Modelling of Escherichia coli removal by a low-cost combined drinking water treatment system

Availability of safe drinking water is a major challenge in many rural and suburban areas of developing countries (Pandit & Kumar 2019; Treacy 2019). Globally, around 780 million rural and 136 million urban dwellers lack access to an improved drinking water supply (RWSN 2010). In sub-Saharan Africa, the discrepancy is even bigger with 272 million rural population lacking access to safe water, compared to 54 million in urban areas (RWSN 2010). Consumption of contaminated water results in waterborne disease outbreaks. Safe drinking water provision through point-of-use (PoU) water treatment is among the key measures required to prevent such outbreaks. While centralized piped water supply is the ideal solution for meeting drinking water needs in many communities worldwide (Lantagne & Yates 2018; Pandit & Kumar 2019), PoU water treatment has been shown by various authors (CAWST 2011; Kausley et al. 2018; Lantagne & Yates 2018; Pandit & Kumar 2019; Treacy 2019) to improve drinking water safety and reduce the burden of waterborne diseases. It is sometimes the only cost-effective option in many rural and suburban areas of developing countries. Although efforts to develop simple yet effective low-cost PoU technologies for rural and suburban areas have intensified globally (Treacy 2019), challenges still exist (Treacy 2019). Therefore, there is still need for development and/or optimization of more PoU techniques appropriate to poor communities. Mathematical modelling may assist in the design and optimization (costs, material combination, etc.) of various PoU systems and can support further research in terms of configuration, flow rate, media combination, and so on while also serving as a decision support tool.

Low-cost PoU water treatment technologies can be broadly classified into five groups (Lantagne & Yates 2018): (1) chemical disinfection (e.g. chlorine disinfection); (2) disinfection by heat (e.g. boiling), ultraviolet or solar radiation; (3) coagulation, flocculation, and sedimentation; (4) filtration (e.g. slow sand filtration); and (5) combined (multi-barrier) systems (CAWST 2011; Pandit & Kumar 2019). The priority of most PoU systems is to make water bacteriologically safe (CAWST 2017, 2011) and aesthetically acceptable (WHO 2017). Good aesthetic quality promotes health gains from drinking safe water (Lantagne & Yates 2018). Water of poor aesthetic quality, although safe, is often avoided (WHO 2017). Additionally, particles that contribute to poor aesthetic quality hinder bacterial inactivation (WHO 2017).

A thorough review of literature showed that modelling of PoU and similar systems for contaminant removal or system optimization has mainly been done on uncombined systems, for example: (i) intermittently operated slow sand filters (Fulazzaky et al. 2009; Jenkins et al. 2011), (ii) disinfection using chlorine (Lee & Nam 2002), (iii) disinfection by natural herbs (Somani & Ingole 2012), (iv) disinfection by silver or silver-coated materials (Chong et al. 2011; Rossainz-Castro et al. 2016; Singh et al. 2019), (v) granular activated carbon filtration (Hijnen et al. 2010), (vi) filtration by geotextile fabrics and other fibrous filter media (Choo & Tien 1991; Siwila & Brink 2018), and (vii) ultraviolet disinfection (Brownell et al. 2008), etc.

This paper presents modelling of Escherichia coli removal by a combined drinking water system developed by the authors (Siwila & Brink 2020) as a contribution to research and development on low-cost PoU drinking water treatment. The combined system consists of three treatment stages: pre-filtration by geotextile fabric followed by filtration and disinfection by SCCGM then GAC filtration (Siwila & Brink 2020). Each of these steps were modelled as a series of compartments by using specialized theoretical removal mechanisms for each barrier. E. coli was modelled as a microbial colloid or particle as proposed in literature (Tufenkji et al. 2003; Hijnen et al. 2010).

