Quantifying the Impact of Climate Change on Household Water Use in Mega Cities: A Case Study of Beijing, China

Abstract

Amid rapid urbanization and climate change, global urban water consumption, particularly household water use, has continuously increased in recent years. However, the impact of climate change on individual and household water use behavior remains insufficiently understood. In this study, we conducted tracking surveys in Beijing, China, to determine the correlation between climatic factors (e.g., temperature, precipitation, and wind) and household water use behaviors and consumption patterns. Furthermore, we proposed a genetic programming-based algorithm to identify and quantify key meteorological factors influencing household and personal water use. The results demonstrated that water use is mainly affected by temperature, particularly the daily maximum (TASMAX) and minimum (TASMIN) near-surface air temperature. In addition, showering and personal cleaning account for the largest proportion of water use and are most affected by meteorological factors. For every 10 °C increase in TASMAX, showering water use nonlinearly increases by 3.46 L/d/person and total water use nonmonotonically increases by 1.14 L/d/person. When TASMIN varies between −10 °C and 0 °C, a significant change in personal cleaning water use is observed. We further employed shared socioeconomic pathway scenarios of the Coupled Model Intercomparison Project 6 to forecast household water use. The results showed that residential water use in Beijing will increase by 21–33% by 2035 compared with 2020. This study offers a groundbreaking perspective and transferable methodology for understanding the effects of climate change on household water use behavior, providing empirical foundations for developing sustainable water resource management strategies.

1. Introduction

Water is essential for human survival [1]. Over the past four decades, global water use has increased by approximately 1% per year, and it is expected to continue rising at a similar rate until 2050 [2]. This upward trend is primarily driven by urban water use, especially domestic consumption. Urban water demand alone is projected to grow by 80% by 2050 [3]. Therefore, ensuring the safety of urban water supply is vital, as it directly influences public health, urban infrastructure, and social stability [4]. However, temporal and spatial variations in water demand pose substantial challenges for forecasting and planning urban water supply [5]. These challenges are further compounded by the increasing frequency of extreme weather events across the world, with reports documenting annual spikes in water demand during heatwaves [6]. In this context, climate change exacerbates urban water stress; however, its effects on household-level water use behavior remain unquantified, particularly in water-scarce megacities. Empirically quantifying the impact of climate change on daily life and urban systems is therefore a critical aspect of sustainable water resources management [7].

Identifying and understanding the factors influencing household water use can reduce uncertainty in demand forecasting and reinforce water security strategies [8]. In densely populated cities, household water use substantially shapes overall urban water demand [9]. Notable regional variations in household water consumption have been observed across countries [10], cities [11], and urban and rural areas [12], as well as among neighborhoods within the same city [13]. Previous studies have categorized the factors influencing household water use, including personal characteristics (e.g., income, age, and gender) [14]; attitudes and habits (e.g., water use behavior) [15]; household attributes (e.g., family size and structure) [16]; environmental values and conservation attitudes (e.g., willingness to protect the environment) [17]; and climate or seasonal variability [18]. Schleich and Hillenbrand [11] found that age affects household water consumption, with usage increasing by 1.8 L per day (L/d) for each additional year of age in Germany. Meanwhile, Yu et al. [19] reported an inverse relationship between family size and household water use, with each additional household member reducing annual per capita cooking water consumption by 10.1%.

Currently, urban water use forecasting primarily relies on analyzing historical consumption patterns, using techniques such as statistical analysis, modeling, and machine learning. For example, Bayesian statistics have been used to forecast urban water consumption based on daily usage data in Canada [20]. Likewise, machine learning techniques has been applied to predict daily residential water demand using metered consumption and billing data information in the United States and Canada [21]. In Australia, singular spectrum analysis and artificial neural networks have been employed to forecast municipal water consumption based on monthly data [22]. Genetic programming (GP), an evolutionary machine learning algorithm and widely used data mining technique, has also been applied in this context. This domain-independent approach enables the generation of optimal mathematical models [23]. In the United Kingdom, GP has been used to develop district water supply models based on nonlinear economic and demographic factors [24]. Recognized for its effectiveness in capturing nonlinear patterns, GP is now being explored in water resource engineering, with early applications in hydraulic and hydrological simulations [25,26].

