SARIMA and neural network models combination for time series forecasting: Application to daily water consumption

Abstract

A water management strategy focused on forecasting of the future demand offers the opportunity to reduce water capture, storage and treatment costs. Consequently, it reduces the consumers billing. In this paper, we propose techniques for monitoring by forecasting water consumption in the context of smart building. We mainly use three individual models: seasonal autoregressive integrated moving average (SARIMA), artificial neural network (ANN) models such as long short-term memory (LSTM) and multilayer perceptron (MLP). We propose several new models of hybrid suits that mix individual models to find the most accurate solution to predict daily water consumption.We concluded that the combination of SARIMA and neural network models MLP seems better than individual models and provides more reliable and accurate results with 4.6% compared to the individual SARIMA model forecasting results which has 7.7%.