Water infrastructure failures under climate change: a meteorological factor analysis using machine learning

“Water distribution networks (WDNs) are critical infrastructures facing challenges that affect their structural integrity due to meteorological hazards, which may be worsened by climate change impacts, particularly through rising soil temperatures, an often overlooked factor in failure prediction models. This study addresses that gap by incorporating soil temperature assessment, derived from meteorological data and climate change scenarios, into the dataset intended for future pipe failure prediction.Two models were evaluated: a Long Short-Term Memory (LSTM) neural network, and the already existing numerical model Soil Temperature Model (BTM) which incorporates soil properties among others factors. Both models were applied to historical and projected hourly meteorological data from The Bilt station in the Netherlands, under two datasets of climate change scenarios; first dataset developed by the Meteorological Institute of the Netherlands (KNMI’23) includes two scenarios: High emissions and dry conditions (Hd) and Low emissions and wet conditions (Ln); second dataset for the Representative Concentration Pathways (RCP) 8.5 (the highest emissions scenario). The LSTM model, trained se-asonally, achieved low validation errors but consistently underestimated observed soil temperatures, while still capturing a warming trend in future projections. In contrast, the BTM model overestimated soil temperatures. Results confirm an upward trend in soil temperatures under climate change, highlighting the need to take this variable into account, which is distinct and higher from air temperature.”

(Citation: Gutierrez Caloir, B., Blokker, E.J.M., Savić, D.A., et.al. – Water infrastructure failures under climate change: a meteorological factor analysis using machine learning – https://doi.org/10.15131/shef.data.29921075.v1 – This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st – 3rd September 2025))