Comparative Study of Different GCM Models for Stream Flow Prediction

Aims: The goal of this work was to provide comparative analysis of different GCM models for stream flow prediction. These models were prepared by training, validation, testing and mean square error. The specific objective of this study was to compare different GCM models for climatic analysis. Future stream flow was predicted by the best one.

Study Design: For the prediction of future flow, an artificial neural network model was developed for down scaling the GCM data. The ANN downscaling model was used to predict the future stream flow of the river.

Place and Duration of Study: This study was conducted in Ranganadi river which originates from the Nilam, Marta and Tapo mountain ranges in Arunachal Pradesh. The Ranganadi sub-basin spreads about 1749 sq. km. across the Lower Subansiri and Papum-Pare districts in Arunachal Pradesh and Lakhimpur district of Assam, where it joins with Subansiri-Brahmaputra river system at khichikagao. The study area was located between 94⁰02’34” E longitude and 27⁰14’01” N latitude in the Brahmaputra River basin of India. For this research, observed stream flow data from 1973-1983 and 2001 to 2009 were used.

Methodology: Neural networks are mathematical representations of a process that operates nerve cells. Each network is made up of nodes and links like nerve cells. In this study the best model was decided by using the different algorithms and varying the number of hidden neuron from 1 to 15 with various combination of learning rate from 0.01 to 0.9 and momentum factor from 0.01 to 0.9. Forecasting was done in three clearly separate stages. They were training mode, validation and testing phase. In training mode, the output was linked to many of the input nodes as desired and the pattern was defined. The network was adjusted according to this error. The validation dataset was used at this stage to ensure that the model was not over trained. In testing phase, the model was tested using the dataset that was not used in training.

Results: In this work proposed the best GCM model for checking the future flow scenario of Ranganadi river using ANN model. For model prediction, stream flow data was used from 1973-1983 and 2001 to 2009. Mean and standard deviation (mapstd) function was used for scaling all input and target data using MATLAB. HadCM3 CGCM2 and GFDL model were used for comparative study of the best model. With each one of the GCM models, we had varied the seven different algorithms for achieving the best ANN model. The ANN model takes into consideration adaptive system with different layer of hidden neurons, so we also varied the number of neuron with each algorithm and each model. The best result was obtained for Levenberg-Marquardt algorithm with number of hidden neuron as 10. The Fig. 6. Showed that the value of correlation coefficient (R2) and Mean square error (MSE) was the best as compared to other GCM models.

Conclusion: The main conclusion was that ANN was optimized in terms of various training algorithm, number of neurons in hidden layer and changes the various combinations of learning rate and momentum coefficient. By using various combinations of algorithm and number of neurons used to minimize the performance error, the best result was obtained for Levenberg-Marquardt algorithm with number of hidden neuron as 10. The Fig. 6. showed that the value of correlation coefficient (R2) and Mean square error (MSE) was the best as compared to other GCM models. According to that the future stream flow was predicted for Ranganadi River which indicated an increasing trend in future.