Machine-learning forecast model for predicting annual water consumption in budget estimation for urban drainage system management

Abstract

Typically, the usable budget for operating the urban drainage network is calculated at the end of the year based on the billed drinking water consumption at customer sites. To estimate the available budget in advance, the network operator uses a simple forecast of annual water consumption, calculated as the average of the past four years. To improve this process, different machine learning based forecasting models were developed with the aim of predicting annual water consumption on a quarterly basis. These models integrate not only historical data but also actual weather conditions and system state measurements with higher temporal resolution. The results showed that Support Vector Machine achieved the highest accuracy across the quarterly forecasting time points, followed by Linear Regression. Consequently, Linear Regression combined with Bayesian statistics was selected as the forecasting model, as it provides the network operator with an uncertainty assessment for the predicted values.