Short-Term Demand Prediction Using an Ensemble of Linearly-Constrained Estimators

Abstract

The benefits of forecasting power demand can bring increased stability to any power grid. Between optimizing the production and control of grid resources and interacting with energy markets, there is a strong motivation for generation, transmission, and distribution grid stakeholders to obtain accurate power demand prediction, which requires more sophisticated prediction methods. We introduce an ensemble of linear predictive nodes called the Ensemble Prediction Network (EPN), which optimizes demand prediction motivated by various microgrid considerations. EPN outputs a nonlinear combination of the individual predictions whose mixing weights are optimized in the least-squares sense. Using a large number of publicly available datasets, we show that on-the-whole, EPN provides substantial improvement relative to each individual predictor. Furthermore, we compare our method with a Long Short-Term Memory (LSTM) neural network and a multi-layer perceptron, and demonstrate the advantages of the proposed method.