While covering only 3% of the earth's land, urban areas accommodate 55% of world population and are responsible for more than two-thirds of global energy consumption and CO₂ emissions (United Nations, 2015; International Energy Agency, 2016). These ever-growing numbers make cities the forefront of the fight against climate change. On the other hand, around 39% of US primary energy is consumed by residential and commercial sectors (Energy Information Administration, 2019): designing more efficient buildings as well as water and energy systems can play an essential role in reducing the energy consumption and environmental footprint of urban districts. Coordinated design of infrastructure has proven to be an effective approach for creating more sustainable urban systems. Such coordination leads to integrated municipal supply systems (e.g., energy and water supply) with improved overall performance when compared to the simple superposition of individual supply systems (Mancarella, 2014). Further, integrated design of infrastructure systems facilitates closed-loop management of resources and can help create “circular economy” in urban areas (World Economic Forum, 2014; Kalmykova et al., 2018).

This study focuses on integrating urban energy and wastewater treatment systems. Typically, urban energy and water infrastructure are designed after the demands of the community are determined. At that late stage, improving the environmental and economic performance of the system is difficult (Best, 2016). However, if the supply and demand of municipal services are optimized simultaneously, the resulting urban system has proven to outperform the isolated design based on social, economic, and environmental sustainability metrics (Keirstead et al., 2012; Allegrini et al., 2015; Best et al., 2015). Since building mix largely determines the consumption profile of urban neighborhoods (Best, 2016; Wu et al., 2018), simultaneously optimizing the building mix and integrating infrastructure represents an opportunity for designing more sustainable urban systems (Keirstead et al., 2012; Best, 2016). Developing a framework to investigate and understand this opportunity is the focus of this research paper.

Factors including urbanization, environmental stresses due to climate change, and demand for more livable built environments are creating challenges for cities to satisfy the needs of an increasing urban population (Bauer et al., 2015; Best, 2016; Khan et al., 2017; O'Neill et al., 2017). This imperative is urging urban and infrastructure planners to design more sustainable infrastructure systems that can maintain and improve the livability of cities (Arup, 2016; SF Environment, 2018). However, concerning the focus of this study, literature shows several obstacles to achieving more sustainable water and energy systems in urban areas;

To address these deficiencies in the context of urban water and energy systems, a framework is needed to optimize the supply and demand of infrastructure systems concurrently (Manfren et al., 2011; Keirstead et al., 2012; Best et al., 2015; Wang et al., 2018; Bakhtiari et al., 2020), integrate the design of urban water and energy infrastructure (Siddiqi and Anadon, 2011; Rodriguez et al., 2013; Bauer et al., 2015; Khan et al., 2017; United Nations, 2019), and provide quantitative feedback on the life-cycle performance of design alternatives (Keirstead et al., 2012; Best et al., 2015; Eicker et al., 2015). This paper develops such a framework to design the supply and demand of integrated wastewater treatment and energy infrastructure for urban neighborhoods. The framework designs infrastructure systems at the neighborhood scale and on an hourly basis following the recommendations by Hawkes and Leach (2005), Jaccard (2006), Evins et al. (2011), and Best et al. (2015).

CCHP plants are a recurrent example of integrated energy infrastructure in practice and research. These plants have shown up to 30% higher overall energy efficiency than comparable segregated energy systems (Rezaie and Rosen, 2012; DOE, n.d). Despite the high efficiency of a CCHP plant, its constant power-to-heat ratio often causes electricity and heating losses (Díaz et al., 2010), as the electricity and heat demanded from the plant to power the network throughout the day do not necessarily align with this fixed ratio. Studies have shown that a wastewater treatment system located within or near the CCHP plant can effectively exploit any “excess” heat and electricity (Elimelech and Phillip, 2011; Zhou et al., 2015), providing an opportunity to design and optimize the water treatment and energy systems concurrently. Given growing water scarcity in many urban areas around the globe, this integration is important for curtailing the fresh water withdrawal used for dissipating the excess heat due to energy generation (Luck et al., 2015; Khan et al., 2017; Energy Information Administration, 2019) and providing decentralized wastewater treatment systems at the neighborhood scale. To realize this opportunity, this paper proposes a framework to design and optimize the demand and supply of integrated energy and water systems (water-energy nexus) for environmental and economic sustainability in an urban neighborhood on an hourly basis.

