The Surface Urban Energy and Water Balance Scheme (SUEWS) is used to investigate the impact of anthropogenic heat flux QF and irrigation on surface energy balance partitioning in a central business district of Shanghai. Diurnal profiles of QF are carefully derived based on city-specific hourly electricity consumption data, hourly traffic data, and dynamic population density. The QF is estimated to be largest in summer (mean daily peak 236 W m−2). When QF is omitted, the SUEWS sensible heat flux QH reproduces the observed diurnal pattern generally well, but the magnitude is underestimated compared to observations for all seasons. When QF is included, the QH estimates are improved in spring, summer, and autumn but are poorer in winter, indicating winter QF is overestimated. Inclusion of QF has little influence on the simulated latent heat flux QE but improves the storage heat flux estimates except in winter. Irrigation, both amount and frequency, has a large impact on QE. When irrigation is not considered, the simulated QE is underestimated for all seasons. The mean summer daytime QE is largely overestimated compared to observations under continuous irrigation conditions. Model results are improved when irrigation occurs with a 3-day frequency, especially in summer. Results are consistent with observed monthly outdoor water use. This study highlights the importance of appropriately including QF and irrigation in urban land surface models—terms not generally considered in many previous studies.
China has experienced unprecedented urban growth in recent decades, with the fraction of city dwellers increasing from 17.9% to 55.6% between 1978 and 2015 (United Nations 2017). If these rates continue, the urban population will exceed 1 billion in China within the next two decades. This rapid urbanization has brought significant economic growth, while at the same time exposing people to urban climatic and environmental risks, such as persistent heat waves, flooding, and air pollution (Jiang et al. 2015; Li et al. 2015; Zhong et al. 2015; Ding et al. 2016; Xu et al. 2016; Yang et al. 2017). Cities are well known to have distinct climatic conditions that result from the alteration of the urban surface–atmosphere energy and water exchanges compared to surrounding rural surfaces (Roth et al. 2017; Zou et al. 2017). Absorption and trapping of incoming shortwave radiation by deep urban canyons leads to greater absorption of energy by the surface and a smaller surface albedo than surrounding environments (Oke 1988; Christen and Vogt 2004; Guo et al. 2016). More heat is stored in high thermal admittance building walls during the daytime, which is then released at night, creating the distinct nocturnal urban heat island (UHI; Grimmond and Oke 1999; Roberts et al. 2006; Wu and Yang 2013; Kotthaus and Grimmond 2014). The replacement of natural vegetative surfaces with impervious paved and built surfaces leads to less energy partitioning into evapotranspiration and reduces the associated cooling effect (Grimmond and Oke 1986; Nakayoshi et al. 2009; Ward and Grimmond 2017). Urban runoff is usually significantly enhanced following rainfall, given the abundance of paved and built surfaces. This rapid rate of runoff also removes a large amount of surface water suppressing evaporation rates (Ragab et al. 2003). Human activities, related to building heating and cooling, vehicles, and human metabolism (Sailor 2011; Lindberg et al. 2013), release extra anthropogenic heat into the urban environment.
Urban land surface models (ULSMs) can be effective tools to investigate and quantify these surface–atmosphere exchanges and interactions to yield insight into the different factors influencing the climate of a city. Numerous ULSMs have been developed over the last few decades, with varying degrees of complexity (e.g., Kusaka et al. 2001; Oleson et al. 2008; Järvi et al. 2011; Chen et al. 2011; Masson et al. 2013; Miao and Chen 2014). Grimmond et al. (2010, 2011), in the first international comparison of ULSMs, found no single model performs best or worst for all fluxes. Considering the implications of this study, Best and Grimmond (2016b) concluded that attention needs to be directed to the modeling of the latent heat flux, inclusion (or not) of vegetation, and calculation of the anthropogenic heat flux by ULSMs. These elements are simulated poorly, yet are key factors impacting overall model performance. They are the focus of this paper.
ULSMs often simulate the latent heat flux separately for natural (vegetation or pervious) and built (road, walls, roof) surfaces, with no interaction between them. Furthermore, ULSMs rarely incorporate any detailed consideration of urban hydrological processes, such as drainage, interception, runoff, or irrigation. For example, the early version of the widely implemented single-layer urban canopy model (SLUCM) system (Kusaka et al. 2001) uses a simplified hydrologic process in which evaporation only occurs after precipitation events, even though SLUCM implements a sophisticated representation of urban canopy geometry. Recently, enhanced hydrological processes including anthropogenic latent heat, urban irrigation, and urban oasis effects have been implemented into the SLUCM system (Miao and Chen 2014; Yang et al. 2015), which improves the model performance substantially especially for the latent heat flux.
Given the large fraction of impervious surfaces in urban areas, city drainage systems are designed to quickly remove runoff. In many settings this gives rise to a deficit of soil moisture in the urban landscape (Coutts et al. 2013) and irrigation is often needed to maintain vegetation health (Grimmond and Oke 1986; Demuzere et al. 2014). This urban irrigation has been shown to be a critical component of the urban water balance, especially in arid and hot regions, and plays a key role in the energy partitioning between latent and sensible heat fluxes and the associated urban cooling efficiency. Vahmani and Hogue (2014, 2015) developed and assessed an irrigation scheme within the framework of the Noah–SLUCM system for the Los Angeles metropolitan area and demonstrated that appropriately incorporating urban irrigation can significantly improve model performance. However, the majority of ULSMs applications still ignore irrigation, especially in subtropical cities, which are considered to have plenty of rainfall to maintain a sufficient water supply. This, however, is not always the case, and with increased frequency of extreme heat waves the potential need for external water supply can be substantial in these cities.
