3.1. Pilot Plant Operation Results
During 114 days of operation of the BF-MBR pilot plant, notable trends in TMP, permeability, permeability slope, MLSS in the membrane separation chamber (MLSS-III) and COD removal were observed (Figure 2
), allowing the development of the qualitative description of the biological activity and its influence on membrane separation process.
The first period (1–20 days) can be described as the period of biological adaptation and biomass development. It is characterized by moderate growth of biomass up to MLSS-III 5–6 g/L and increasing biodegradation of organics in the range of 67–81%, together with a steep TMP growth and a respective decrease of permeability at a relatively high rate of 0.35–0.47 LMH/bar/s. This state can be identified as conditioning fouling.
After reaching the conditionally critical value of 1.7 times permeability decrease, the return of suspended solids from separation chamber (III) to MBBR chamber (II) was doubled, leading to stabilization of permeability and MLSS-III in the next period (21–34 days) and decreasing the membrane fouling rate to 0.25–0.27 LMH/bar/s by permeability, which is considered steady fouling.
In order to increase the system productivity in terms of permeate, membrane flux was increased, entailing the TMP jump during the third period (35–36 days), which indicates a severe fouling. Following that, backwash and relaxation times were adjusted in order to stabilize rapid fouling development during 37–44 days.
Chemical cleaning (CIP), applied in the fifth period, exhibited relatively high values of the recovered membrane permeability. While recovery of the permeability between the backwashes performed at the end of every filtration cycle was in the range 88–126%, recovery of the permeability after CIP was in the range of 158–182%.
The sixth period (48–77 days) was another steady fouling state. It reproduced the same trends from the second period (21–34 days), except for a more stable COD degradation due to well-developed biofilms in MBBR part and on carriers in the separation chamber (III). After reaching 400 mbar of TMP, a second chemical cleaning was provided, applying higher backwash pressure with the subsequent soaking of the membrane elements in the cleaning solutions, which caused the permeability to recover to the initial value.
The last, eighth period of system operation is a control period which is characterized by both conditional and steady fouling in the permeability pattern.
In general, in the way described above, the operation of the BF-MBR pilot plant was observed during all the states, which is important for the determination of membrane fouling patterns: Conditional fouling, steady fouling, and TMP jump at different fluxes. Two chemical cleaning procedures were conducted to estimate the recovery of permeability. Data, which were recorded during these states, were taken as the basis for further statistical analysis.
3.2. Statistical Determination of Membrane Fouling Patterns
According to the literature, the influence of the mixed liquor parameters (i.e., MLSS, SVI (DSVI), COD, and RH) on the filtration performance and fouling intensity is controversial. Indeed, a positive impact of higher MLSS concentration on MBR hydraulic performance has been indicated [15
]. On the contrary, Chang et al. [46
] observed a positive link between the MLSS increase and the flux decline, which is the opposite of its effect on the specific cake resistance, while Brookes et al. [107
] and Jefferson et al. [108
] showed that MLSS concentration is not a governing factor influencing the overall membrane fouling, and no consistent correlation was observed between MLSS and fouling intensity.
The influence of the relative hydrophobicity on system performance is also not fully comprehended. According to the findings by Deng et al. [40
] and Huang et al. [109
], high RH fosters the mitigation of fouling due to the weaker interactions of hydrophobic flocs with a hydrophilic membrane. In addition, lower RH values entail floc deterioration and the consequent increase of cake layer resistance [29
], whereas higher RH values are associated with better flocculation [60
]. Meanwhile, as specified by Meng et al. [36
] and Tian et al. [64
], higher RH of sludge causes the formation of a more dense cake layer on the membrane surface, resulting in a greater TMP rise.
There is a lack of data on the correlation between SVI and membrane fouling intensity. Chae et al. [110
] stated that high SVI values corresponded to severe membrane fouling in an MBR system. Ng et al. [111
] linked the increased SVI values to the higher ratio of non-flocculating components of the activated sludge but did not mention if this affected the fouling intensity. In contrast, according to Fan et al. [112
] and Wu and Huang [113
], this parameter is not a reliable indicator to predict the membrane fouling potential for MBR systems and has no effect on membrane filterability.