The contributions of this paper are as follows:

This work enables the demand module from Best et al. (2015) to model the hourly wastewater treatment demands for the 21 building archetypes originally discussed by Best et al. (2015) which represent more than 70% of the buildings in the US (Deru et al., 2011). Table 1 shows a summary of the specifications for these 21 building archetypes. The hourly wastewater outflow for these building archetypes were modeled by converting the potable water demand for each building into the wastewater generated by the building. Table A2 in Supplemenatry Material lists the references used to create the profiles of hourly water use for these 21 building archetypes over the 12 months of the year. As mentioned later in this article, San Francisco, CA, is selected as the location of the test case study; hence, studies on buildings located in San Francisco or in areas of California with similar climates to San Francisco, e.g., Santa Barbara (“IECC Climate Zone Map | Building America Solution Center”, 2012), were used for synthesizing the water demand profiles for the 21 building archetypes. The water demand was considered only for indoor use—outdoor use, comprised mostly of irrigation use, was disregarded in this study.

In converting water use to wastewater outflow, the consumptive use of water in the buildings was conservatively assumed negligible, i.e., the outflow of wastewater was taken equal to the inflow of potable water. The average daily water use of the four “residential” building types were calculated using Equation (1).

here Qwd is the average daily indoor water use (m³/h), GFA is the total gross floor area of the building (in m²), RES^a is the average number of residents per unit area (person/m²), and Qwc is the average daily indoor water use per capita for the building (m³/h/person). For this equation, for each different building type, the average number of residents per unit area was extracted from Deru et al. (2011) and the average daily indoor water use per capita was extracted from Carollo Engineers (2015). The average daily water use (Qwd) for all the other building types, except for the mixed-use buildings, were calculated using Equation (2).

here Qwa is the average daily indoor water use per unit area of the building (m³/h/m²) based on US EIA (2017).

The average daily water use (Qwd) of mixed-use building M1 (symbol explained in Table 1) was computed by combining the average daily water uses of buildings R1 and C2 proportional to the gross-floor-area assigned to each building type, i.e., 8,621 m² of R1 and 784 m² of C2, based on the definition of this mixed-use building archetype (Best et al., 2015). The same approach was followed for computing the average daily water use for mixed-use building M2: the average daily water uses of buildings O1 and C2 were combined proportional to the gross-floor-area assigned to each, i.e., 8,621 m² of O1 and 784 m² of C2.

Assuming the same water demand profile for every day in a month, the wastewater demand of each building type at each hour was calculated using Equation (3).

here Qwwh, m is the wastewater outflow at hour h during month m (m³/h), Qwd is the average daily indoor water use (m³/h). C^h and C^m are the hourly and monthly coefficients for converting the average daily water use to water used at hour h and in month m, respectively. The assumption of constant daily water demand profiles for each day in a month enabled us to construct hourly profiles of water demand over the entire year across several building types within the same climatic region. We recognize that such profiles do change but found that this information is lacking in the technical and academic literature. Tables 2–4 list Qwd, C^h and C^m, respectively, for all the 21 building archetypes.

This work enables the framework proposed by Best et al. (2015) to simulate wastewater treatment processes at the neighborhood scale. For this purpose, three wastewater treatment technologies are modeled and added to the supply module of the original framework: two membrane-based treatment technologies, namely FO-RO and FO-MD, and one conventional central treatment plant (CTP). These treatment systems augmented with pre-treatment and post-treatment processes can treat municipal wastewater to a drinking quality. “Appendix B: Wastewater Characteristics” in Supplementary Material shows the assumed composition of the input wastewater, the required quality of the output drinking water, and discusses the treatment capabilities of FO-RO and FO-MD systems.

