مدل شبکه عصبی مصنوعی برای اثرات کود مرغی بر روی آب زیر زمینی An artificial neural network model for the effects of chicken manure on ground water

۱٫ Introduction Chicken farms, amounting to nearly 400, widely exist in the province of Corum and have become an important source of ground water pollution in the area. In these farms the manure is transferred by means of pressurized water to the manure pool. In the course of this transfer and following operations, chicken manure penetrates into the ground water by runoff, flooding and diffusion. Furthermore farms get their water supply from 20 to 90 m deep wells. For predicting the degree of pollution for major pollutant constituents in ground water wells in poultry farms, one approach could be the identification of an input–output relationship between the involved variables based on the field measurements. From this perspective, artificial neural networks (ANNs) are powerful tools that have the abilities to recognize underlying complex relationships from input–output data only [1]. ANN models have been widely used tools in the field of water quality prediction [2–۶]. An artificial neural network is an information processing system that imitates the behavior of a human brain by emulating the operations and connectivity of biological neurons [7]. It performs a human-like reasoning, learns the attitude and stores the relationship of the processes on the basis of a representative data set that already exists. In general, the neural networks do not need much of a detailed description or formulation of the underlying process, and thus appeal to practicing engineers who tend to rely on their own data [1]. 1.1. ANN modeling Depending on the structure of the network, usually a series of connecting neuron weights are adjusted in order to fit a series of inputs to another series of known outputs [1]. When the weight of a particular neuron is updated it is said that the neuron is learning. The training is the process that neural network learns. The adaptability, reliability and robustness of an ANN depend upon the source, range, quantity and quality of the data set. The feed forward neural networks consist of three or more layers of nodes: one input layer, one output layer and one or more hidden layers. The input vector passed to the network is directly passed to the node activation output of input layer without any computation. One or more hidden layers of nodes between input and output layers provide additional computations. Then the output layer generates the mapping output vector. Each of the hidden and output layers has a set of connections, with a corresponding strength-weight, between itself and each node of preceding layer. Such structure of a network is called a multi-layer perceptron (MLP) [1]. A feed-forward back-propagation artificial neural network (BPNN) is chosen in the present study since it is the most prevalent and generalized neural network currently in use, and straightforward to implement. Fig. 1 illustrates the basic configuration of the network model. Each interconnection in the model has a scalar weight associated with it, which modifies the strength of the signal. The function of the neuron is to sum the weighted inputs to the neuron and pass the summation through a non-linear transfer function. In addition, a bias can also be used, which is another neuron parameter that is summed with the neuron s weighted inputs. Back-propagationrefers to the way the training is implementedand involves using a generalized delta rule [1]. A learning rate parameter influences the rate of weight and bias adjustment, and is the basis of the back-propagation algorithm [8]. The set of input data is propagated through the network to give a prediction of the output. The error in the prediction is used to systematically update the weights based upon gradient information [9]. The network is trained by altering the weights until the error between the training data outputs and the network predicted outputs is small enough. There are many back-propagation training algorithms available. The choice of algorithm depends on the type of problem and may require experimentation of different algorithms. The algorithms have different computation and storage requirements, and train data at different speeds [10]. The goal of selection is to efficiently and accurately train the network while keeping the speed of training relatively fast. After generating sets of training patterns, appropriate NN architecture andassociatedparametersmust be chosenfor theparticular application. The main design parameters are the number of hidden layers, number of neurons in each layer, and the neuron processing functions. The choice of these parameters will depend on the complexity of the system being modeled and they will affect the accuracy of the model. If the number of hidden neurons is too high, the network may over fit the data. On the other side if the number of hidden neurons is small, network may not have sufficient degrees of freedom to learn the process correctly [11]. There is no exact guide for the choice of the numbers. The architecture of most ANN model is designed by trial and error [12]. In this work, we have investigated the modeling of the effects of chicken manure on ground water by artificial neural networks. An ANN model was developed for predicting the total coliform in the ground water well in poultry farms. The backpropagation algorithm was applied to training and testing the network. Levenberg–Marquardt algorithm [13] was used for optimization. The model holds promise for use in future in order to predict the degree of ground water pollution from nearby chicken farms.