Application of neural networks to predict ice jam occurrence

Artificial neural networks show potential for modeling the behavior of complex nonlinear processes, such as those involved in the occurrence of breakup ice jams. Because breakup ice jams and related flooding occur suddenly, ice jam prediction methods are desirable to provide early warning and to allow rapid, effective ice jam mitigation. Unlike open-water flooding, however, an analytical description of all the complex physical processes involved is not available. As a result, breakup ice jam prediction models have historically been limited to classical empirical single-variable threshold-type analyses to statistical methods such as logistic regression and discriminant function analysis. A neural network is shown to improve the error rates of ice jam prediction at Oil City, PA. The neural network input vector is determined and the methods used to appropriately account for the relatively low occurrence of jams are addressed. The neural network prediction proves to be more accurate than the current method used at this site, with a false positive error rate of 5.9% and a false negative error rate of 7.4%.