Climate change is a potential determinant of household water use and could increase the demand for urban water resources [27]. With global warming and the resultant high frequency of extreme weather events such as droughts and heatwaves [28], the effects of climate change on daily life have become increasingly evident [29]. The uncertainty surrounding climate change has a widespread impact on daily human life, influencing travel choices, water use, and related behaviors [30]. Climate change also places increasing demands on national policy responses.

Although previous studies have linked climate to regional water use, the relationship between climate and household-level water use behavior remains insufficiently quantified. Research in this field has mainly focused on the relationship between climate factors and urban water consumption. Balling and Gober [31] conducted a correlation analysis in Phoenix, United States, and found that total municipal water consumption is positively related to average temperature and negatively related to precipitation at the urban level. Meanwhile, a panel regression analysis of household- and individual-level data by Qin et al. [6] revealed a positive correlation between the average temperature and daily household water consumption in southern China, with water consumption increasing by approximately 3.3 L/d per 1 °C. Studies have widely used questionnaires to effectively determine the relationship between climate factors and household water use; for instance, Aydamo et al. [32] used stepwise multiple linear regression to analyze 288 household questionnaires in southern Ethiopia, finding that each additional family member reduced per capita water consumption by 6.63 and 7.00 L/d in the dry and rainy seasons, respectively. Similarly, Ibrahim A et al. [33] administered questionnaires to 550 individuals in Freetown, Sierra Leone, to quantify differences in water use between rainy and dry seasons; they found that the relationship between personal water use and climate is complex. Overall, existing studies on the relationship between climate factors and household water use have primarily focused on precipitation and temperature, with limited attention to other meteorological factors.

The combined effects of climate change and complex household water use patterns exacerbate the challenges of urban water management. With rising global temperatures, the variability in urban water use may become more severe and difficult to predict [34]. These trends highlight the urgent need to analyze variations in household water use patterns with climate change and predict urban domestic water use [35]. Therefore, this study investigates the effects of climate change on individual lifestyles and water use. Moreover, it quantifies the relationship between climate and household water use behavior and proposes relevant policy recommendations. Specifically, a habit-tracking survey was conducted to build a database covering individual water use behavior, habits, water use volume, and relevant climate factors. A GP-based algorithm was then developed to analyze how climate factors influence individual water use behaviors under varying conditions. Beijing residents were taken as the sample population to identify and quantitatively analyze the key climate factors influencing household water use. The study also predicts changes in personal water use under various climate change scenarios. Overall, by focusing on the challenges of climate change in the context of household water use, this study offers valuable quantitative insights to support sustainable urban water management and mitigate the rapid growth of urban resource consumption.

2. Materials and Methods

With increasing urbanization, multi-story and high-rise apartments have become common in Chinese cities, meeting the housing demands of densely populated regions. In such cities, most domestic water consumption occurs indoors, primarily through the use of faucets, drinking water purifiers, dishwashers, showers, toilets, and washing machines. Residential water use can be broadly categorized into the following six types based on use behaviors and appliances: personal cleaning, showering, flushing, laundry, environmental cleaning, and culinary activities. Other indoor water uses, such as for indoor plant irrigation and pet care, account for a relatively small proportion [36] and are not included in this study.

Data on individual and household attributes, water use behaviors, and water use patterns were collected via a face-to-face social survey. Principal component analysis (PCA) and kernel density estimation were applied to determine the data structure of household water use under different climate conditions. GP was used to develop interpretable mathematical models linking water use in various aspects of daily life to potential influencing factors.

