An Extensible Framework for Short-Term Holiday Load Forecasting Combining Dynamic Time Warping and LSTM Network

Abstract

Due to the extreme change of human behavior, the load consumption in public holidays fluctuates more significantly compared to general weekdays resulting in the difficulty of hourly holiday load forecasting. The holiday load forecasting is even challenging because the forecast is practically predicted on the nearest workday which might be more than one days prior to the public holiday. In this paper, we propose a Joint Dynamic time warping and LSTM (JDL) framework, to predict the hourly holiday load consumption on the nearest workday which is at least one day before the incoming holiday. The proposed JDL is a hybrid short-term holiday forecasting framework which combines dynamic time warping (DTW) and long-short term memory (LSTM) network. The DTW predicts the load consumption of the nearest workday and any preceding compensatory holiday(s), if any, based on the similar holiday occurrence pattern. LSTM predicts the highly unpredictable load consumption of the target holiday by univariate and multivariate models. Current results show the proposed JDL outperforms others.

Keywords

• Holiday load forecasting
• short-term load forecasting (STLF)
• long-short term memory (LSTM)
• dynamic time warping (DTW)