Univariate water consumption time series prediction using deep learning in neural network (DLNN)

The ability to predict water consumption benefits in terms of future temporal water patterns, operational reasons in distribution and decision making related to the development of water supply systems. Although, there have been numerous research work on time series forecasting, the methodology may still need further development in the context of non-linear modeling and analysis. There is many unpredictability that results in complex temporal dependence with traditional time series approaches. But now, artificial neural networks are available to help resolve these issues can improve model prediction. Therefore, this paper presents an experimental study to perform univariate water consumption for time series prediction processes using Deep Learning in Neural Network (DLNN). DLNN has abilities to learn the algorithm from previous data. It also able to process an explosion of data with a nonlinear approach, which the univariate water consumption time series prediction problem was modelled into artificial neural networks. The vast majority of learning algorithms along with DLNN are used in this study called Multilayer Perceptron (DLNN-MLP), Convolutional Neural Networks (DLNN-CNN) and Long Short-Term Memory (DLNN-LSTM) for the model prediction to generate prediction with the lowest error. Dataset from the residential water usage in SIBU Division was chosen for the experiments of this study. The performance of the DLNNs' models is measured using RMSE and state-of-the-art algorithms. The value of RMSE for a problem relates to the calculation of the lowest error. Based on the experiments result indicates that the DLNN-LSTM (RMSE:0.051) can make decent predictions for water consumption time series despite being inferior to SARIMA (RMSE:0.183). Hence, this finding implies that DLNN-LSTM can be implemented on a univariate water consumption time series prediction problem given that the problem is modelled as supervised learning.