The anthropogenic heat flux QF also plays a critical role in ULSMs and has been the focus of significant attention (Grimmond 1992; Sailor and Lu 2004; Allen et al. 2011; Zhang et al. 2016; Best and Grimmond 2016a). The QF, the additional energy produced by human activities released into the environment, can be a significant component of the urban energy balance with distinct seasonal and diurnal variations. For example, the estimated daytime QF in central Tokyo (Ichinose et al. 1999) exceeded 400 W m−2 at 1400 local standard time (LST) on average and reached 1590 W m−2 in winter (25-m resolution), enhancing the UHI by 1°–2.5°C. The magnitude of QF is scale and location dependent. Typically, it is highest in central urbanized areas and much less when averaged over the entire city. Incorporating QF into mesoscale weather forecast models has been shown to have a significant impact on model predictions when city-specific QF profiles and magnitudes are provided (Salamanca et al. 2014). However, such city-specific QF diurnal profiles are often very difficult to obtain given the lack of detailed local energy consumption data. As a result, most urban modelers simply use a default fixed QF diurnal profile (e.g., two diurnal peaks at 0800 and 1700 LST is the default in WRF–UCM), fixed values regardless of the city (e.g., Wang et al. 2015; L. Chen et al. 2016), or turn anthropogenic heating off (e.g., Zhang et al. 2010; Loridan et al. 2013; Wang et al. 2014; S. Zhong et al. 2017). This diversity of approaches has contributed to contradictory results on the impact of QF on local climate. For example, two recent studies have shown inconsistent impacts of QF with the WRF–UCM default QF on precipitation: F. Chen et al. (2016) report it results in increases in precipitation, while Feng et al. (2012) report a decrease of precipitation in the same region (Hangzhou, China). Others, however, have made significant advances in this realm. Sailor et al. (2015) developed a national database of anthropogenic heat profiles for the United States and extended this, by simple adjustments, for a range of international megacities. Adopting a different approach, Nie et al. (2017) used WRF–building effect parameterization plus building energy model (BEP+BEM) to estimate spatially and diurnally varying QF in Beijing. However, given the vast diversity of cities in China, there is an urgent need to develop datasets and models that simulate the spatial and temporal variability of QF across cities.
The Surface Urban Energy and Water Balance Scheme (SUEWS) is a local-scale urban land surface model of moderate complexity (Järvi et al. 2011; Ward et al. 2016). SUEWS has the advantage that it simulates the urban surface energy balance in combination with the complete urban hydrological cycle, considering irrigation and runoff processes. The urban water balance interacts with the energy balance through evaporation E, as QE = LVE, where QE is the latent heat flux and LV is the latent heat of vaporization. Moreover, SUEWS requires only commonly available meteorological input data and detailed information about the urban surface. The urban surface is split into seven land cover types (buildings, paved surfaces, coniferous trees and shrubs, deciduous trees and shrubs, grass, bare soil, and water) with integrated urban vegetation effects, a previously highlighted key factor for improving the accuracy of ULSMs (Grimmond et al. 2010, 2011). These characteristics of SUEWS have enabled the model to be used widely as an effective tool for climate (water) sensitive urban design and urban climate disaster and mitigation strategy assessment (Mitchell et al. 2008; Järvi et al. 2017; Ward and Grimmond 2017; Ward et al. 2017; Rafael et al. 2017).
The SUEWS model was originally tested using data collected from a midlatitude suburb in Vancouver, Canada (Grimmond and Oke 1986, 1991; Loridan et al. 2011; Järvi et al. 2011). It has been evaluated extensively in North American and European cities and shown to produce realistic and robust results (Järvi et al. 2014; Alexander et al. 2015; Karsisto et al. 2016; Ward and Grimmond 2017; Kokkonen et al. 2018). However, evaluation of SUEWS in rapidly urbanizing subtropical (or tropical) cities is still lacking, with the exception of recent work in tropical Singapore (Demuzere et al. 2017). Given the vast diversity of climatic settings and urban geometrical structures [cf. local climate zones in Stewart and Oke (2012)] of different cities, further evaluation of the model in subtropical cities is of paramount importance. Shanghai, the largest subtropical city in China, characterized by numerous skyscrapers and dense urbanization, provides a test bed to evaluate SUEWS.
The objective of this study is to evaluate the performance of SUEWS in a central business site of Shanghai for one year using directly measured surface energy flux observations (Ao et al. 2016a,b). Special attention is directed to the impact of the seasonally varying diurnal profiles of QF derived from city-scale annual energy consumption data, hourly electricity power load data, and traffic count data. The impact of urban irrigation on the simulation of latent heat flux (evaporation) is also addressed using the empirical irrigation scheme in SUEWS. This study provides insights into the performance of SUEWS and its potential to investigate strategies to mitigate urban heat stress and create resilient and sustainable urban environments.
a. Site and observations
The evaluation of SUEWS uses data (December 2012–November 2013) observed over a dense urban site (XJH) in Shanghai, China (Fig. 1). The four components of net all-wave radiation, the turbulent sensible and latent heat fluxes, and basic meteorological variables (air temperature, relative humidity, and pressure) are directly measured at this site on a tall tower (full details are provided in Ao et al. 2016a,b). Precipitation is measured nearby (60 m away) with an automatic weather station (AWS).
Annual and seasonal performance of SUEWS is considered using carefully quality controlled data (see further details in Ao et al. 2016a,b). Data from sectors strongly influenced by a tall building (210°–247°) and the tower itself (320°–337°) are excluded. Wet conditions (within 1 day after rain) are excluded as rain drops on open-path sensors generate errors.
Surface cover parameters needed for SUEWS are retrieved from GIS data and a ground survey for a 500-m radius around the site. The Kljun et al. (2004) flux source area model suggests that the 80% source area extends to about 600 m from the site.
