Development of a Lightweight Model for Short-term Prediction of COD Concentration under Limited Data Conditions

Abstract

As a core indicator of the degree of organic pollution in water bodies, chemical oxygen demand (COD) monitoring results directly affect the identification of pollution events and the response efficiency of water quality management. Traditional detection methods are difficult to realize continuous high-frequency COD monitoring, resulting in limited timely response to pollution dynamics, which seriously impedes the initiative of water quality management. To address this problem, this paper explores the construction of a lightweight COD short-term prediction model with simple structure, low computational overhead and basic predictive ability based on real water quality monitoring data of the Weihe River Tieqiao section in Xianyang City, China, using three models: multiple linear regression, Lasso regression, and shallow decision tree. Through pre-processing and feature selection of water quality-related auxiliary variables, redundant information was eliminated, and finally pH, total nitrogen (TN), total phosphorus (TP), ammonia nitrogen (NH₃-N), and water temperature were identified as modeling variables. The results showed that the multiple linear regression model performed well in the medium concentration interval, but systematic bias existed in the high and low concentration intervals, reflecting the limitations of linear models for nonlinear features. The study in this paper provides a simple and effective modeling idea for the prediction of water quality indicators under resource-constrained conditions, which helps to improve the early warning capability and response efficiency in environmental management.