E. coli inactivation by SCCGM was modelled using Chicks, Chick-Watson, complete mix system and Collins-Selleck disinfection models (MWH 2012; Metcalf & Eddy 2014; Qasim & Zhu 2018), which were considered sufficiently representative to describe the removal. The Chick's and Chick-Watson models have been applied by various authors (Rossainz-Castro et al. 2016; Shimabuku et al. 2018; Singh et al. 2019) to model bacterial removal by silver and other metals. Additionally, Chick worked with silver nitrate among other disinfectants and E. coli among other organisms (MWH 2012). Geotextile removal was modelled using colloidal filtration theory (CFT) models for removal of hydrosols by fibrous media developed by Guzy et al. (1983) and Choo & Tien (1991) as presented by Tien (2012). The filtration removals by the SCCGM and GAC were modelled using the Yao CFT model for removal of colloidal particles from liquids by granular media developed by Yao et al. (1971) then refined by Rajagopalan & Tien (1976) (the RT model) and expanded further by Tufenkji & Elimelech (2004) (the TE model) (MWH 2012). The highlighted removal theories, governing equations and respective modelling procedure are explained in the methodology section of this paper.

Applied modelling of multi-barrier systems such as presented here can be helpful to system design and optimization (MWH 2012). This can help engineers understand the governing characteristics and contribution of each barrier to the effluent quality, subsequently enabling them to make informed decisions on the appropriate optimization measures. The importance of each treatment stage and associated removal mechanisms can be assessed as a function of system parameters (MWH 2012). This can allow engineers to vary design parameters until the desired system effectiveness and cost are achieved. For example, the modelling may assist in minimizing cost while ensuring the system is effective while deciding on which component needs more attention to increase removal efficiency of the water quality parameter of interest.

Since testing for every possible pathogen in water is difficult, time consuming, and expensive, indicator organisms such as Escherichia coli are often employed to assess bacteriological safety of drinking water (CAWST 2017). The presence of E. coli is definitive evidence of fecal contamination (Horan 2003; Brandt et al. 2017). Thus, if E. coli is detected in treated water, it indicates the presence of fecal matter and potentially pathogens (CAWST 2017). This signals potential malfunctioning in the responsible water treatment system posing a health risk requiring urgent action. In addition, E. coli has been indicated to be a better indicator for predicting diarrhoeal and gastrointestinal disease-causing pathogens than fecal coliforms particularly when detected in tropical drinking waters (Horan 2003; Brandt et al. 2017; Qasim & Zhu 2018). Therefore, bacterial removal by use of E. coli was chosen for modelling in this research. However, future research should expand to include the other pathogen classes (i.e. viruses, protozoa and helminths).

The general experimental methodology aspects (study setting, design aspects, set up, sampling, testing methods, etc.) are presented in Siwila & Brink (2020). For instance, the measured E. coli log removal values (Table 4 and Figure 4) and contact time (Table 4) and resulting Ct values were calculated based on the work done in Siwila & Brink (2020). Furthermore, flow rate measurement, sample collection, contact time estimation, etc. as well as E. coli testing and enumeration procedures are given in Siwila & Brink (2020). The methodology for the present work primarily presents the mathematical modelling approach for prediction of E. coli removal by the modelled system. The schematic diagram of the combined system that is modelled is given in Figure 1. The system consisted of geotextile fabric for pre-filtration, SCCGM for filtration and disinfection, GAC filtration and a safe storage compartment for treated water. The key system parameters, particularly, those applied directly to the modelling in this paper are included in Table 1.

Sampling was done after at least 7.5 liters of water was passed through the system and at varied flow rates for the first 9 runs and at 2 L/h for the last 3 runs (Siwila & Brink 2020). The first four filtration runs were done at the maximum obtainable flow rate of 10 L/h (Table 4), while subsequent flow rates were varied from 8 L/h to 2 L/h (Table 4). Varying the flow rate was done to arrive at an optimal flow rate and produce varied contact time, and provide data for the modelling done here. Thus, flow rates were staggered from the highest obtainable to an optimal 2 L/h where 0 CFU/100 ml for E. coli and fecal coliform in the effluent (>99.99% removal) were consistently achieved.

