Detection of water level anomaly values based on different machine learning algorithms

To enhance the accuracy and reliability of hydrological monitoring, this study employed different machine learning algorithms, Isolation Forest, Random Forest, and k-Nearest Neighbors, to detect anomalies of water level data from Dongyulin, Hanjialou, Xinjiang, and Zhaocheng Hydrological Stations in Shanxi Province. By comprehensively considering characteristics such as time series and seasonal fluctuations of water level data, the algorithms were finely optimized according to the data characteristics and anomaly distribution patterns of different hydrological stations. The performance of each algorithm was evaluated using the F₁ score and ROC curve. The results indicated that all three algorithms demonstrated certain classification capabilities in detecting water level anomalies across different hydrological stations. Notably, the Random Forest algorithm, leveraging its ensemble learning properties, achieved F₁ scores ranging from 0.923 to 0.950 and AUC values to 1.0 across all four stations, exhibiting the best overall detection performance. The research results can provide references for the scientific management of water resources and flood prevention and disaster reduction.