Usage of Multisource Data through Transfer Learning to Improve Short-Term Water Demand Forecasting

Accurate short-term water demand forecasting is essential for the operational management of water distribution systems. Previous research on the use of machine learning approaches to improve predictive accuracy has tackled issues of model selection, data quality, and feature input reduction. Limited effort, however, has been placed on tackling data availability in district metered areas (DMAs), which is a major barrier to practical application of machine learning approaches in the water industry. This study investigates the performance of transfer learning and its ability to mitigate the impact of data unavailability in DMAs. This is achieved by incorporating external data from other DMAs to facilitate demand forecasting. Two forecasting models—extreme gradient boosting and long short-term memory—are used for water demand forecasting at two temporal resolutions—15-min and hourly demands. The results demonstrate that transfer learning significantly enhances forecasting accuracy by using external data sets, providing reliable forecasts even with a limited amount of target data. Additionally, correlation-based data selection marginally outperforms quality-based selection, highlighting the importance of choosing relevant external data for optimal performance. These findings emphasize the potential of transfer learning in urban water management, particularly for newly established or poorly monitored DMAs in water distribution networks, providing a practical solution to accurately forecast water demands and enhance operational efficiency.