E. coli removal by the combined system was modelled as a series of three compartments. The models were coupled as depicted in Figure 2, whereby the effluent from the geotextile was modelled as the influent to the SCCGM and effluent from the SCCGM was modelled as influent to the GAC. Thus, the effluent of one compartment was modelled as influent to the next (Masters & Ela 2014). This modelling approach was derived from the works of (i) Metcalf & Eddy (2014), who modelled a number of wastewater reactors in series for pollutant removal, (ii) Rietveld (2019), who modelled a large-scale multi-barrier water treatment plant comprising ozone and sand filtration for E. coli reduction as units in series, as well as (iii) Masters & Ela (2014), who modelled a four-chamber tank for large scale drinking water disinfection as tanks in series.

A total of eight combined mathematical models (see Table 3) were tested using combinations of disinfection and filtration modelling approaches as given below. The respective disinfection and filtration modelling approaches alongside the various E. coli removal mechanisms and parameter equations are explained below. Thereafter, the eight combined models as were used in the numerical calculations of this study are summarized in Table 3 and the associated text just above Table 3.

E. coli removal by SCCGM was first modelled using disinfection kinetics in the first four combined models (see Table 3). The removal was thereby modelled as being only due to bacterial inactivation by the silver coating of the media. The inactivation by silver was modelled using Chick's, Collins-Selleck, complete-mix system (CMS) model and the Chick-Watson bacterial inactivation models (de Moel et al. 2007; Metcalf & Eddy 2014) explained below. Thereafter, E. coli removal contribution by SCCGM filtration (Table 3) was included in the last four combined models using the colloidal filtration theory (CFT) numerical modelling procedure explained under E. coli removal by GAC filtration, but using appropriate SCCGM characteristics.

Assuming that for any length, dx, and throughout the corresponding cross section, (i) mixing of the microbial particles is ideal (Figure 3), (ii) flow rate is constant, and (iii) no storage exists in the SCCGM filter bed, mass balance was done as follows (de Moel et al. 2007):

Since the form of Equation (7) is exactly like the Chick's model (Chick 1908; MWH 2012; Metcalf & Eddy 2014; Qasim & Zhu 2018) for disinfection, it was handled as such to simplify the theoretical approach. The Chick's model (Equation (7)) and Chick-Watson model (Equation (11)) were also used by Rossainz-Castro et al. (2016) to model E. coli and Candida albicans inactivation by silver and copper-coated granular zeolite, by Singh et al. (2019) to model E. coli inactivation by silver and other metals as well as by Somani & Ingole (2012) for kinetic modelling of water disinfection by natural herbs.

The SCCGM bed was modelled as having a volume V (m³) fed by a flow rate Q and an E. coli count N_o and with effluent E. coli count of N_e and flow rate Q same as the influent flow rate. Steady state mass balance for the filter bed was as follows (de Moel et al. 2007; Masters & Ela 2014):

Theoretically α ranges between 0 and 1 from poor to optimal sticking conditions respectively (Tufenkji & Elimelech 2004; Tobiason et al. 2011).

Overall, eight combined mathematical models as defined below and summarized in Table 3 were used in the numerical calculations of the present study. It is worth noting here that most of the model equations used are associated with the various removal mechanisms and parameter equations explained above.

