Author(s): Claudio Mineo; Stefania Passaretti; Eleonora Boscariol; Anna Varriale

Linked Author(s): Claudio Mineo, Stefania Passaretti, Eleonora Boscariol

Keywords: Machine learning; Random forest; Water resources management; Sustainability of withdrawals; Water utility

Abstract: A reliable assessment of the water-bodies quantitative status over an area is of fundamental importance in evaluating the potential water availability, and in turn this, in the adoption of management strategies aimed at maintaining both the water resource balance and the health of several natural ecosystems. As well known, building a model without uncertainties is impossible by definition. It is certainly important to assess them, so that the model's evaluations remain adequately circumscribed within the assumptions made. In this respect, in the application of water resources management model, its performances can be defined a priori as a compromise between the responses to which one wishes to strive and the degree of reliability achievable. In this context, the scientific community has always continued to propose different modeling approaches to describe (or predict) the phenomena related to the dynamics of the aquifers. Traditionally, in the field of hydro-environmental disciplines, mechanistic (physically-based) models are implemented, although more recently the application of models based on machine learning techniques (ML) are favouring a wealth of opportunities and benefits for water management practitioners. The main reasons why ML approaches are increasing their appealing to hydro-environment practitioners are likely referable to (i) the greater accessibility of data (freely available); (ii) the practicability of computational techniques made increasingly accessible through open-source programming environments; (iii) the prediction and inference capabilities for complex and partially unknown hydrological contexts. For what concern the water resource management, it becomes essential for water managers to adopt straightforward methods to solve problems of practical interest, in order to rationalize interventions aimed at minimizing disruption to users and, at the same time to maintain the environmental flows to support stream ecology (i.e. sustainability of withdrawals). In this context, ACEA ATO2, a water utility that has been supplying fresh water to the city of Rome (Italy) and to the surrounding areas for over 100 years, in order to plan a sustainable water resources exploitation, here proposes the application of a random forest (RF) machine learning approach for the evaluation of the water resources status throughout its managed area (in Central Italy). According to literature studies, preliminary results highlight that although the RF performance is strictly affected by the accessibility of exhaustive datasets available, they seem to integrate the understanding of complex phenomena offered by traditional models. In fact, in the present investigation an high rate of variability on RF performance have been found related to the specific water bodies (i.e. lake, well field, natural spring). Actually, at this stage of the study the original purpose of the investigation is considered not to be satisfactory covered yet, about the real applicability of RF for practical uses in water management resources.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221804

Year: 2022