Characteristics of recursive and non-recursive adaptive network-based fuzzy inference system models for the forecast of typhoon inundation levels

Abstract

The accurate forecasting of typhoon inundation levels is vital for damage mitigation actions during such an event. The objective of this paper is to investigate the characteristics of adaptive network-based fuzzy inference system models for the forecasting of typhoon inundation levels. A novel approach of recursively using the model to achieve higher prediction lead times is proposed. The approach is advantageous in conducting water level forecasts for various prediction lead times using a single model, whereas common non-recursive models are only applicable for the designed prediction leads. In this study, a total of 6 models, with various configurations and types of recursions, are constructed based on the cross-correlations between rainfall and water level records. The performance of each model is evaluated and compared using three indices: coefficient of efficiency, relative time shift, and threshold statistics. The best recursive and non-recursive models are selected and compared with traditional approaches based on autoregressive models with exogenous input. The results show that although the recursive models display somewhat lesser but comparable forecasting capacities compared to the non-recursive models, the former models have achieved forecasts single handedly for all the prediction leads using single models only. On the other hand, although the non-recursive models exhibit better forecasting capacities, this is at the cost of using multiple models, with each designed for a specific prediction lead time. In comparison with other traditional approaches, both the recursive and non-recursive types of models demonstrate superior performance on all the aspects inspected.