Stormwater Treatment in Future Tropical and Sub-Tropical Climates

Abstract

Stormwater treatment systems play an integral part in achieving sustainable urban development. The performance of these systems is likely to be impacted by potential changes in climatic patterns, including precipitation. This project investigates the simulated impacts of climate change on the performance of stormwater treatment systems used as a part of Water-Sensitive Urban Design (WSUD). Townsville and the Gold Coast of Queensland, Australia, were selected for the study to investigate tropical and sub-tropical climates experienced by cities across the globe adjoining sensitive coastal environments such as wetlands and coral reefs. The daily precipitation output projected by Coupled Model Intercomparison Project Phase 5 (CMIP5) climate models was downscaled to pluviograph input into the Model for Urban Improvement Conceptualisation (MUSIC). The treatment performance of bioretention systems and constructed wetlands was variable across both locations, with some models showing little to no change or improvement. Worsening of treatment performance was more prominent in the tropical climate, with numerous models reaching a decline of up to 16%. However, the highest observed reduction from a single model output occurred in the sub-tropical climate location. To make the WSUD treatment system effective under the future climate scenarios, physical modification is necessary to increase the treatment area or depth. Increasing the area in the worst-case scenario could incur a cost increase of 20% to 30% and present challenges due to development constraints. Increasing the depth could be a viable alternative for bioretention systems but is likely impractical for constructed wetlands.

1. Introduction

Stormwater treatment systems play an integral part in achieving sustainable urban development. In Australia, Water-Sensitive Urban Design (WSUD) describes design practices that attenuate stormwater by maintaining the natural hydrological balance with enhanced storage, infiltration and evaporation. Examples of such WSUD assets include bioretention basins, vegetative swales, infiltration basins and constructed wetlands [1]. The effectiveness of these WSUD assets depends on several climatic conditions, including rainfall characteristics. The impacts of higher intensity rainfall on pollutant generation and stormwater treatment system performance are significant; so, this field is becoming an emerging area of research. Alamdari et al. [2] applied a calibrated model to assess the potential impacts of increased precipitation and temperature on pollutant loading in northern Virginia. An annual runoff volume increase by 6.5% resulted in increases of Total Suspended Solids (TSSs), Total Nitrogen (TN) and Total Phosphorus (TP) by 7.6%, 7.1% and 8.1%, respectively. A subsequent investigation for this same catchment used an ensemble of five Global Climate Models (GCMs) and two emission scenarios [3]. Similarly, under a medium emission scenario, the median TSS, TN and TP increased by 3.1%, 2.5% and 9.9%, respectively, and by 3.8%, 3.1% and 10.4% under a high-emission scenario. Also, the treatment efficiency of retention ponds was reduced by more than 10% for TP, indicating a requirement to increase the size of the treatment systems.

The extent of climate change impacts on pollutant loads and treatment efficiencies may vary by geographic location. Sharma et al. [4] reported that in Denmark higher intensity rainfall events could impact the TSS loading for rainfall with average recurrence intervals (ARIs) above 0.5 years, as the increased runoff volume interrupted the particle settling process. However, the decrease in treatment efficiency was considered relatively minor. Changes in the intensity of rarer ARIs, will not necessarily result in higher annual pollutant loading. Liu et al. [5] compared the influence of high-intensity short-duration events, high-intensity long-duration events and low-intensity long-duration events on TSS runoff loading in a WSUD treatment system for various ARIs. Whilst the high-intensity short-duration events were the most important contributor to the annual TSS loading, this generally resulted from less frequent ARIs of this storm event category. Borris et al. [6] found that the pollutant loading from relatively frequent storms with low-to-intermediate depths and intensities was most sensitive to change. This is because whilst higher intensity storms will effectively remove pollutants from the surface, a pollutant supply limitation will ultimately restrict the quantity of pollutants transported in a single event.

