China’s adaptive response to climate change through air-conditioning

Summary

Studies have shown that the soaring demand for air conditioners in recent years is closely related to the worsening global warming; however, little evidence has been provided for China. This study uses weekly data of 343 Chinese cities to investigate how air conditioner sales respond to climate variability. We detected a U-shaped relationship between air-conditioning and temperature. An additional day with average temperature above 30°C increases weekly sales by 16.2%. Heterogeneity analysis shows that the adoption of air-conditioning is different for south and north China. By combining our estimates with shared socioeconomic pathway scenarios, we project China’s mid-century air conditioner sales and the resulting electricity demand. Under the fossil-fueled development scenario, air conditioner sales in the Pearl River Delta would rise by 71% (65.7%–87.6%) in summer. On average, the per capita electricity demand for air-conditioning will surge by 28% (23.2%–35.4%) in China by mid-century.

Graphical abstract

Highlights

• •We find a U-shaped relationship between air-conditioning and temperature in China
• •The pattern of air-conditioning adoption is different for south and north China
• •Air conditioner sales in the Pearl River Delta would increase by 71% in summer
• •Per capita electricity demand for air-conditioning surges by 28% by mid-century

Global change; Environmental policy; Urban planning

Introduction

The latest data shows that the increase in the average global surface temperature has exceeded 1.1°C, compared to the pre-industrial level.¹ In the absence of stringent greenhouse gas control efforts, global temperatures could rise by 3–6°C by the end of this century.² Global warming has a significantly negative impact on ecology, economy, and society, with potentially severer impacts in developing countries that are more vulnerable to climate change.^3,4,5 Taking China as an example, high temperatures greatly reduce labor and capital productivity, and this effect is particularly notable in the manufacturing industry, resulting in a corresponding output drop of 12%.⁶ Extreme weather, such as heat waves, may exacerbate the local air pollution situation, which in turn will increase household energy demand.⁷ Adaptation is an effective means of mitigating climate damage.^8,9 Under the optimal adaptation strategy, adaptation can attenuate climate loss by an average of 28%.¹⁰

Active air-conditioning, which is one of the most common adaptation or coping strategies to a warming world, is one way to reduce indoor damage resulting from temperature increase. In recent years, the peak load of power demand has been closely related to more air conditioner (AC) ownership.¹¹ For instance, AC sales in China rose from 51.5 million units in 2010 to nearly 100 million units in 2020,¹² which doubled the average number of ACs per household, from 0.63 units to 1.2 units. Correspondingly, the proportion of China’s AC stock in the world rose from 27% in 2010 to 37% in 2020.¹³ Meanwhile, the energy demand for space cooling is growing at an average annual rate of 13%. For instance, in Beijing, ACs contribute to more than 50% of the peak residential power load in summer.¹³ Therefore, it is of great importance to understand how climate change will affect AC adoption and the associated energy demand, especially for developing countries like China.

The current research on adaptation to climate change through air-conditioning can be divided into two different aspects, namely the intensive margin and the extensive margin.^14,15 Intensive margin means that consumers can respond to extreme temperature by making more use of existing air-conditioning appliances and thus increasing energy consumption. Extensive margin means that consumers buy new air-conditioning appliances to cope with climate change. The existing literature on the intensive margin typically finds that high temperature leads to more electricity consumption or AC usage time,^{14,16,17,18,19} despite some differences in the frequency of data that are used (hourly, monthly, or annual) and in the magnitude of the impacts found in different regions.^20,21 For instance, Deschênes and Greenstone (2011)¹⁶ used a state-level panel of annual energy consumption data and daily weather data over several decades to explore the relationship between temperature and residential energy consumption and found that one extra day where the temperature exceeds 90°F (relatively to a day in the 50–60°F range) is associated with a 0.4 percent increase in annual energy consumption. Based on the daily electricity consumption data for 1.8 million metered residents in Pudong, Shanghai, over a two-year period, Li et al. (2019)¹⁸ estimated the impact of climate change on residential electricity consumption and found that a 32°C day (compared with a 20°C day) would lead to a 170% increase in daily electricity consumption.

The research on the extensive margin mostly focused on how AC ownership responds to climate change, which is generally based on annual household survey data and concentrated on industrialized countries.^22,23,24 The annual data of AC ownership lacks the information on the purchase date of the ACs, which restricts the possibility of exploring how the extensive margin of climate change may differ in different time of the year (summer vs. winter). This is one of the knowledge gaps this study aims to fill in.

Furthermore, the literature on the extensive margin of climate change in developing economy is relatively thin due to data constraints. As a matter of fact, there exists significant inequality in the usage and adoption of air-conditioning between rich and poor countries due to their differences in affordability.²⁵ Generally, developing regions could encounter higher climate vulnerability but have lower AC penetration rates, and thus more attention should be paid to these regions.¹⁷ Two typical studies^17,26 investigated the air-conditioning and adaptation cooling deficit in Brazil, India, Indonesia, and Mexico, by employing cross-sectional survey data. However, a drawback of the cross-sectional approach is that one cannot econometrically control for unobservable differences across households or regions, which may be correlated with weather/climate and lead to biased estimation¹⁵; the unobserved elements, such as culture and institutions, can always bias cross-sectional estimates.²⁶ As a result, the studies on the extensive margin impacts based on panel data estimates are highly desirable, especially for developing countries. This is another knowledge gap this study aims to close.

In order to fill in the aforementioned knowledge gaps, this paper attempts to examine the extensive margin impacts of climate change by constructing a unique panel data of weekly AC sales in 343 cities across China. Specifically, this study contributes to the extant literature in at least three aspects. First, this study provides new insights into the extensive margin impacts with estimations and projections regarding the AC adoption across typical seasons of the year, which is underexplored in the literature due to the annual survey data of AC ownership that is commonly used. This is important since ACs may be used for different purposes in summer/winter, which cannot be characterized by annual survey data without knowing the adoption dates of ACs. Second, there would be a non-linear (U-shaped) relationship between climate change and AC adoption if more AC adoption occurs on hot or cold days (compared to the mild ones). The high-frequency (weekly) AC data, which can be better matched with the daily weather data, enables us to employ a non-linear and nonparametric approach (with temperature bins) to validate the non-linear (U-shaped) impacts of climate change on AC adoption, in contrast to most existing studies on extensive margin impacts where the cooling degree days (continuous variable) are used to proxy the temperature and thus not able to characterize the potential non-linear impacts. Third, this study uses panel data with weekly city AC sales where the unobservable elements can be better dealt with to get reliable estimates, whereas most studies on the extensive margin impacts rely on cross-sectional or repeated cross-sectional data.