Model 1 refers to the modelled combined removals, starting with geotextile filtration governed by Equation (23) coupled to SCCGM disinfection removals modelled by Chick's model (Equation (7)) followed by GAC filtration removals modelled by Equation (13). Model 2 refers to the modelled combined removals starting with geotextile filtration governed by Equation (23) coupled to SCCGM disinfection removals modelled by the complete mixing system model (Equation (9)) followed by GAC filtration removals modelled by Equation (13). Model 3 refers to the modelled combined removals starting with geotextile filtration governed by Equation (23) coupled to SCCGM disinfection removals modelled by the Collins-Selleck model (Equation 10) followed by GAC filtration removals modelled by Equation (13). Model 4 refers to the modelled combined removals starting with geotextile filtration governed by Equation (23) coupled to SCCGM disinfection removals modelled by Chick-Watson model (Equation 11) followed by GAC filtration removals modelled by Equation (13). Model 5 refers to the modelled combined removals of Model 1 plus SCCGM filtration contribution modelled by Equation (13). Model 6 refers to the modelled combined removals of Model 2 plus SCCGM filtration contribution modelled by Equation (13). Model 7 refers to the modelled combined removals of Model 3 plus SCCGM filtration contribution modelled by Equation (13). Model 8 refers to the modelled combined removals of Model 4 plus SCCGM filtration contribution modelled by Equation (13). Model calculations for each run began with E. coli counts in the influent then integrated removal efficiencies (Equation (3)) by each stage (Figure 2 and Table 3) were applied successively to the influent counts to get effluent E. coli counts and respective LRVs (Table 4).

Sensitivity analysis was carried to test (i) the effect of neglecting the filtration removal component by the SCCGM to the models, (ii) the effect of modelling E. coli removal by disinfection alone, and (iii) the sensitivity of contact time, filtration rate, collector diameter and microbial particle (E. coli) size to the models. Condition (i) was assessed by essentially comparing the predicted E. coli removals by models 1–4 with the corresponding removals by models 5–8 (Figure 4), while conditions (ii) and (iii) were tested using models 3 and 8 to test the sensitivity of simulated E. coli removal to each condition or parameter (Figures 5 and 6). The results of the sensitivity analysis are given and explained below under results and discussion.

Predicted E. coli removals were calculated using the coupled models presented above, which are based on the removal mechanisms elaborated on earlier. Figure 4 gives comparative plots of theoretically predicted and measured E. coli Log Removal Values (LRVs) for each run and model respectively. The results (Figure 4 and Tables 4 and 5) show that the coupled models – except for models 2 and 6 – reasonably described the combined E. coli removals by the multi-barrier system. Although models 1, 4, 5, and 8 gave slight underestimations for runs with lower contact time, their predictions were considered satisfactory as also shown by the model performance criteria in Table 5. Models 3 and 7 gave the closest predictions of E. coli removal values with respect to the measured values, but generally overestimated the LRVs. The appreciable performance by models 1, 3, 4, 5, 7 and 8 signifies they simulated the combined physical and chemical E. coli disinfection mechanisms by the multi-barrier system relatively well.

The silver disinfection component by these models was theoretically within the findings by Singh et al. 2019, who indicated that about 24 mg·min/L of Ct (silver concentration multiplied by time) value is required to eliminate 99% of E. coli from natural stream waters. The calculated Ct values in this study, which corresponded to approximately 99% (≈2 LRV) and higher, were >20 mg·min/L. These were from runs 8 to 12 (Table 4 and Figure 4). Further improvement of the models, particularly in terms of the input parameter values obtained or calculated from literature, could conceivably minimize the underestimations (by models 1, 4, 5 and 8) and overestimation (by models 3 and 7) of E. coli removals. Long term experimentation is therefore recommended to improve or calibrate the model input parameter values (especially those obtained from literature).

Models 3 and 7 may prompt the design engineer to under design the system, since the expected removals were higher than those measured. Conversely, models 1, 4, 5 and 8 may prompt the design engineer to over design the system. This shows that different model and removal mechanism combinations can produce different bacterial removal predictions. Therefore, using an array of models coupled with larger experimental data sets may help minimize under and over predictions and correspondingly minimize over and under designs of multi-barrier PoU water treatment systems such as the modelled system.