The effects of climate change on the selection and design of stormwater treatment systems have been investigated by numerous authors. Hatheway et al. [7] investigated changes in bioretention basin hydrologic function under two emission scenarios in North Carolina, USA. With the modelled increase in overflow, between 90 mm and 310 mm of additional storage was required to restrict the volumes to the baseflow. Tirpak et al. [8] utilized ten downscaled projections to investigate the design modifications required for bioretention basins in Tennessee, USA. The greatest improvement with respect to annual volume of infiltration and surface overflow was achieved by increasing the ponding depths, thickness of media layer, media conductivity rates, and surface area of the current design standards by 307%, 200%, 200% and 300%, respectively. Conversely, increasing ponding zone depth, media layer thickness and media conductivity alone resulted in lower performance, with 13% to 82% and 77% to 100% of the models falling below the historical annual volumes of infiltration and surface overflow. Increasing the surface area relative to the contributing catchment area was considered a more favourable option, especially for locations with low in situ soil drainage rates. The local site conditions is likely a critical factor in designing for climate change resiliency.

To date, only a select few studies have investigated the impacts of climate change on WSUD performance in Australia, and none have concerned the tropical and sub-tropical climates in Queensland [9]. Burge et al., Lam and Gribler and Zhang et al. [10,11,12] all used Victorian case studies. They generally found that decreases in WSUD treatment efficiency will be minor under a worst-case scenario. Burge et al. [10] applied season-specific stochastically generated changes to each rainfall day in a baseline series whilst also reducing the number of rainy days. The rainfall volume extracted from the deducted rainfall days was also distributed to the future daily totals to simulate a drier future with more intense rainfall. Overall, there was little change in WSUD performance besides a 6% improvement in constructed wetlands treatment efficiency. This was attributed to the existing interannual climate variability in a temperate climate. Similarly, Zhang et al. [12] found that the only significant impact on constructed wetlands performance was a 3% decrease in harvesting reliability, attributed to a decrease in runoff caused by an overall drier future. The overall treatment area of constructed wetlands required to achieve pollutant reduction targets remained similar to that in the base-case scenario, and in some cases, an improvement was even observed.

Climate change impacts across the globe are likely to be highly variable in nature, influenced largely by different climatic classifications and phenomena. Likewise, the impacts on stormwater treatment performance are expected to vary throughout Australia and beyond. This study aimed to investigate how predicted precipitation changes in the tropical and sub-tropical climates of Queensland, Australia, may impact the effectiveness of WSUD treatment devices. Townsville and the Gold Coast were selected as case study locations to represent a wide spatial extent of climate change projections in tropical and sub-tropical climates experienced by cities adjoining sensitive coastal environments such as wetlands and coral reefs.

2. Materials and Methods

The Model for Urban Stormwater Improvement Conceptualization (MUSIC) (EWATER Melbourne) is the industry standard tool for modelling WSUD in Australia. MUSIC uses annual time series precipitation data over a historical period to model annual the treatment efficiencies for TSS, TP and TN. It should be noted that MUSIC takes a stochastic approach to generating pollutant loads, as evidently the most significant portion of pollutants is attributed to frequent low-magnitude rain events [13]. Hence, this study may slightly underestimate the full effects of altered precipitation where increased intensities are expected.

GCMs obtained from the Coupled Model for Intercomparison Project Phase 5 (CMIP5) were applied to project future rainfall predictions at each location. As GCMs operate at a relatively coarse spatial resolution, spatial downscaling was required to generate local rainfall data. Two downscaling techniques were selected to account for inter-method variability, both of which were selected for their availability in the public domain.

The first method used was the Statistical Downscaling Model (SDSM), developed by Wilby et al. [14], which uses regression-based analysis to downscale GCM output directly to the station-level scale. The SDSM uses an inbuilt statistical significance test function to identify strong predictor–predictand relationships [14,15]. A statistical significance value of p < 0.5 was used to identify atmospheric predictors for each site. Surface humidity was selected for Townsville, as this demonstrated statistical significance for all months of the year except for August and September, for which no significant predictors were identified. For the Gold Coast site, a consistent atmospheric variable that demonstrated statistical significance throughout all months was unavailable. Therefore, temperature, surface humidity, geostrophic airflow velocity and zonal velocity were all selected as predictors for the analysis.

