Abstract
Forecasting of river ice breakup timing is directly related to the local ice-caused flooding management. However, river ice forecasting using k-nearest neighbor (kNN) algorithms is limited. Thus, a kNN stacking ensemble learning (KSEL) method was developed and applied to forecasting breakup dates (BDs) for the Athabasca River at Fort McMurray in Canada. The kNN base models with diverse inputs and distance functions were developed and their outputs were further combined. The performance of these models was examined using the leave-one-out cross validation method based on the historical BDs and corresponding climate and river conditions in 1980–2015. The results indicated that the kNN with the Chebychev distance functions generally outperformed other kNN base models. Through the simple average methods, the ensemble kNN models using multiple-type (Mahalanobis and Chebychev) distance functions had the overall optimal performance among all models. The improved performance indicates that the kNN ensemble is a promising tool for river ice forecasting. The structure of optimal models also implies that the breakup timing is mainly linked with temperature and water flow conditions before breakup as well as during and just after freeze up.
Highlights
As an important annual event in northern high-latitude regions, general public and water resources managers are concerned about the river ice breakup every spring [1,2,3,4]
In terms of its functions, the k-nearest neighbor (kNN) base models link the breakup dates (BDs) with their affecting indicators; the simple average method (SAM) ensemble models describe the relationship between the predicted BDs by each base model and the observed BDs
From the data-driven perspective, the inclusion of Figure 8 illustrates the structure of the SAM-MC4, which combines the outputs from five base models (y2, y45, y5, y8 and y9 )
Summary
As an important annual event in northern high-latitude regions, general public and water resources managers are concerned about the river ice breakup every spring [1,2,3,4]. KSEL method to a representative in Alberta, developing kNN base models with diverse features (inputs)unregulated and distance river functions for river ice breakup dates;. Data (3) comparing the structures and performance of both base and ensemble kNN models; and (4) applying the proposed KSEL method to a representative unregulated river in Alberta, The. glacial meltwater in the Columbia Icefields in Canada, which River is frequently prone from to riversnow ice related the southwestern Alberta, Canada [31,32,33,34]. The water flows and levels were collected at the Water Survey Canada gauge of Athabasca River below McMurray (Station Number: 07DA001) These indicators were pre-screened as the candidate inputs of the models based on two criteria: availability before breakup and correlation with BDs
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