As found, COD is indirectly related to the fouling intensity. COD is linked to the concentration of soluble foulants which have a negative effect on membrane filterability [114
]. In addition, COD in the effluent from aerobic and anaerobic biological systems is encountered in the form of soluble microbial products which are among the foulants in MBRs [115
]. Meanwhile, Lesjean et al. [116
] found no clear correlation between COD and the fouling intensity.
Hence, to gain a deeper understanding of the role of the mixed liquor characteristics in the filtration performance of the investigated system, it was decided to monitor these parameters and their variation over time in the separation chamber (Table 3
) and to process the collected data statistically.
Since the operating conditions varied significantly throughout the whole filtration period (Table 2
), which influenced both the activated sludge parameters and the fouling indicators, it was decided to split the whole data range into its characteristic phases and statistically analyze them separately from each other, excluding the data which covered the chemical cleanings. Hence, three basic periods were established: period A (days 3–34), period B (days 49–77) and period C (days 86–114).
PLS regression (also known as a projection of latent structures) was used as an advanced mathematical and statistical tool to model the relations between the X variables and the Y responses within every single period (Table 4
The X- and Y-matrices were modelled simultaneously to find the latent variables in input X parameters that best predicted the latent variables in the corresponding Y responses (i.e., PCAs on the X- and Y-data were performed with the subsequent acquisition of the relative scores). Then, the plotting of two sets of the scores (those related to X and Y) against each other was conducted, maximizing the covariance between X and Y [117
The obtained model was validated by applying a random cross-validation in PLS. The number of PLS components (factors), was chosen according to the explained variance.
The results of the performed analyses of the data from the initial period of the system performance (Period A) are shown below (Figure 3
The correlation loadings plot is computed by accounting for each variable for the displayed latent variables (factors). From the loadings plot, Factor-1 clearly describes DSVI, dDSVI/dt, TMP, COD, dMLSS/dt, permeability, Pn, and its slope, dPn/dt, since the first three variables are located at the far left, and the rest at the far right along the Factor-1 axis. Factor-1 also accounts for dCOD/dt, while MLSS and dRH/dt mainly contribute to Factor-2. According to the PLS loadings plot, COD and DSVI explain more than 50% of the variance and are probably the most important variables. DSVI has a negative correlation with both permeability and permeability slope, but is positively linked to TMP. Particularly in this case, COD has a negative correlation with the variables DSVI, dDSVI/dt, MLSS and dMLSS/dt, and is negatively linked to the average normalized permeability (nP). Although the rest of the variables are located in the inner ellipse, which indicates up to 50% of the explained variance and thus does not contain enough structured variation to discriminate between the mixed liquor samples, it was decided to keep them to make the model more reliable.
The analysis of the scores and loadings plot and the bi-plot demonstrates that the samples from days 1–20 are mostly characterized by higher RH, dRH/dt, MLSS, dMLSS/dt, COD, and dCOD/dt, while the samples taken during the period 22–34 day have higher DSVI and dDSVI/dt values.
As demonstrated by the graph of explained variance (Figure 3
c), it is preferable to use five components, since this number gives a lower residual variance.
According to the Figure 3
d (the validation graph), the developed model is linear (R-squared = 0.73) and with a reasonable fit to the majority of data: Slope = 0.81, offset 0.07 and the dispersion of the validation samples around the regression line (Root Mean Square Error of Cross Validation–RMSEV) and the standard error of cross-validation (SECV) are approximately 0.036. Consequently, the model is reliable and can be used for future predictions for the defined number of factors under the operational conditions applied during the period A.
Relative hydrophobicity and its change required much more effort and time to be experimentally determined than other variables. In addition, RH and dRH/dt are characterized by relatively low-weighted regression coefficients: 0.02 and −0.086, respectively (Factor-2); and, 0.07 and 0.04, respectively (Factor-1) (i.e., these variables are not well explained by the components). Considering the above-mentioned aspects, it was decided to exclude RH and dRH/dt from further monitoring and analysis.
The second period, B, covers the filtration performance data collected between the first and the second chemical cleanings of the system. Obtained results of the PLS analysis are represented below (Figure 4
According to the bi-plot (Figure 4
b), the majority of the samples within period B are characterized by higher dCOD/dt values. Meanwhile, the samples taken on days 49–50 are characterized by higher COD values; on days 51, 57 and 68 by relatively high dMLSS/dt, DSVI, and dDSVI/dt values; on day 72 by comparatively high dCOD/dt values; and on days 76 and 77 by more significant MLSS values.