The four seasons of a year are defined based on the commonly used classification in China: winter [December–February (DJF)], spring [March–May (MAM)], summer [June–August (JJA)], and autumn [September–November (SON)]. Local standard time is used (China does not use daylight saving time).
b. Estimation of anthropogenic heat flux QF
In SUEWS, the daily anthropogenic heat flux QF,S,daily is calculated adopting the Sailor and Vasireddy (2006) based approach. This is a function of population density ρpop and heating and cooling degree days (HDD and CDD; Fig. 2a):
with daily HDD and CDD defined based on the hourly air temperature Ti and a balance point temperature Tb for human comfort, as
Logic variable l equals 1 when (Tb − Ti) > 0°C and equals 0 when (Tb − Ti) < 0°C for HDD, and vice versa for CDD, and aF0 is the base QF,S,daily from all sources at the balance point temperature. The slopes aF1 and aF2 differ (Fig. 2a). Three coefficients (aF0, aF1, aF2) need to be specified for a study site. Here daily results (QF,L,daily) from LQF (see appendix A) are used to obtain these three coefficients. The fitted cooling slope (aF1) is 0.0181 W m−2 K−1 (capita ha−1)−1, the heating slope (aF2) is 0.0035 W m−2 K−1 (capita ha−1)−1, and aF0 = 0.3963 W m−2 (capita ha−1)−1 with the single Tb of 20°C.
The diurnal profiles of the building QFB for weekday, weekend, and holidays (Fig. 3a) are mainly based on diurnal variations of the city-scale electricity consumption data and further scaled by the electricity fraction of Shanghai (14%; Table 1), industry fraction of the XJH site (10%), and diurnal variation of population density (Yu and Wen 2016; W. Zhong et al. 2017). The diurnal profiles of the vehicle based QFB (Fig. 3b) are derived from hourly highway traffic data for the inner ring of Shanghai in 2011 (see section 3a and appendix A; Su et al. 2014).
c. Estimation of irrigation
Irrigation and street cleaning provides additional water to precipitation for runoff, evaporation, and soil moisture storage. In areas dominated by impervious surfaces, with limited available water for evaporation, the impact can be significant (Best and Grimmond 2016b; Sun et al. 2014). In SUEWS, the probable daily external water use Ie is calculated as a function of the daily mean air temperature Td and number of days after rain Dar (Järvi et al. 2011):
For each of the iυ vegetation types the irrigation fraction (fiυ) and its method (automatic faut or manual) are needed. The site specific coefficients (b0,a, b1,a, b2,a and b0,m, b1,m, b2,m) characterize the behavior of automated a and manual m irrigation during the period of external water use (Ie,start and Ie,end, days of year). The parameters used are given in Table 1. Additionally, the days of week when irrigation is permitted need to be specified (Shanghai: no restricted days of week). With the user-provided diurnal profiles of water use [see section 3d(1)], the daily irrigation is downscaled to hourly and then 5 min.
d. SUEWS model setup and parameters
Version v2017a of SUEWS (Ward et al. 2016, 2017; Ward and Grimmond 2017) is used with a 5-min time step but forced by 60-min meteorological input data. The incoming shortwave radiation flux K↓, air temperature Ta, pressure p, relative humidity (RH), and precipitation P are linearly interpolated to 5 min by SUEWS. The 60-min averaged (by SUEWS) outputs are used in the comparison with observations. The parameter values used to characterize the study site (XJH) are given in Table 1.
Surface cover and mean building height zH parameters are derived from GIS and ground survey data (Ao et al. 2016a). The roughness length z0 and the zero-plane displacement height zd method selected are the simple fixed fraction of the mean building height zH or rule of thumb (0.1zH and 0.7zH; Grimmond and Oke 1999). Given the conclusions of Kent et al. (2017) and Tang et al. (2016), runs are also undertaken using the Kanda et al. (2013) method. Mean values calculated around the site (1° interval) gave z0 = 7.6 and zd = 55 m. The model result difference from that of rule-of-thumb method is small (not shown).
The leaf area index (LAI) or phenology has a critical influence on the latent heat flux QE and timing of appropriate seasons. It dynamically responds to growing degree days (GDD) and senescence degree days (SDD). Phenology parameters are determined from observations and photographs with local daily mean air temperatures [leaf-on (Tbase-on), leaf-off (Tbase-off)], growing degree days until full LAI (GDDfull), and senescence degree days that initiate leaf-off (SDDfull) are calculated as cumulative values of GDD and SDD during growing and senescence periods in Shanghai (around 15 March–20 May and 15 October–20 December, respectively). The equation used for GDD (SDD) in SUEWS is (McMaster and Wilhelm 1997):
where Tmax and Tmin are daily maximum and minimum air temperature, respectively. GDD (SDD) must be greater (smaller) than 0, and if not, they are forced to 0. As expected, the Shanghai Tbase-on and Tbase-off (10° and 20°C; Table 1) are much warmer than for Helsinki and Montreal (5° and 10°C; Järvi et al. 2014) and London (6° and 11°C; Ward et al. 2016). Latitude strongly impacts the length of growing (senescence) season. The daily LAI for each vegetation type iυ is a function of the previous day LAI (LAId−1,iυ), daily mean air temperature Td, and day length td (Järvi et al. 2014). For leaf-on, Td > Tbase-on:
For leaf-off, it may be initiated by a thermal condition only (Td < Tbase-off):
or also account for day length (td < 12 h)
From sensitivity experiments (not shown), it is found that accounting for day length [Eq. (5c)] is unsuitable at this subtropical site (small variation of td) as simulated leaf-off occurs too early and the senescence is too fast. Hence, Eq. (5b) is adopted in this study. The parameters c1,2,3,4 control the changing rate of LAI. Parameter c1,2,3 values are the same in Järvi et al. (2014) for Helsinki, while for c4 for evergreen trees, a smaller value is used based on local visual/photograph surveys that leaves are still partially active to late December. The subtropical climate and careful maintenance (e.g., turf replacement if it turns yellow) at the Shanghai site results in a longer growing season for grass than higher latitudes. Based on photographs taken around the XJH site in winter and spring (not shown), the grass remains green in winter months. Therefore, the minimum LAI for grass is increased from 1.6 in Järvi et al. (2014) to 3.2 at our site (Table 1), which results in a better model performance, especially in winter and spring seasons (not shown).
