Author(s): Jose Antonio Bueno Uceta; Knut Tore Alfredsen; Wenjun Lu; Knut Vilhelm Hoyland; Raed Khalil Lubbad; Sveinung Loset
Linked Author(s): Knut Hoyland, Knut Alfredsen, Wenjun Lu, Knut Vilhelm Høyland, Raed Lubbad
Keywords: River ice breakup; Mechanical breakup; LSTM; Norwegian rivers; Class imbalance
Abstract: Mechanical river ice breakups pose a significant hazard due to their sudden nature and potential to cause flooding, infrastructure damage, and safety risks. Despite their significance, there is currently no reliable tool for forecasting these events. An accurate predictive model could support early warning systems and help mitigate their negative consequences. In this work, a Long Short-Term Memory model is applied to forecast mechanical river ice breakups by estimating the probability that a breakup event will occur on a given day based on the hydrometeorological conditions observed in the preceding days. The strong class imbalance due to the scarcity of ice breakups is handled by using a weighted binary focal loss function. Model hyperparameters are tuned using the Tree-structured Parzen Estimator (TPE) algorithm and 3-fold cross-validation. The results indicate that the proposed approach can successfully predict breakup events in two of the three rivers considered. However, the model performance is sensitive to data quality, highlighting the importance of reliable input data.
Year: 2026