Comparative study of week-ahead forecasting of daily gas consumption in buildings using regression ARMA/SARMA and genetic-algorithm-optimized regression wavelet neural network models

Abstract

Forecasting energy consumption in buildings is crucial for achieving effective energy management as well as reducing environmental impacts. With the availability of large amounts of relevant data through smart metering, gas consumption forecasting is becoming an integral part of smart building design so that these requirements are met. In this study, we investigate week-ahead forecasting of daily gas consumption in three types of buildings characterized by different gas consumption profiles during a five-year period. As gas consumption in buildings is highly correlated with the average outdoor temperature, regression models with additional residual modeling are used for forecasting. However, conventional regression models with autoregressive moving averages (ARMA) errors (regARMA) perform poorly when the temperature forecasts are inaccurate. To address this, a new forecasting model termed genetic-algorithm-optimized regression wavelet neural network (GA-optimized regWANN) is proposed. It uses the wavelet decomposition of the residuals of temperature regression time-series, which are modeled by multiple nonlinear autoregressive (NAR) models based on sigmoid neural networks. The appropriate delays in the regression vectors of the NAR models are selected using a binary GA. Compared with regARMA and seasonal regARMA, the GA-optimized regWANN model achieved in the three buildings a reduction of 22.6%, 17.7%, and 57% in the mean absolute error (MAE) values in ex post forecasting with recorded temperatures, and a 52.5%, 27%, and 43.6% reduction in the MAE values in ex ante forecasting with week-ahead forecasted temperatures, even under conditions of relatively significant errors in the forecasted temperature.