The advanced water/wastewater treatment technologies Forward Osmosis (FO), Reverse Osmosis (RO), and Membrane Distillation (MD) have recently received considerable attention from researchers and practitioners for water reuse applications (Alkhudhiri et al., 2012; Council, 2012; McCutcheon and Huang, 2013; Coday et al., 2014; Blandin et al., 2015; Shaffer et al., 2015; Guizania et al., 2019). This is partly because these technologies are more compact, require lighter maintenance, are easier to automate, and can more effectively remove pathogens compared to traditional treatment systems (Council, 2012). These systems have consistent effluent quality and can potentially reduce the chemical demand of the treatment process (Council, 2012). Such qualities make these technologies ideal for implementing at a combined cooling, heating, and power (CCHP) plant with limited space, and for treating domestic wastewater to drinking water quality—a procedure which requires robust treatment processes.

RO and MD membranes suffer irreversible fouling (Husnain et al., 2015a,b; Kim et al., 2018). This phenomenon gradually reduces the water flux, demanding frequent replacement of the membrane. FO can largely mitigate this problem when used as a pretreatment step for RO (Cath et al., 2005; Zaviska and Zou, 2014; Kim et al., 2018) and MD (Husnain et al., 2015a,b). FO prevents the foulants from reaching the RO or MD membrane, whereas the resulting fouling of the FO membrane is mostly reversible and can be resolved with backwashing and chemical cleaning (Husnain et al., 2015a,b). Further, FO-pretreatment filters the bacteria, viruses, and organic and inorganic compounds (Zaviska and Zou, 2014; Akther et al., 2015) making for a more robust treatment.

“FO-RO” (Figure 2) and “FO-MD” (Figure 3) are the hybrid treatment processes resulting from combining FO process with RO and MD. The wastewater treatment module in the framework introduced here incorporates these two systems as the main “integrated” wastewater treatment technologies, i.e., treatment technologies collocated with the energy supply equipment. This module takes as its input the volume of hourly wastewater produced by the neighborhood and the ambient temperature (taken as the temperature of the wastewater). Assuming a 24-h storage capacity for the input wastewater, the module calculates the maximum membrane area required to treat the 24-h moving-average of the inflow over the entire year. This number is used to calculate the capital and operational expenses of the treatment systems. The treatment module also outputs (1) the hourly electricity and heat demand of the selected treatment models (which will be supplied by the CHP engines), and (2) the operational electricity requirement as well as the initial and operational expenses associated with the pre-treatment processes, e.g., screening, and the post-treatment processes, e.g., disinfection and pH adjustment. The calculation of life-cycle expenses for the treatment modules also considers the costs associated with chemicals, labor, retentate handling, and all the other expenses required for normal operation of a treatment plant (Al-Obaidani et al., 2008; McGivney and Kawamura, 2008).

Both treatment models consume electricity while the FO-MD also consumes low-grade heat to treat the wastewater. This characteristic was the other reason for selecting these systems to be integrated with the CCHP plant: these wastewater treatment technologies can consume the unused heat and electricity generated by the CHP engine, and thus mitigate the loss of primary energy and increase the efficiency of the integrated CCHP-wastewater treatment plant over the levels achievable using segregated systems. FO-RO only requires electricity to operate while FO-MD mostly requires heat: FO-RO consumes 3.34 kWh of electricity, while FO-MD consumes 33.2 kWh of heat and 0.55 kWh of electricity to treat 1 m³ of input wastewater at 20°C. The optimization algorithm is expected to favor FO-RO for treating the wastewater in cases with large amounts of extra electricity (generated by the CHP engine), and to favor FO-MD in cases with large quantities of extra heat. As Figure 3 shows, the excess heat produced by the CHP engine will be used to heat the feed water to the MD process, while the ambient air will be used to cool the product water and concentrated draw solution. The induced temperature difference will provide the vapor pressure difference required to run the MD process. “Appendix C: FO-RO” and “Appendix D: FO-MD” in Supplementary Material detail the mathematical models developed for simulating the FO-RO and FO-MD systems, respectively, with overall water recovery ratios (RR) of 50%.

