Short-Term Electric Load Forecasting Via State-State Approach with a Mixture of Long-Lag Vector Autoregressive Models

Abstract

This paper proposes a novel day-ahead electric load forecasting model for buildings and regions. We explore the problem that the daily load curves of electric load data in buildings and regions have periodicity and follow some patterns. We use long-lag terms of historical load data to capture periodicity and clustering to figure out daily load curve patterns. We propose a forecast model called a state-space model with a mixture of long-lag vector autoregressive models (SSM-MLVAR), which combines a state-space model and long-lag vector autoregressive models into a single model. The proposed model ensures feasibility by reducing the number of parameters of the long-lag vector autoregressive model. The state-space model alleviates forecasting errors that cannot be fine-tuned with the vector autoregressive model. With electric load data from a building and 10 regions, we compare the proposed model with existing forecasting methods such as persistence, multiple-output support vector machine (M-SVM), cluster model (CM), logistic mixture vector autoregressive model (LMVAR), Bayesian mixture density network (BayMDN), and variational mode decomposition with multiple wavelet decomposition network (VMD-mWDN). It turns out that the proposed model achieves significantly higher prediction accuracy than these models on the datasets.

Keywords: Electric load, State-space model, Mixture of autoregressive models, Long lag, Short-term forecasting, Building and regions