Author(s): Amin Talei; Lloyd Hock Chye Chua; Chai Quek
Linked Author(s):
Keywords: Sediment concentration; Real time forecasting; Neuro-fuzzy; ANFIS
Abstract: Neuro-Fuzzy Systems (NFS) are computational intelligence tools that have recently been employed in hydrological modeling. The learning algorithms used are based on batch learning where all the parameters of the fuzzy system are optimized off-line. Although these models have frequently been used, one criticism of batch learning is its inability to react to changes in the system. For a model to be able to react to changes in the problem characteristics without the need for re-training, two conditions have to be satisfied. Firstly, the model must be capable of online or incremental learning. This means that learning takes place with the presentation of each input data, rather than the entire dataset. Secondly, the model should be capable of learning in real-time. To address these issues, Real Time Dynamic Evolving Neural Fuzzy Inference System (RT-DENFIS) has been developed. RT-DENFIS is a Takagi-Sugeno-type fuzzy inference system and utilizes online learning. In the present study, updating capabilities of RT-DENFIS is compared with the Adaptive Network-based Neuro-Fuzzy Inference System (ANFIS) which is a common batch (offline) NFS model for sediment concentration forecasting in Bhakra catchment, India, with the area of 56, 980 km2. The discharge and sediment concentration time series from 1987 to 2004 have been used in this study. Different performance evaluation measures including coefficient of efficiency (CE), coefficient of determination (r2), root mean square error (RMSE), mean absolute error (MAE), and relative peak estimation error (RPE) are used in this study. Results show that RT-DENFIS gave consistently better results without the need for retraining; however, the batch model (ANFIS) trained with historical data had to be retrained periodically in order to achieve similar results. Moreover, RT-DENFIS results are also compared with the results obtained by an autoregressive model with exogenous inputs (ARX) and a rating curve (based on a nonlinear regression) as bench marks. It was revealed that RT-DENFIS is also superior to both regression-based models.
Year: 2013