In addition to the integrated treatment systems, a conventional wastewater treatment plant was modeled according to the treatment plant serving the location of the case study in San Francisco. “Appendix E: CWWTP” in Supplementary Material describes the model developed for computing the electricity requirements and expenses associated with this central plant. The wastewater treatment module takes as its input the aggregated hourly demands of the neighborhood for wastewater treatment, and computes the required cost and electricity (1.1 kWh per 1 m³ of input wastewater) to process this wastewater as well as the emissions due to the treatment process. The resulting heating and electric loads are added to the rest of the heating and electric loads imposed on the CHP engine. This module also considers the costs of chemicals, labor, bio-solid and retentate handling, and all the other expenses required for normal operation of a treatment plant (McGivney and Kawamura, 2008).

This study enables the supply-side modeler from Best et al. (2015) to model Sensible Thermal Energy Storage (TES). Hot water storage tanks are selected and modeled for this purpose: hot water generated by the CHP engine for heating with a source temperature of 85°C (Best, 2016) is assumed to get stored in the TES and discharged to satisfy the heating loads when necessary. Four primary factors contribute to the heat loss in a storage tank (Dincer, 2002): heat loss to the surrounding environment, heat conduction between the segments of the storage fluid with higher and lower temperatures (due to the stratified storage of hot water), vertical conduction in the tank wall, and mixing of water during charging and discharging. “Capacity model,” a common loss model (Schütz et al., 2015), assumes the hot water tank to contain a homogenous fluid. With this assumption, the thermal storage at each time-step can be calculated using Equation (4):

where TES is the energy stored in the TES at time t [in kW], k_v is a temporal loss factor (assumed 0.0534% based on Schütz et al., 2015), Δt is the time interval (1 h), and HTES,tin [in kW] is the net heat flow into the TES tank at time-step t (positive for net charging and negative for net discharging of the tank).

Equation (5) was extracted from IEA-ETSAP and IRENA (2013) to calculate the capital costs of a TES tank.

here, TES_Cap is the capacity of the TES [in MWh], and 1.17 is a factor converting 2008 USD to 2019 USD (CPI, 2020). The TES is sized to store enough hot water to satisfy 6 h of maximum aggregate heating demand from the buildings that is computed for each designed neighborhood. The unused heating, in the form of hot water, at each hour gets stored in the TES tank, and the stored hot water in the tank is the first resource responding to heating demand at each hour, i.e., TES has priority over the CHP engine for satisfying the heating demand.

The original objective functions defined by Best et al. (2015) comprised undiscounted life-cycle cost of equipment, annual CO₂ emissions of equipment, and total-fuel-cycle energy efficiency. In this study, the energy efficiency metric was removed from the set of objectives, as it was considered intangible and with no direct monetary or environmental consequences which were not captured by the other two objectives. Reducing the number of optimization objectives lowered the complexity of the optimization algorithm: now, the algorithm needs to inspect only two objectives rather than three when comparing different neighborhoods and finding the most optimal ones.

Further, the two optimization objectives for carbon emissions and life-cycle cost were normalized by the GFA of the neighborhood. Without this normalization, the optimization algorithm would only favor neighborhoods with low emissions or low operational costs regardless of their size: e.g., a small neighborhood with 10,000 ft² of total GFA, would be equivalent to a 1,000,000 ft² neighborhood if they both have the same values of annual carbon emissions and life-cycle cost. However, the larger neighborhood is satisfying a much larger demand in a far more sustainable way than the smaller neighborhood. Normalization addresses this error, enabling the optimization method to achieve optimal (sustainable) neighborhoods with a variety of sizes.

