A new machine learning approach to support asset management in water distribution networks

One of the main causes of the widespread problem of freshwater scarcity lies in unfruitful maintenance of distribution infrastructure, leading to failures with consequent waste of precious resources. It is estimated that more than 25% of the annual loss of water is due to poor conditions of the distribution networks and, in a scenario of continuously increasing demand for water, effects of such inefficiency might be even more dramatic, beyond the merely economic aspect. However, with the rise of data analysis, the awareness of the power of predictive technologies and machine learning techniques, the opportunity to make use of these tools to support decision making has become more than a hope. With this study, the author attempts to address the problem of usage of historical data of pipes and their failures in the Spanish city of Manresa to deduce conclusions on how to conduct maintenance interventions. After conducting an explorative study on how pipes intrinsic factors may have reflections on breakages, machine learning algorithms (Logistic Regression and Random Forest have been chosen in this thesis) are used to predict pipe failures over time. Lately, results from predictions will be used to take out conclusions from two different assessment models. The first method, given the structure of cost of a general distribution company, tries to establish the optimal ratio between sensitivity and sensibility of a predictive model to return the best economic benefit from the predictive maintenance. The second approach wants to assess how the uptime of the service level can be improved whether relying on prediction to replace pipes, given a certain agreed investment budget. In an old industry such as water distribution, difficulties come up not only during the development of predictive models but also during the reconstruction of the data on which training and testing models, since they can suffer from inconsistencies. Indeed, data gathering has not unique and standardized methodologies and time and people take-over have changed procedures during the data collection, making the whole work harder.