Previous evaluation of the radiation components of SUEWS (Ao et al. 2016b) found good performance. The net all-wave radiation flux Q* modeled with downward longwave radiation flux L↓ estimated as a function of RH and Ta. The storage heat flux is calculated using the objective hysteresis model (OHM; Grimmond et al. 1991), with Q* + QF (rather than only Q*) used:
where fi is the fraction for the ith surface type, and t is the time. The three coefficients (a1, a2, a3) of OHM are from the literature (Table 1).
The latent heat flux QE is modeled using the modified Penman–Monteith equation (Grimmond and Oke 1991). More detailed description of the parameterization of QE is given in section 3c. The sensible heat flux QH is calculated as a residual of the surface energy balance. The soil layer underneath each surface type (except water surface) is assumed to be 350 mm, with a maximum water capacity of 150 mm. To obtain appropriate initial conditions, SUEWS is run for one year with the 2012/13 forcing to get probable initial state of soil moisture storage and leaf area index.
a. Anthropogenic heat fluxes
The diurnal profiles of the building anthropogenic heat flux for weekdays, weekends, and holidays are similar (Fig. 3a): all are low from 0000 to 0600 LST, then increase gradually until 1100 LST, with a small decrease at 1200 LST. Thereafter, the three diurnal patterns begin to diverge. During weekdays values remain relatively constant from 1200 to 1700 LST, then decrease steadily. During weekends the timing of this decrease lags by 2 h (i.e., from 1900 LST), whereas on holidays there is a stronger evening peak around 1900 LST.
The diurnal profiles that account for population variations (Fig. 3a) have much larger amplitudes than the original profiles and are similar in amplitude to the default LQF and SUEWS Vancouver 1987 (Vs87) profiles (Grimmond 1992; Järvi et al. 2011) and to other studies, for example, in Japan (Takane et al. 2017). The ratios of the maximum and minimum value for scaled weekday, weekend, and holiday are 4.7, 3.2, and 3.4, respectively. The corresponding values for LQF and SUEWS QF schemes are 5.3 and 3.8, respectively.
Seasonal differences in diurnal patterns and magnitudes of the vehicle heat emissions are relatively small (not shown). The weekday morning peak (at 0800 LST) is distinct whereas the evening peak (at 1700 LST; Fig. 3b) is unlike North American cities, where the evening peak is generally stronger than the morning peak (Grimmond 1992; Hallenbeck et al. 1997; Chow et al. 2014). This may be because the end of work varies between companies; while most institutions or government offices finish between 1700 and 1800 LST, many people often stay at the office in the evening. Additionally, shopping malls and restaurants are open until 2100–2200 LST. The weekend pattern, without distinct peaks, slowly increases in the morning then stays flat from about 1000 to 1700 LST. Holidays have a similar pattern to the weekend but with smaller magnitudes.
The LQF and SUEWS QF results are very similar (Fig. 4). As the three SUEWS coefficients (aF0, aF1, aF2) are derived using the LQF results, this is expected. The larger summertime results are a function of the larger aF2 slope and therefore dependence on CDD, as expected. The peak mean daily summer QF,L (Fig. 4) is around 236 W m−2; winter and autumn mean fluxes peak are similar (190 W m−2) and spring is slightly smaller (180 W m−2). These values, using the new response function, give a slightly bigger seasonal variation than the original function (not shown). The building heat emission is the major subcomponent, accounting for about 95% of the total QF,L. The modeled metabolic heat emission QFM is about 3 W m−2 at night and 7 W m−2 during daytime and does not show seasonal variations. Seasonal differences in vehicle heat emissions also are very small, with QFV around 3 W m−2 in the daytime. The small (or no) seasonal variation for QFV and QFM is because the same parameter settings are assumed for each season, as there is a lack of information to suggest otherwise. The difference between winter and the other three seasons is because more holidays occur in winter. The magnitudes of QFM and QFV estimated here are similar to previous studies (Chow et al. 2014; Lu et al. 2016; Stewart and Kennedy 2017). There is a high correlation coefficient (0.97) between hourly QF,S and QF,L. Given the simplicity of SUEWS QF, the dynamic temperature response, and comparable results to LQF, the SUEWS QF is regarded as an appropriate method to use after careful determination of the parameters. As the purpose of this study is the evaluation of SUEWS, the SUEWS QF are used in the following sections.
b. Impact of QF on surface energy fluxes
Two simulations to examine the impact of QF on the surface energy balance fluxes in Shanghai are conducted (Figs. 5, C1–C3): 1) assuming QF = 0 W m−2 (hereafter QF,0 Noirr) and 2) using SUEWS determined flux (QF,S Noirr). No irrigation is considered in these two simulations (Noirr).
When QF,0 is assumed, the turbulent sensible heat flux QH is generally underestimated [negative mean bias error (MBE)] for the four seasons, especially during afternoon and midnight. Including QF leads to a large increase in QH. The seasonal mean diurnal QH is overestimated through the entire day in winter, spring, and autumn, while in summer daytime QH is slightly underestimated and nocturnal QH is overestimated (Fig. 5). The MBE for the four seasons are all positive when QF,S is used (Table 2). The overall performance for QH is improved in spring, summer, and autumn with root-mean-square error (RMSE) decreased (−7, −33, and −12 W m−2, respectively). The coefficient of determination R2 increases slightly in spring and summer for the QF,S case. But the RMSE and MBE increase in winter, suggesting winter QF may be overestimated.
