STREAMFLOW FORECASTING USING LINEAR PERTURBATION MODEL & ANN APPROACH

Abstract

Forecasting of streamflow is important for public safety, power, inland navigation, environmental issues, irrigation and water management. For this purpose, various methods have been developed based on physical and statistical considerations. Physically based hydrologic models require diversified data in large quantity. Such type of problems, however, can effectively be tackled by black-box models like Linear Perturbation Model (LPM), Artificial Neural Network (ANN) models, etc. These models are capable of adopting the complexity of input-output patterns and accuracy goes on increasing as more and more of data are available to them. In this study, a Linear Perturbation Model, described by Nash and Barsi (1983) and Cascade Correlation Neural Network (CCNN) are used for daily flow forecasting at Amuntai gauging site on Negara river basin, Indonesia. The rainfall-runoff records for the year 1978 to 82 are used for calibration and the data for 1983 to 84 is used for forecasting of flows. A comparison of the forecast of both models has also been made. In LPM, the daily mean rainfall and runoff series are represented by 36 and 2 harmonics respectively. The LPM results show that the overall performance of the model in fitting the daily flow values for five years (1978 to 1982) is about 93 percent which might be difficult to achieve even by using a very complex daily rainfall-runoff model. The results also indicate that efficiency of the model in forecasting the daily flow values for the year 1983 is 98 percent. For the year 1984, the efficiency of the model is computed as 95 percent. In this study, a 2-32-1 architectural cascade correlation neural network consisting one input layer with two units, one hidden layer with 32 hidden units and one output layer with one output unit is used. The input layer consists of rainfall and runoff and output layer consists of next day runoff. After training, the final network with trained Xi weights is used for forecasting of flows for the year 1983 and 1984. The results of CCNN forecast show that the efficiency of the model in forecasting the daily flow values for the year 1983 is 95 percent. For the year 1984, the efficiency of the model is computed as 91 percent. From these results, it is observed that the efficiencies of both the models are comparable and LPM gives slightly better results. The comparison of two approaches on a more lengthy data is suggested before deciding the supremacy of one approach over the other. xii