Ensemble flood forecasting by neuro-fuzzy inferencing
Date of Issue2016-08-08
School of Civil and Environmental Engineering
Physically-based models, regressive models and data-driven models have been applied to flood forecasting to produce accurate predictions in many studies. However, there remain drawbacks in individual models, whether physically based or otherwise. For example, not all the phases of the hydrograph can be predicted well by any model, even though the global optimum may be reached. Therefore, in order to exploit the strengths of different models, the ensemble model approach can be used to improve forecast accuracy. This thesis began with a review on the research carried out on flood forecasting conducted at two levels: (i) flood forecasting by individual models (including physically-based models, statistical models and data-driven models) and (ii) methods used to develop ensemble flood forecasts by combining component model results. For the former, much work has been done by applying conceptual, distributed or lumped models, time series models and black box models to flood forecasting. For the latter, only limited studies have been attempted with some simple statistical methods and data driven models. Although limited in scope, these studies indicate that ensemble flood forecasting show improved accuracy over the individual models. In addition, when multiple predictions are available, it is common to calculate an average of the different models’ results. However, there is often no basis to use an averaging procedure, and therefore, a better approach is needed. Current ensemble methodologies adopted in flood forecast studies include simple average method (SAM), the weighted average method (WAM), Bayesian Model Averaging (BMA) and fuzzy logic. These ensemble models are applied to the component models with arbitrary weights or fixed weight allocation strategy and are not considering the performance of the component models at different stages of the hydrograph. This thesis presents the use of neuro-fuzzy inference system (NFIS) as an ensemble methodology exploiting the parameter learning from neural networks and interpretation from fuzzy logic. In particular, the Dynamic Evolving Neural-Fuzzy Inference System (DENFIS) was used in this study with the clustering algorithm for its weight allocation strategies and online learning for model adaptation during testing stage. Here, two general cases of ensembling are investigated: (i) Different rainfall-runoff models with the same rainfall inputs and (ii) Different rainfall inputs but with the same rainfall runoff model. For the former, data from the Lower Mekong River was analyzed. For the Lower Mekong, flood forecasts are from two independent models, Adaptive-Network-Based Fuzzy Inference System (ANFIS) and the Unified River Basin Simulator (URBS) hydrological model. A real time updating ensemble model based on the online learning ability of DENFIS was proposed to provide an ensemble forecast. With the proposed ensemble model using real time updating, the ensemble model can adapt to higher water levels during testing stage than those in the training stage. By continuously updating the model, the model is able to better adapt to changes in the forecast by reducing the spikes from the component URBS model and the time shift error from the ANFIS component model. For the Taiwan case study, data from a catchment in Taiwan was analyzed. The Taiwan data includes runoff predictions based on 15 rainfall inputs, obtained from 15 different perturbations of an atmospheric model. A data processing procedure is suggested as a preliminary step to form a truncated input space for the ensemble model. The modified offline models which impose weight constraints and consider the effects of the slopes were proposed to highlight the interpretation of the ensemble process. Not only the peak and the shape of the hydrograph was better predicted compared with the benchmark SAM model, the fuzzy rules of the weight allocation were interpreted to show the mechanism of the ensemble approach based on NFIS model. The results from the two catchments show the possible ensemble solutions to optimizing the water level estimations for different cases of flood forecasting in practice.