The second method involved perturbing the stochastic weather generator output, in line with delta factors produced from a Regional Climate Model (RCM). Previous studies using this method generally used the Stochastic Weather Generator (WGEN) which is limited to 1st-order Markov chains [16,17,18]. Due to the limited number of transition states, 1st-order Markov chains tend to underestimate extensive dry periods. This study utilised the MATLAB (Version 9.13) based Weather Generator of the École de Technologie Supérieure (WeaGETS) which has the addition of 2nd- and 3rd-order Markov chains. Percentage change in consecutive dry and wet days and seasonal rainfall amounts were obtained from the Conformal Cubic Atmospheric Model (CCAM), available from the Queensland Future Climate Dashboard website [19]. Regression- and weather generator-based downscaling methods are subject to numerous advantages and disadvantages. Hashmi et al. [20] compared the SDSM to the Long Arm Research Station Weather Generator (LARS-WG), which incorporates climate change factors directly from a GCM. LARS-WG was found to be superior at replicating the daily means from the base scenario, although only predicted an increase for low ARIs. The SDSM conversely predicted that the 100-year ARI magnitude event would become a 20-year ARI event. The LARS-WG was considered less reliable as it incorporates relative precipitation changes directly from a GCM which operates at a relatively coarse spatial resolution. Hassan et al. [19] found that the LARS-WG was less accurate than the SDSM at replicating the mean daily precipitation on a monthly basis but was superior at replicating the mean dry and wet periods throughout all months of the year. SDSM generally predicted higher annual rainfall. The use of both types of methods in this study was considered important for accounting for a wide range of uncertainties associated with climate change projections and downscaling.

The CCAM standard baseline period from 1986 to 2005 was applied in both spatial downscaling methods to represent a present-day rainfall scenario. CCAM percent change factors from the Queensland Future Climate Dashboard [9] website are given relative to the historical period from 1986 to 2005. This period was used as the baseline period for both downscaling methods to maintain consistency.

The selected GCMs included the Norwegian Earth System Model (NOR-ESM) and the Max Plank Institute Earth System Model (MPI-ESM). These GCMs were not selected using specific criteria because the focus of the study was assessing the effectiveness of stormwater treatment across different climate scenarios, rather than a detailed examination of the climate models themselves. However, for future work, a more systematic approach to GCM selection could be adopted. The WeaGETS-RCM method also used percent changes from an ensemble of 11 CMIP5 models available from the CCAM. Ensemble predictor–predictand data sets were not readily available for use within the SDSM, and this data set could therefore not be replicated for the SDSM method. The Intergovernmental Panel on Climate Change (IPCC) [21] has adopted measures known as Representative Concentration Pathways (RCPs) which describe the level or radiative forcing arising from different emission scenario. The scale of the RCPs includes the mitigation scenario corresponding to the very low forcing scenario RCP 2.6 and the medium-emission scenarios RCP 4.5 and RCP 6.0. The highest emission scenario is the RCP 8.5, which describes a future in which very little to no action is taken to manage the impacts of climate change [15]. For this investigation, two (2) RCPs were assessed for each GCM data set, being the medium-emission scenario RCP4.5 and the high-emission scenario RCP8.5.

Temporal downscaling of the daily precipitation to a 6 min pluviograph time series was achieved via the K-nearest neighbour algorithm [22]. The daily sub-hourly totals over the historical precipitation record were factored with climate change projections, and each day in the generated time series was paired with the closest match, i.e., the “nearest neighbour” for the relevant month. Conventionally, the corresponding hourly distribution would be adopted, and the process would be repeated to obtain the sub-hourly distribution for each hour. However, for this investigation, the daily precipitation was matched directly with the relevant 6 min distribution for the selected nearest neighbour.

For the modelling phase, a baseline scenario was first established by applying historical pluviograph data to two WSUD treatment scenarios designed and modelled in MUSIC to achieve the minimum pollutant reduction targets for TSS, TP and TN relevant to the local government stormwater design guidelines outlined in

Table 1. Pollutant reduction targets.

