Using multiple machine learning algorithms to optimize the water quality index model and their applicability

Water quality assessment model and spatiotemporal heterogeneity pose challenges to the uncertainty of water quality assessment. To improve the accuracy of the water quality index (WQI) model, multiple machine learning algorithms (CatBoost, SVM, LR, XGBoost, LightGBM) and entropy weight method (EWM) were introduced to determine the objective weight. Six combined weights were determined by game theory combining objective and subjective weights (AHP). Three aggregation functions were established, including a new function proposed based on the sigmoid function and two existing functions. Based on the six combined weights and three aggregation functions, eighteen WQI models were developed. To reduce the influence of spatiotemporal heterogeneity, the assessment models for the different water quality characteristics were proposed respectively. To validate the performance of improved model, the monthly water quality monitoring data of 16 sampling sites in Chaohu Lake during 2016–2020 was used. Among them, totally 10 water quality indicators were selected, including TN, TP, etc. The results showed high accuracy and reliability of the improved WQI assessment models. The model improved by CatBoost and EWM had low uncertainty (0.559–0.903) than SVM and LR (0.576–1.034). The sensitivity of the models improved by six combined weights is ranked as W_AE > W_AC > W_AS > W_AX > W_ALGB > W_AL. The uncertainty of the models improved by the three aggregation functions were ranked as SGM > SWM > WQM and the sensitivity were ranked as WQM > SWM > SGM. Compared with WQM and SWM, SGM could reflect the water quality spatiotemporal heterogeneity more accurately. The WQM_AE, SGM_AC and SWM_AC models were recommended for assessing water bodies with good quality, poor quality and heterogeneity respectively. Chaohu Lake was mainly Class II and Class III water. East had better water quality than the west. Water quality in summer and autumn was better than in spring and winter. This study can provide theoretical support for related water quality assessment work.