Research On Soft Sensing Method For Wastewater Treatment Process

With the development of productivity, the amount of human’s daily live wastewater and industrial wastewater increase rapidly, wastewater treatment is becoming highly valued, it is very important to improve the level of wastewater treatment. In order to achieve the purpose of efficient wastewater treatment we must detect some of the key water quality parameters in the process of wastewater treatment, and implement timely and effective regulation and control. However, some parameters cannot be measured directly, this is why we must study on soft sensor methods.The first part of this paper has studied some variable selection methods, they are variable importance of projection(VIP), genetic algorithm(GA) and moving window(WM).and discuss the advantages and disadvantages of them. Then combine these three methods with partial least square regression model, the simulation platform is MATLAB, in order to forecast the amount of BOD. The results show that, after apply on variable selection, although the running time of model has increased slightly, model’s ability of prediction has improved significant.The second part this paper has studied two online models: sliding time window(STW) and just in time(JIT) and also the offline model: time difference(TD). Discuss the advantages and disadvantages of each method, make models using the variable selection methods, simulation results show that the accuracy of the online models are good. In addition, when compare the online model with offline model, we find that the latter one’s prediction accuracy is much worse. That is to say, when establish an online model, we give full considering the characteristics of the working point of the system, as time goes by, the online model can find data which most relevant to the working point of the system, Thus the model can reflect the state of the system, as a result, the prediction becomes more accurate.Finally, this paper study the method based on subspace identification. First, the algorithm is deduced in detail. Then combine the algorithm with online model(JIT) and offline model(TD). The simulation results show that not all the online model predict better than the offline model, because the data of this chapter is sampled every 15 minutes, the dynamic characteristics are very obvious, it just find few data when searching for relevant data, so it get less information about the system and the accuracy of the model is pool. While the model based on the TD method, it can extract the dynamic part of the system through differential operation of the sample data. That is why the offline model plays better than the online model. As a result, we should examine a plant’s circumstances before we built our models.