Location	Design Guideline	TSS (%)	TP (%)	TN (%)
Townsville	Design Objectives for Stormwater Management in the Coastal Dry Tropics (2011) [23]	80	60	45
Gold Coast	Section 13 of Policy 11: Land Development Guidelines (2005) [24]	80	65	40

.

The catchment parameters used in MUSIC modelling are summarised in

Table 2. Catchment node properties.

	Surface Type	Area (Ha)	Fraction Imperviousness (%)
Sub-Catchment 1
Catchment Node 1	Sealed Road	0.045	70
Catchment Node 2	Roof	0.067	100
Catchment Node 3	Residential	0.132	30
Sub-Catchment 2
Catchment Node 1	Sealed Road	0.157	70
Catchment Node 2	Roof	0.116	100
Catchment Node 3	Residential	0.245	30
Total			57

. The rainfall-runoff parameters and pollutant export parameters are shown in

Table 3. Rainfall-runoff properties.

Impervious Area Properties
Rainfall Threshold (mm/day)	1	1
Pervious Area Properties
Soil Storage Capacity (mm)	120	500
Initial Storage (% of Capacity)	25	10
Field Capacity (mm)	80	200
Infiltration Capacity Coefficient (a)	200	211
Infiltration Capacity Coefficient (b)	1	5
Groundwater Properties
Initial Depth (mm)	10	50
Daily Recharge (%)	25	28
Daily Baseflow (%)	5	27
Daily Seepage Rate (%)	0	0

and

Table 4. Pollutant export parameters.

	Sealed Road		Roof		Residential
	Mean	Standard Deviation	Mean	Standard	Mean	Standard
TSS
Baseflow (mg/L)	1.00	0.34	N/A	N/A	1.00	0.34
Stormflow Concentration (mg/L)	2.43	0.39	1.30	0.39	2.18	0.39
TP
Baseflow (mg/L)	−0.97	0.31	N/A	N/A	−0.97	0.31
Stormflow Concentration (mg/L)	−0.30	0.31	−0.89	0.31	−0.47	0.31
TN
Baseflow (mg/L)	0.20	0.20	N/A	N/A	0.20	0.20
Stormflow Concentration (mg/L)	0.26	0.23	0.26	0.23	0.26	0.23

, respectively. One treatment train featured a bioretention basin as the tertiary treatment device, while the second scenario incorporated a constructed wetland. The treatment trains were then modelled by applying each climate change projected rainfall to investigate the change in treatment performance. Rainfall data sets that exhibited a substantial decline in treatment performance provided insights for a sensitivity analysis on WSUD design compliance. For both constructed wetlands and bioretention basins, the change in area and depth required to achieve compliance were identified, with each parameter examined separately. The inbuilt MUSIC life cycle costing module provided the percentage cost increase from the baseline scenario to meet the reduction targets under the prospective future scenario. This function calculates the overall lifecycle cost based on the treatment surface area; however, it does not incorporate depth. The rainfall-runoff and pollutant export parameters were adopted from the MUSIC modelling guidelines [1].

The base parameters for bioretention basins and constructed wetlands are shown in

Table 5. Bioretention basin modelling parameters.

	Townsville		Gold Coast
Inlet Properties
Low-Flow Bypass (m3/s)	0	0	0	0
High-Flow Bypass (m3/s)	100	100	100	100
Storage Properties
Extended Detention Depth (m)	0.30	0.30	0.30	0.30
Surface Area (m2)	75	58	68	60
Filter and Media Properties
Filter Area (m2)	65	48	58	50
Unlined Filter Media Perimeter (m)	20	20	20	20
Saturated Hydraulic Conductivity (mm/h)	200	200	200	200
Filter Depth (m)	0.50	0.50	0.50	0.50
TN Content of Filter Media (mg/kg)	400	400	400	400
Orthophosphate Content of Filter Media (mg/h)	30	30	30	30
Infiltration Properties
Exfiltration Rate (mm/h)	0	0	0	0

and

Table 6. Constructed wetlands modelling parameters.

