The limitation of machine learning methods for water supply and demand forecasting: A case study for Greater Melbourne, Australia

Abstract

Many cities around the world have faced water scarcity due to climate change, population growth, and urbanization. Accurate water supply and demand forecasting is critical for sustainable urban water management. Machine learning (ML) models provide new possibilities for forecasting compared with traditional models in handling non-linearity. This study aims to address the efficacy of ML models, long short-term memory (LSTM), stacked LSTM, bidirectional LSTM (Bi-LSTM), and multilayer perceptron (MLP), for forecasting water supply and demand in Greater Melbourne, Australia. The ML modelling utilized daily water supply and demand, and climatic variables (rainfall and maximum temperature) recorded by Melbourne Water and the Bureau of Meteorology from 1990 to 2019. The stacked LSTM performs better than other models in forecasting with R2, RMSE, and MAPE values of 0.908, 335.74 ML, and 23.5% for water supply and 0.791, 94.88 ML, and 5.3% for water demand, respectively. The inclusion of climatic variables enhanced the accuracy of forecasting by improving R2 and reducing RMSE and MAPE. The results indicate effectiveness of ML models, particularly LSTM-based architectures, in forecasting water supply and demand. However, these models have limitations, particularly in forecasting extreme values, emphasizing the need to improve ML models for more reliable and accurate water management.