Time series forecasting of hourly water consumption with combinations of deterministic and learning models in the context of a tertiary building

The search for a model to provide an accurate prediction of water consumption is one of the major challenges in water supply systems. Auto-Regressive Integrated Moving Average (ARIMA) with and without seasonality combined with an Artificial Neural Networks (ANN) represent one of the most popular hybrid models for time-series forecasting. Actually, these models have recently demonstrated success in water consumption forecast. However, each rely on assumptions and show some limitations. So, this study proposes several new hybrid models that combine the ARIMA with seasonality i.e., SARIMA, neural approaches like the Long Short-Term Memory (LSTM) or the Multi Layer Perceptron (MLP), and a deterministic model based on a time function. These different hybrid models that combine individual models are used to predict hourly water consumption of a tertiary building. The experiments show that the resulting hybrid model with the time series deterministic model, the ANN and the SARIMA is efficient by improving the accuracy of the water consumption prediction. Indeed, this hybrid model allows to reduce the error of 8.24 % and 5.53 % in the mean of training and testing errors respectively, compared to all other individual or combined models.