	Townsville		Gold Coast
Inlet Properties
Low-Flow Bypass (m3/s)	0	0	0	0
High-Flow Bypass (m3/s)	100	100	100	100
Inlet Pond Volume (m3)	0	0	0	0
Storage Properties
Extended Detention Depth (m)	0.50	0.50	0.50	0.50
Surface Area (m2)	350	250	325	325
Permanent Pool Volume (m3)	105	75	97.5	97.5
Initial Volume (m3)	105	75	97.5	97.5
Exfiltration Rate (mm/h)	0	0	0	0
Evaporative Loss as % of PET	125	125	125	125
Outlet Properties
Equivalent Pipe Diameter (mm)	25	21	24	24
Overflow Weir Width (m)	35	25	32.5	32.5
Notational Detention (hrs)	47.2	47.8	47.6	47.6

, respectively. The surface area was modified to achieve the reduction targets shown in Table 1 for the baseline time series data described in Section 3.1. For bioretention basins, the filter area was assumed to be 10 m² less than the surface area. For constructed wetlands, the permanent pool volume and initial volume were assumed to be 30% of the treatment surface area, whilst the overflow weir width was assumed to be 10% of the treatment area. The equivalent pipe diameter was modified to ensure that the notational detention time was as close to 48 h as possible. All other parameters were adopted from the MUSIC modelling guidelines [1].

3. Results

3.1. Generation of Baseline Rainfall Data

The performance of each baseline time series generated by the SDSM and WeaGETS was compared with the observed precipitation record from 1986 to 2005. The monthly mean precipitation, period of consecutive dry days, and period of consecutive wet days are illustrated in Figure 1a,b, for both Townsville and Gold Coast. The WeaGETS baseline data were generally more accurate at replicating the observed statistics. Notably, the mean monthly dry period modelled using the WeaGETS was generally within one day of the observed record.

For Townsville, the WeaGETS method demonstrated a higher level of consistency with the baseline period, notably for precipitation amounts between 50 and 100 mm. the SDSM was more variable in its ability to replicate the monthly mean throughout the year. Similarly to what observed for Townsville, the WeaGETS was also more accurate at replicating the monthly mean at the Gold Coast. However, it clearly overestimated the rainfall amounts slightly for the drier months of the year, whilst for some of the wetter months, it underestimated them. The SDSM was generally more variable throughout the entire year.

For Townsville, the WeaGETS was also more accurate at replicating the historical mean monthly wet period, with nine months being modelled within 0.5 days of the observed data. The WeaGETS method was generally less accurate at replicating the mean monthly wet period for the Gold Coast than for Townsville. Whilst the results for three months were very close to the observed values, the WeaGETS generally overestimated the wet period by 0.3 to 0.5 days. Conversely, the SDSM tended to underestimate the mean monthly wet period for 11 out of 12 months. The WeaGETS was reasonably accurate at modelling the mean monthly dry period for most months of the year at the Townsville location. The SDSM tended to underestimate this parameter for all months and was less consistent with the historical record. For the Gold Coast location, the WeaGETS output was also more consistent with the observed mean monthly dry period. However, considerably more variance was evident for this location. For the drier periods in the range of 7 to 10 days, the WeaGETS and SDSM performed on a similar level in replicating the mean monthly dry period. In some cases, the SDSM data were closer to the historical record.

3.2. Downscale of GCM Forecasted Data

Statistics for the GCM forecasted rainfall including annual rainfall (mm), annual maximum dry period (days) and annual maximum wet period (days) are presented in Figure 2a for Townsville and Figure 2b for Gold Coast.

All SDSM models tended to predict an increase in annual rainfall from the respective base time series. The Ensemble RCP 4.5 was the only WeaGETS-based model to predict an increase in annual rainfall. This model also predicted an increase in annual dry periods along with the SDSM MPI RCP 4.5. The WeaGETS-RCM-based MPI-RCP 4.5, NOR RCP 4.5 and NOR RCP 8.5 all predicted an increase in wet periods, which did not correlate with higher annual rainfall. Notably, NOR RCP 8.5 predicted a substantial increase in the maximum annual wet period despite projecting an increase in annual dry periods and lower annual rainfall.

