(54e) A Statistical Approach to Prediction of Transmembrane Pressure in Membrane Bioreactors

In water treatment fields such as sewage treatment and industrial liquid waste treatment, the membrane bioreactor (MBR) has been widely used to purify wastewater for reuse [1,2]. MBRs combine biological treatment with membrane filtration. First, bacteria within activated sludge metabolize the organic pollutants and produce environmentally-acceptable metabolites, then a microfiltration (MF) or ultrafiltration membrane separates liquids from solids, i.e. the clean water from the sludge.

However, MBRs suffer from some practical difficulties. One of the critical difficulties is membrane fouling [3,4]. Membrane fouling is a phenomenon wherein foulants, such as activated sludge, sparingly-soluble compounds, high molecular weight solutes and colloids, absorb or deposit on the membrane surface and absorb onto and block the membrane pores. For example, in cases where the MBR is operated under constant-rate filtration, significant energy is required to achieve constant permeate flow rate due to membrane fouling.

To reduce membrane fouling, physical cleaning, for example, aeration, backwashing, or back-pulsing, is performed periodically during MBR operations. Furthermore, chemical cleaning must be carried out with chemical reagents after a given period of processing time, when the transmembrane pressure (TMP) exceeds a given value, because some foulants cannot be removed by physical cleaning and these residual foulants will prevent the recovery of membrane performance with time. Frequent chemical cleaning and replacement of membranes are both expensive.

Hence, to enable chemical cleaning to be performed at an appropriate time [5], membrane fouling must be predicted in the long term. When an MBR plant is operated under constant-rate filtration, this corresponds to prediction of the TMP. Moreover, to enable the distributed MBR systems mentioned above, each MBR must be operated automatically and controlled remotely. Therefore, the TMP must be predicted a priori and a schedule of chemical cleaning must be created in advance.

Geng et al. assumed that fouling that could not be reduced by aeration and backwashing was deep pore clogging and that an increase of TMP could be represented as an exponential function of processing time, then fitted parameters of the exponential equation with experimental data [6]. Kim et al. attempted to adapt changes of states in an MBR plant by recursively identifying the parameters of the exponential equation [7]. Additionally, Chen et al. focused on a critical flux [8], below which membrane fouling can be neglected in the short term, and a TMP jump [9], which is a rapid increase in TMP after a period of processing, and constructed two models that predicted the initial increase and rapid increase of TMP, respectively [10]. However, the predictive performance of these models for new data were not confirmed, although the fitting accuracy of the models was high. Furthermore, water quality variables were not considered in the models and hence it is conceivable that they cannot predict TMP well in other MBR plants and lack broad utility.

On the other hand, Meng et al. studied the relationships between membrane fouling resistance and water quality variables [11]. By investigating correlations among the water quality variables and membrane fouling resistance, the results suggested that there should be a direct influence of MLSS concentration, sludge particle size distribution and extracellular polymeric substances and an indirect influence of the other variables on membrane fouling. Thus, water quality variables have multiple effects on MBR fouling phenomena and the simultaneous consideration of these effects is necessary to predict membrane fouling resistance and TMP in the long term in various MBR plants.

Therefore, in this study, we use chemoinformatics [12] to construct statistical models to predict TMP from other MBR parameters, X, such as water quality variables and operating conditions. Chemoinformatics uses informatics methods to solve chemistry problems and is used to study a variety of topics, such as quantitative structure–property relationships, quantitative structure–activity relationships, reaction design, drug design, and process systems.

In this paper, we target an MBR plant operated under constant-rate filtration, construct prediction models for the increase in TMP per unit time, DTMP, from X, and predict TMP using the models. First, a statistical model is constructed between X and DTMP by using databases of variables that have been measured in the past. Next, the constructed model predicts the increase in TMP from time n to time n+1, DTMP_pred(n+1), by inputting MBR parameters in new data x(n). The target value of TMP at time n+1, TMP_pred(n+1), can be calculated by adding DTMP_pred(n+1) to TMP(n), which is measured previously.

When long term prediction of TMP is carried out, the constructed model predicts DTMP_pred(n+2), DTMP_pred(n+3), ... by inputting expected values or setting values of MBR parameters at times n+1, n+2, ..., respectively. We then use these predicted values as the baseline and can then predict TMP_pred(n+2), TMP_pred(n+3), ... by adding DTMP_pred(n+2), DTMP_pred(n+3), ... to TMP_pred(n+1), TMP_pred(n+2), ..., respectively.