2.1. Study Area

Beijing, a densely populated megacity, was selected as a case study to investigate the impact of climatic factors on personal household water use. Beijing (39.4°–41.6° N, 115.7°–117.4° E), the capital of China and a densely populated megacity, has a total area of 16,410 km². By the end of 2020, its permanent population reached 21.89 million, with a population density of approximately 1334 people/km². The per capita housing area in Beijing is 34 m². Beijing’s climate exhibits pronounced seasonal and interannual variability, making urban water use in the city particularly sensitive to climatic conditions. The city receives an annual average precipitation of approximately 600 mm and experiences a warm temperate, semi-humid, semi-arid continental monsoon climate with four distinct seasons. Summer and winter are long, while spring and autumn are brief. Temperatures in spring and autumn range from 10 to 25 °C. Summers are hot and rainy, with average temperatures of 20–34 °C. Notably, 70% of the annual rainfall occurs during this season. Winters are cold and dry, with average temperatures ranging from −7 °C to 2 °C. In 2020, Beijing experienced notable climate variability, including intense rainstorms and extreme high temperatures. The questionnaire survey incorporated weather samples that effectively covered the analysis of individual and combined climate factors.

Beijing experiences considerable water scarcity, with an annual per capita water volume of only approximately 150 m³, much below the internationally recognized warning threshold of 500 m³. Furthermore, the imbalance between water supply and demand in Beijing exceeds that of other Chinese megacities, such as Shanghai and Guangzhou [37]. This issue is exacerbated by the high population density and insufficient local water supply, requiring approximately 70% of the city’s urban water to be sourced from the South-to-North Water Diversion Project [38]. To address these challenges, city officials must urgently improve water resource management and system efficiency and support sustainable socioeconomic development.

2.2. Research Data

2.2.1. Social Survey and Questionnaire Design

Face-to-face social surveys are effective tools for collecting data on residents’ lifestyle and water use behaviors [39]. Notably, face-to-face surveys have some key advantages [40], including a well-defined structure, flexibility, and high adaptability [41]. To determine how climate influences household water use trends, the questionnaire designed herein focused on the following two aspects: (1) basic family and personal information and (2) water use behaviors and household appliances. The survey recorded the frequency of water usage and the climate conditions at the time of the survey for comparative analysis (Figure 1).

Acknowledging the influence of familial and personal characteristics on domestic water use, the questionnaire included 12 questions on “family and personal information” related to family size, family member structure, dwelling ownership, housing area, age, gender, occupation, and average monthly income. Furthermore, 29 questions under the section “personal behavior and water appliances” addressed the frequency, duration, type, and water efficiency of appliances used. Finally, 11 questions concerned meteorological factors.

Sampling methods are critical for ensuring the representativeness of survey results. Random sampling can ensure the internal validity of questionnaire data. This study employed stratified multistage random sampling to improve survey accuracy. Survey respondents were randomly selected for continuous face-to-face interviews conducted over one year. Questionnaires were distributed on randomly selected days across all four seasons, with additional surveys administered to assess changes in residents’ water use behaviors during periods of high temperatures, precipitation, and other weather variations. After filtering for anomalies, the final dataset comprised 711 valid participants and 9576 questionnaire samples collected between March 2020 and February 2021. Three normal-weather days and one special-weather day were randomly selected in each season. The participants reported their water use behavior from the previous day to align their water use with corresponding climatic conditions. A hybrid data collection approach was employed, and the questionnaire content remained consistent throughout the study. Among the respondents, 54% were male. Age distribution was as follows: age <20 years (6.6%), 20–39 (38.3%), 40–59 (44.1%), and 60 (11%). Families with 1–2 members accounted for 29.2% of the sample size, those with 3–4 members represented 57.3%, and those with 5–6 members represented 13.5%. The average housing area of respondents was 37 m², closely aligning with the corresponding average for Beijing. Personal household water use was calculated based on the questionnaire. The average annual personal water use was 141.2 L, which is 4.6% higher than the per capita annual household water use of 135 L reported for Beijing residents in urban statistical water use from the 2020 China Water Resources Bulletin (http://www.mwr.gov.cn/sj/tjgb/szygb/202107/t20210709_1528208.html (accessed on 23 December 2024)). To improve data reliability, monthly water bills were collected from respondents’ households and used to cross-validate reported use against local water pricing data.

