Comparison of ANN, Fuzzy Logic and Regression Tree Models for Reservoir Inflow Forecasting

Abstract

In water resources planning and management the forecasting of reservoir inflow for various time steps is one of the important components which can help in decision making regarding allocation of water for various sectorial demands and the quantity of water to be reserved for catering to the future demand. The present study demonstrates the capability of three forecasting techniques namely Artificial Neural Network consisting of two layer feed forward neural network with back propagation algorithm, Mamdani type Fuzzy Model and Regression Tree Method consisting of coarse, medium and fine tree models applied in the prediction of weekly inflow values of the Deer Creek Reservoir in Utah, United States. The data used for forecasting is daily data for a span of 60 years comprising of a total of nine parameters. ANN model has been created using 45 year data for training and the remaining 15 year data is used for testing and validation. Optimum architectures have been chosen based on the performance criteria of Root Mean Square Error, Coefficient of Correlation and the Coefficient of Determination. The result of this study has also been compared by evaluating the performance criteria of models created from the three methods of forecasting to choose the best one. The final results concluded that ANN to be an effective tool when compared to the other two methods when forecasting inflow in Deer Creek Reservoir using atmospheric data.