The higher removals by models 3 and 7 could be attributed to the silver disinfection by the Collins-Selleck model, which generally gave higher predictions for E. coli inactivation rates. On the other hand, silver disinfection by models 1 and 5 was simulated by the Chick's model while silver disinfection by models 4 and 8 was simulated by the Chick-Watson model. The Chick's and Chick-Watson models generally under predicted the E. coli inactivation rate. However, with further refinement, models 1, 4, 5 and 8 are tentatively expected to give a better match since their removals showed a more realistic gradual increment in E. coli removals, which closely corresponded to the contact time, as is theoretically supposed to be the case.

Figure 4 shows that models 2 and 6 were essentially unable to predict the E. coli removals. The weak predictions of the two models could be attributed to the CMS model (Equation (9)) being weak at simulating E. coli inactivation by silver in the SCCGM, which is the main disinfection step in the multi-barrier system. The CMS model predicts microbial inactivation by using natural die-off kinetics assuming microbial die-off with time (de Moel et al. 2007; Qasim & Zhu 2018). It may thus need higher contact time for substantial removals, and is principally applicable to bacterial die-off in non-disinfection treatment processes like natural treatment reservoirs (Qasim & Zhu 2018).

The effect of neglecting filtration contribution by the SCCGM was assessed by essentially comparing the predicted removals by models 1–4 with the corresponding removals by models 5–8 (Figure 4). Modelling the additional removal contribution by SCCGM had minimal effect on the predicted overall removal (Figure 4 and Tables 4 and 5). This was not surprising because, except for advanced technologies such as reverse osmosis and iodine resin filters (Backer 2000), bacterial removal by fabric and granular filtration alone is primarily most efficient for suspended solids removal, not for removal of bacteria, due to large pore sizes (Backer 2002, 2000; Kausley et al. 2018; Siwila & Brink 2018). Therefore, fabric and granular filtration are normally used as a first stage before other treatment steps (Kausley et al. 2018). The need for an in-built disinfection step by silver in the SCCGM is therefore indicated. This finding is supported by literature from various authors where bacterial removal by silver-impregnated filter media (Chong et al. 2011; Rossainz-Castro et al. 2016; Shimabuku et al. 2018) was satisfactorily modelled by considering silver disinfection only without consideration of physical removal by filtration. Thus, porous media impregnated with silver or other metal disinfectants have been shown to be efficient at bacterial inactivation (Chong et al. 2011; Rossainz-Castro et al. 2016). However, disinfection may also be improved if fabric pre-filtration is provided to increase bacterial contact with the metal disinfectant (Tobiason et al. 2011) while GAC post filtration was provided to make the water more acceptable (Backer 2002; Tobiason et al. 2011).

It was assessed whether E. coli removal could be modelled by disinfection kinetics only (Figure 5) using two models (models 3 and 8) representing possible overestimation and underestimation. Thus, the removal of the coupled models was contrasted with removal by SCCGM disinfection alone. Model 3 was selected for this purpose over model 7 because it showed better performance statistics than model 7 (Table 5). Similarly, model 8 was chosen over models 1, 2, 4, 5 and 6 since it also depicted better statistics (Table 5). It can be seen from Figure 5 that, although E. coli removal prediction by SCCGM alone seemed to be a good representation, modelling additional removal by other treatment steps (i.e. geotextile and GAC removals) was still important for the models to be fully representative of the multi-barrier system. Thus, from the results shown in Figure 5, it can be seen that disinfection removal alone could not fully describe the E. coli removals, giving predicted LRVs below measured values for both models.