2.2.2. Water Use Accounting Method

This study developed a systematic household water accounting method based on the water use behaviors of urban residents in China, covering the following six water uses: personal cleaning, showering, flushing, laundry, environmental cleaning, and culinary activities. The total quantity of water used by residents (Q_sum, L/d) was calculated as

$$Q_{sum}=\sum _{i=1}^6{Q}_i$$

(1)

where Q_i represents the quantity of water used for different activities, with i = 1, 2, 3, 4, 5, and 6 representing personal cleaning, showering, flushing, laundry, environmental cleaning, and culinary activities, respectively. Among individual water uses, the quantity of water used for personal cleaning (Q₁) was calculated as

$$Q_1=\sum _{i=1}^4\left(E_{1i}\times Q_{1i}\times W\right)$$

(2)

where Q₁ includes water used for small-scale cleaning of personal body parts; E_1i represents the frequency of personal cleaning activities (times/day), with i = 1, 2, 3, and 4 denoting face washing, tooth brushing, hand washing, and other body part cleaning; Q_1i represents the duration of water use per activity (min/time); and W denotes the water efficiency of faucets (L/min), the primary water appliance used by the respondents. Obtaining accurate and comprehensive information about the water use efficiency of appliances through questionnaires is challenging; therefore, this study measured the time of purchase of appliances to uniformly determine their water efficiency. Similarly, the quantity of water used for showering or bathing (Q₂) was calculated as

$$Q_2=E_2\times T_2\times W$$

(3)

where E₂ denotes the frequency of showering or bathing (times/day), T₂ represents the duration of showering or bathing (min/time), and W represents the water efficiency of showerheads (L/min). Notably, respondents reported minimal water use for bathing. Next, the amount of water used for toilet flushing (Q₃) was calculated as

$$Q_3=E_3\times Q_{31}\times \left[1-K_2\times \left(1-K_1\right)\right]$$

(4)

where E₃ represents the frequency of residential toilet flushing (times/day) and Q₃₁ represents the water efficiency of flushing equipment (L/time). The questionnaire also assessed the use of wastewater or reclaimed water for toilet flushing and calculated the percentage of such reuse. Specifically, K₁ in the above equation indicates whether wastewater is used for toilet flushing (1 = no, 0 = yes) and K₂ represents the percentage of wastewater reuse (%). The amount of water used for laundry (Q₄) was calculated as

$$Q_4=\alpha \times \theta \times \left(E_{41}\times Q_{41}\times W+E_{42}\times Q_{42}\times F\right)$$

(5)

where E₄₁ and E₄₂ represent the frequencies of hand washing and machine washing (times/week), respectively; Q₄₁ denotes the duration of hand washing per water use (min/time); Q₄₂ represents the weight of laundry per machine wash (kg/time); F denotes water use per kilogram of machine-washed clothes (L/kg); and W represents the water efficiency of washing basin taps (L/min). Some domestic water uses are common household activities, such as laundry and environmental cleaning; however, it is difficult to accurately determine the frequency of water use for such activities on a daily basis. Therefore, we measured this frequency on a weekly basis. Specifically, a family size coefficient $\left(\alpha \right)$ was added to determine individual water usage, and a time coefficient ($\theta$) was applied to convert the data to a daily scale. The quantity of water used for environmental cleaning (Q₅) was calculated as follows:

$$Q_5=\alpha \times \theta \times E_5\times Q_{51}$$

(6)

where E₅ represents the frequency of environmental cleaning activities (times/week) and Q₅₁ represents the amount of water used per activity (L/time).

