Forecasting Groundwater Consumption for an Urban Environment Using Machine Learning Techniques

Abstract

Poor water management systems in major cities can be addressed by water consumption forecasting using multiple factors e.g., climatic data, population count, and water requirement. In this study, a primary dataset is obtained including water consumption of a 3 sq. kilometres urban site in Pakistan over the course of 7 years along with environmental variables like temperature, precipitation, humidity, wind speed, and population. The results indicate that time-series modelling is the best approach for forecasting problems that include environmental variables like temperature, precipitation, humidity, wind speed, population, and water consumption. Three distinct machine learning models, namely artificial neural network, Long Short-Term Memory (LSTM) models, and transformers, were rigorously evaluated. In terms of accurately forecasting urban water demand and supply, the proposed architectural framework of transformer models outperformed the other models, according to the evaluation results. The LSTM model has an R2 score of 0.31 for predicting monthly water consumption, whereas the transformer performed exceptionally well with an R2 score of 0.98. For further substantiation, annual water consumption forecasts are made for the transformer whose R2 score was 0.917. The proposed model has been successfully employed to forecast water consumption in all four seasons indicating that it is impactful for sustainable water resource management in an urban environment.