Prediction of Water Consumption Using Machine Learning Approach

Abstract

Water is essential and the main element for any living thing. Water consumption is the portion of water that is used and not returned to the original water source after being used. As population size increases, the demand and consumption of water also increase. Proper management and understanding of how much water is consumed in a specific city as well as in one’s community is very important to ensure the sustainable growth of the economy. This is not only used to affirm development but also to understand the scarcity of water. As far as the researcher's knowledge of Water Sewerage Corporation, there is a lack of decision making to forecast water demand and consumption for the future from an existing set of data. Furthermore, no study has been conducted to predict monthly water consumption in Ethiopia, as well as predictions of very short time, such as hourly and daily predictions. The data for this study was collected from Wolaita Sodo town, and it covers a four-year water consumption record from 2012–2015 (up to May) E.C. It is important to develop a predictive model of water consumption by using machine learning techniques. This study has been conducted using an experimental approach to determine the best-performing model and used the Python programming language for implementation and analysis purposes with its libraries. The developed prediction model is used to reduce the burden of estimation tasks for water supply officers, to get adequate water service, and for other concerned parties. For this study, the researchers developed both multivariate and univariate time series prediction models. The data used for this study was time series data, and after preprocessing, the number of instances used for model development is 7134 with 45 columns (months). The model for time series prediction was done using a percentage split of 90% for training and 10% for testing. The random forest performed better for multivariate time series prediction with an R-squared of 0.944, MSE of 0.093, and MAE of 0.169, while exponential smoothing better performed in univariate time series prediction with an MSE of 0.22, MAE of 0.188, and an R-squared of 0.29 on test time series data