Given the complex sampling issues, direct measurements of the storage heat flux ΔQS in an urban area with a wide range of tall 3D volumes has not been undertaken. Instead, the “observed” ΔQS is estimated as the residual of the surface energy balance (ΔQS = Q* + QF − QH − QE), therefore including considerable uncertainties. As a result, the observed ΔQS differs with QF used. If QF is assumed to be 0 W m−2, SUEWS generally performs well in winter, spring, and autumn with RMSEs of 51, 91, and 94 W m−2, respectively (Fig. 5, Table 2). The R2 is above 0.7 both in winter and spring but considerably lower (0.45) in autumn. Although the shape of the diurnal pattern including sign transition is well replicated, the nocturnal ΔQS is slightly underestimated. In summer, ΔQS is substantially overestimated during the daytime and underestimated at night (absolute values). When QF is included, a large portion goes into the storage heat flux. There is a large improvement in summertime RMSE (39 W m−2 decrease; Table 2) and smaller in spring and autumn (9 and 14 W m−2 decreases, respectively). However, wintertime RMSE deteriorates (+26 W m−2). The coefficients of determination have substantial increases for all seasons (0.09, 0.07, 0.19, and 0.18 increases, respectively). From the above analysis, it appears the simulated wintertime QF may have greater uncertainty than other seasons.
c. Latent heat flux
The underestimation of QE, especially in summer and autumn, is explored by considering the surface resistance rs (or its reciprocal, surface conductance gs) as evapotranspiration is very sensitive to it. The overall capability of water to be transported through the surface (soil, leaves, etc.) to the lower atmosphere is captured by gs. The Jarvis–Stewart model (Jarvis 1976; Stewart 1988), which has been used extensively in previous studies (e.g., Grimmond and Oke 1991; Ogink-Hendriks 1995; Matsumoto et al. 2008; Järvi et al. 2011; Ward et al. 2016), is employed in SUEWS to estimate gs. The response of stomata opening and closing, and other surface controls are included as a synergistic function of the LAI, incoming solar radiation K↓, specific humidity deficit Δq, air temperature Ta, and soil moisture deficit Δθ:
Although the specific mathematical formulations of each subfunction may differ among studies, the general form is similar. The five subfunctions are presumed to be independent from each other. Although Δq and Ta are usually highly correlated, some studies found that incorporating both functions provides better results [e.g., Khatun et al. (2011) for East Asian forests]. Moreover, the shapes of the gs dependence curves to Δq and Ta are very different (Ward et al. 2016). Low Ta often coincides with low Δq, while low Ta constrains gs but low Δq favors gs. The subfunctions range from 0 to 1, to reduce the maximum surface conductance, except for g(LAI):
where for the three vegetation types (evergreen, deciduous, grass) iυ the fraction of area fiυ, maximum conductance gmax,iυ, and the maximum LAI (LAImax,iυ ) are needed. It should be noted gmax,iυ is usually larger for irrigated than unirrigated vegetation. For radiative control, the observed maximum incoming solar radiation K↓max is used (here, set to 1200 W m−2)
For air temperature, lower TL and upper TH temperature limits are used when evaporation turns off [g(Ta) = 0]. Here TL and TH are set to a relatively wide range as −10° and 55°C, respectively:
The normalized form of g(Δθ) enables g(Δθ) = 1 when there is no soil moisture deficit and g(Δθ) = 0 when the soil moisture deficit equals to the wilting point (Δθwp, here set to 120 mm):
The hourly time series for the whole year of each subcomponent of gs [Eqs. (8)–(12)] and for gs itself [Eq. (7)] without irrigation are shown in Figs. 6 and 7. In July and August, modeled gs is smaller (<1 mm s−1) than other months, explaining the underestimation of QE. As gs (K↓), gs (Ta), and gs (Δθ) all have relatively high values from July to September, K↓, Ta, and Δθ are not key factors for the underestimation of QE. July and August 2013 gs(Δq) values are quite small as there are very large specific humidity deficits, associated with the unusually hot and dry conditions (Ao et al. 2016a). Despite this the observed QE in these months remained relatively large, probably maintained by irrigation (section 3d).
1) Irrigation scheme evaluation
In Shanghai, it is not easy to obtain accurate external water use data. The behaviors are characterized from field observations, including direct conversations with those undertaking irrigation, and a literature review. At the Shanghai Meteorological Service (150-m radius of site) irrigation is conducted throughout the year on any day of the week. It occurs most intensively in July and August when it is hot. Generally, only grass is irrigated. The mostly manual irrigation usually occurs twice on hot summer days (0600–1100 LST and 1500–1900 LST) in different areas, so by the end of the day the whole area is irrigated. In winter irrigation occurs once per day (morning or evening) every 3–4 days. Street cleaning and park irrigation are sometimes observed. Irrigation of private gardens may be extremely variable (Mitchell et al. 2001). Based on this, a diurnal profile is assumed. The modeled latent heat flux using this field-based diurnal profile when compared with an evenly distributed profile has very small differences (seasonal ΔRMSE differences < 0.5 W m−2, not shown). Variables fgrass and faut are set to 0.4 and 0, respectively. For b0,m, b1,m, and b2,m, the same values as Järvi et al. (2011) are used (Table 1), which are based on Vancouver (Grimmond and Oke 1986; Grimmond and Oke 1991). The water from these are assumed to be included in the coefficients used, as no additional detailed information is available.
Previous studies show that evaporation is very sensitive to both the amount and frequency of irrigation (Grimmond and Oke 1986; Vahmani and Hogue 2014). To test the influence of irrigation frequency on QE, the following scenarios are tested: 1) SUEWS irrigation [Eq. (3)] is used with parameter that set irrigation to 0 within 6 h of rain and with the other parameters as specified in Table 2 (hereafter QF,S Irr), and 2) as in scenario 1, but with irrigation every 3 days independent of weather conditions (Table 2; hereafter QF,S Irr 3dGap).
