ANN AND FUZZY LOGIC IN HTDROLOGICAL MODELLING AND FLOW FORECASTING

Abstract

During last few decades, scientific and engineering community has acquired extensive experience in developing and using soft computing techniques. Artificial neural networks (ANNs) and fuzzy logic based systems have emerged as potential soft computing techniques. In hydrological literature, a number of studies based on ANN have been reported. However, use of fuzzy logic is a relatively new area of research in the field of hydrology and water resources. In areas like stage-discharge relationship, stagedischarge- sediment relationship, rainfall-runoff modeling and hydrological forecasting etc., the fuzzy logic approach remains almost unattempted. Therefore, the present study has been undertaken to explore the potential of fuzzy logic based approaches in these areas and compare their performance with artificial neural network models. The objectives of the present study can be summarized as follows: i. To develop ANN, fiizzy rule-based and regression models for stage-discharge relationships, andcompare their performance in modeling hysteresis, ii. To develop and test ANN, fuzzy rule-based and regression models for deriving stage-discharge-sediment relationships, and compare their performance for estimation of river sediment load, iii. To investigate potential of ANN and fuzzy rule-based approaches for modeling rainfall-runoff relationships using different model structures and compare their performance with linear transfer function based models. iv. To develop fuzzy rule-based models for flood forecasting in order to provide accurate enough forecasts for very short lead periods and compare their performance with ANN models. River discharge is one of the most important inputs in various hydrological models and it is very important to estimate the discharge in a river reliably. Traditionally, hydrologists use regression equation based rating curves for flow estimates. However, this approach fails to model the non-linearity in the relationship and particularly in the cases where hysteresis is present in the data. Focusing on the ANN and fuzzy rule based models, different stage-discharge relationships were developed and compared using data of Jamtara, Manot, Mandla, Satarana and Hirday Nagar gauging stations lying in Narmada basin. Suitability of fuzzy modelling for substantially less data was also verified. Furthermore, hypothetical data for loop rating curve were used to explore hysteresis modelling capabilities of the models. The results show that the fuzzy modeling approach is superior than the conventional and artificial neural network (ANN) based approaches. Comparison of the models on hypothetical data set also reveals that the fuzzy logic based approach models the hysteresis effect (loop rating curve) more accurately than the ANN approach. In order to estimate bias of the fuzzy, ANN and curve fitting models for different output ranges the testing data sets of all the gauging sites were scaled so as to lie in the range of zero to ,one and than poled together. The average underestimation and overestimation errors were computed and plotted for different discharge ranges. The results indicates that fuzzy models provide a very accurate estimation in all ranges of river discharges in the study area. n Many practical problems in water resources require knowledge of the sediment load carried by the rivers or the load the rivers can carry without danger of aggragadation or degradation. Hence, the measurement of sediments being transported by a river is of vital interest for planning and designing of various water resources projects. The conventional methods available for sediment load estimation are largely empirical, with sediment rating curves being the most widely used. The rating relationships based on regression technique are generally not adequate in view of the inherent complexity of the problem. ANN and fuzzy logic algorithm were developed using available data of two gauging sites in the Narmada basin in India. The results suggests that the fuzzy model is able to capture the inherent nonlinearity in the river gauge, discharge and sediment relationship better than the ANN and conventional regression method, and is able to estimate sediment concentration in the rivers more accurately. A comparative analysis of predictive ability of these models in different ranges of flow indicates that the fuzzy modeling approach is slightly better than the ANN. The models were also compared to each other in estimation of total sediment load since it is important in water resources management. It was found that the curve fitting approach poorly estimates the total sediment load. While, the fuzzy logic model estimates were considerably better than the ANN model. Comparison of results showed that the fuzzy rule based model could be successfully applied for sediment concentration prediction as it significantly improves the magnitude of prediction accuracy. More applications and research is needed to support the utility ofANN and fuzzy logic technique in the area of rainfall-runoff modelling and to help in establishing their full practical use in dealing the real world problems. Therefore, in this study ANN, fuzzy 111 rule based and linear transfer function models were constructed for estimating catchment discharge by developing rainfall-runoff models for Manot sub-basin of Narmada River system. Different model structures were constructed by considering eleven combinations of input data vectors under four different categories: (i) only rainfall as input, (ii) rainfall and antecedent moisture content as input, (iii) rainfall and runoff as input, and (iv) rainfall, runoff and antecedent moisture content as input. The performance of the models were examined using the model performances indices such as: root mean square error, the correlation coefficient, model efficiency and volumetric error. The results indicate that the fuzzy logic based approach is capable of modelling the rainfall-runoff process more accurately in comparison to ANN and linear transfer function based modeling approaches. Another fundamental aspect of many hydrological studies is the problem of forecasting the flow of a river in a given point of its course. Therefore, real time flood forecasting models were developed using ANN and fuzzy logic methods. Finally, to it' improve the real time forecasting of floods, a modified Takagi Sugeno fuzzy inference system termed as threshold subtractive clustering based Takagi Sugeno (TSC-T-S) fiizzy inference system has been introduced using the concept of rare and frequent hydrological situations. The proposed fuzzy inference systems provide an option of analyzing and computing cluster centers and membership functions for two different hydrological situations generally encountered in real time flood forecasting. Accurate forecasting of floods at shorter lead periods is a very important task for flood management in Narmada basin, Central India. The methodology has been tested on hypothetical data set and than applied for flood forecasting using the hourly rainfall and river flow data of upper iv Narmada basin upto Mandla gauging site. The available rainfall-runoff data has been classified in frequent and rare events and suitable TSC-T-S fuzzy model structures were suggested for better forecasting of river flows. The performance of the model during calibration and validation was evaluated by model performance indices such as root mean square error, NS model efficiency and coefficient of correlation. A new performance index termed as peak percent threshold statistics was proposed to evaluate the performance of flood forecasting model. The developed model was tested for different lead periods using hourly rainfall and discharge data. Further, the proposed fuzzy model results were compared with artificial neural network (ANN) and subtractive clustering based T-S fuzzy model (SC-T-S fuzzy model). It was concluded from the study that the proposed TSC-T-S fuzzy model provide reasonably accurate forecast with sufficient leadtime. The results presented in this thesis are highly promising and suggest that fuzzy modeling is a more versatile and improved alternative to ANN approach. Furthermore, fuzzy logic algorithm has the ability to describe the knowledge in a descriptive human like manner in the form of simple rules using linguistic variables. The ANN and fuzzy logic methodology presented in this thesis can provide a promising solution to various hydrological modeling and forecasting problems. However, the analysis of the results reported in this work leave sufficient scope and opens new dimensions for further investigations, which couldnot be taken up owing to time constraint. w