Less variability was noted between the two downscaling techniques across the climate models, when applied to the sub-tropical climate of the Gold Coast compared to tropical Townsville climate. Figure 2b shows that the WeaGETS-RCM-based MPI RCP 4.5 model predicted the most substantial increase in annual rainfall from the base data, along with increased annual dry periods. Model data at this location were generally orientated towards an overall drier future, although some models, such as the WeaGETS-RCM-based NOR RCP4.5 and NOR RCP 8.5, projected increased consecutive wet days.

3.3. Generation of Future Rainfall Predictions Using GCMs

Figure 3a illustrates the TN reduction simulated across all models and emission scenarios generated using the WeaGETS-RCM and SDSM for Townsville. There was considerable variability among the climate scenarios for each treatment system, as some treatment systems improved under one climate scenario, whilst others experienced a performance reduction. For the SDSM models, the TN pollutant reduction generally declined by as much as 16%, and only the bioretention basin modelled under the MPI RCP 4.5 scenario was close to the target reduction. The WeaGETS-RCM model predictions showed more variability. The NOR RCP 4.5 and NOR RCP 8.5 improved the TN reduction, whilst the ENSEMBLE RCP 4.5 showed a decrease close to the maximum decline of 16%. Constructed wetlands were generally more sensitive to changes in precipitation.

For the Gold Coast (Figure 3b), the decline in treatment efficiency was generally quite small compared to Townsville. The MPI RCP4.5 scenario modelled using the WeaGETS predicted the most significant performance decrease of up to 20% from the target. Unlike for Townsville, the ENSEMBLE RCP 4.5 scenario was much closer to the target reduction, even showing slight improvement for constructed wetlands. The model sets that exhibited the most substantial decline in treatment performance were selected for the sensitivity and cost analysis. These model sets were compared with their respective changes in annual rainfall statistics, as shown in

Table 7. Townsville annual rainfall and treatment summary.

Scenario	Change in Median Annual Rainfall (%)	Change in 75% Quartile Rainfall (%)	Change in Maximum Annual Rainfall (%)	Change in TN Reduction (%) Bioretention	Change in TN Reduction (%) Constructed Wetlands
Ensemble RCP 4.5—WeaGETS (Location—Townsville)	+17%	+17%	+17%	−10%	−10%
NOR RCP—4.5 SDSM (Location—Townsville)	+34%	+28%	+43%	−10%	−15%
MPI RCP 4.5—WeaGETS (Location—Gold Coast)	+6%	+10%	+12%	−22%	−13%

.

The results showed that reductions in treatment efficiency were correlated with larger maximum annual rainfalls. This was most evident for the most critical scenarios for each location, such as NOR RCP 4.5—SDSM for Townsville and MPI RCP 4.5—WeaGETS for the Gold Coast, which both corresponded to the largest increase in annual maximum rainfall for the respective locations. Both critical scenarios exhibited a substantial increase in the annual maximum dry spells from their respective baseline scenarios. This suggests that these climate scenarios predicted a generally drier future interspersed with short periods of heavy rainfall.

3.4. Sensitivity and Cost Analysis

The results of the sensitivity analysis are presented in

Table 8. Compliance sensitivity of treatment systems.

	Treatment System	Baseline Area (m2)	Amended Area (m2) and % Change	Baseline Depth (m)	Amended Depth (m) and % Change
Townsville—ENSEMBLE RCP 4.5 WeaGETS	Bioretention	75	110 (+147%)	0.30	0.55 (183%)
Townsville—NOR RCP 4.5 SDSM	Bioretention	58	78 (+135%)	0.30	0.55 (183%)
Gold Coast—MPI RCP 4.5 WeaGETS	Bioretention	68	110 (+162%)	0.30	0.75 (250%)
Townsville—ENSEMBLE RCP 4.5 WeaGETS	Constructed Wetlands	350	450 (+129%)	0.50	1.4 (280%)
Townsville—NOR RCP 4.5 SDSM	Constructed Wetlands	250	360 (144%)	0.50	0.55 (320%)
Gold Coast—MPI RCP 4.5 WeaGETS	Constructed Wetlands	575	760 (132%)	0.50	0.75 (300%)

