مقایسه کارآیی شبکه‌های عصبی CANFIS، MLP و MLP بهینه ‌شده به روش ژنتیک در شبیه‌سازی رسوب معلق رودخانه (مطالعه موردی: حوزه آبخیز زشک- ابرده، شاندیز)

     In this study, the predictive performance of three Artificial Neural Networks (ANNs), i.e. Co-Active NeuroFuzzy Inference System (CANFIS), Multi-Layer Perceptron (MLP) and MLP integrated with Genetic Algorithm (GA) in the Zoshk-Abardeh watershed were compared. In this study, three scenarios were considered and simulated in each model. In order to simulate the scenario S1 water flow were fed into the network as input. Daily water discharge and rainfall depth were considered as the input for the scenario S2. The scenario S3 was simulated based on the water discharge, daily rainfall and temperature as the inputs. In all scenarios daily sediment load was considered as the network output. Results showed that the optimum architecture for the S3_CANFIS (as the best network) was based on the Bell membership function, hyperbolic tangent transfer function and the Levenberg-Marquardt training algorithm. The S3_CANFIS with the lower MSE and NMSE acted better as compared with other scenarios during the testing process. This scenario based on the NSE equal to 0.743 and the AM equal to 0.806 showed better performance, as well. The results also suggest that the S2_MLP with 5 neurons in two hidden layers, sigmoid transfer function and the momentum learning algorithm with NSE and AM equal to 0.604 and 0.626, respectively acted better as compared with other MLP scenarios. Since the MLP network compared with CANFIS showed weaker performance in sediment yield simulation, the GA was integrated with MLP to determine the optimal network architecture parameters for the S2_MLP. Results showed that GA-MLP with NSE and AM equal to 0.658 and 0.655, respectively led to a higher capability for sediment load simulation in comparison with MLP network. Totally, the S3_CANFIS according to the criteria MB equal to -0.043, NSE equal to 0.743 and AM equal to 0.806 showed better performances in predicting sediment yield than the other networks in the studied watershed. However, both networks did not show a satisfactory power in sediment load simulation which could be arisen from the lack of data (especially extreme data) in the training series and also the existence of systematic error in observed records.