Automatic selection of fittest energy demand predictors based on cyber swarm optimization and reinforcement learning

Effective demand forecasting is essential for regulating power distribution, scheduling production, and initiating new energy projects. Existing forecasting models have contrasting features and manifest various types of errors. This paper proposes a multi-predictor approach which applies reinforcement learning for selecting the fittest predictors to enhance the collaborative performance. A new univariate predictor is developed based on the Gaussian mixture and phase shifting and rescaling techniques, and two multivariate predictors are developed from a landscape analysis with potential econometrics. Each individual predictor is trained by the cyber swarm algorithm (CSA) to find the optimal parameter values. The 10-fold cross validation for regression parameter optimization by using CSA and the constriction factor particle swarm optimization (CFPSO) shows the effectiveness of the former against the latter. Our reinforcement learning forecasting method is able to automatically select the best predictor to perform at various time instances and allow the embedding predictors to complement one another. Our experimental results experimented with Taiwan's electricity demand time series during 2001-2014 show that the prediction improvement contributed by the proposed approach over the original individual predictors is significant in terms of the mean absolute percentage error (MAPE) and the mean square error (MSE).