Our results show a statistically significant U-shaped relationship between temperature and air-conditioning adoption, with an additional day in the 25–30°C and >30°C bins associated with 9% and 16.2% increases in weekly sales, respectively. Furthermore, we find a larger increase in AC demand when the temperature is below 10°C in southern China than in the north. This could be explained largely by the offsetting effect of the Huai River heating policy.

By combining the estimates with climate forecasts under typical shared socioeconomic pathway (SSP) scenarios, we project a pronounced increase in air-conditioning adoption. Specifically, AC sales in the Pearl River Delta would rise by up to 71% in summer and 31% in winter under the SSP5 scenario, compared with the current level. On average, the per capita electricity demand for air-conditioning in China will surge by approximately 28% under the SSP5 scenario by the middle of this century.

Results

Extensive margin of air conditioner adoption

We examine the possible relationship between climate change and AC adoption. To this end, we built a unique data panel by combining (1) weekly AC sales for 343 cities at the prefecture level from 2008 to 2019 with (2) station-level weather data drawn from the National Meteorological Information Center of China (more details are provided in the supplemental information).

Changes in air-conditioning and temperature

The proposed model was estimated by using weekly AC sales and station-level weather data (details are included in the method details section). As shown in Figure 1, the background color of the map displays the geographical coverage of AC sales, and different shades of color indicate average weekly AC sales. It can be seen that the AC sales vary significantly across different cities, and Chongqing, Beijing, Shanghai, Xi’an, Chengdu, and Guangzhou are among the cities with the highest sales. Chongqing had the highest average weekly AC sales. Compared with cities in southern China, cities in the north had relatively lower weekly sales.

Figure 1.

Weather monitoring stations and geographic coverage of air-conditioning

The purple background color indicates cities with AC data. Different shades of purple indicate the average weekly AC sales in different cities in 2019. The yellow dots reflect the distribution of meteorological stations. The green line depicts the Huai River-Qin Mountains line, dividing northern and southern China.

Figure 1 also illustrates the geographic distribution of the weather stations, and the resulting temperature distribution is shown in Figure 2 (or Figure S1). We observed a significant variation in the daily average temperature from 2008 to 2019, given that China’s territory spans five temperature zones from southern to northern China (the northernmost region is the cold temperate zone, and the southernmost region has a tropical climate). Each bin describes the number of days in a year that meet the daily average temperature in the corresponding bin. Residents in China experience (daily average) temperatures below 10°C for approximately 120 days in a year. As more than half of the area in China is in the temperate zone, it is unsurprising that the largest number of days fall in the <10°C temperature bin. Meanwhile, only 7 days in a year on average are in the above 30°C bin; however, this figure is expected to grow with the warming climate. Additional summary statistics are presented in Table S1.

Figure 2.

Effect of temperature on air conditioner adoption

On the y axis on the left side, we plot the natural log of AC sales. The area shaded light blue is a 95% confidence interval. On the y axis on the right side, we plot the frequency of each temperature bin. We add up all days for each city from 2008 to 2019 to the corresponding temperature intervals and calculate the frequency of each interval (See the full results in Table S2).

Main estimates

Based on the high-frequency AC sales panel data, we can investigate the nonparametric relationship between temperature and AC sales using the high-dimensional fixed effects panel model (details are included in the method details section). The estimated effect of climate change on AC sales is plotted in Figure 2 (detailed estimates are provided in Table S2). We used the ${\beta }_{1m}$ estimates and 95% confidence intervals to plot the upper half of the figure, where the parameter ${\beta }_{1m}$ measures the percentage change in weekly AC sales associated with one additional day in the $m$th temperature bin relative to the reference temperature bin [15°C, 20°C) (the reference bin is omitted in the regression to avoid perfect collinearity).

The results show that there is a non-linear (U-shaped) relationship between AC sales and temperature, with more ACs being sold for high- or low-temperature bins (compared with mild-temperature bins). When the temperature was lower than 10°C, we observed a moderate increase in AC sales. This is because that the vast majority of ACs sold in China can also be used for heating (given that they are generally equipped with electric auxiliary heating or heat pump function) and people buy them for heating purpose to keep indoors warm for low temperatures, especially for people in southern China, where there is no central heating. Meanwhile, it is observed that AC sales rise sharply for the 25–30°C and above 30°C temperature bins. To be more specific, each additional day at 25–30°C and above 30°C temperature bins would lead to an increase of 9% and 16.2%, respectively, in weekly sales, as compared to the reference bin [15°C, 20°C).

Given the important role of household income in purchasing decisions, the aforementioned results reveal a positive effect (though insignificant) of per capita disposable income on AC sales, which is consistent with existing findings¹⁷ on the role of income in AC adoption. Precipitation and humidity have negative and positive effects, respectively, on air-conditioning, albeit small in magnitude. Substantial rainfall helps to lower the temperature, leading to a decline in people’s demand for cooling, whereas higher relative humidity may result in a higher apparent temperature, which may increase the cooling demand to some extent.

Heterogeneity in climate response

Due to differences in geographical location and socioeconomic characteristics, there are substantial discrepancies in how temperature affects air-conditioning adoption. Based on this, we conduct several dimensions of heterogeneity analysis in terms of geographical locations, sizes of households, levels of disposable income per capita, and AC types and brands to see the possible impact on the relationship between temperature and AC sales.