According to the correlation loadings plot, Factor-1 apparently describes TMP, MLSS, COD, average permeability (avPn), dPn/dt, DSVI and dDSVI/dt. Factor-2 is related to dCOD/dt and dMLSS/dt. All the variables were marked as significant according to the plot of correlation loadings, even though the MLSS variable gives slightly less than 50% of the explained variance. MLSS and dCOD/dt are positively linked to the TMP response, in contrast to dMLSS/dt, DSVI, dDSVI/dt, which have a negative correlation with TMP and the permeability slope (dPn/dt). The COD variable has a high positive correlation with dPn/dt and is positively linked to the average permeability (avPn).
demonstrates that the optimum number of factors is five, which provides more than 57% of the explained Y-variance.
An analysis of the validation plot shows that the developed model is linear, having R-squared = 0.71 and with a good fit to the majority of data (i.e., slope = 0.64). RMSEV and SECV are approximately 10, but it is essential to acknowledge that the mentioned errors have the same units as the reference Y (in this case, average normalized permeability, avPn). R-squared (Pearson) is close to R-squared correlation (0.68 vs. 0.82), which indicates the reliability of the model. Consequently, a good prediction is attained with the developed model, which proves that the model is reliable and can be used during further stages when the operating conditions applied in the period B are replicated.
The output from the PLS modelling of the data acquired after the second CIP (the period C) is demonstrated below (Figure 5
The bi-plot shows that the samples from day 89 have a higher DSVI value, while dMLSS/dt and dCOD/dt are the most distinctive parameters for days 91 and 96. Days 100, 107 and 110 are characterized by higher COD content, whereas days 103, 105 and 114 have higher MLSS values. Day 112 is characterized by a higher dDSVI/dt.
From the correlation loadings plot (Figure 5
b), COD, MLSS, TMP, dDSVI/dt, DSVI, avPn
/dt contribute to Factor-1, while Factor-2 describes dMLSS/dt and dCOD/dt. All the specified variables explain more than 50% of the variance and thus have high importance in relation to Factor-1 and Factor-2. MLSS and dDSVI/dt are positively linked to TMP and have a negative correlation with the permeability indicators, avPn
/dt. DSVI is positively correlated to dPn
/dt, while dMLSS/dt and dCOD/dt have a negative correlation with the permeability slope.
The explained variance plot indicates that the optimum number of factors is four, which provides more than 70% of explained Y-variance.
The points of the validation graph in Figure 5
d have a linear trend (R-squared = 0.8), having a good fit to the majority of data (slope = 0.93). R-squared (Pearson) is close to R-squared correlation (0.79 vs. 0.89), which indicates the reliability of the model. Only the errors RMSEV and SECV are higher than in previous cases, but this can be explained by the higher values of the response function (average permeability) in this particular case.
Since the higher amount of data was available to be collected during the last period C (Table 5
) in comparison to the previous modes, it was decided to apply the predict function to new data.
Full prediction with the identification of outliers was used. The following results were obtained (Figure 6
The deviation between the predicted and the reference values is in the range 0.01–0.034, which demonstrates the reliability of the applied model.
Consequently, a good prediction is attained by applying the developed model, which proves that the model is reliable and can be used during further stages under the operating conditions that were applied during period C.
In addition, MLR was performed using leverage correction. However, obtained results are unreliable since the same data was validated and used for the prediction, which provided overly optimistic results. The application of the test matrix in MLR would merely copy the PLS strategy but do so in a more difficult way. MLR is a simpler way of doing the calculations, but PLS is much more advanced due to the applied validation techniques.
SRT and permeate flux are among the key operating parameters controlling fouling intensity in MBR.
In order to estimate the influence of SRT on the performance of the current system, this parameter was included in the models as an additional variable. The acquired results are represented in Figure 7
According to the correlation loadings plot related to period A, SRT explains less than 50% of the variance and thus has relatively little influence. In this particular case, SRT exhibits an independent variation in relation to other variables, except for COD, which has a weak positive link with SRT. Meanwhile, SRT exhibits a slightly negative correlation with the normalized permeability and permeability slope for period A. Concerning the model enhancement, the introduction of the new variable did not entail any significant improvement: RMSECV was just 0.002 less than its value in the initial model, while the bias, on the contrary, showed an order of magnitude increase in absolute value.
