Forecasting rare hydrological events by machine learning methods: case study of ice jams on the Pechora river

Rare hydrological events, as the name suggests, occur quite infrequently, but are often catastrophic for humans. They are also inadequately provided with measurements (the so-called class imbalance). In its turn, this hinders the creation of reliable models for predicting such processes. This is especially evident when constructing models of natural processes using machine learning algorithms, which are particularly sensitive to class-imbalanced samples. The study attempts to overcome the above-mentioned limitations by supplementing a series for model training with artificially generated events. The subject and object of the study were long-term forecasts of ice jams occurring at the mouth of the Pechora River in the Arctic area of the European Russia. Data on ice jams were collected over a long period of observations, and applicable predictors and models were selected. The following machine learning algorithms were used: k-nearest neighbors (KNN), logistic regression, gradient boosting (CatBoost), and multilayer perceptron (MLP). As a result all models demonstrated higher quality of modeling after supplementing artificial events to a series. This confirms the prospects of the method of series supplementing for training models of rarely occurring processes.