In 1992, the Chinese government introduced a residential heating policy in which local governments offered central heating systems to city dwellers. The provision of central heating applies to areas north of a line from the Qinling Mountains to the Huai River (the Huai River-Qin Mountains line, depicted in Figure 1). Thus, cities located north of the Huai River-Qin Mountains line are provided with central heating systems, which are an efficient solution for heating, while cities south of this line have to rely heavily on individual heating appliances, such as ACs, to keep indoors warm. In addition, most buildings in the north have double-glazed windows to keep heat from escaping, in contrast to single-glass windows that are common for the buildings in the south.²¹ These building characteristics also make indoor heating in the south less efficient.

This difference in central heating policy for northern and southern China is clearly visible in the different effects of climate change on air-conditioning, as shown in Figure 3A. While the north and south basically maintain a U-shaped relationship between AC sales and average temperature, the coefficients are significantly higher in southern cities than those in northern cities for temperature bins with average temperatures above 25°C, indicating that the south is more inclined to use ACs to cope with extreme heat (detailed estimates are given in Table S3 Columns 1 and 2). However, the coefficients of the lower-temperature bins indicate a divergence in air-conditioning demand between northern and southern China. More specifically, when the temperature is lower than 10°C, cold weather increases the air-conditioning demand for heating (most ACs sold in China have a heating function in addition to the primary cooling function), and this effect is more profound for southern China. This is largely explained by the availability of central heating in northern China.²⁷

Figure 3.

Heterogeneity in the temperature-air-conditioning adoption relationship

(A) Air-conditioning adoption responses to climate change were found to differ by geographical location.

(B) The heterogeneous effects of household size on the climate-air-conditioning adoption relationship. The gray and blue shaded areas represent 95% confidence intervals.

We then examine the heterogeneous impact of household features, such as size and family membership. Extreme weather has sizable impacts on the well-being and state of health of the household members, and ACs have proven to be an effective adaptation strategy for mitigating possible detrimental effects. Compared with other family members, children and the elderly are more susceptible to the adverse effects of extreme weather owing to their physical conditions.²⁷ We use the average household size in a city to create subsamples to examine whether this holds true for air-conditioning adoption. The number of households and population data used to calculate the average household size in a city are taken from the China City Statistical Yearbook. We divided the dataset into two subsamples: cities below and above the median (average) household size. As shown in Figure 3B, compared with cities whose average household size is below the median value, those with above-median size are more sensitive to extreme weather, with greater increases in AC demand when the average temperature is less than 10°C or more than 30°C (see Table S3 Columns 3 and 4).

Furthermore, we explored how the temperature-cooling response could be affected by diverse income levels (proxied by average disposable income per capita) and types of ACs (i.e., wall-mounted type or floor-standing type), as well as brand attributes, which have not been examined in the literature.²⁸ The details are presented in Tables S4 and S5. Given the important role of heat waves,²⁸ we also examined their effect on the sales of ACs. The results revealed that compared to days without prolonged heat waves, AC sales increased by 0.8% on the first day of a heat wave, 1.6% on the second day, and so on. The time-lagged effect of temperature on air-conditioning purchases was also statistically confirmed, particularly on hot summer days (see Tables S6 and S7 for detailed estimates and analysis).

Climate-air-conditioning response forecasts

In this section, we make medium-term forecasts of AC sales and the resulting electricity demand by combining our estimates of the temperature-response curve with three typical climate change scenarios (SSPs), SSP1, SSP2, and SSP5, formally defined in the assessment report (AR5) of the Intergovernmental Panel on Climate Change (IPCC). More specifically, SSP1 takes the green road, with low challenges to mitigation and adaptation, SSP2 is a path with medium challenges to mitigation and adaptation, while SSP5 represents a continued drastic development of fossil fuel resources and adoption of resource and energy-intensive lifestyles worldwide.²⁹

The predicted data of climate change, including the daily near-surface temperature for each grid point from 2015 to 2100, is obtained from the NEX-GDDP-CMIP6 dataset of the NASA Center for Climate Simulation in the United States. The NEX-GDDP-CMIP6 dataset contains bias-corrected global downscaled climate projections under SSPs, which are derived from the General Circulation Model (GCM) runs conducted under the Coupled Model Intercomparison Project Phase 6 (CMIP6).³⁰ The grid points are separated into 0.25° latitude and 0.25° longitude.

Climatic models may vary on the amount and direction of future changes in climate variables, and point estimates generated from only one climate model’s predicted data may be deceptive.³¹ Although more than 20 such climate prediction models exist in climate science, studies have relied primarily on several models.^32,33,34 Therefore, we utilize six climate models (CanESM5, CESM2, CMCC-ESM2, CNRM-ESM2-1, GFDL-ESM4, and IPSL-CM6A-LR) to represent the uncertainty of climate model predictions and display projections for AC sales and per capita annual power demand of AC use, based on the arithmetic mean of the results across different climate models. We also discuss the impact of climate model uncertainty in the last part of this section.

Climate change prediction

There is a clear increase in days in the 25–30°C and higher than 30°C temperature bins, under all scenarios (see method details for detailed calculations). Specifically, the medium-term forecast indicates that the number of days with an average temperature higher than 25°C under the SSP1 scenario increases to approximately 23 days. The increments under the SSP2 scenarios are slightly greater than those under the SSP1 scenarios. For the SSP5 scenario, the number of days on which the average temperature is above 25°C increases to 32 by mid-century (Figure 4).

Figure 4.

Distribution of changes in daily average temperature

This figure displays the distribution of changes in mid-term temperature distribution based on the historical temperature distribution (the average temperature in 2008–2019) under the SSP1, SSP2, and SSP5 scenarios.

Air-conditioning forecasts

We discuss the changes in AC sales during summer and winter separately (see Method details for detailed calculations). The average temperature in summer is generally the highest throughout the year; thus, ACs are adopted to meet the increasing cooling demand. The forecasts demonstrate that as the temperature rises and income grows, the future demand for air-conditioning in summer will increase in most cities. The results of the different scenarios show that the air-conditioning demand under the SSP1 scenario is slightly higher than that under the SSP2 scenario in 2040–2059 (Figures 5A and 5C). Because of the hotter weather and larger income growth under the SSP5 scenario, AC sales in most cities are much higher under this scenario than under the former two scenarios (Figure 5E). Furthermore, AC sales increase more in most southern cities than in northern cities under all scenarios. For instance, the increase in AC sales (by mid-century) in the Beijing-Tianjin-Hebei region is approximately 37%, 35%, and 54% under the SSP1, SSP2, and SSP5 scenarios and about 50%, 46%, and 71% in the Pearl River Delta, respectively.

