PREDICTION OF TRAYELTIME AND LONGITUDINAL DISPERSION OF POLLUTANTS IN STREAMS USING ANN

Abstract

Streams serve as source of water for various uses like water supply, generation of hydroelectric power, recreation, navigation etc., and also as sink of waste substances resulting from human activities. Wherever streams act as sinks for intentional or accidental spillage of effluents resulting from the activities of society, it becomes necessary to know the concentration of pollutants downstream of its point of injection into the stream to check whether the required environmental standards are being satisfied or not. After sufficient time of the pollutant injection, the pollutant is mixed completely across the stream and onwards-primary variation of concentration is in one direction and dispersion there is known as the longitudinal dispersion. Although many excellent models are available to make predictions for travel time and concentration of pollutants in process of longitudinal dispersion none can be used with confidence before calibration and verification to the particular river reach in question. The review of literature revealed that the potential of ANN could be used in many processes related to the water resources systems. It was therefore aimed in the present study to apply the ANN technique for prediction of travel time and the longitudinal dispersion of pollutants in streams. The time of travel and longitudinal dispersion of pollutants data were collected from the experimental work of Singh (1987) and U.S Geological Survey Report (Jobson 1996). Using these data the existing models for prediction of unit peak concentration C, and peak time of travel Tp were checked for their accuracy. It was noticed that the existing models do not predict Cup and Tp satisfactorily. Therefore, an ANN model is utilised in the present study for prediction of Cup and Tp . Three layers of ANN were used in which one was the hidden layer and the other two were the input and the output layers. The available data were split into three different sets, first set of which was used for the training of ANN , the second was used for validation while the third independent set was made use of for the testing of the proposed methodology. A number of iterations with different number of neurons and transfer functions were done. The transfer function and the number of neurons which gave the best predictions for the training were chosen. Comparison of predicted C„~, and Tp by proposed model with observed CUP and T~, revealed that the proposed model gives better results than the existing models. The results obtained from ANN model had a maximum error of ± 20 per cent for most of the laboratory data and ± 50 per cent for most of the field data.