Desenvolvimento de um sensor virtual de vazão utilizando redes neurais para macromedição de uma estação de distribuição de água

Abstract

Law No. 14,026/2020, which established the new legal framework for sanitation in Brazil, introduced significant changes for water and sewage distribution companies by imposing stricter regulations and increasing competition in the sector. One of its key objectives is to reduce water losses and improve macromeasurement rates. According to the National Sanitation Information System (SNIS), the average water loss in Brazil is around 40%, while only 70% of the distributed water is properly measured. However, the high cost of sensors, especially flow macrometers, limits their widespread deployment, leading many companies to focus monitoring efforts on metropolitan areas. To address this issue, indirect measurement methods allow flow estimation based on network pressure and reservoir level data, reducing the need for physical sensors across the entire distribution system. In this context, artificial intelligence (AI) has been increasingly applied to improve the efficiency of water supply systems, enabling the reconstruction of missing data, fault detection, and forecasting of operational variables. Soft sensors based on neural networks allow continuous system monitoring and anomaly detection, enhancing operational decision-making and minimizing resource waste. This study presents the development and implementation of a soft sensor based on Long Short-Term Memory (LSTM) neural networks to estimate flow rates in a water supply system. The model was trained using flow, pressure, and reservoir level data collected by the Companhia de Água e Esgotos da Paraíba (CAGEPA) in the municipality of Salgado de São Félix, through an MQTT broker and cloud-based monitoring. The model validation was performed by comparing the estimated values with real measurements from physical sensors, resulting in a mean absolute percentage error lower than 5%, demonstrating the reliability of the proposed approach.