Ice-Jam Forecasting during River Breakup Based on Neural Network Theory

Abstract

Forecasting of ice jams and their breakup is crucial to prevent or reduce flooding risk in cold regions. A back propagation (BP) neural network model improved by the Levenberg-Marquardt clustering method has been developed with air temperatures and precipitation as inputs and applied for ice-jam forecasting in a given year in the upper reaches of the Heilongjiang River (Amur River), where ice flooding occurs frequently during spring. The accuracy rate achieved was 85%, higher than that obtained using the conventional statistical method (62% accuracy), for ice-jam breakup forecasting. The BP model has a forecast period of 1 days with a maximum error of two days and a qualified rate of 1% for national standards breakup date forecasting. The forecast on the ice-jam breakup, which was released 24 days ahead, provided accurate results for the breakup date and the occurrence of ice jams in the spring of 217.