The monthly water use data from January 2013 to December 2013 for the entire Shanghai area (Chang et al. 2015) provided by the Shanghai Water Authority (http://www.shanghaiwater.gov.cn/) are used to evaluate the SUEWS irrigation scheme. As December 2012 water use data are unavailable, December 2013 data are used. The water use data are split between indoor and outdoor, assuming the minimum month value is the indoor water use (minimum month method; Vahmani and Hogue 2014). Differences from this minimum are considered to be outdoor water use, and indicative of irrigation and/or street cleaning, etc. Uncertainties arise from using city-level data given land use and land cover variations around this large city. Total city water use distributed to the Xuhui district (area of 54.76 km2) where the study site is located is based on the annual water use fraction. Further, it is assumed that irrigation is applied to all vegetation and half the road (street cleaning) areas. The annual water use fraction, vegetation, and road cover fractions for the Xuhui district are 0.017, 0.232, and 0.522, respectively (Shanghai Municipal Statistics Bureau 2016). The modeled monthly cumulative results from the irrigation scenarios are evaluated against the estimated outdoor water use (Fig. 8). The peak outdoor water use months occurred in July and August (around 18.5 mm month−1), with annual total (97 mm yr−1) corresponding to 9% of the annual rainfall amount. The monthly trend for the Irr 3dGap scenario matches the outdoor water use estimates relatively well, with a bit larger annual irrigation amount (140 mm). However, the scenario with only a 6-h gap after rain (Irr) results in much larger irrigation rates than suggested from the city outdoor water use (Fig. 8). Therefore, the Irr 3dGap scenario is considered more appropriate for the study area.
2) Impact of Irrigation on simulated surface energy fluxes
For the first irrigation scenario (QF,S Irr; Figs. 9, C1), the modeled QE have small differences compared to QF,S Noirr in winter (ΔRMSE = −0.3 W m−2, ΔMBE = −1.6 W m−2) and spring QE (ΔRMSE = +0.8 W m−2, ΔMBE = −5.8 W m−2; Table 2). The summer QE increased the RMSE (+4.5 W m−2) while the R2 improved (from 0.08 to 0.15) and MBE changed sign (from negative to positive; Table 2). The seasonal mean diurnal cycle shows that the daytime summer QE is largely overestimated under this irrigation condition (Fig. 9). Irrigation has a very positive impact on modeled QE in autumn (ΔRMSE = −7.7 W m−2, ΔR2 = +0.14). A sensitivity test of the coefficients b1,m and b2,m (changing from 3 and 1.1 to 2 and 2) amplifies the relative importance of the days after rain, causing a slight decrease in RMSE (not shown).
The trade-off between QH and QE is obvious as an overestimation of QE in summer leads to an underestimation of QH (Fig. 9). The summer and autumn QH have an increase in RMSE (+22.7 and +0.5 W m−2, respectively), but a slight decrease in the other two seasons. The modeled total QH and QE remain constant between the Noirr and Irr cases. As the options selected for ΔQS coefficients do not change with soil moisture, irrigation also has no impact on modeled ΔQS (not shown).
As the second irrigation scenario (QF,S Irr 3dGap) is less frequent than when irrigation is permitted almost every day except in winter (Fig. 7), the latter annual total irrigation of about 380 mm (~30% of annual rainfall) is much larger than the former (~140 mm).
From the seasonal mean diurnal cycles of observed and simulated latent heat flux (Fig. 9), the extreme overestimation of QE under the first irrigation scenario in summer is largely improved, although the diurnal peak is still overestimated with 3-day frequency. The modeled seasonal mean diurnal curves in autumn and spring agree well with the observed curves. The modeled wintertime QE changes little, as there is only a small amount of irrigation. The QE RMSE has the largest decrease in summer (−7.7 W m−2). The QH RMSE in spring, summer, and autumn also decreased (−0.3, −12.1, and −1.5 W m−2) for the second scenario, but increases slightly (<1 W m−2) in winter.
e. Comparison of surface conductance gs among scenarios
With irrigation turned on, the surface conductance has a substantial increase in July, August, September, and November (Figs. 7d–f). These four months have the least rainfall in summer and autumn (Ao et al. 2016a). When irrigation is reduced to every 3 days, the surface conductance in these months still has an obvious increase compared to the no-irrigation scenario, but less than the more frequent irrigation scenario.
where β is the Bowen ratio (β = QH/QE); s is the slope of the saturation vapor pressure curve (Pa °C−1) and is a function of the air temperature; γ is the psychrometer constant (Pa °C−1) and is determined by air pressure, temperature, and humidity; ρ is the density of air; cp is the specific heat of air at constant pressure; VPD is the vapor pressure deficit which is a function of the air temperature and relative humidity; and ga is the aerodynamic conductance which describes rate of water transport from the air above leaves to the atmosphere at a certain reference height. The term ga is calculated assuming a logarithmic wind profile and therefore is primarily influenced by the wind speed, atmospheric stability, and roughness length for momentum (and the boundary layer resistance is impacted by the roughness length for heat).
Similar to central London (Ward et al. 2016), the diurnal cycle of the observed gs at XJH fluctuates with variable patterns. The monthly median diurnal maximum gs is around 2–4 mm s−1. The relatively large gs (both daytime and nighttime) in winter months is somewhat unexpected. This may be caused by the wet winter with still active grass cover and street cleaning activities. The monthly median diurnal aerodynamic conductance ga is regularly sinusoidal in shape with relatively small monthly variations between scenarios. The monthly median ga is much larger than gs (around 12–25 mm s−1), indicating that evaporation is limited by gs rather than ga.
The modeled gs are very different between scenarios. Without irrigation (QF,S Noirr), the gs is totally constrained (near 0 mm s−1) in July and August. The modeled gs is consistent with observed gs in May, June, and October, but is largely underestimated in winter months. The modeled nocturnal gs is forced to a constant of 0.1 mm s−1, which is underestimated through the year. When irrigation is supplied continuously (QF,S Irr), the modeled gs has a substantial increase from May to December and is overestimated during daytime in July and August, which causes the overestimation of QE in summer. When the irrigation frequency is decreased to every 3 days (QF,S Irr 3dGap), the overestimation of gs is improved. The modeled gs during January–April is almost unimpacted by the three scenarios as the irrigation amount is very tiny during this period.
4. Discussion and conclusions
The performance of the urban land surface model SUEWS driven by one year of field measurements is evaluated at a central urban site (XJH) in Shanghai focusing on the estimation and impact of the anthropogenic heat flux QF and irrigation on the surface energy flux components.
SUEWS estimates QF as a function of heating and cooling degree days, and scheme coefficients are fitted by results from the inventory-based LQF model. As such, QF estimates from SUEWS are almost the same as LQF. LQF estimates building QFB, vehicle QFV, and metabolism QFM components based on city-level hourly electricity consumption data, air temperature, and population density. A new building heat emission–air temperature response function using two balance points made the seasonal variation of the building QFB more distinct.