The amount of water used for culinary activities (Q₆), including drinking water (Q₆₁) and cooking water (Q₆₂), was calculated as follows:

$$Q_6=Q_{61}+Q_{62}$$

(7)

$$Q_{61}=\alpha \times \left(E_{61}\times Q_I\times W+Q_z+Q_T\right)$$

(8)

$$Q_{62}=\alpha \times E_{62}\times \left(Q_K+Q_H+Q_M\right)\times W$$

(9)

where Q₆₁ includes tap water (Q_I), water from community machines (Q_z), and bottled water (Q_T); E₆₁ represents the frequency of tap water use (times/day); and W represents faucet efficiency. Moreover, Q₆₂ accounts for water used before, during, and after cooking, including that used for food washing (Q_K), cooking (Q_H), and dishwashing (Q_M). Further, E_6i represents the frequency of cooking activities (times/day).

2.2.3. CanESM5 Model and Weather Station Data in Household Water Use Prediction

This study incorporates both historical meteorological data and projected climate data. The historical data were sourced from the National Meteorological Science Data Centre and covered daily observations from 2416 weather stations across China from 1960 to 2015. The following meteorological variables were analyzed: near-surface air temperature (TAS), maximum and minimum near-surface air temperatures (TASMAX and TASMIN, respectively), ground temperature (GT), relative humidity (HURS), surface pressure (P), surface wind speed (WIND), eastward and northward near-surface wind components (UAS and VAS, respectively), surface evaporation (EVS), and precipitation (PR).

Estimates of meteorological data under various future scenarios were obtained using output data from general circulation models provided by Coupled Model Intercomparison Project (CMIP) 6 (https://esgf-node.llnl.gov/projects/cmip6/ (accessed on 12 December 2023)). CanESM5 was selected for its strong simulation performance in East Asia [42] and high equilibrium climate sensitivity [43]. CanESM5 demonstrates strong efficiency in reproducing the climatological spatial distribution of precipitation and exhibits a pronounced response to humidity changes [44,45]. Three widely used shared socioeconomic pathway (SSP) scenarios (SSP1-2.6, SSP3-7.0, and SSP5-8.5) from CanESM5 were selected for analysis. Scenario SSP1-2.6 represents a sustainable development pathway, limiting global warming to <1.5 °C. SSP3-7.0 describes a moderate pathway with projected CO₂ emissions of 60–80 billion tons and warming of approximately 4.5 °C by the end of the 21st century. SSP5-8.5 represents the worst-case scenario, assuming that no climate policies are implemented, with CO₂ emissions exceeding 120 billion tons and a temperature rise of 5–6 °C by the end of the 21st century [46].

To ensure consistency across historical and future climate data, the Anuspline interpolation method was employed to downscale all data into a 25 × 25 km spatial grid. Systematic errors in climate predictions were addressed using the non-stationary cumulative distribution function matching technique, which corrected biases in future estimates based on historical data [47].

2.3. Data Analysis and Processing

PCA is commonly used to group related variables, reduce data dimensionality [48], and enhance data interpretation. In PCA, a two-dimensional projection of samples is typically constructed, with the principal components (PCs) serving as the axes [49]. The original PCA data in this study are structured as follows:

$$\left(\begin{array}{cccc}{T}_{\mathrm{1,1}}& y_{\mathrm{1,1}}& \dots & y_{\mathrm{1,6}}\\ ⋮& ⋮& \ddots & ⋮\\ T_{n,1}& y_{n,1}& \dots & y_{n,6}\end{array}\right)$$

where each row represents a questionnaire, with n = 9576 denoting the total number of questionnaires. The first column of the matrix ${\left(\begin{array}{ccc}{T}_{\mathrm{1,1}}& \dots & T_{n,1}\end{array}\right)}^T$ represents the survey date, and columns 2–7 ${\left(\begin{array}{ccc}{y}_{1,j}& \dots & y_{n,j}\end{array}\right)}^T$ denote different water uses (j = 1 to 6). By standardizing the original data, PCA constructs a covariance matrix and performs eigenvalue decomposition to extract eigenvalues and their associated eigenvectors. The eigenvectors corresponding to the two largest eigenvalues are then selected as the PCs, PC1 and PC2. To analyze the relationship between water use and seasons, the questionnaire dates were categorized into four seasons and used as attributes in the PCA. In this context, PCA used behavioral water use to differentiate household water use across seasons, which will be reflected in the distinct distributions along the PC axes. It reduced the dimensionality of household water use data and highlighted key differences across seasons.