The results related to period B demonstrate that SRT is an important variable which explains more than 50% of the variance in the dataset. It has a strong negative correlation with COD and the normalized permeability. In addition, SRT is positively correlated with MLSS along Factor-1. The negative correlation between SRT and COD during this period can be attributed to the higher treatment performance of the biomass, which becomes well-developed at SRT up to 40 days and thus is capable of a more efficient biodegradation of organic contaminants, particularly SMPs, causing the decrease of COD values [118
]. Meanwhile, the increase in SRT promotes the development of higher MLSS concentrations [120
], thus inducing membrane fouling.
The introduction of the new variable into the existing model decreased its linearity R-squared = 0.65 vs. R-squared = 0.71 (values in the new model vs. values characteristic for the basic model related to period B), with a slightly worse fit to the majority of data (slope = 0.52 vs. slope = 0.64), RMSEV 10.9 vs. 9.9, SECV 10.97 vs. 10.1, bias 2.73 vs. 1.7. In addition, the new model underestimated a sample from day 72 (marked with the blue circle).
The modelling of the dataset from period C demonstrates the importance of the SRT variable. SRT is highly positively correlated with MLSS and TMP, and is negatively linked to normalized permeability and its slope, hence indicating the fouling enhancement through the increase of MLSS at higher SRTs, which agrees with the previous findings by Le-Clech et al. [29
], Van den Broeck et al. [120
], Yigit et al. [121
]. The positive link between SRT and COD along Factor-1 during this period can be attributed to the accumulation of small microbial by-products (SMP with the molecular weight (MW) < 1 kDa), which contribute to fouling through deflocculation at high SRTs (>31 days) [118
]. However, further studies are required to confirm the presence of different groups of microorganisms at various SRTs in this system (for example, tightly and loosely bound EPS, small SMP, etc.), since the deep investigation of the biomass content was not in the scope of the current research.
The new model exhibits higher linearity (R-squared = 0.89 vs. R-squared = 0.80) and a slightly higher accuracy (RMSEV = 20.9 vs. RMSEV = 28.7; SECV = 21.8 vs. SECV = 29.6; and, bias = −3.4 vs. bias = −6.0) than the initial model.
It is noteworthy that the purpose of including SRT in modelling was not to improve the models for the relevant periods developed earlier in this work, since the inclusion of a new variable is undesirable as it could complicate the model (i.e., it is preferable to use as low a number of variables as possible) [123
]. Besides, the introduction of the SRT variable to the model covering period C barely decreased the deviation in the prediction of the new dataset (Table 4
; 0.016–0.0261 vs. 0.011–0.034), making the extension of the model size unreasonable for its further use in the system controller. The scope was to show the influence of SRT on the operational parameters and fouling intensity in the current system to achieve the highest possible fouling inhibition.
As discovered, SRT should be less than 31 days to avoid a severe membrane fouling. This can be called the critical SRT. The SRT that can be applied without a sharp decrease in permeability is 20 days for the current BF-MBR system. In the studied pilot plant, SRT was adjusted by changing the frequency of sludge removal and the volume of the removed sludge per batch.
Concerning the permeate flux, it can be decreased in order to minimize the filtration resistance if the biomass exhibits high fouling propensity. The current system worked at a constant permeate flux, which varied depending on the monitoring period (Table 2
). In general, all the applied fluxes were below the critical flux value to avoid a severe membrane fouling [124
]. The critical net flux was determined by the flux-step method, described by Miller et al. [127
], and was in a range of 12–15 LMH.
In addition to the desludging option, the concentration of the mixed liquor in the separation and biological chambers was regulated by adjusting the RAS pumping intensity (i.e., pulse length and frequency). The introduction of the RAS line made it possible to build up the desired level of biomass in biological and separation chambers, and to adjust the endogenous decay of the biomass, thus providing sufficient COD and NH4+ removal.
To summarize, the monitored mixed liquor characteristics allowed the controlling of the fouling intensity by adjusting the operating conditions which helped to maintain the stability of the system performance and, hence, the permeate quality: BF-MBR installation assured 100% MLSS elimination and 67–90% treatment efficiency in terms of COD removal, keeping the TMP below 500 mbar.