The diurnal patterns of QFB for weekday, weekend, and holidays derived using local electricity data are similar with a peak around 1100 LST. On holidays there is a larger evening peak around 1900 LST. A weekday diurnal profile of QFV derived from local traffic data has two peaks associated with rush hours. The morning peak is more distinct (at 0800 LST in all seasons) than the evening peak (at 1700 LST). Weekends have no distinct peaks. The largest QF (estimated by LQF) is in summer with seasonal mean daily peak around 236 W m−2. Winter and autumn have similar mean daily peaks (~190 W m−2), and spring is the smallest (~180 W m−2). Building heat emission is the largest subcomponent (~95% of the total QF).
The impact of QF on surface energy fluxes is explored with (SUEWS, QF,S) and without QF (QF,0). Ignoring QF, the seasonal diurnal pattern of sensible heat flux QH is reproduced well generally, but the magnitude of QH is underestimated for all seasons. When QF,S is used, the seasonal mean diurnal QH is overestimated throughout the day in winter, spring, and autumn. In summer, daytime (nighttime) QH is slightly underestimated (overestimated). Overall performance for QH is improved in spring, summer, and autumn (RMSE decreased), but not in winter. For QF,0, SUEWS summer daytime (nighttime) storage heat flux ΔQS is overestimated (underestimated) whereas QF,S is improved (RMSE decreases by 39 W m−2). Spring and autumn have improvements (RMSE decreases of 9 and 14 W m−2, respectively). But winter does not (RMSE increase of 26 W m−2). This indicates winter QF may be overestimated.
Underestimation of QE is associated with underestimation of the surface conductance gs in summer, mainly caused by large specific humidity deficits. External water supply may maintain evaporation rates. Having the appropriate seasonal cycle of the LAI in winter and spring improve the QE model performance. Irrigation amount and frequency have a large impact on QE. Seasonal mean summer daytime QE is largely overestimated if continuous irrigation is permitted, indicating an overestimation of irrigation. In autumn irrigation improves QE (RMSE decreased, R2 increased). Overestimation of QE with too frequent irrigation in summer is improved when reduced to every 3 days (RMSE decreased), and slightly improved in spring. Reducing irrigation frequency to 3 days also improves summer QH (RMSE decreased).
This study emphasizes the importance of appropriately estimating the anthropogenic heat flux and external water use in dry and hot seasons in urban land surface models. Previous studies have evaluated SUEWS at two sites (urban and suburban) in the same city or nearby cities with contrasting surface characteristics (Karsisto et al. 2016; Ward et al. 2016). Results suggest that the surface cover, especially the vegetated versus impervious proportion along with the anthropogenic heat emission, has the largest impact factor on model performances. The magnitude of QF may be substantially smaller at suburban sites because of much lower population densities. The difference of building heights at urban and suburban sites will also influence where QF is released into the atmosphere. Larger vegetated fractions in suburban areas may also have more intensive irrigation. Therefore, future work is inevitably needed to compare simulation results of this central urban site with suburban sites in Shanghai to improve understanding of potential sources of model biases.
Future SUEWS evaluation should consider seasonal variability in the OHM coefficients for the simulation of the storage heat flux ΔQS. Ward et al. (2016) found adjusting the OHM coefficients for a specific site can significantly improve model performance both for ΔQS and for other terms, most notably QH as the residual term. Seasonal variations of surface properties such as albedo, Bowen ratio, wind speed, and soil moisture (Arnfield and Grimmond 1998; Sun et al. 2017) have critical impacts on ΔQS. Adjusting OHM coefficients should analyze more observations and use the recently developed analytical objective hysteresis model (AnOHM; Sun et al. 2017) to determine a wider range of parameters.
This work was supported by the National Natural Science Foundation of China (Grants 41775019, 41675008), Newton Fund/Met Office Climate Science for Service Partnership (CSSP) China (SG), the Project of Science and Technology Commission of Shanghai Municipality (Grant 17DZ1205300), the Project of Scientific and Technological Development of the Shanghai Meteorological Service (Grant MS201803), and the fund from China Scholarship Council (CSC). We thank everyone who contributed to instrument maintenance, data collection, and model development.
Anthropogenic Heat Flux
LQF (Gabey et al. 2018) is a new implementation of the Large Scale Urban Consumption of Energy (LUCY) model (Allen et al. 2011; Lindberg et al. 2013). LQF (short for LUCY QF) is embedded in the Urban Multiscale Environmental Predictor (UMEP), which is an open-source, city-based climate service tool that combines models and tools for climate simulations (Lindberg et al. 2018). LQF takes a “top down” approach, using publicly available annual energy consumption data for a large area (e.g., country, province, city) with high-resolution population density data to distribute the energy consumption across the area of interest. It separately considers three emission sources: buildings QFB, traffic QFV, and metabolism QFM (Grimmond 1992; Sailor 2011). Like the SUEWS method, the daily totals are a function of temperatures (temperature response function) and subdaily patterns are based on diurnal use profiles.
a. LQF temperature response function
Variations of energy consumption with air temperature can be modeled with a single balance point temperature Tb [Eq. (2)] obtained from when the energy consumption is lowest. Consequently, Tb varies with climate (Amato et al. 2005) and/or with building type. Therefore, it is preferable to have the appropriate local Tb as it has a large impact on seasonal variations of the building anthropogenic heat emissions. This approach is used in QF,S (SUEWS) and was originally used in LQF (QF,L). In this work we introduced a new LQF temperature response function (Fig. 2) with two balance point temperatures, that is, threshold temperatures when heating Th and cooling Tc commence.