Household water consumption is a complex nonlinear process influenced by both natural and human factors [50,51]. To address the complexities in this process, GP was used to model nonlinear relationships and construct interpretable mathematical models. GP leverages its capabilities to reduce uncertainties in climate prediction [52]. Collecting accurate data on household water use behavior is challenging because of its multifaceted influencing factors. Although the dataset used herein is small for machine learning applications, it is substantial for the household water use behavior research field. Additionally, the proposed fitting equation demonstrates GP’s ability to model nonlinear relationships and better captures interactions between variables compared with common methods such as ordinary least squares. The fitting equation can also directly distinguish the effect of a single variable on water use. Consequently, GP was used to model the interaction between household water use and its influencing factors.

The operational framework of GP can be visualized as a tree structure with nodes and branches, optimized using natural selection principles. The GP model’s structure was determined based on the specific structure of the household water use behavior and informed by previous literature [53,54], as illustrated in Figure 2. For accurate GP training, the dataset was divided into training and verification sets in a 4:1 ratio. To ensure generalizability, cross-validation was employed to assess model performance. During five-fold cross-validation, the seasonal sample proportions (spring/autumn, summer, and winter) were maintained to match those of the original dataset, to reflect the temporal characteristics of climate change. Derived from questionnaires and continuous observation experiments, the dataset for GP training includes independent variables (x_n,t) and dependent variables (y_n,j), as follows:

$$\left(\begin{array}{ccc}{y}_{\mathrm{1,1}}& \dots & y_{1,j}\\ ⋮& \ddots & ⋮\\ y_{n,1}& \dots & y_{n,j}\end{array}\right)=f\left(\begin{array}{ccc}{x}_{\mathrm{1,1}}& \dots & x_{1,t}\\ ⋮& \ddots & ⋮\\ x_{n,1}& \dots & x_{n,t}\end{array}\right)$$

where ${\left(\begin{array}{ccc}{y}_{1,j}& \dots & y_{n,j}\end{array}\right)}^T$ represents a dependent variable, with j = 1–6 denoting different water uses and j = 7–16 representing water use frequency and duration. ${\left(\begin{array}{ccc}{x}_{1,t}& \dots & x_{n,t}\end{array}\right)}^T$ represents influencing factors, including meteorological data, personal and family characteristics, and water use appliances, encompassing a total of 42 variables (t = 42). The GP model was trained to separately analyze the relationship between water use behavior and climate factors for each water use type. During training, all independent variables and the corresponding dependent variable were input simultaneously to generate a fitting equation for the dependent variable. Within the GP framework, f() represents a function combining independent variables using logical operations. During iterative processing, independent variables and the function set were combined to form equations, with their fitness verified against actual values (dependent variables) using a fitness function. Once the behavior fitting equation was obtained, the influence of meteorological factors on water use behavior was analyzed by controlling for other variables.

The first step in the GP process involved providing the independent variables and the function set as input. The initial population in the first generation was randomly generated to begin the iterative process. Each generation contained the same number of individuals, set to 5000 in this study. Each solution was represented as an equation in a tree structure (Figure 2). Fitness verification was performed at each step. If the fitness criteria were not met, offspring were generated through a reproduction process within the same generation. Offspring were produced via crossover, mutation, and copying. To reduce computational burden and accelerate iteration, a tournament size of 20 was randomly selected. The solution with the highest score in the tournament was recognized as the winner and advanced to the next generation. In the next generation, all solutions except the winner were randomly regenerated. Key parameters included a crossover rate of 0.85, a mutation rate of 0.1, and the following constraints: maximum depth = 12, parsimony coefficient = 0.01, stopping criterion = 0.01, and 200 generations. The function set was defined as (+, −, ×, ÷, √, log, abs, max, min) to support robust symbolic regression model development. This process was repeated until a solution meeting the fitness requirements was identified. Given the complexity and practicality of behavior fitting, the Pearson correlation coefficient (r) was selected as the fitness function. This coefficient is defined as

