River ice breakup forecasting using artificial neural networks and fuzzy logic systems

Abstract

Due to the complexity of breakup ice jam processes deterministic modelling cannot yet forecast every aspect of the timing and severity of possible consequent flooding, especially when some lead-time is needed. In most northern regions, the sparse network and short record of data have impeded the successful development of empirical and statistical models. In this study, a multi-layer modeling approach was investigated for forecasting breakup ice jam flooding using the two soft computing techniques: artificial neural networks and fuzzy logic systems. The Town of Hay River in NWT, Canada was chosen as the case study site, where the breakup ice jam flooding is an annual threat. This thesis first presents the development of index variables as potential predictors to breakup severity and timing. For the case study site, it was found that water level at the onset of freeze-up and accumulated degree-days of freezing during the winter could be potential predictors for breakup severity. The indicator variable of the timing of the onset of breakup was found to be completely nonlinear with respect to any of the index variables. Then the feed-forward artificial neural network (ANN) modeling technique was assessed for its applicability in forecasting of onset of breakup. Detailed results of the ANN model calibration and validation are presented and discussed. It was found from the calibration results, that the ANN model has greater potential for successfully forecasting the onset of river ice breakup (i.e. the first transverse cracking of the ice cover) compared to the conventional multiple linear regression technique. However, rigorous validation also indicated that the accuracy of such ANN models can be optimistically overestimated by looking only at the calibration results. Finally, the applicability of a Mamdani-type fuzzy logic system to forecast the peak snowmelt runoff during breakup for a long lead-time of ~3 to 4 weeks prior to breakup was assessed, and was found to be a good predictor of breakup flood severity at the Town of Hay River. In particular, it was found that the fuzzy logic model could predict most of the high flow, the exception being those that were triggered by short intense rainfall events during the breakup period (a factor that cannot be included in a long lead-time forecast). This study contributes new knowledge and techniques, advancing the breakup ice jam flood forecasting capabilities for the northern communities. The two most common soft computing techniques (e.g. ANN and fuzzy logic system) were studied comprehensively for their potential in river ice breakup forecasting and demonstrated step by step at the case study site. A hydrometeorological data base for the Town of Hay River was also established for the further research.