Also, to make more tangible the environmental impacts of building and operating the neighborhood, Social Cost of Carbon (SCC) was used to convert the annual carbon emissions into dollar values representing the social, environmental, and economic damage resulting from CO₂ emissions (EPA, 2016). Among the different models used to calculate SCC, the version created by US Environmental Protection Agency (EPA) was chosen to be used here (EPA, 2016). “Appendix F: Social Cost of Carbon” in Supplementary Material presents the regression function used to calculate the SCC for each year over the design period. In this new objective function, the carbon emissions associated with the construction of the equipment were also accounted for in addition to the operational emissions. The construction costs were converted into equivalent carbon emissions using conversion factors from EIO_LCA database (Carnegie Mellon University, 2007) for the construction of utilities: 243 Tons of CO_2e was assumed to result from 1 million 2007 USD worth of economic activity.

To showcase the capabilities of the proposed framework in co-optimization of supply and demand as well as integrated design of water and energy infrastructure, a case study was conducted based on that proposed by Best et al. (2015). Best et al. modeled a 160-acre greenfield development in Civic Center, San Francisco, and simulated the electricity, cooling, and heating demands of 21 building archetypes (presented in Table 1) using EnergyPlus ver. 8.0.0 (Crawley et al., 2001). Best et al. also modeled 32 CHP engines and 16 chillers as the set of supply equipment. The same 160-acre development, building archetypes, CHP engines, and chillers are considered in this study with the following modifications: (i) Hourly wastewater treatment demands were added to the set of energy demands of the buildings, (ii) Thermal Energy Storage and three wastewater treatment technologies (namely, FO-RO, FO-MD, and CWWTP) were added to the original set of supply models, and (iii) the optimization objectives were changed to GFA-normalized life-cycle cost of equipment and life-cycle social cost of carbon.

The proposed model was run for two scenarios: (i) In the first one, the building mix and energy system were designed and optimized together, while a conventional central wastewater treatment plant (CWWTP) treated the wastewater from the buildings, and (ii) in the seconds scenario, the building mix, energy system, and wastewater treatment system were optimized and designed concurrently. In the first scenario, the grid was assumed to supply the electricity required by the CWWTP, and grid emissions for generating this electricity were attributed to the CWWTP. The emissions and expenses of the CWWTP were in turn attributed to the neighborhood in proportion to the ratio of the “average daily wastewater output of the neighborhood” to “57 MGD” (i.e., the actual average daily input of the treatment plant). Since the existing CWWTP serving the location of the case study, i.e., Southeast Treatment Plant (SF PUC, 2019), does not produce potable water, it was arbitrarily replaced by a conventional treatment plant of the same size but capable of producing potable water—as also explained in Appendix E: CWWTP Model in Supplementary Material. This modification was necessary to make the two scenarios comparable. In the seconds scenario, a combined CCHP-wastewater treatment plant (CCHP-WWTP) supplied the energy and wastewater treatment demand of the buildings.

Both scenarios were designed for minimum life cycle cost of equipment (LCC) and minimum social cost of carbon (SCC). The LCC optimization objective incorporated the construction and discounted operation and maintenance costs of only the CCHP plant for the first scenario, and the entire CCHP-WWTP for the seconds. The SCC optimization objective considered the construction and discounted operation costs of carbon emissions associated with only the CCHP plant for the first scenario, and the entire CCHP-WWTP for the seconds. Both scenarios were simulated for 20 year of operation with a 3.5% discount rate. The neighborhoods generated by the optimization algorithm were compared in terms of their total (i.e., including those associated with the CWWTP) life cycle cost of equipment, operational and construction GHG emissions, types of supply technologies, and building mix among other metrics. The LCC and SCC calculations excluded the reuse and redistribution of the recovered water.

The excess electricity generated during each hour was assumed to be sold back to the grid at 80% of the purchase price at that hour which was extracted from CAISO for year 2018 (California ISO, 2019). The hourly ambient temperature for the case study was extracted from TMY3 dataset for the San Francisco International Airport (Wilcox and Marion, 2008). Further, the grid emissions for each hour were extracted from Beyond Efficiency (Beyond Efficiency, 2017) for year 2016.