$$r={\displaystyle \frac{{\sum }_{i=1}^n\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{{\sum }_{i=1}^n{\left(X_i-\overline{X}\right)}^2}\sqrt{{\sum }_{i=1}^n{\left(Y_i-\overline{Y}\right)}^2}}}$$

(10)

where X_i and Y_i represent the equation’s fitted value and the corresponding dependent value, respectively; $\overline{X}$ and $\overline{Y}$ denote the mean values of the fitted values and corresponding dependent values; and n is the number of samples. The training process concludes when the fitness function r exceeds 0.95 for both the training and verification sets. Each water use type was simulated at least 50 times. Upon completion of training, the output was expressed as an equation combining the independent variables and the function. Fitting equations for water use behaviors (including meteorological factors) were developed, and by controlling for individual factors while considering the range and probability of other variables from the questionnaires, the specific impacts of meteorological factors were analyzed using the display equation.

GP and PCA were implemented in Python 3.10 using the gplearn and scikit-learn libraries within the Jupyter Notebook (Version: 7.2.1) environment.

3. Results

3.1. Household Water Use Under Varying Climatic Conditions

Based on seasonal classification, the influence of climatic factors on household water use patterns was first analyzed. Beijing experiences a temperate monsoon climate with notable seasonal differences, including cold, dry winters and hot, rainy summers. To investigate how climatic conditions influence household water use, PCA was conducted on daily water use data across the four seasons (Figure 3). Figure 3 includes the following two sections: the upper section presents kernel density curves, while the lower section illustrates the PCA results. Data points represent individual water use data collected via the questionnaire surveys, categorized by seasonal climate conditions (spring, summer, autumn, and winter). Circles denote the 95% confidence intervals (CIs) of the corresponding data. The X-axis and Y-axis correspond to the two PCs, with PC1 and PC2 explaining 30.0% and 18.2% of the total data variation, respectively. The curves reflect probability density estimates, with the X-axis presenting the range of household water use data and the Y-axis indicating the density estimates at these values.

Overall, 83% of the household water use data are relatively concentrated. Seasonal patterns highlight a clear separation in water use distribution between summer and the other seasons, while winter, spring, and autumn display certain similarities. In the kernel density curve, the summer data point with the highest density deviates by 108 L from the data points of other seasons. Meanwhile, the maximum water use at the highest density point in winter is 122 L. During summer, 25–75% of the data fall within 98.57–175.38 L, with values above this range being more dispersed. Individual water use exhibits the greatest variation in summer, followed by winter, whereas spring and autumn demonstrate relatively stable patterns. The CIs of spring and autumn water use overlap in the PCA results, indicating similar overall water use distributions. Therefore, in subsequent analyses, spring and autumn water use are treated as a single group.

Further analysis examined water usage across various water use types. Figure 4 illustrates seasonal variations in household water usage, including the contribution of different water use types. The top line of each column represents the mean value, the box indicates the 5–95% range of household water use, and the curve denotes data density.

Average total household water use is highest in summer, at 149.35 L/d, while the average annual personal water use is 141.2 L/d. According to the 2020 China Water Resources Bulletin, per capita water use was 135 L/d in that year. The error between the calculated water use and real data is less than 5%. The average water use difference between winter and the combined seasons of spring and autumn is approximately 4%. Of the water use types, the amount of water used for personal cleaning and showering demonstrates the greatest seasonal variation. In summer, 91% of people take showers more frequently than in winter, using an additional 19 L/d of water on average. Conversely, 98% of people perform personal cleaning more frequently in winter, averaging 8.6 times per day, resulting in additional consumption of 2.8 L/d compared with that in summer. Personal cleaning constitutes the highest proportion of water use in winter, while showering dominates in summer. The proportions of other water use types remain stable. The 5–95% water use range in Figure 4 indicates that summer demonstrates the largest variation, whereas winter presents the smallest variation, with a difference of nearly 50 L/d.

