Sahu, R.K.Sinha, Jitendra2016-02-052016-02-052011http://krishikosh.egranth.ac.in/handle/1/64169Artificial intelligence such as artificial neural network (ANN) has been proved to be an excellent tool for rainfall runoff modelling with reasonable accuracy. The advantage is that, even if the 'exact' input-output relationship is not known, the network can be 'trained' to ‘learn' that relationship, not requiring a prior knowledge of the basin characteristics. Comparative studies have demonstrated that neural solutions can be more efficient to implement in comparison to conventional approaches. The gauged data in watersheds are meagre in the country which is the biggest constraint in planning water resources. Even in gauged watersheds, the maintenance, operations and up keeping of records are not an easy task. The Central Water Commission took initiative and is engaged in gauging important rivers/tributaries. Upper Kharun Catchment (Gauging site: Pathardih) is one such catchment in the state of Chhattisgarh, selected as study area (2511 km2). River Kharun is a tributary of river Shivnath, which itself is a tributary of river Mahanadi. The present study has been planned to develop ANN runoff simulation models for the study area. Reported work using ANN for rainfall – runoff modelling is lacking in Chhattisgarh. However, the importance of this aspect for prediction of stream flow is strongly felt in water resources studies. Since this technique has the potential to uncover the non-linear rainfall-runoff relationship, it can be expected to result in reliable and accurate estimates of stream flow. Three layered ANN models to simulate runoff on yearly and active period database with daily, weekly and monthly time bases have been developed for the study area. A total of 18 models were developed for simulating runoff, 6 each on 3 family of methodology viz. Multiple Linear Regression (MLR), Back Propagation ANN (BPANN) and Radial Basis Function ANN (RBFANN). The daily rainfall and gauge – discharge data for past 20 years (1990 to 2009) were used. The catchment behaviour to infiltration and other losses was found to be variable with the average runoff-rainfall ratio of 0.3444. First 15 years data (1990-2004) was used for calibration/training of the model while last 5 years data (2005-2009) was used for model verification. A correlation matrix followed by student’s ‘t’ test and thereafter stepwise regression was carried out to finalize the number of significant inputs. Input data was filed in MS excel and subsequently MLR models were developed (through stepwise regression and statistical category function ‘linest’). The ANN based analysis was carried out in MATLAB. BPANN training was conducted using the Levenberg Marquardt back propagation algorithm. The ‘logsig' activation function was used for both hidden and output layer nodes. RBFANN models were developed by activating ‘newrb’ code in MATLAB. Gaussian activation function was used between input and hidden layer for non linear transformation, while a linear function was used between hidden and output layer. Input and output dataset was presented to the neural network as series of learning sets. Separate programmes for BPANN and RBFANN were developed in MATLAB. The performance function chosen for both the family of models was sum squared error. These programmes normalized the data in the range 0 to 1 before training. BPANN models were developed by network growing technique adopting trial and error. Best results were arrived only at an optimum number of iterations, beyond which the performance deteriorated. A trial was made on the combination of goal and spread for RBFANN models. The number of hidden nodes was determined automatically. The best model was selected based on the combination of parameters (goal, spread and corresponding number of hidden neurons), that resulted in best performance in both calibration and verification. The best combination of number of input nodes, hidden nodes and output nodes for Back Propagation models ( BPYD, BPYW, BPYM, BPAPD, BPAPW and BPAPM ) were found to be 10-4-1, 4-3-1, 3-3-1, 10-4-1, 2-2-1 and 2-5-1 respectively. The best combination for Radial Basis Function models was arrived at 10-63-1, 4-120- 1, 3-5-1, 10-62-1, 3-21-1 and 2-2-1 in case of RBFYD, RBFYW, RBFYM, RBFAPD, RBFAPW and RBFAPM respectively. The performance of the models was tested through statistical tools such as MAD, RMSE, CC, CE and EV. The best performing models for different datasets, based on the performance in the verification period (unknown output dataset), were as below: • The RBF model was found best in case of yearly data set on daily basis (MAD: 11.57, RMSE: 49.11, CC: 91.67%, CE: 84.02%, & EV: -3.14% ). • The BP model was found best in case of yearly data set on weekly basis (MAD: 106.02, RMSE: 287.69, CC: 89.96%, CE: 80.5%, & EV: -9.33%) • The RBF model was found best in case of yearly data set on monthly basis (MAD: 419.88, RMSE: 810.75, CC: 92.67%, CE: 81.91%, & EV: - 18.89%), • The RBF model was found best in case of active period data set on daily basis (MAD: 26.89, RMSE: 76.38, CC: 90.6%, CE: 81.99%, & EV: - 3.62%), • The RBF model was found best in case of active period data set on weekly basis (MAD: 232.83, RMSE: 435.36, CC: 88.08%, CE: 76.45%, & EV: - 11.72%) • The RBF model was found best in case of active period data set on weekly basis (MAD:888.0, RMSE: 1260.2, CC: 87.48%, CE: 70.25%, & EV: - 20.93%) Validation of these models by t- test revealed that there was no significant difference in the means of the observed flow and the simulated flow. In general, it was found that Yearly database models performed better than Active Period models and Daily models performed better than weekly and monthly models. Comparison of the 3 family of models on the basis of MAD, RMSE, CC and CE showed that RBFANN models performed better than BPANN and MLR models. It was realized that RBFAPW model is a better choice for simulating water availability during Kharif season. In view of above performance evaluation, based on the statistical tests, it can be concluded that use of ANN models (particularly RBFANN) is certainly a much better choice as compared to MLR models, for rainfall – runoff modelling of the study area.ensurface water, runoff, precipitation, irrigation, layering, biological phenomena, sets, fruits, area, research methodsThesis