The two scenarios, namely scenario 1 (conventional design) and scenario 2 (integrated design), were simulated each using an Intel Xeon E5-2640v4 processor (10-core Broadwell, 2.40 GHz) with 800 MBs of effective memory use over 145 mins of CPU wall-clock time. The map method from Scoop library (Hold-Geoffroy et al., 2014) was used to parallelize the map calls to Genetic Algorithm (GA) methods from Deap library (Fortin et al., 2012). The advantage of using a GA for solving this problem is that it does not require assignment of weights to either of the two objective functions, and instead enables simultaneous optimization of both objectives (Deb et al., 2002). The GA algorithm, NSGA-II, used the following optimization parameters based on the recommendations of Best et al. (2015):

Number of generations = 1,000; Population = 200; Mutation Probability = 0.05; Similarity Parameter = 2.5; Crossover Probability = 0.75.

Substituting series data processing with parallelized Numpy (Oliphant, 2006) operations, reducing the number of objective functions from three to two, and other minor technical upgrades cut the optimization runtime by 97% compared to the optimization with the same configurations done using the original framework in Best et al. (2015). Of the 200,000 neighborhoods produced by the GA algorithm for each scenario, 93,210 for scenario 1 and 87,644 for scenario 2 had admissible total GFA (i.e., less than 3,237,485 m²) and site GFA (i.e., less than 647,497 m²). These neighborhoods are called the “admissible neighborhoods” from this point on.

Figure 4 shows the two GFA-normalized objective functions of the optimization, social cost of carbon vs. life-cycle cost, for all the admissible neighborhoods with SCC values less than 0.10 k$/m², i.e., the 99.8th and 99.7th percentile of SCC across all neighborhoods in the first and the seconds scenario, respectively. Earlier generations of neighborhoods produced by the GA are plotted using symbols that are more transparent than the later ones. One outstanding observation from Figure 4 is that the optimization algorithm has successfully identified paths to designing neighborhoods with minimum LCC and SCC over consecutive generations: the dots (neighborhoods) closer to the origin (which has the ideal lowest SCC and LCC, i.e., 0 kg of lifetime CO₂ emissions and $0 of life-cycle cost) are more opaque (belong to later generations) than the dots farther away from the origin. Figure 4 also offers an insight on the comparison between the two scenarios: the blue dots in the figure are generally located lower than their adjacent red dots. This implies that the integrated design scenario, scenario 2, has resulted in neighborhoods with generally lower SCC than those of conventional design, scenario 1, for the same values of LCC.

Figure 5 shows the approximate Pareto fronts of the two scenarios for the objective functions, life-cycle cost and life-cycle social cost of carbon, highlighted in yellow. In the Pareto optimal solutions, the SCC values of the integrated approach seem to be about 1/3 the SCC values of the segregated approach, while the LCC values of these solutions are similar between the two scenarios. These two observations about the Pareto fronts comply with the general trends deduced in the previous paragraph.

To get a better understanding of the distribution of performance metrics across the designs produced by the two scenarios, standard deviation (SD), mean (average), 25th percentile, 50th percentile (median), and 75th percentile of LCC, SCC, annual aggregated energy demand, and annual wastewater treatment demand for all the neighborhoods resulting from the two design scenarios were calculated—Table 5 shows the results. The annual aggregated energy demand for each neighborhood was calculated by summing the total electricity and heating demand of the buildings and supply equipment (including chillers and treatment systems) over 8,760 h of the year. In this table, “CCHP | CWWTP” denotes scenario 1, and “CCHP+WWT” denotes scenario 2. The numbers in Table 5 are unitless since the values of each metric for each variable, e.g., 25th percentile of LCC, for both scenarios are normalized by the value of the scenario 1. This facilitates comparing the statistical metrics of the seconds scenario against those of the first scenario by converting the metrics of the first scenario into unity.