. The selected data sets for the cost analysis were the ENSEMBLE RCP 4.5 (WeaGETS) and NOR RCP 4.5 (SDSM) for Townsville and the MPI RCP 4.5 (WeaGETS) for the Gold Coast. This analysis included the determination of the increase in surface area and the increase in depth required to achieve compliance with the pollutant reduction targets under the worst-case future climate scenario.

The increase in surface area for bioretention basins was generally higher, with two cases requiring increases by a factor of 1.5 to 1.6. Two cases for constructed wetlands required an increase of approximately 1.3, and the maximum required was 1.4. The opposite was observed for depth, as constructed wetlands required the largest change by a factor of 3, while bioretention basins required a change by a factor of 2.5.

The life cycle cost analysis results for both treatment systems are shown in

Table 9. Costs incurred to account for climate change impacts.

	Treatment System	Baseline Area (m2)	Amended Area (m2)	Baseline Life Cycle Cost ($)	Amended Life Cycle Cost ($)
Townsville—ENSEMBLE RCP 4.5 WeaGETS	Bioretention	75	110	$98,821	$119,150
Townsville—NOR RCP 4.5 SDSM	Bioretention	58	78	$86,925	$100,749
Gold Coast—MPI RCP 4.5 WeaGETS	Bioretention	68	110	$94,141	$119,150
Townsville—ENSEMBLE RCP 4.5 WeaGETS	Constructed Wetlands	350	450	$163,627	$194,057
Townsville—NOR RCP 4.5 SDSM	Constructed Wetlands	250	360	$130,305	$166,782
Gold Coast—MPI RCP 4.5 WeaGETS	Constructed Wetlands	575	760	$229,273	$277,328

. Overall, the increase in costs required ranged between 20% and 30% for both bioretention basins and constructed wetlands. No clear trend was observed for either location, although constructed wetlands were more costly.

4. Discussion

The WeaGETS model was able to more accurately replicate historical precipitation amounts and dry periods. This is likely because the method uses observed precipitation characteristics directly in the time series output. Nonetheless, Zhang et al. [12] and Tousi et al. [25] indicated that the impacts should not be solely considered based on the best fit model and that even methods that poorly replicate the current climate can help quantify the full range of uncertainties.

The findings indicate that those climate models that simulate long dry periods followed by short periods of highly intense rainfall are most likely to experience a decline in treatment efficiency. During periods of high rainfall, it is likely that MUSIC generates a higher flow of pollutants to the treatment device, which cannot be adequately handled by the standard first-order decay function within the time it occurs. It is evident that the MUSIC model is most sensitive to changes in maximum annual rainfall occurring from less frequent but more extreme wet days rather than to changes in rainfall frequency. For instance, Figure 2a shows that NOR RCP 8.5 WeaGETS predicted higher median but lower maximum annual rainfall. Subsequently, the treatment performance shown in Figure 3b improved rather than declined.

There are, of course, some inherent uncertainties with each projection, which may not be accurately represented by the MUSIC model itself. The results infer that larger masses of pollutants are being continually transported during sequences of high runoff volumes, assuming a constant supply of pollutants. However, the pollutant wash-off generally declines as the supply is diminished over time. Borris et al. [6] inferred that increases in storm magnitude of rare ARIs because of climate change may not necessarily result in higher pollutant discharges. This is because high-intensity ARIs have already reached a pollutant threshold, whereby any increases in storm intensity will generally not cause significantly higher pollutant discharges.

