Development a Proper Model for Online Estimation of the 5-Days Biochemical Oxygen Demand Based on Artificial Neural Network and Support Vector Machine

IntroductionFive-day biochemical oxygen demand (BOD5) is one of the key parameters which is widely applied to evaluate the biological and chemical conditions of rivers, and to study the effects of releases in rivers from wastewater treatment plants, factories, farms, etc. Measurement of this parameter requires about 5 days. However, 5 days is a long time for the determination of BOD5. Therefore, in the present research two software sensors, i.e. artificial neural network (ANN) and support vector machine (SVM) are developed for the online estimation of BOD5. To achieve this goal, the Sefidrood River in Iran was selected. It is pointed out that although ANN models have been used to predict BOD5, but using SVM techniques for the online estimation of this parameter have never been reported.Materials and Methods ANN and SVM models are two well known artificial intelligent techniques which have frequently been used for modeling nonlinear parameters in engineering problems. In this research, a feed-forward ANN technique and a regression model is considered for online prediction of BOD5. To optimize ANN parameters, three optimization algorithms, i.e. Levenberg-Marquardt (LM), resilient back-propagation (RP), and scaled conjugate gradient (SCG) are used. Besides, two-steps grid search algorithm is considered for the optimization of SVM parameters, i.e. ε, C and γ.Results and DiscussionTo achieve the best network geometry according to the highest correlation (R), and the lowest root mean square error (RMSE) of the testing sets for online estimation of BOD5, various structures of NN model with different training functions are investigated. Finally, the network with LM algorithm (ANN (LM) model) is selected as the best ANN model. The results of training and testing steps for developed ANN models with different training functions are given in Table 1.In addition, for SVM model development, first, ε, C and γ should be determined. In this research, the best fitting value was obtained by trial and error procedure. Therefore, this parameter was set to several values. Thus, 14 models were developed (ε and C are considered to be varied). The RMSE value has been assessed on different values in the different -SVM regression models. Obtained results indicated that the SVM model including the value equal to 1.472 has the least RMSE. Therefore, it can be chosen as the best SVM model for the online prediction of BOD5. However, optimal values for C and using two-step grid search were 13 and 0.037, respectively. Finally SVM model is trained and tested using the determined best fitted values.