Modelagem do decaimento de cloro residual livre em redes de distribuição de água de comunidades rurais por meio de redes neurais artificiais.

Abstract

Disinfecting water with chlorine aims to guarantee the sanitary integrity of the water and the prevention of waterborne diseases. However, maintaining free residual chlorine (FRC) concentrations is challenging in the management of water supply systems, as its concentration can decline throughout the distribution network, harming water quality and the health of the population. To understand the behavior of FRC throughout distribution networks, ensuring satisfactory disinfection performance, it is necessary to model its decay. Using water quality data from the Sistema de Informação de Vigilância da Qualidade da Água para Consumo Humano (Sisagua) and the t-Distributed Stochastic Neighbor Embedding (t-SNE) variable selection method, an artificial neural network (ANN) was developed, to evaluate the decay of FRC in water distribution networks in Brazilian rural communities. The developed model was validated with data from samples from nine Simplified Water Supply Systems (SWSS) from rural communities in the state of Ceará. The distribution networks of these locations were evaluated regarding FRC compliance with Ordinance No. 888/2021, based on performance indicators and the Sustainability Index (SI). The results obtained highlight the relevance of the FRC parameters in the reservoir, turbidity, pH, pipe length and daytime temperature in predicting chlorine in the distribution system. The ANN developed from these input variables showed high accuracy rates ranging from 85.45% to 99.84%, for the accuracy thresholds of the colorimetric method, commonly used in rural supply systems. Compared to numerical and computational methods, the model can be applied to any rural supply system, without the need to determine the kinetic order of the reaction or estimate the FRC decay coefficient. The application of ANN in the SWSS of rural communities indicated that the robustness of the model is reduced for systems with capture from surface water sources, the existence of a filtration stage and chlorine reinforcements, while obtaining satisfactory results for the SWSS with capture from wells, reaching 100 .00% accuracy for the 0.5 mg/L FRC threshold. Analysis of the performance of water distribution networks revealed the predominance of inadequate performance in terms of FRC concentration (55.56% of systems), indicating the need to improve service management strategies. In general, the approach developed offers an early prediction of nonconformities related to free residual chlorine, in addition to assisting in adjustments in chlorine dosage.