Short-term electric load forecasting for buildings using logistic mixture vector autoregressive model with curve registration

This paper proposes a day-ahead electric load forecasting model for buildings where daily load curves follow a few distinctive patterns. A pattern lasts for several days before changing into another. We particularly explore the problem that the day-ahead curve mostly depends on the load pattern history and is relatively insensitive to external environments such as weather conditions. The problem considers clusters for daily curve patterns and a day-ahead electric curve forecast from previous electric load and pattern history. We propose a model called the logistic mixture vector autoregressive model (LMVAR) that combines both clustering and forecasting in a single model through the expectation–maximization (EM) algorithm. To improve internal clustering performance, we apply the curve registration technique to the model. We test two models (the models with/without curve registration) with electric load data sets collected from a library and a grocery store. We then compare them with existing forecasting methods such as persistence, sequence-to-sequence long short-term memory (S2S LSTM), seasonal autoregressive (SAR), multiple-output support vector machine (M-SVM), multilayer perceptron (MLP), and a cluster-based model. The result shows that the proposed models outperform the benchmark methods.