To quantify Tb for Shanghai, the whole city electricity consumption (Liu and Cao 2013) and XJH air temperature data are analyzed (Fig. 2b). Ideally, all sources of energy would be analyzed, but electricity consumption data are often used as a proxy for building heat release. For example, Kikegawa et al. (2014) estimate that in Tokyo ~80% of office building energy demand in summer is from electricity consumption. The buildings in the XJH area are predominately office and residential buildings, with very little industry. Shanghai Municipal Statistics Bureau (2016) indicates industry around the dense urban XJH site (finu) accounts for about 10% of total energy consumption. In contrast, suburban districts such as Jiading have a much higher industry fraction.
Given the similar climatic regime, building energy consumption at our study site is assumed to be like Tokyo. However, Shanghai’s electricity consumption is only 14% of the total energy consumption (Table A1). As industry consumption is relatively insensitive to weather conditions (Sailor 2011), it has a different profile to the commercial study site of interest, with reduced daily amplitude and seasonal variations. Given the difficulty of accessing details of industrial consumption patterns, for simplicity the industrial load is assumed to be uniform through the day (Sailor 2011).
The relation between energy consumption and air temperature is analyzed. In Shanghai, the energy consumption rises almost linearly when the daily mean air temperature is warmer than 21°C or cooler than 15°C, providing evidence that cooling or heating systems are operating in these temperature ranges (Fig. 2b). In the “comfortable” range (15°–21°C) energy consumption stays nearly constant. The energy consumption for cooling increases more rapidly than for heating as central heating systems are absent south of an east–west (Qin-Huai) line near 33°N (Makinen 2014; Shi et al. 2016). Given this in SUEWS, a single Tb of 20°C is used.
In LQF the three components of QF (building, transport, metabolism) are treated separately. A temperature response function fb modifies the base daily building energy consumption EB,b (kWh day−1 per capita):
where ρpop is the population density of XJH site (Npop = 261.62 capita ha−1). Here the original LQF one balance point temperature function (Allen et al. 2011; Lindberg et al. 2013; Fig. 2b, dotted line) is modified to allow minimum energy consumption to occur over a range of temperatures. Threshold temperatures when heating Th and cooling Tc) commence, and when saturation energy use occurs as additional energy consumption is minimal (Fig. 2a, solid line); that is, <Tmin no additional heating occurs (>Tmax for cooling). Therefore, logic variables (lc1, lh1) are 0 except when the air temperature is with a critical range (Fig. 2b):
Or more completely,
with two logic variables (lc2, lh2) which are nonzero when
The three nondimensional coefficients bb, bc, and bh are the base fraction of energy use when small changes in daily air temperature have no impact. The comfortable air temperature ranges (i.e., between Th and Tc) for bb and then once all heating bh or cooling bc are on. Variables Ac and Ah are the building energy consumption thermal response slopes for cooling (heating) when daily air temperature warmer (colder) than Tc (Th).
To determine the parameter values for Shanghai, the 2005–09 city-wide electricity consumption data (Liu and Cao 2013) are used (Fig. 2b). The resulting parameters are bb = 0.88, Ac = 0.04°C−1, and Ah = 0.01°C−1. In this subtropical city, the larger cooling coefficient Ac reflects the absence of a centralized heating system in Shanghai but extensive use of air conditioning in summer.
b. Vehicle based heat emissions
The heat released by motor vehicles QFV from combustion of petrol or diesel fuel (Sailor 2011) generally does not have seasonal variations (Sailor and Lu 2004). In LQF, QFV is calculated as a function of vehicle numbers (cars = 89, motorcycles = 17.2, freight vehicles = 8.8 per 1000 capita; Shanghai Municipal Statistics Bureau 2016), traffic speed, and time (days, hour). The fuel type is assumed to be petrol (Zhao 2007). The mean vehicle speed is set to 48 km h−1 (Su et al. 2014).
c. Population density
In LQF, the default population is from the Gridded Population of the World, version 4 (GPWv4; CIESIN 2017), with estimates for 2005, 2010, and 2015. The 2010 30-arc-s (~1 km) population density around the XJH site is about 261.62 capita ha−1 (CIESIN 2017), whereas the statistics in 2013 (Shanghai Municipal Statistics Bureau of Xuhui District 2013) for the XJH site neighborhood (4.07 km2) have a permanent resident population of 92 764 (i.e., 227.92 capita ha−1). Here the population density of 261.62 capita ha−1 is used as the resolution is closer to the source area of XJH site.
Urban population density varies significantly through the course of a day and from working days to nonworking days (Gabey et al. 2018), particularly in areas such as XJH with a mix of permanent residents, shoppers, tourists, hospital visitors and patients, etc. who come and go. The publicly available data from the national census of Shanghai do not capture these dynamics. Yu and Wen (2016) estimate daytime and nighttime population for the Jing’an district using land use and population age structure data. They suggest the daytime population is 39.2% higher than at night for the district as a whole (and 147.7% higher in the busiest subdistrict). W. Zhong et al.’s (2017) analysis of cell phone signals found from 0600 to 1000 LST people move into the center of Shanghai from outer areas, with a peak at 1000 LST that is sustained until 1800 LST, when the return to suburban areas occurs. The day-to-night population density ratio in central Shanghai is about 1.5 (W. Zhong et al. 2017, their Fig. 7). Based on these two studies, the daytime (1000–1800 LST) population is assumed to be 1.5 times the nocturnal population at XJH site, with the periods 0600–1000 and 1800–2200 LST being transition periods when the population moves between work, leisure, and residential sites [a linear increase (decrease) is assumed, following Sailor and Lu (2004)]. Using this dynamic ratio, and assuming the daytime and nocturnal population for the whole Shanghai is roughly conserved, the diurnal variation of the electricity consumption is further scaled.
Statistical Evaluation Techniques
Common statistical metrics are used to assess model performance. The modeled Mi and observed Oi values are used with their corresponding mean values to calculate the coefficient of determination R2,
the root-mean-square error (RMSE),
mean bias error (MBE),
and the mean absolute error (MAE),
The RMSE, MBE, and MAE all have units of the variable analyzed and an ideal value of 0; whereas R2 varies between 0 and 1 with an ideal value of 1.