Similarity based neuro-fuzzy systems for tunoff forecasting in a rural catchment

Talei A., Chua L.H.C, Quek C., Proceedings ITISE2015, 1-3 Jul 2015, Granada, Spain, 2015

The Neuro-Fuzzy Systems (NFS) are computational intelligence tools that have recently been employed in hydrological modeling. Approximate Analogical Reasoning Schema (AARS) is one of the approaches which employs a similarity measure (SM) to determine whether a rule should be fired for a specific observation in the pattern matching phase. In the present study, an NFS model with rule pruning capability is developed based on AARS and is tested for a runoff forecasting problem in Klippan_2 a sub-catchment of Rönne watershed. For this study, the mean daily rainfall, runoff, and temperature time series from 1961 to 2003 were used. An input selection method based on correlation and mutual information analyses were used to identify the proper input variables of the proposed model. Results showed that a combination of rainfall, temperature, and discharge antecedents are required to be used as inputs. Temperature was found to be a necessary input to the model to take into account the effect of snowmelt on the catchment discharge. The training data set was chosen through a cross validation procedure. The trained model then was tested for the testing data set (roughly 30% of total available data) and the results were compared with the ones obtained from Autoregressive Regression with exogenous inputs (ARX) and Adaptive Network-based Fuzzy Inference System (ANFIS). The results showed that the similarity-based NFS model can successfully predict discharge up to 3 days ahead. Model performance in terms of several statistics including Coefficient of Efficiency, r2, RMSE, and MAE outperformed the ones obtained by ANFIS and ARX models. The proposed model was found to have a more flexible structure compared to ANFIS due to its rule pruning mechanism.