Neuro-fuzzy river ice breakup forecasting system

Despite the serious threat posed to communities by ice during spring river ice breakup, there are no reliable means to predict the severity of breakup with a significant lead time. Building on previous data collection and regression analyses for the Athabasca River at Fort McMurray, this paper evaluates the application of soft computing through fuzzy logic and artificial neural networks for modeling the maximum water level during river ice breakup for both flood and non-flood event years. A prototype fuzzy logic model is presented, based on four input variables, each available with a lead time of several weeks prior to river breakup. The performance of the model was evaluated for several designs including a neuro-fuzzy model created to reduce the subjectivity of expert knowledge for rule base definition. It was found that a simple fuzzy expert system, based exclusively on expert experience, could qualitatively distinguish years when flooding occurred but produced poor quantitative results. A neuro-fuzzy model was able to simulate water levels with an R2 of 0.88, performing equally in comparison to a multiple linear regression model based on twice as many input variables, some with much less lead time. The performance of this neuro-fuzzy model with relatively few input variables holds promise for modeling sites where the volume of available data is limited.