Comparison of Short-Term Weather Forecasting Models for Model Predictive Control

Abstract

Model predictive control applied to commercial buildings requires short-term weather forecasts to optimally adjust setpoints in a supervisory control environment. Review of the literature reveals that many researchers are convinced that nonlinear forecasting models based on neural networks (NNs) provide superior performance over traditional time series analysis. This paper seeks to identify the complexity required for short-term weather forecasting in the context of a model predictive control environment. Moving average models with various enhancements and (NN) models are used to predict weather variables seasonally in numerous geographic locations. Their performance is statistically assessed using coefficient-of-variation and mean bias error values. When used in a cyclical two-stage model predictive control process of policy planning followed by execution, the results show that even the most complicated nonlinear autoregressive neural network with exogenous input does not appear to warrant the additional efforts in forecasting model development and training in comparison to the simpler MA models.