Figure 5.

Projected air-conditioning response to climate change

The figures display the average change in weekly AC sales from 2040 to 2059 considering both climate change and income growth. Specifically, in (A), (C), and (E), we plot the rate of change in average weekly ACs sales in the summer of mid-century (2040-2059) under the SSP1, SSP2, and SSP5 scenarios, compared with the current sales level, and (B), (D), and (F) show the corresponding forecasts for winter.

In winter, ACs are primarily used for heating. In contrast to the estimates based only on the marginal impact of climate change, where we freeze the income at the current level (Figure S2), the demand for air-conditioning in most cities under the SSP1 and SSP2 scenarios will still rise modestly due to income growth (Figures 5B and 5D). Under the SSP5 scenario, the demand for ACs increases more in both the south and north, with AC sales in the Beijing-Tianjin-Hebei region and the Pearl River Delta rising by 34% and 31%, respectively. It is worth noting that with either the marginal impact of climate change or the combined impact of climate change and income growth, AC sales in the Pearl River Delta and Hainan Province would increase, especially under the SSP5 scenario.

Electricity demand forecasts

The increased use of ACs under global warming will increase peak power demand,³⁵ which in turn poses challenges to decarbonization goals, such as carbon neutrality. By calculating the mid-century AC holdings and multiplying them by the corresponding operating hours (see method details) and the average AC power, we obtained the mid-century per capita electricity demand for air-conditioning (see method details). According to the International Energy Agency (IEA) (2018),³⁶ the average energy efficiency of the stock of ACs worldwide could more than double between the present day and 2050. To account for energy efficiency improvements driven by technological changes, we therefore assume that the average power demand per unit time for air-conditioning will be reduced by half by mid-century.

From 2008 to 2019, the per capita electricity consumption for air-conditioning use in most southern cities was higher than the median of all cities. To be specific, Shaoxing has the largest electricity consumption of ACs, reaching 2213.3 kWh per capita, and people in Leshan only use approximately 469.8 kWh of electricity. The situation changes when we move to the north of China, and the per capita electricity consumption for air-conditioning use is lower than the median value. In the mid-term, there are minor differences in the air-conditioning-related power demand between the SSP1 and SSP2 scenarios. Most cities would consume more electricity than the current level, especially Chengdu, Nanchang, Hefei, Nanjing, and Beijing (Figures 6A and 6B). The diversity in air-conditioning power consumption across cities expands under the SSP5 scenario. On average, the per capita electricity demand for air-conditioning will surge by approximately 28% under the SSP5 scenario by the middle of this century.

Figure 6.

Projected per capita annual power demand of air-conditioning use

The internal scatter charts show the per capita annual electricity demand resulting from the use of air-conditioning from 2008 to 2019, where the dotted line represents the median (1289.1 kWh) of all cities. Blue circles indicate that the per capita electricity consumption of the corresponding cities is below the median, while red circles indicate that the values are above the median. The outer circular heatmaps show the rate of change in per capita annual electricity consumption under different scenarios by mid-century, compared with the current levels. From the inside to the outside, the circular heatmaps represent the results under the SSP1, SSP2, and SSP5 scenarios, respectively.

(A) and (B) plot the rate of change in per capita electricity demand by mid-century (2040–2059) in southern cities and northern cities, compared with the current level.

Implications of climate uncertainty

Projections under six climate models consistently show that seasonal AC sales and annual power demand in the mid-term are higher than the current level. All models show that, during the summer, weekly AC sales growth rate in the Pearl River Delta is larger than that in the Beijing-Tianjin-Hebei region under the corresponding scenario (Figure 7A). The reverse is true in winter (Figure 7B). Considering both climate change and income growth, the per capita electricity demand for air-conditioning will increase significantly cross scenarios, with a range of [23.2%, 35.4%] under the SSP5 scenario (Figure 7C). Taking only the marginal effect of climate change into account, six models consistently conclude that electricity demand under the SSP5 scenario is positive, with a range of [6.1%, 16.5%] (Figure S6C).

Figure 7.

Projected AC sales and annual power demand of AC use under six climate models

The six climate models are CanESM5, CESM2, CMCC-ESM2, CNRM-ESM2-1, GFDL-ESM4, and IPSL-CM6A-LR. All subfigures show the rate of change compared to the current level.

(A) and (B) display the average change in weekly AC sales in summer and in winter from 2040 to 2059, considering both climate change and income growth. The upper half of (A) and (B) represents the average results for the Beijing-Tianjin-Hebei region, while the bottom half depicts results of the Pearl River Delta.

(C) shows the change in per capita annual power demand of air-conditioning use across the country by mid-century.

In addition to consistent findings, differences exist among the projections based on six climate models. Because CanESM5 model predicts a higher temperature rise compared to other models (Figure S5), AC sales in summer in this model are almost the greatest (except for the results in Beijing-Tianjin-Hebei region under the SSP2 scenario) (Figure 7A). For example, under the SSP5 scenario, per capita annual power demand of air-conditioning use under the CanESM5 model and the IPSL-CM6A-LR model increases by over 30%, much higher than the results under the remaining four models (Figure 7C).

Discussion and conclusion

Concluding remarks

Estimating the impact of climate change on air-conditioning is of great importance for energy load management, carbon mitigation, and long-term adaptation strategies.³⁷ This is especially true for developing regions where climate vulnerability could be higher but the air conditioner penetration rate is lower. Therefore, in this study, we construct unique panel data of weekly AC sales in 343 cities across China to investigate China’s adaptive response to climate change through air-conditioning. As mentioned in the Introduction, this study contributes to the existing studies on the extensive margin effect of climate change in several aspects, from providing new insights on AC adoption across typical seasons of the year to taking nonparametric approach to validate the non-linear impacts of climate change on AC adoption with panel data. We find a statistically significant U-shaped relationship between temperature and air-conditioning adoption, which is robust across various specifications. Specifically, each additional day in the 25–30°C and >30°C bins was associated with a 9% and 16.2% increase in weekly sales, respectively.

