Machine Learning-based Irrigation Control Optimization

Abstract

Irrigation schedules on traditional irrigation controllers tend to disperse too much water by design and cause runoff, which results in wastage of water and pollution of water sources. Previous attempts at tackling this problem either used expensive sensors or ignored site-specific factors. In this paper, we propose Weather-aware Runoff Prevention Irrigation Control (WaRPIC), a low-cost, practical solution that optimally applies water, while preventing runoff for each sprinkler zone. WaRPIC involves homeowner-assisted data collection on the landscape. The gathered data is used to build site-specific machine learning models that can accurately predict the Maximum Allowable Runtime (MAR) for each sprinkler zone given weather data obtained from the nearest weather station. We have also developed a low-cost module that can retrofit irrigation controllers in order to modify its irrigation schedule. We built a neural network-based model that predicts the MAR for any set of antecedent conditions. The model's prediction is compared with a state-of-the-art irrigation controller and the volume of water wasted by WaRPIC is only 2.6% of that of the state-of-the-art. We have deployed our modules at residences and estimate that the average homeowner can save 38,826 gallons of water over the course of May-Oct 2019, resulting in savings of $192.