The Chemical Oxygen Demand Modelling Based on a Dynamic Structure Neural Network

Wastewater treatment process aims at achieve the purpose of purification by degradation of organic matter in water. To ensure the effluent water quality, some indicators should be measured, including chemical oxygen demand (COD), Biochemical oxygen demand (BOD), etc. Through the prediction on effluent indicators can provide effective guidance for the operation of wastewater treatment plant. Wastewater treatment process itself is a nonlinear, time-delay process with complex reactions. Thus, when using traditional mathematical model there is often a lack of accuracy, large amount of calculation and lack of flexibility in system simulation, while the prediction based on neural network model can effectively eliminate these disadvantages because of its learning mechanism. Nowadays, applying neural network in wastewater treatment process has become a research hotpot, and some breakthroughs were achieved in terms of algorithms or modelling. zhu, et al. used MLP model to reduce the data dimension, then used the time-delay neural network to predict the effluent BOD online. Chang, et al. reduced data dimension through principal component analysis(PCA), then used extracted system inherent characteristics from data by fuzzy C clustering, at last, TSK-type fuzzy inference system was used to predict the effluent COD. Chai et al. proposed a activated sludge process mechanism model based on hierarchical neural network, connecting the mechanism model and neural network in cascade way and with the neural network identifying the reaction rate of nonlinear components in activated sludge process model to predict the effluent COD. The above evidences show that artificial neural networks can directly establish the model according to the input / output data without prior knowledge of the condition object, and has strong online correction ability. For the process with a large amount of data information which can not be described by rules or formulas, the artificial neural network shows great flexibility and adaptability which is ideal for wastewater treatment systems. However, these network models have the same shortcomings that the network structure would no longer able to modify after finalized in the early stage of designing. For the different cases of wastewater treatment process, the re-design of neural network prediction model is necessary. To solve this problem, meet the needs of the object by dynamically adjusting the neural network structure is an available approach. Huang et al. proposed a simple sequential learning algorithm called the “RBF growing and pruning algorithm” (GAP-RBF), which was later developed into the GGAP-RBF algorithm.