Online neuro-fuzzy models for real time flow forecasting
Date of Issue2017-03-31
School of Civil and Environmental Engineering
Forecasting with limited data or sparse data are two main challenges needed to be addressed. Data should be representative of the system under consideration when forecasting with traditional neuro-fuzzy models (NFMs); the condition which is not met in case of forecasting with limited data. Also traditional NFMs cannot handle sparse rule base; the condition which is not suitable for modeling the low-frequency events. Generic Self-Evolving Takagi-Sugeno-Kang or GSETSK is a state-of-the-art NFM with specific capabilities due to its specific clustering, rule pruning and optimization mechanisms. The fully online property allows the model to be used independent of historical data. Also its dynamic expanding-shrinking structure makes its rule base dynamic, compact, up-to-date and interpretable. GSETSK was run fully-online form the 1st sample onwards for two real time hydrologic forecasting problems in this study. The two commonly encountered problems; rainfall-runoff and flow routing modeling represent different modelling complexities. The data used belongs to two totally different datasets including a catchment in Sweden and Lower Mekong River in South-east Asia. In this research firstly the capability of GSETSK in forecasting with large datasets were critically examined. GSETSK was evaluated against traditional NFMs including the local NFM, DENFIS (Dynamic Evolving Neural-Fuzzy Inference System) and global NFM, ANFIS (Adaptive-Network-Based Fuzzy Inference System). Also the rule base dynamics of GSETSK were evaluated in relation to the complexity of the problem. Secondly the fully-online property of GSETSK was examined in forecasting with limited data e.g. forecasting in basins with limited data. Eventually for handling the sparse data, the Fuzzy Interpolative-Extrapolative Reasoning (FIER) was developed and imbedded in GSETSK. The overall findings of this research are as follows. i) In case of large datasets, GSETSK was able to provide comparable results to traditional NFMs, ANFIS and DENFIS. However GSETSK does not need any knowledge of data upper/lower bound plus that its rule base was more compact, up-to-date and interpretable comparing to DENFIS. Also GSETSK needs less training data and time compared to ANFIS. GSETSK was found to be more suited for the complex problem of rainfall-runoff forecasting comparing to river routing problem. ii) In case of limited data, e.g. forecasting in basins with limited data, GSETSK was found to be a good candidate. In addition it provided a compact, up-to-date rule base which was easily interpretable and could be related to the physical process under consideration. iii) The sparse data was handled by proposed FIER algorithm. The new model, FIER-GSETSK appeared to model the infrequent events better than GSETSK when evaluated based on threshold-based indices. In all experiments, the results of GSETSK were comparable to the physically-based models currently in use for each dataset. The two datasets used in this study are representative of rainfall-runoff and river routing problem. Extensive datasets should be used in other researches in future and also other hydrologic applications to examine its generalization ability. Also boosting the model with online input selection and online optimization of the clusters are recommended as the future research direction.