In our study, two methods are used to construct regression models, the partial least-squares (PLS) [13] method and the support vector regression (SVR) [14] method. PLS models are based on linear regression and we can obtain semi-quantitative regression coefficients (or contribution ratios) for each considered parameter to predict the increase in TMP. On the other hand, SVR models are based on nonlinear regression and are reported to have more predictive accuracy than that of neural networks in some cases.

We verified the usefulness of the proposed method with real industrial data obtained from operation of a MBR plant. The data were obtained from a MBR plant operated by Japan Sewage Works Agency in Moka, Tochigi Prefecture, Japan. In this MBR plant, sewage water flows into two lines of the MBR, A and B. MBR system A comprises anoxic and aerobic tanks and B comprises anaerobic, anoxic and aerobic tanks. A membrane module is immersed in each aerobic tank. Flocculants are added in MBR A. Values of process variables measured online, such as TMP and temperature, were averaged daily while those of other variables, measured offline, such as MLSS concentration and soluble total organic carbon, could be obtained only once every few days. Values of X that were not measured were interpolated from measured data. We used a simple interpolation method, where the newest value was used as a missing value, and the Hermite interpolation method,

The analysis showed that the measurement frequency of the process variables, interpolation techniques, and regression methods did not affect the modeling and prediction results. However, it is necessary to pay close attention to the use of Hermite interpolation for variables with low measurement frequency. A linear relationship between the increase of TMP and MBR parameters was indicated because there was little difference between the results of PLS and SVR approaches. Furthermore, the effects of the water quality variables and operating conditions on TMP were analyzed using standard regression coefficients obtained by the PLS method.

Our results suggest that the statistical approach used in this study can be a practical tool for the analysis and prediction of TMP. In addition, we believe that by applying our method to process control, MBR plants could be operated effectively and stably.

References

[1]    W.B. Yang, N. Cicek, J. Ilg, State-of-the-art of membrane bioreactors: Worldwide research and commercial applications in North America, J. Membr. Sci. 270 (2006) 201–211.

[2]    F. G. Meng, S.R. Chae, A. Drews, M. Kraume, H.S. Shin, F.L. Yang, Recent advances in membrane bioreactors (MBRs): Membrane fouling and membrane material, Water Res. 43 (2009)1489–1512.

[3]    P.L. Clech, V. Chen, T.A.G. Fane, Fouling in membrane bioreactors used in wastewater treatment, J. Membr. Sci. 48 (2008) 534–541.

[4]    M. Kraume, D. Wedi, J. Schaller, V. Iversen, A. Drews, Fouling in MBR: What use are lab investigations for full scale operation?, Desalination 236 (2009) 94–103.

[5]    M.J. Kim, B. Sankararao, C.K. Yoo, Determination of MBR fouling and chemical cleaning interval using statistical methods applied on dynamic index data, J. Membr. Sci. 375 (2011) 345–353.

[6]    Z. Geng, E.R. Hall, P.R. Berube, Membrane fouling mechanisms of a membrane enhanced biological phosphorus removal process, J. Membr. Sci. 296 (2007) 93–101.

[7]    M.J. Kim, B. Sankararao, C.K. Yoo, Determination of MBR fouling and chemical cleaning interval using statistical methods applied on dynamic index data, J. Membr. Sci. 375 (2011) 345–353.

[8]    R.W. Field, D. Wu, J.A. Howell, B.B. Gupta, Critical flux concept for microfiltration fouling, J. Membr. Sci. 100 (1995) 259–272.

[9]    B.D. Cho, A.G. Fane, Fouling transients in nominally sub-critical flux operation of a membrane bioreactor, J. Membr. Sci. 209 (2002) 391–403.

[10]  Y. Ye, V. Chen,  A.G. Fane, Modeling long-term subcritical filtration of model EPS solutions, Desalination 191 (2006) 318–327.

[11]  F.G. Meng, H.M. Zhang, F.L. Yang, S.T. Zhang, Y.S. Li, X.W. Zhang, Identification of activated sludge properties affecting membrane fouling in submerged membrane bioreactors, Sep. Purif. Technol. 51 (2006) 95–103.

[12]  J. Gasteiger, T. Engel, Chemoinformatics – A Textbook, first ed., Wiley-VCH, Weinheim, 2003.

[13]  S. Wold, M. Sjöström, L. Eriksson, PLS-regression: a basic tool of chemometrics, Chemom. Intell. Lab. Syst. 58 (2001) 109–130.

[14]  V.N. Vapnik, The Nature of Statistical Learning Theory, second ed., Springer, New York, 1999.