by NNs. Instead of using point forecasts, this study employs the delta and bootstrap methods for construction of prediction intervals (PIs) for uncertainty quantification. The confidence level of PIs is changed between 50% and 90% to check how their quality is affected. Experiments are conducted with Australian electricity price datasets for three different months. Demonstrated results indicate that while NN forecasting errors are large, constructed prediction intervals efficiently and effectively quantify uncertainties coupled with forecasting results. It is also found that while the delta PIs have a coverage probability always greater than the nominal confidence level, the bootstrap PIs are narrower, and by that, more informative.