Predicting Water Temperature Inside Water Distribution Systems Using Deep Learning

Abstract Submission: Water Distribution Systems (WDS) play a crucial role in delivering clean water to consumers, but climate change is posing new challenges by raising water temperatures. These higher temperatures degrade water quality by speeding up chlorine decay, encouraging microbial growth, and increasing pipe corrosion. Many existing models for predicting WDS temperatures are often not well-calibrated or validated, and machine learning models have not been widely used in this context. This paper introduces and evaluates a Long Short-Term Memory (LSTM) model that uses direct measurements to provide valuable insights for water utilities dealing with rising temperatures. The initial findings demonstrate that the LSTM model is highly accurate in predicting hourly water temperatures using historical water and air temperature data. Three prediction scenarios were tested, varying by the number of future hours predicted and the amount of past data used, along with two dropout techniques (fixed rate and Monte Carlo). The accuracy of the models ranged from 95% to 99% at the first hour of prediction, and they were able to forecast water temperatures up to ten hours ahead, achieving over 91% accuracy at the tenth hour. The LSTM model performed better at predicting summer water temperatures (which is more important from operational point of view) compared to winter temperatures, as winter water temperatures generally stay above freezing while air temperatures can drop well below zero.

Learning Objectives/Expected Outcome (Optional) : Using a deep learning technique in predicting water temperature.