Adaptability of Machine Learning Prediction Models for Chlorine Consumption to Monitoring Frequency of Residual Chlorine in Wastewater Treatment Plants

• • At a 1-h monitoring interval, LSTM achieved the highest prediction accuracy for chlorine consumption, leveraging its temporal modeling capability to capture high-frequency dynamics; at 2–4-h intervals, RF outperformed others due to ensemble robustness; at intervals of 6 h or lower, BP was optimal, showing lower time dependence and better static nonlinear mapping. • • SVR consistently performed worst across all monitoring frequencies (1, 2, 4, 6, 8 h), indicating its limited suitability for this application. • • Chlorine consumption exhibited more extensive significant correlations (P<0.05) with online indicators (NH3-N, CODCr, TP, TN, temperature, flow, and chlorine dose) than residual chlorine itself, making it a more appropriate model output. • • For low-frequency manual monitoring (6–8 h intervals), a BP model optimized by particle swarm optimization (PSO) significantly improved prediction accuracy, enabling precise disinfectant dosing control using high-frequency online water quality data.