Children and the elderly are generally less tolerant of extreme weather; consequently, the estimates indicate a greater increase in AC demand in households with a polarized population structure when the temperature is less than 10°C or greater than 30°C. Household income growth is positively related to high saturation and ownership of ACs.^17,38 Consistent with existing studies, our results also show that cities with lower per capita disposable income generate an increasing demand for air-conditioning on hot days.

By combining the main estimates with SSP scenarios running up to the middle of this century, we predict a pronounced increase in air-conditioning adoption with global warming and income growth, particularly in southern cities and in summer. Under the fossil-fueled development scenario (i.e., the SSP5 scenario), AC sales in most cities would increase, with the increase in AC sales (by mid-century) in the Beijing-Tianjin-Hebei region and the Pearl River Delta reaching 54% (43.2%–65.6%) and 71% (65.7%–87.6%) in summer and 34% (33.9%–34.7%) and 31% (29.3%–31.9%) in winter, respectively, compared with the current level. By considering both extensive and intensive margin effects, we project the city-level per capita electricity demand for AC use by mid-century. On average, the per capita electricity demand for air-conditioning in China will surge by approximately 28% (23.2%–35.4%) under the SSP5 scenario by the middle of this century.

Policy insights and discussions

Our findings have critical policy implications. First, although a massive increase in the sales of ACs would provide some primitive adaptation to a warmer climate, this may enhance the need for energy saving and emission control. For instance, the manufacture of ACs consumes copper, steel, aluminum, and other metals, as well as a great deal of energy, such as electricity and petroleum, all of which result in an increase in carbon emissions. Besides, the increase in AC use would drive up peak power demand and pose great challenges to the grid load, while the upgrading of the power grid will lead to an increase in carbon emissions associated with material production. Furthermore, the rising electricity demand due to more ACs would directly increase carbon emissions, particularly when the electricity grid is not green.³⁹ Therefore, the government needs to formulate higher energy efficiency standards for ACs, encourage the adoption of super-efficient technologies and building- or city block-wide AC facilities, and promote the recycling and reuse of air-conditioning components and materials.

Second, the rise in temperature and popularity of air-conditioning will reshape the future lifestyles of residents. For example, in hot summer months, families will shift outdoor activities indoors, watch TV, or use electronic products in air-conditioned rooms. This will further increase the power demand. Differentiated subsidies are, therefore, needed to purchase energy-efficient ACs by considering regional climate and income differences. Third, as a typical alternative option for air-conditioning, green buildings have proven to be effective in reducing peak power demand and carbon emissions,^40,41 given increasing global warming. The government can clearly formulate green building standards and provide tax reductions and R&D subsidies for companies in developing green building technologies.⁴² Meanwhile, emergency centers for the poor, elderly, and other vulnerable groups should be provided, in case of extreme situations (e.g., when the grid fails).

Adaptation has long been a critical part of global climate strategy and sustainable development^43,44 and has increasingly attracted more attention to cope with the possible risk of climate damage, particularly when facing great difficulty in worldwide cooperation on substantial emission mitigation.¹⁰ As a critical adaptation or coping mechanism in response to global warming, air-conditioning demand in developing economies is poised to increase dramatically with climate change and income growth.¹⁷ Actually, the use of ACs can improve human comfort, lessen the physical and mental damage, and lower the probability of mortality caused by extreme temperatures.^16,36,43 High dependence on air-conditioning, however, also triggers higher energy consumption and worsens outdoor heat stress,⁴⁵ which may in turn exacerbate energy poverty and inequality and increase carbon emissions, jeopardizing the attainment of carbon neutrality and sustainable development.^26,44 Estimating the impact of climate change on air-conditioning usage can help with grid planning, investment rationalization in power system, and long-term balance between mitigation and adaptation. In this regard, our research assists the government in designing adaptation plans early on and aligning investments in adequate adaption strategies.

Although our work is an important complement to the limited research on the impact of climate change on air-conditioning demand, particularly from the perspective of developing countries, there are several directions for further research. First, owing to the lack of electricity load data, we were not able to empirically investigate the effect of climate change on electricity demand⁴⁶ together with air conditioner sales and thus rely on several assumptions to project the future electricity demand. Combining data on AC sales and electricity load in the future would generate more insights into the adaptive response to climate change in China. Second, more alternative heating and cooling services to cope with climate change, such as fans, electric heaters, and water usage,⁴⁷ need to be considered to explore the differential impacts of climate change on diverse services. Lastly, there are several other factors that could affect the AC adoption in China, such as population change and changes in people's lifestyle and habits, which were not included in this study due to data unavailability but could be a direction for further research in the future.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Weekly air conditioner data in China	China Market Monitor Co., LTD	From China Market Monitor Co., LTD directly
Historical weather data	the National Meteorological Information Center of China	http://data.cma.cn/
Historical annual disposable income per capita	the National Bureau of Statistics in China	http://www.stats.gov.cn/tjsj/zxfb/202102/t20210227_1814154.html
Climate projection data	the NEX-GDDP-CMIP6 dataset	https://www.nccs.nasa.gov/services/data-collections/land-based-products/nex-gddp-cmip6
Projections of future economic growth	SSP database	https://tntcat.iiasa.ac.at/SspDb/
Software and algorithms
Stata	Stata 17	https://www.stata.com/new-in-stata/
Python	Python 3.7.6	https://www.python.org/downloads/release/python-376/
R	R 4.2.2	https://cran.r-project.org/bin/windows/base/
ArcGIS	ArcGIS 10.8	https://www.esri.com/en-us/home

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Xiaobing Zhang (xbzhmail@gmail.com).

Materials availability

The study did not generate new materials.

Method details

Air conditioner data

We used air conditioner sales data collected by China Market Monitor Co., LTD, from 2008 to 2019. The company, which is engaged in market research on home appliances in China, collected offline air conditioner sales data from a representative set of over 10,000 stores across China, covering all first- and second-tier markets and a typical set of lower-tier markets. We were provided with weekly AC sales for 343 cities at the prefecture level, covering 30 provinces and municipalities. Specifically, if the first day of a year in the dataset is not Monday, the AC sales in the first week of that year include sales in the same calendar week of the previous year.