The results in Table 5 show that integrated design of energy and wastewater treatment systems has resulted in lower life-cycle carbon emissions and cost as well as lower annual energy and wastewater treatment demands compared to a separate approach to the design of these two infrastructure systems. The difference in life-cycle social cost of carbon is the most noticeable, with scenario 2 having a mean value 56% lower than scenario 1. The reduction in mean values of annual demands from scenario 1 to 2 is ca. 40%. Comparing the SDs of the two scenarios also offers another observation: the optimal neighborhoods produced by the GA under scenario 2 are closer (more uniform) in terms of the four parameters inspected in Table 5. This can imply the existence of some neighborhoods with dominant genes (i.e., configurations resulting in very low SCC and LCC) in the population of scenario 2 which have been identified by the GA and sustained across multiple generations of the analysis.

To further inspect the differences between the designs resulting from the two scenarios, K-Means Clustering (KMC) was used to create clusters of similar individuals among the admissible neighborhoods based on eleven features: the seven building-type ratios (i.e., GFA associated with each building type divided by the total GFA of the neighborhood), GFA-normalized LCC and SCC, and GFA-normalized annual energy and wastewater treatment demands. The features were standardized before clustering, and K-Means Clustering method from Scikit-Learn library (Pedregosa et al., 2011) was used to conduct the clustering with the following variables:

Initialization_method: “k-means++”; max_iterations: 300; n_init: 10; random_state: 42; tol: 0.0001.

Euclidean distance was selected as the measure of proximity for clustering. The number of clusters was varied between 8 and 40 to find the highest average Silhouette Score (Rousseeuw, 1987) and lowest cost function across all clusters: for both scenarios, the highest Silhouette scores (0.53 for scenario 1, and 0.32 for scenario 2) were obtained for 10 clusters. However, the cost function (i.e., sum of squared distances to centroids) for the 10-cluster configuration was high −233,834 (2.51 per individual) and 257,098 (2.93 per individual) for scenarios 1 and 2, respectively—and the clusters were unevenly populated. To solve this issue, the number of clusters was set to 26 (scenario 1) and 40 (scenario 2) which had yielded other local optima of the Silhouette Score among the tested number of clusters. These new cluster counts resulted in acceptably reduced cost functions of 105,064 (i.e., 1.1 per each neighborhood) for scenario 1 and 107,026 (i.e., 1.2 per each neighborhood) for scenario 2 with Silhouette scores of 0.36 and 0.31, respectively.

After the KMC algorithm converged, clusters with less than 500 members were removed to prune out the outliers, and the remaining centroids (19 centroids for scenario 1, and 27 for scenario 2) were plotted in Figure 6. Each cluster centroid in this figure is shown as a continuous, unicolor line across the eleven variables along the x-axis, and the thickness of each line is proportional to the number of neighborhoods represented by that centroid. SCC, LCC, and the annual energy and wastewater demands for both scenarios in Figure 6 are normalized by the maximum values of each variable from both scenarios, and are presented in percent.

Figure 6 shows that the building mix and normalized annual energy and wastewater demand for the most populous clusters of optimal neighborhoods for both design approaches cover roughly the same ranges of values. However, the most populated clusters in scenario 2 have lower GFA-normalized SCC and LCC than those of scenario 1, corroborating the statistical data from Table 5. Also implied by the low standard deviation values of scenario 2 in Table 5, the most populous centroids of scenario 2 in Figure 6 are spread over smaller ranges of the eleven parameters plotted in the figure. The GFA ratios of individual building types show that (i) the percentage of residential buildings ranges from 0 to 60% for both scenarios, (ii) the office buildings range from 30 to 95% in scenario 1 and from 15 to 100% in scenario 2 with a high concentration around 80 to 100%, (iii) industrial and educational buildings are mostly negligible in scenario 1, while they range from 0 to 10% in scenario 2, and (iv) hospitality and medical buildings are mostly negligible in scenario 1 with two exceptional clusters showing ~30 and ~60% hospitality and medical type-ratios, while they range from 0 to 20% in scenario 2. These ranges show that the integrated design has resulted in a more diverse building mix among the optimal designs.

Figure 6 offers another interesting observation: the building mix in the majority of optimal neighborhoods from both scenarios have a low share of residential buildings, a medium to high share of office buildings, low shares of educational and industrial buildings (in scenario 2), and trivial shares of commercial, hospitality, and medical buildings. This mix of building types seems to provide the lowest LCC and SCC per unit area among all possible building mixes inspected by the optimization algorithm and is hence predominant among the optimal building mixes for both scenarios.

