Predictive Water Usage Optimization with IoT and Regression Models

Abstract

Water scarcity and inefficient water distribution systems continue to pose significant global challenges, particularly in regions with limited infrastructure. This study proposes an IoT-enabled water management system that integrates regression-based predictive modeling to optimize consumption and reduce wastage. The system employs ultrasonic sensors to continuously monitor water tank levels and gathers real-time usage data. Using this data, multiple machine learning algorithms including Linear Regression, Polynomial Regression, and Random Forest Regression were evaluated to forecast short-term water demand. Among them, the Polynomial Regression model of degree 4 achieved the highest accuracy with an R2 score of 0.727.

One of the system’s key innovations lies in its ability to locally update the predictive model on the NodeMCU microcontroller. A weighted update mechanism is employed, where the stored model retains a cumulative weight representing the past 30 days, and each new data point is added with a weight of 1. This approach ensures the model remains responsive to recent consumption patterns while maintaining historical trends—allowing lightweight, on-device forecasting without needing constant cloud retraining.

The predicted consumption values feed into a smart refilling mechanism, automating water distribution based on anticipated demand. This scalable and cost-effective solution supports sustainable water management by embedding real-time intelligence directly into low-power devices.