There are 20,332 models of ACs sold by 91 manufacturers, including domestic and foreign companies. They are sold through four marketing channels: local chain stores, home appliance stores, department stores and others. The AC types in our dataset include window, wall-mounted, and floor-standing types. Thus, the raw sales data is at the product-store year-week level. As the exogenous variation in climate is given at the city level, the data company aggregates the AC sales data to the same scale for our empirical analysis.

The background color in Figure 1 depicts the geographical coverage of AC sales data and different shades of color represent the average weekly AC sales. It is observed that AC sales vary significantly across various cities and more ACs are sold in southern cities, such as Chongqing, Shanghai, Chengdu, and Guangzhou. Chongqing has the highest average weekly AC sales. Compared with cities in south China, cities in the north have relatively lower weekly sales, except for Beijing and Xi’an.

Weather data

Weather data was obtained from the National Meteorological Information Center of China (CMA Meteorological Data Center). The raw data encompassing information on daily average, highest and lowest temperature, precipitation, and relative humidity spans from December 31, 2007 to December 26, 2019, to match the AC data. We included 699 meteorological stations and adopted the inverse distance weighting method for all stations to derive daily weather data for each city in our datasets. The first step was to identify the geometric centers of each prefecture city. Next, we calculated the distance between each monitoring station and the geometric center of each city. Third, we chose a circle with a radius of 200 km for each prefecture city’s centroid and kept the same distance between all stations within the circle.⁶ Based on the inverse distance to the city center, we took the weighted average of the daily records.

The yellow dots in Figure 1 depict the geographic distribution of weather stations that satisfy the distance to the corresponding city of less than 200 km. Owing to the 200 km radius restriction, we excluded some stations, such as the Maqu station in Gansu province. As a result, 687 stations were used and the distance between 343 cities was calculated. Overall, we obtained relatively comprehensive geographic coverage. Considering the climate conditions in China and referring to existing literature,^6,17,18 we consider six bins in this analysis, corresponding to temperature intervals of <10°C, 10–15°C, 15–20°C, 20–25°C, 25–30°C, and >=30°C, respectively. The most comfortable temperature bin [15°C, 20°C) was set as the baseline group, implying the assumption that there is no observable effect of temperature on air conditioning adoption within this bin. Consequently, all estimates were interpreted relative to the baseline. The distribution of days across bins is shown in Figure S1. To merge with the AC datasets, we similarly transformed the weather data from daily frequency to weekly data. Heatwaves describe the number of days that daily average temperature is above 30°C for at least three days. We define CDDs as the weekly total of the cooling degree days with daily average temperature above 18.3°C and HDDs as the weekly total of the heating degree days with daily average temperature below 18.3°C. Maximum temperature is the highest temperature in a week. Precipitation is the average of 24-hour cumulative rainfall in a week, and relative humidity is also the weekly average of daily data.

Economic data

Annual disposable income per capita in each city was obtained from the National Bureau of Statistics in China¹² and deflated to the price level in 2008. Overall, we acquired data for 296 cities.

Econometric model

Benefitting from high-quality AC sales panel data, we can assess the nonlinear relationship between temperature and AC sales. Our empirical approach is fixed-effects panel regression using a simple log-linear equation, which has been widely used in previous studies.^6,18,38 Specifically, the model can be described as follows:

$$\mathit{ln}\left(y_{ijk}\right)={\beta }_0+\sum _m{\beta }_{1m}{f}_m\left(tem{p}_{ijk}\right)+X_{ijk}+{province}_i·{year}_k+{week}_j·{year}_k+{city}_i+{\epsilon }_{ijk}$$ (Equation 1)

The dependent variable $\mathit{ln}\left(y_{ijk}\right)$ is AC sales in city $i$ at week $j$ in year $k$, measured in logs. In our data, we had 343 cities and 52 or 53 weeks per year for most cities. $f_m\left({temp}_{ijk}\right)$ is a function that reveals the nonlinear relationship between the daily average temperature and air conditioning. It denotes the number of days for city $i$ at week $j$ in year $k$ on which the average temperature was in the $m$ th temperature bin. As mentioned earlier, six bins are considered here, corresponding to temperature bins less than 10°C, from 10°C to 15°C, 15°C to 20°C, 20°C to 25°C, 25°C to 30°C, and above 30°C, respectively. Parameter ${\beta }_{1m}$ is a vector of coefficients that interprets the change in AC sales associated with an additional day in each specific bin. There is an implicit assumption that the warming effects do not vary within each bin and are the same across all cities. The most comfortable temperature bin [15°C, 20°C) was set as the baseline group, with which all estimates should be interpreted relatively.

$X_{ijk}$ is a vector of control variables including per capita disposable income, precipitation, and relative humidity. Instead of separately controlling for province fixed effects and year fixed effects, we include province-year fixed effects to control for the impact of unobservable factors that change over time at the provincial level on AC sales. We also control for week-year fixed effects ${week}_j\cdot {year}_k$ , which represent the time-variant factors across weeks and years, such as the Home Appliances Going to the Countryside Policy and other AC promotion policies. In addition, we add city-fixed effects ${city}_i$ to control for city-level characteristics that do not change over time. The error term ${\epsilon }_{ijk}$ captures the differences in AC sales that are not expressed by temperature, control variables, or fixed effects. For the basic results, we cluster the standard errors at the city level to allow autocorrelation within each city. We also perform robustness checks by clustering at the province level.

Heterogeneity analysis

We investigate how the temperature-air conditioning response could be affected by diverse income levels that are proxied by the average per capita disposable income from 2008 to 2019. For cities with below-median per capita disposable income, high temperature leads to more increase of AC demand than that of the above-median group. Generally, there is a positive relationship between annual income and air conditioning saturation,¹⁷ implying that cities with below-median per capita disposable income have lower saturation than those with above-median per capita disposable income. In this circumstance, the lower-income cities increase more demand for air conditioning, confronting hotter weather (Table S4 Columns 1 and 2). As for the response to relatively lower temperature (e.g., <10°C), the difference between the below-median and above-median groups is less noticeable.