The types of selected supply equipment -were also analyzed for the admissible neighborhoods from both scenarios to identify the connections between the choices of supply equipment choices and metrics of life-cycle performance. Figure 7 shows for the two scenarios the normalized SCC against the types of CHP engines selected in neighborhoods with SCCs less than 0.1 k$/m², i.e., the 99.8th percentile of the first scenario and 99.7th percentile of the seconds. In both cases, the neighborhoods with smallest SCCs use biomass/biogas, or fuel-cell CHP engines, which is consistent with the fact that the operation of these engines is deemed carbon-neutral since they are assumed to use renewable fuels (Best et al., 2015). The Pareto optimal solutions marked in Figure 5 are expected to be among these neighborhoods with carbon-neutral CHP engines, since the Pareto optimal solutions also have very low SCC values compared to the rest of the solutions. The neighborhoods designed with carbon-neutral CHP engines in scenario 1 seem to have a lower SCC compared to neighborhoods using the same engines in scenario 2. Further, since a large portion of the CHP engines selected in both scenarios for the local energy system are those with negligible operating emissions, supplying the wastewater treatment equipment with the local energy system in the integrated scenario would expectedly result in a much smaller life-cycle emission (SCC) compared to supplying the treatment equipment with the electric grid in the segregated scenario. These observations justify the lower SCC values of scenario 2 compared to scenario 1. Figure 7 also shows that the Genetic Algorithm has favored gas engines (i.e., gas turbines, microturbines, reciprocating engines, and steam turbines) much less than the carbon-neutral CHP engines in both scenarios given the significantly higher emission rates of the gas engines.

Figure 8 also shows the normalized SCC vs. the types of chillers chosen for those neighborhoods in the analyzed scenarios which had SCCs less than 0.1 k$/m². Low-emission neighborhoods have been possible to create using all types of the available chillers.

Figure 9 shows the wastewater treatment (WWT) systems selected by the optimization algorithm for scenario 2. FO-MD has resulted in more low-emission neighborhoods than the other WWT system. Before, Figure 7 indicated that biomass engines were popular with low-emission neighborhoods. Most of the available biomass engines in the simulation framework have power-to-heat ratios lower than 0.42, which means they generate more heat than electricity in response to the energy demand. Part of the reason for popularity of FO-MD among the low-emission neighborhoods seems to be that this treatment system can “recycle” the excess heat generated by the carbon-neutral biomass engines selected for these neighborhoods. This excess heat can also be used for dewatering the solid residue from the treatment process (Scherson and Criddle, 2014) (discussed in more detail in Section Conclusions and Future Work).

Figure 6 suggests a relationship between the oscillations in ratios of different building types with the values of SCC, LCC, and the annual demands. In a final analysis, the Pearson correlation matrices between the eleven variables plotted in Figure 6 were computed to inspect the linear relationships between these variables. Figure 10 shows the resulting correlation matrices for both scenarios. Normalized LCC and SCC are not strongly related to any other variables than each other, while normalized annual energy and wastewater demand seem to be negatively correlated with the ratio of office buildings in the neighborhoods. The latter is rooted in the low energy- and water-use intensities of the office buildings. The normalized annual demands also show a strong direct linear relationship with each other, meaning that neighborhoods with high annual wastewater treatment demands tend to have high annual energy demands and vice versa. Further, annual energy demand seems to be directly proportional to the percentage of medical buildings. This is caused by the high energy-use intensity of the medical buildings. The annual wastewater demand in scenario 2 also shows a strong collinearity with the percentage of residential buildings, which is caused by the high water-use intensity of these building types. Among the building type ratios, percentages of hospitality and medical buildings in both cases and commercial buildings in scenario 2 are negatively collinear with that of office buildings. The optimization algorithm seems to have strongly preferred office buildings in reverse proportion to the residentials, especially in the integrated design scenario.