عنوان مقاله [English]
Flooding behavior of rivers of Iran, lack of water and necessity of surface water controlling detect the importance of riverâs behavior simulating and modeling. In this way we can have a long term plan for proper and reasonable operating from potential of rivers. Rain âRunoff water simulation is main step for managing of basins. This process is one of the complicated nonlinear phenomenons in water engineering. Most of calculation and designing in water engineering need a proper evaluation of quantity and quality of running water that comes from a determined rain. There are common and various methods for evaluating of basinâs runoff. Nowadays, using of Artificial Neural Networks (ANN) in various branch of hydrology engineering is acceptance because this method is capable with good accuracy simulate and predict the nonlinear functions. This research tries to predict runoff in southern sub-basin of Gharasoo river in Ardabil province by Artificial Neural Networks (ANNs). This research based on 5 climatic parameters (2007-2010) that affects runoff. These data is obtained from Hydrometric station that located in the end of this basin.
Artificial Neural Networks (ANNs) is a simple model from humanâs brain that with a special mathematic structure in each system is able to clarify the process and nonlinear relation between inputs and outputs. These networks during teaching process are teached and are used for feature predicting. For best designing of ANN model in this research to predict runoff in basin under study, first correlation of humidity average, rain average, monthly temperature average and evaporation average with the flow of basin is obtained. Then effective parameter with more correlation for multi-layer perceptron network is selected. Data matrix with following input and output is made. Inputs: monthly rain average (mm), monthly rain average (million m3), monthly relative humidity average (percent) and monthly temperature average (c) Outputs: Predicting monthly runoff in next yearâs. From existing 39 year statistical period, 90 percent of them is used for net teaching and other 10 percent for test step is used. After selecting input and output data of net and defining net structure (stimulator function, number of neurons ,hidden layers ,number of cycle, amount of educational parameters) Net teaching by program teaching algorithm, first with one hidden neuron is began and with increasing that up to all neuron number is continuing. After each teaching, net is tested via regression analyzing and correlation coefficient between input and output data (in teaching step) and error percent (in test step). Basis of neuron numbers and cycles was maximum correlation and errors less than 5 percent. By detected number of hidden optimum neurons and cycles, to reach an optimum network several times the value of teaching parameters and the number of hidden layers is changing. For this reason network is designed in a way that by entering last years information (rain, relative humidity, temperature, evaporation, runoff) is able to predict next yearâs flow with error less than 5 percent. After designing of 12 various networks for predicting of basin flow, various structure of percepetron is selected to reach an optimum network. For evaluating of ANN function, amount of R2, RMSE, MAE and R are used
The result of this research show that in all months there is a high correlation (more than 93%) between runoff and average of rain, humidity, temperature, evaporation and monthly flow. Minimum correlation coefficient in teaching step belongs to April (93%) and Maximum belongs to Jun (98%). For this research, Marcoart-Levenberg algorithm is the best algorithm because of more correlation in Teaching step and lower error in test step. For defining proper number of hidden neurons, maximum number of neurons for all 12 networks is 10 neurons. In the selected network, most of neuron number belongs to January with10 neuron in the first- layer and 2neuron in the second hidden layer and minimum of them is related to October with 3neuron in the first hidden layer. For defining hidden layer number, some of the networks (except February, March, May, July, October and December) with one hidden layer have a good result and some other with 2 hidden layers have a good result. The primary number of teaching cycles of network for each month in Marcoart-Lonberg algorithm first with 10 cycle for each neuron in hidden layer and with initial error value (=0.005) starts and maximum up to 700 continuing and in the end network with minimum cycle (10 cycle) in July and maximum cycle (700 cycle) in December reached to its goal.
The result of this research show that one model with 5 parameter including(monthly rain average, monthly runoff average, monthly relative humidity average, monthly evaporation average and monthly temperature average) is the best ANN for predicting the flow of river because with error less than 5 percent and high correlation can predict the runoff level. Number examination of various neuron in hidden layers show that one model with 4 neuron in the first hidden layer and 3neuron in second hidden layer sigmoid tangent stimulator function in the first hidden layer and 10 cycle, has best accuracy. The best ANN model in this research is one perceptron model with 3 layers and 4neuron in the first hidden layer and 4 neuron in the second hidden layer a 2 hidden layer stimulator function an one output and Marcoart-Lonberg teaching algorithm. The result of this research show that ANN model with low error and proper capability for predicting of basin rivers flow is a good model for evaluating of this parameter in future.