Short-term interval-valued load forecasting with a combined strategy of iHW and multioutput machine learning

Abstract

Interval-valued load forecasting is an important risk management tool for the utility companies and can provide more comprehensive and richer information to assist in decision-making. However, the existing literature mainly focused on point-valued load forecasting, neglecting the significance of interval-valued load forecasting. In this paper, we propose a combined framework based on interval Holt-Winters and multioutput machine leaning method to predict daily interval-valued load. Firstly, we improve the traditional Holt-Winters and propose interval Holt-Winters that takes account of the seasonal characteristics of daily load. Secondly, interval Holt-Winters is applied to predict daily interval-valued load series and obtain the forecasting results and residual series. Thirdly, multioutput machine learning models including multioutput support vector regression, interval multilayer perceptron and interval long short-term memory are employed to predict residual series and obtain the forecasting results of residual series, respectively. Finally, the final forecasting results of the daily interval-valued load are obtained by summing the forecasting results of interval Holt-Winters and residual series. Empirical results show that the proposed combined interval model outperforms the corresponding single interval model and has excellent robustness. Besides, compared with point forecasting models, the interval models have better performance.