Due to the uncertainty of predicting climate change impacts, the scenarios projecting the largest decline in treatment performance should only be considered as a possible worst-case scenario. For some cases in Townsville, the higher emission scenarios, such as the ENSEMBLE RCP 8.5, predicted either no change in treatment performance for bioretention basins or an improvement for constructed wetlands compared to the lower emission counterparts. The CCIA technical report [26] inferred that variations in the direction of projected change may be influenced by competition between increasing atmospheric moisture content owing to higher temperatures and a slowing down of tropical circulation. Models also differ in their ability to capture the Madden–Jullian oscillation (MJO), diurnal evolution of rainfall and tropical cyclone contribution. Consequently, there is a low confidence to predict the direction of annual rainfall change.

The models used in the ENSEMBLE RCP 8.5 may be biased towards predicting a drier future and may not accurately account for the levels of extreme precipitation caused by less frequent but more severe tropical cyclones. Interannual variations with prolonged droughts could cause a higher pollutant build-up to be washed off during shorter periods of highly intense cyclones and storms. Sediments could also be transported through other means, such as cyclonic winds. More research could be undertaken to investigate the influence of extended pollutant build-up on the supply of pollutants available during wash-off from these events.

The findings for the sub-tropical Gold Coast region which suggest an overall drier future are largely consistent with the findings from Dowdy et al. [27], especially in the use of regression-based models. The only model for the Gold Coast that predicted a significantly wetter future was the RCM-based MPI RCP 4.5 (WeaGETS). This is consistent with the enhanced abilities of the CCAM to replicate the occurrence of East Coast lows [28].

The cost analysis results suggest that the surface area required to achieve the reduction targets under the worst-case scenario must increase by 50% to 60% for bioretention basins and by 30% to 40% for constructed wetlands. The subsequent increases in life cycle costs range from 20% to 30%. Increasing the available depth of the WSUD system is one alternative that could be considered. This could be a more sustainable option as it would decrease the land area requirements and minimise evaporative losses. The sensitivity analysis indicated that the depth of constructed wetlands must increase by a factor of 3 to achieve the reduction targets for the given catchment size. Adding 1.0 m to the design depth is likely impractical in most situations. This solution may be more viable for bioretention basins, as they must increase by smaller factors of 1.8 to 2.5, which equates to 250 mm to 450 mm, with respect to the current design value. Unfortunately, the life cycle costing module is limited to data relative to the treatment system surface area. This is because the system’s depth is often standardised, and the surface area is the modifiable parameter for each design case.

More work should be undertaken to determine the costs associated with deeper treatment systems. Also, the feasibility of increasing the depth may vary with the geotechnical and topographical conditions.

If site constraints limit the available area, then increasing the base depth by a factor of 2 could be considered for bioretention basins. Bioretention basins should be prioritised where possible, as increasing the surface area of constructed wetlands could lead to higher evaporative losses, especially under a drier future. The results were based on the coarse climate projections generated by the chosen climate models. For more accurate outcomes to be used to decide on infrastructure designs, higher resolution climate models should be employed.

5. Conclusions

The ability of WSUD treatment systems to meet water quality performance targets in future climate scenarios was examined for tropical and sub-tropical climates in Queensland, Australia. The performance of bioretention systems and constructed wetlands varied considerably according to each method, model and emission scenario. The findings indicated that WSUD treatment systems are unlikely to be significantly impacted by climate change. The worst-case scenarios suggested a performance decline of up to 16% for tropical climates, increasing to 20% in the sub-tropical regions, although most cases here performed closer to the targets. Modifying the treatment area for the worst-case scenario could raise the life cycle costs by up to 30%, for both treatment options. Alternative solutions could explore increasing the depth of the treatment system, although this is far less practical for constructed wetlands and will vary depending on local geotechnical and topographical constraints. Future research should explore the impacts of increased antecedent dry period pollutant build-up and wash-off under extreme weather events such as tropical cyclones, which are projected to increase in intensity under any climate change outcome. These factors would need to be analysed on an individual-event basis, as they are unlikely to be modelled adequately by the current version of MUSIC software. The overall findings indicate that applying a factor of 50% to the standard bioretention area and of 35% to constructed wetlands could be a viable option to meet the reduction targets under a potential worst-case scenario.