Input optimization of ANFIS typhoon inundation forecast models using a Multi-Objective Genetic Algorithm

Inundation forecast models based on a subset of data inputs are advantageous in their efficiency due to the rather fewer inputs needed to be processed. This type of models is nevertheless rare mainly because of the difficulty in selecting the appropriate combination of input variables. In this study, an innovative methodology is proposed to overcome this difficulty by integrating Adaptive Network-based Fuzzy Inference System (ANFIS) and Multi-Objective Genetic Algorithm (MOGA). The three indices of the coefficient of efficiency (CE), relative peak error (RPE) and relative time shift error (RTS) are used to assess the performances of the models from various perspectives. Optimal combinations of data inputs for the ANFIS models are searched for by MOGA, and the three models with the best performances for each index are selected. Comparisons show that the optimal models obtained by the proposed methodology have better overall performances than the ARX and Nonlinear ARX models. Test results reveal that the optimal models preferably selected for a designated prediction lead time cannot maintain optimality under different prediction leads. This problem has been resolved by the use of a series of models, each optimized with respect to a designated prediction lead time using the same methodology. Test results show that the model series exhibits significant improvements under various prediction leads and thus effectively improves the accuracy of inundation level prediction.