Then, we examine whether the relationship between temperature and air conditioning adoption is impacted by the types of ACs. The estimates show that households tend to increase more demand for wall-mounted and window ACs than that of the floor-standing ones on hot days (Table S4 Columns 3 and 4). Wall-mounted ACs are usually used in bedrooms, while floor-standing types are mainly placed in the living room. Since people take most of their time at home sleep (particularly for business days), wall-mounted ACs with higher mute levels and energy savings are therefore intensively used during hot summer days. On the contrary, floor-standing air conditioning may not be necessarily adopted, particularly for the less affluent family.

Further, we attempt to explore if there are different impacts of temperature variations on the sales of ACs with different brands. We aggregate AC sales data for all brands from 2008 to 2019 to count the market shares. The four brands with the largest market share are GREE, Midea, Haier and Hisense, accounting for 68% of the total sale. The dominant-brand group contains sales of the aforementioned four air conditioning brands, with all the rest of the brands categorized as the other-brand group. However, we cannot get any clear or consistent conclusions, given the current estimates shown in Table S5.

Heatwave effect and time-lagged effect

To explore the effects of heatwave on the sales of ACs, we first define the occurrence of a heatwave when the daily average temperature exceeds 30°C for at least 3 consecutive days. Specifically, the first day of a heatwave counts one, followed by two for the second day, and so on.^43,47 Days without heatwaves are assigned as zero. Then we aggregate the total value to each week. Results show that there is a significantly extensive effect on air conditioning adoption as a heatwave prolongs (Table S6). The coefficient, i.e., 0.008, can be interpreted that as compared to days without a prolonged heatwave, AC sales increase by 0.8% on the first day of a heatwave, 1.6% on the second day, and so on.

To further examine whether there are any time-lagged effects of temperature on air conditioning adoption, particularly on hot summer days, we take the maximum temperature during one week and its one-week lagged term into account. To be specific, we focus on the possible time-lagged effect in hot summer, including May, June, July and August. According to the estimate, we can conclude that higher temperature remarkably increases air conditioning needs during the current week and the next week (Table S7).

Robustness check

We conduct several robust checks to see if our results are driven by variable setting, model selection and basic assumptions. First, it is difficult to compare our estimates with those in previous literatures, as very few studies use air conditioning panel data to explore the extensive margin of climate change, particularly for China. The most relevant literatures are Sailor and Pavlova (2003)⁴⁸ and Davis and Gertler (2015),¹⁷ in which the authors used annual data of AC saturation, ranging from zero to one, to calculate annual CDDs and explore the effect of warming. They found that increases in CDDs could give rise to saturation increase. Benefiting from the availability of AC data, we could expand the research by considering non-linear effect of temperature on air conditioning adoption. For the convenience of comparison, we also conduct a regression using week-level CDDs. We define CDDs as the weekly total of the cooling degree days with daily average temperature greater than 18.3°C.^28,49 The result shows that the coefficient of CDDs is positive at the 1% statistical significance level, implying that warmer climate does promote people to purchase ACs. Besides, we also run the regression with HDDs, which is defined as the weekly total of the heating degree days with daily average temperature lower than 18.3°C, and the coefficient is found positive (though insignificant), which is consistent with our conclusion regarding the U-shaped effect of temperature on air conditioning (cold weather could also lead to more AC sales).

In addition, we substitute week-year fixed effects with month-year fixed effects in the econometric model to check if the former one is absorbing the impact of time-variant factors. Results reveal that the estimates become smaller for temperature lower than 15°C, while for temperature higher than 20°C all coefficients are larger than the baseline estimates. The disparity in estimates means that month-year fixed effect could not capture enough impact of time-variant factors. As a result, it is much better to choose the week-year fixed effect in the analysis, as we have done in the main regression model.

Furthermore, we test the robustness of estimates from the data cleaning process. The dataset is trimmed at the 1^th and 99^th percentile and the statistical significance is invariant for coefficients of all temperature bins. Our baseline estimates are also tested to be robust by clustering at the province level, as shown in Column 5 of Table S8.

Climate change prediction

The predicted data of climate change, including the daily near-surface temperature for each grid point from 2015 to 2100, is obtained from the NEX-GDDP-CMIP6 dataset of the NASA Center for Climate Simulation in the United States. The NEX-GDDP-CMIP6 dataset contains bias-corrected global downscaled climate projections under Shared Socioeconomic Pathways (SSPs), which is derived from the General Circulation Model (GCM) runs conducted under the Coupled Model Intercomparison Project Phase 6 (CMIP6).³⁰ The grid points are separated into 0.25° latitude and 0.25° longitude. We use projection average from 6 climate models (CanESM5, CESM2, CMCC-ESM2, CNRM-ESM2-1, GFDL-ESM4 and IPSL-CM6A-LR) to reduce model uncertainty and further analyze the impact of model uncertainty on the AC sales and related electricity demand. To avoid the abnormal effects of short-term weather fluctuations, we choose 20-year weather averages.¹⁸ We consider the average temperature of 2040–2059 to predict the mid-term impacts of climate change.

Air conditioning forecasts

Consistent with the literature,^14,17 we investigate the combined effect of climate change and income growth on mid-century AC sales (2040–2059). With the baseline estimates in hand, we replace the number of each city’s temperature bin during 2008–2019 and the current per capita disposable income with mid-century values under the SSP1, SSP2, and SSP5 scenarios, respectively, and then project the rate of change in AC sales under the given scenarios, compared with the current level. We use the nationwide growth rate of China’s per capita gross domestic product (GDP) from the SSP scenarios, and assume that disposable income per capita will increase at the same rate. In China, the future growth rate was particularly high under the SSP5 scenario (3.1% annually), followed by that in the SSP1 and SSP2 scenarios (2.5% and 2.3%, respectively). We also explore the marginal impact of climate change on air conditioning by freezing the per capita disposable income at the current level (detailed results can be found in Figure S2).