The effect of contact time, filtration rate, collector diameter and microbial particle size on E. coli removal was assessed using models 3 and 8 (Figure 6). Both models indicated that larger contact time (Figure 6(i)) resulted in higher E. coli removal. Since contact time is dependent on filter media depth and filtration rate (see Equation (4)), optimizing either or both of the parameters optimizes contact time and subsequently enhances E. coli removals. Each of the models used in this study was affected by contact time and hence by media depth (h) and filtration rate (v). For instance, as filtration rate increases (Figure 6(ii)), E. coli removal by both models decreases. Therefore, careful optimization of these parameters is expected to enhance E. coli removal performance. It is worth noting that filtration rate is affected by various factors, of which particle size distribution is the key factor (Siwila & Brink 2020). Therefore, to optimise contact time it is necessary to not only look at filtration rate but also at factors affecting it, such as particle size distribution. If fine granules are insufficient in a filter media the filtration rate is very high, leading to lower contact time, while if coarse granules are insufficient the filtration rate is too low, leading to higher contact time. The collector diameter (d_c) depends on particle size distribution and subsequently affects E. coli removal prediction.

The sensitivity of models 3 and 8 to collector and fiber diameter is shown in Figure 6(iii) and 6(v). The smaller the collector or fiber diameter, the higher the removals (Figure 6(iii) and 6(v)). The sensitivity of microbial particle (E. coli) size was also assessed (Figure 6(iv) and 6(vi)). The effect of the collector/fiber diameter and microbial particle diameter were assessed by applying the varied particle sizes on the geotextile and GAC CFT models, which are directly affected by particle size. This analysis was done using the input parameters listed in Table 1 but keeping the optimal flow rate (2 L/h) constant. It can be seen from Figure 6(iv) and 6(vi) that the microbial particle diameter having the least removal efficiency by GAC and geotextile in both models is somewhere between 1 and 2 μm. Removal of microbial particles below this range increases with decreasing particle diameter because removal is primarily by diffusion (Yao et al. 1971; Tufenkji & Elimelech 2004), while removal of bacteria with diameters larger than 2 μm increases with particle diameter and removal is mainly by sedimentation and interception (Yao et al. 1971; Tufenkji & Elimelech 2004). This explanation consequently entails that removal of microbial particles by porous media filtration alone is a huge challenge. This finding is important because it further supports the need for a carefully optimized inbuilt disinfection step to ensure continued safety of the produced water. Overall, the sensitivity analysis results of predicted E. coli removals by models 3 and 8 were significantly similar for each parameter assessed (Figure 6).

The modelling exercise has demonstrated that suitable removal mechanisms can be integrally used to model a combined PoU system to predict the overall effluent bacterial quality. This kind of modelling can be used to optimize system design by allowing the engineer to systematically vary design parameters until the desired system effectiveness is attained. This research has also indicated that each barrier or treatment stage contributes to the overall E. coli removal. Therefore, the bacterial load on the SCCGM (which is the main disinfection stage) can be significantly reduced by optimizing all components of the multi-barrier (combined) system, especially the pre-filtration stage. Some reasons for differences between predicted E. coli inactivation and actual inactivation by models such as the Chick's and Chick-Watson models include (Qasim & Zhu 2018): (i) disinfectant residue may not be constant or uniform throughout the system and filter runs, (ii) pH changes may affect the inactivation rate, (iii) variations in the incoming suspended particle loads of the water being treated, (iv) varying temperature, and (v) the disinfectant may be consumed by other competitive reactions.

It is recommended that future research should keep the obtained optimal flow constant then model the breakthrough of E. coli for several runs, ensuring water is passed in triplicate for each run. Furthermore, modelling of data obtained from field testing to assess possible applicability of the mathematical models on field data is proposed. Also, concurrent modelling of E. coli and turbidity is proposed, since performance of filter systems is usually monitored by measuring effluent turbidity (MWH 2012). Additionally, since the proposed multi-barrier water treatment design is scalable such that the capacity is flexible and can be increased to serve more consumers, modelling the effect of scalability is proposed. Long term experimentation is also recommended, and may help in calibrating model parameters to achieve the best fit between the modelled and measured values. Quantification of measured influent and effluent E. coli counts using particle counting techniques is also recommended. This may help characterize the modelled microbial particle diameter (d_p) better.