Long-term forecast of the ice-free period on arctic shipping rivers (a case study of the Pur River)

Abstract

Forecasting the onset and end dates of ice phenomena on Arctic shipping rivers, such as the Pur River, is essential for navigation planning and ensuring transport accessibility under changing climatic conditions. The development of reliable forecasting models that outperform traditional averaged approaches is of considerable scientific and practical importance. This study presents the development and comparative analysis of five machine learning models for forecasting the dates of ice formation and ice clearance on the Pur River at the Samburg gauging station: a convolutional neural network (CNN), a fully connected neural network (Dense), a multilayer perceptron (MLP), a support vector regression model (SVR), and a random forest model. Twelve hydrometeorological parameters were used as input variables. Forecast performance was quantitatively evaluated using the mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²). A comparison with an averaged baseline model showed that, for predicting ice-clearance dates, the fully connected neural network (Dense) demonstrated the best overall performance (MAE = 5.18 days, RMSE = 5.97 days, R² = 0.387). For predicting ice-formation dates, the multilayer perceptron (MLP) model exhibited the lowest prediction error and the highest explanatory power (MAE = 3.59 days, RMSE = 4.21 days, R² = 0.315). The results indicate that more complex machine learning models do not necessarily provide superior performance when forecasting complex hydrological events such as ice phenomena. Consequently, the optimal forecasting model should be selected individually for each predicted date.