Explore an evolutionary recurrent ANFIS for modelling multi-step-ahead flood forecasts

Abstract

Reliable and precise multi-step-ahead flood forecasts are crucial and beneficial to decision makers for mitigating flooding risks. For a river basin undergoing fast urban development, its regional meteorological condition interacts frequently with intensive human activities and climate change, which gives rise to the non-stationary process between rainfall and runoff whose non-stationary features is difficult to be captured by a non-recurrent data-driven model with a static learning mechanism. This study proposes a recurrent Adaptive-Network-based Fuzzy Inference System (R-ANFIS) embedded with Genetic Algorithm and Least Square Estimator (GL) that optimize model parameters for making multi-step-ahead forecasts. The main merit of the proposed method (R-ANFIS(GL)) lies in capturing the features of the non-stationary process between rainfall and runoff series as well as in alleviating time-lag effects encountered in multi-step-ahead flood forecasting. To demonstrate model reliability and effectiveness, the R-ANFIS(GL) model was implemented to make multi-step-ahead forecasts from horizons t + 1 up to t + 8 for a famous benchmark chaotic time series and a flood inflow series of the Three Gorges Reservoir (TGR) in China. For comparison purpose, two ANFIS neural networks of different structures (one dynamic and one static neural networks) were also implemented. Numerical and experimental results indicated that the R-ANFIS(GL) model not only outperformed the two comparative networks but significantly enhanced the accuracy of multi-step-ahead forecasts for both chaotic time series and the reservoir inflow case during flood seasons, where effective mitigation of time-lag bottlenecks was achieved. We demonstrated that the R-ANFIS(GL) model could suitably configure the complex non-stationary rainfall-runoff process and effectively integrate the monitored rainfall and discharge data with the latest outputs of the model so that the time shift problem could be alleviated and model reliability as well as forecast accuracy for future horizons could be significantly improved.