Evaluation of artificial neural network and multiple regression model for Cd ( II ) sorption on activated carbons

Introduction Industrial effluents of metallurgical alloying, ceramics, electroplating, photography, pigment works, textile, printing, chemical industries, Cd/Ni batteries and lead mine drainage containing toxic metal such as cadmium (Wesley Eckenfelder Jr, 2000; Cheremisinoff, 1995). Higher concentration of cadmium causes gastrointestinal discomforts, kidney damage, nausea, vomiting, and diarrhea, destruction of red blood cells, renal disorder, itai-itai disease and high blood pressure (Drush, 1993). Stricter environmental regulations on cadmium discharge proclaim due to their toxicological and bioaccumulation tendency of cadmium. In India, the permissible limit for cadmium discharge into inland surface (IS: 1991) is 2.0 mg/L and for drinking water (ISI: 1982) is 0.01 mg/L. This necessitates the removal of cadmium from wastewater before its discharge into the environment. The removal of cadmium from aqueous system is effectively done by adsorption technique using activated carbon as an adsorbent (Srinivasan, and Balasubramanian, 2003). Selecting suitable operating conditions for adsorption of Cd(II) from wastewater requires many experimental studies involving many functional parameters. Factors influencing the Cd(II) removal efficiency of activated carbon is quite complex in nature. But once a model is developed, it simplifies the tremendous work. Numerous reports are available with respect to multiple regression equation (Manju, and Anirudhan, 1997; Raji, and Anirudhan, 1997; Babu, and Ramakrishna, 2002) and artificial neural network (ANN), applications for wide range of problems in water and wastewater treatment purposes (Prakash, et.al., 2008; Fagundes-Klen, et.al., 2007; Salari, et.al., 2005; Abbaspour, and Baramakeh, 2005; Arzu Sencan, et.al., 2006; Ivo M. Raimundo Jr. and Narayanaswamy, 2003). ANN has already been used to develop a model for complex adsorption system of metal removal using activated carbon (Alireza Khataee, and Ali Khani, 2005). A neural network is a powerful data modeling tool that is able to capture and represent complex input/output relationships. An artificial neural network is similar to biological neural system. ANN is composed by simple elements operating in parallel. The unit element of a network is the neuron; a simple mathematical function is associated with the so-called intraneural connections. The ANN mathematically transforms an input vector in to an output vector through a suitable transfer functions (Simon Haykin, 2008). The model prediction with ANN is made by learning of the experimentally generated data by changing the connection weights. Each neuron receives the information in the form of inputs that occurred at the previous neurons. This information is processed together with the weights values of each connection of this neuron with previous one and the transfer function. The most common neural network model is the multilayer perceptron (MLP) is known as a supervised network (Rojas, 1996). MLP network requires a desired output along with input in order to learn. The predicted model correctly maps the input to the output using historical data so that the model can then be used to produce the output, when the desired output is unknown. The back-propagation algorithm is used in layered feed-forward ANN. The artificial neurons are organized in layers and signals, are forwarded and then the errors are propagated backwards. There may be one or more intermediate hidden layers. The backpropagation algorithm uses supervised learning (http://rxiv.org) and it computes the network with the given input and output data. The algorithm comparing the simulated output data with the actual/given output data and the error is calculated. The error is reduced by the algorithm until the ANN learns training data by changing the connection weights between layers.