Projections of marginal air conditioning response to climate change

By freezing the per capita disposable income at the current level, we are able to explore the marginal impact of climate change on air conditioning. With the baseline estimates in hand, we replace the number of each city’s temperature bin during 2008–2019 with the values in mid-century, under the SSP1, SSP2 and SSP5 scenarios, then we project the rate of change in AC sales under the given scenarios, compared with the current level.

Model average forecasts reveal that the future air conditioning demand in summer will increase in most cities. Results under different scenarios display that there is no remarkable difference in air conditioning between the SSP1 scenario and SSP2 scenario in 2040–2059 (Figures S2A and S2C), while AC sales in most cities under the SSP5 scenario are relatively higher (Figure S2E). Furthermore, AC sales generally increase more in most southern cites than in northern cities. For example, the mid-century growth rate of AC sales in the Beijing-Tianjin-Hebei Region is about 7.7%, 8.1% and 13.8% under the SSP1, SSP2, SSP5 scenario, versus about 18%, 17% and 26.5% in the Pearl River Delta, respectively.

In winter, ACs are mainly adopted for heating, particularly in southern China where no central heating is provided. As a result, the projected changes of ACs sales in northern cities are small, given the broad adoption of central heating systems, and the declining rates in most northern cities are less than 1% under all scenarios (Figures S2B, S2D, and S2F). The situation in the south is rather different. Given a warmer climate, air conditioning demand for heating in most southern cities will decrease, and the declines in air-conditioning sales under the SSP5 scenario is greater than those under the SSP1 and SSP2 scenarios.

Impact of demographic change

Looking into the future, according to World Bank projections, China’s population will peak in 2030 and then begin a steady decrease if no effective policies are implemented. Under this circumstance, aging and low birth rate will coexist, resulting in a decrease in the average size of Chinese families, and the demand for air conditioning under different temperatures would be closer to the outcome of a lower than median household size (see Column 3 of Table S3). Moreover, on July 20, 2021, China issued a comprehensive package of guideline measures for the three-child policy, with the goal of reversing the trend of population aging and declining birth rate. With increasing longevity and the new three-child policy, the family structure might shift to accommodate more members including children and elder people.⁵⁰ Therefore, if the policy is effective, the demand for air conditioning at various temperatures should be closer to result of households with memberships above the median (see Column 4 of Table S3), which implies that the demand for air conditioning in the future could increase even more.

Annual hours of air conditioning use

To calculate the current and future operating hours of ACs, we first clarify the conditions under which ACs are used. Generally, we assume that when the daily average temperature is higher than 25°C, households turn on ACs for cooling, and when the temperature is below 10°C, ACs are used for heating.⁵¹ On the days when the above conditions are met, families use six hours of air conditioning on average per day.⁵² Simultaneously, we consider the difference in heating between the northern and southern cities. We assume that households in northern cities do not use ACs during the central heating period.⁵¹ The central heating period in Beijing, Tianjin, and cities in Hebei, Shanxi, Inner Mongolia, Henan, and Shandong provinces is 4 months, starting from November 15 to March 15 of the following year.^53,54 Central heating starts on November 1 and lasts until March 31 of the following year (five months) in Shaanxi, Gansu, Qinghai, Ningxia, Xinjiang, and Tibet.⁵⁵ Heilongjiang, Jilin, and Liaoning provinces have the longest central heating time of six months (from October 20 to April 20 of the following year).⁵⁶ Without a central heating system in southern cities, households will turn on ACs when the temperature conditions are met.

Annual electricity demand of air conditioning use

The increased use of ACs will increase peak power demand, which in turn will pose challenges to grid loads. To determine the per capita electricity consumption of air conditioning use now and in the future, we need to estimate current AC holdings. Based on the number of AC units in 100 households and the number of households, collected from the statistical yearbooks of cities, we obtain the current AC holdings. For cities that provide the number of ACs owned by 100 households in urban and rural areas, we use the following method:

$$\left({AC}_{urban}\ast W_{urban}+{AC}_{rural}\ast W_{rural}\right)\ast H_{total}$$ (Equation 2)

$$W_{.}=\left(\frac{P_{.}}{S_{.}}\right)/\sum _{.}\left(\frac{P_{.}}{S_{.}}\right),.=urban,rural$$ (Equation 3)

where ${AC}_{urban}$, ${AC}_{rural}$ are the numbers of ACs owned by 100 urban and 100 rural households, respectively. $W_{urban}$ and $W_{rural}$ are the urban and rural weights, respectively. As cities do not provide the number of urban and rural households, we divide the urban population by urban household size to obtain the number of urban households. The same is true for rural households as well. We assume that cities located in a province have the same household size (for both urban and rural areas), where the household size data is acquired from the sixth census. Subsequently, we divide them by the total number of households to obtain urban and rural weights. $P$ is the population and $S$ is the family size obtained from the sixth census. $H_{total}$ is the total number of households.

To calculate AC ownership, we need to estimate the annual sales of ACs from the present to mid-century. We estimate mid-century AC sales in the previous section. Assuming that the annual growth rates of AC sales are the same from the present to mid-century, and a 5% AC obsolescence rate, we obtain mid-century AC ownership. By multiplying the ownership with the operating hours and the average unit power (our sample median), we finally obtain the electricity consumption of air conditioning use.

Published: February 10, 2023

Data and code availability

• •Data: Weekly air conditioner data in China is from China Market Monitor Co., LTD, and they are available from the authors upon reasonable request. Historical weather data is obtained from the National Meteorological Information Center of China at http://data.cma.cn/. Historical annual disposable income per capita is from the National Bureau of Statistics in China at http://www.stats.gov.cn/tjsj/zxfb/202102/t20210227_1814154.html. Climate projection data is obtained from the NEX-GDDP-CMIP6 dataset at https://www.nccs.nasa.gov/services/data-collections/land-based-products/nex-gddp-cmip6. Projections of future economic growth is available from the Shared Socioeconomic Pathways (SSP) database at https://tntcat.iiasa.ac.at/SspDb/. All data and models are processed in Stata 17.0, Python and ArcGIS. The figures are produced in Python and RStudio.
• •Code: All custom code can be available on request from the lead contact.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.