3.2. Correlation Analysis of Meteorological Factors and Water Use Behavior

Through repeated iterations and training of GP, fitting equations were developed for each water use type and meteorological factor. Notably, the frequency of occurrence of meteorological factors during the training process reflects their importance in shaping water use patterns. The five-fold cross-validation revealed that the average R² values of the GP model was 0.957 and 0.944 for the training and validation sets, respectively. The R² values for all validation folds exceeded 0.94, satisfying the predefined performance threshold.

Table 1. Frequency of occurrence for meteorological factors during genetic programming training.

Meteorological Factors	Frequency of Occurrence of Water Use Types						Comprehensive Occurrence Frequency (%)
	Personal Cleaning	Showering	Culinary Activities	Environmental Cleaning	Laundry	Flushing
WIND	100.00%	100.00%	100.00%	0.00%	100.00%	0.00%	66.67%
TASMIN	100.00%	100.00%	0.00%	0.00%	2.00%	54.00%	42.67%
TAS	100.00%	86.00%	2.00%	0.00%	4.00%	44.00%	39.33%
PR	70.00%	46.00%	0.00%	0.00%	12.00%	92.00%	36.67%
HURS	100.00%	12.00%	0.00%	100.00%	0.00%	6.00%	36.33%
TASMAX	2.00%	34.00%	4.00%	0.00%	2.00%	92.00%	22.33%
GT	14.00%	2.00%	0.00%	0.00%	2.00%	2.00%	3.33%

presents meteorological factors with occurrence frequencies exceeding 3%. Among these, temperature emerges as the most influential factor, with high frequencies observed for TASMIN, TAS, and TASMAX. WIND also demonstrates a strong association with several water use types, including personal cleaning, showering, culinary activities, laundry, and flushing. As the frequency of influencing factors and variations in water use may not align, a correlation analysis was performed between household water use types and major influencing factors with a frequency greater than 3% (Figure 5). In Figure 5, correlation coefficients are shown in the upper right of the diagonal, while the lower left uses color coding to depict the correlation strength and significance test results. W1, W2, W3, W4, W5, and W6 represent water use for personal cleaning, showering, flushing, laundry, environmental cleaning, and culinary activities, respectively, and W is the total household water use.

As depicted in Figure 5, the correlation coefficients between meteorological factors and water use types are all <0.5, likely owing to the nonlinear relationship between water use and meteorological factors. While the correlation coefficients between meteorological factors and total household water use are generally small, certain water use types exhibit strong correlations with specific meteorological factors. The strongest correlation was observed between TAS and personal cleaning (−0.48), with temperature emerging as the most influential meteorological factor overall, particularly TAS and TASMIN. TAS demonstrated the strongest correlations with personal cleaning, showering, and laundry, with correlation coefficients of −0.48, 0.19, and 0.17, respectively. TASMIN exhibited the highest correlation with environmental cleaning, with a correlation coefficient of −0.18. TAS and TASMIN demonstrated the strongest correlations with flushing and culinary activities, with correlation coefficients of 0.06 and 0.20, respectively. Most significance test results were satisfactory, except for HURS. The variability of HURS on a daily scale, which tends to be higher in winter and lower in summer, may explain the unsatisfactory results. In contrast, the significance test results for TAS and TASMIN were satisfactory.

3.3. Impact of Meteorological Factors on Household Water Use

The GP model was trained to elucidate the relationship between the use trends of different water use types and climatic factors (

Table 2. Quantification of water use trends for different water use types under varying meteorological factors.

Meteorological Factor	Variation Range	Change Trend Statistical Step Size	Changes in Daily Water Use Behavior Under Varying Meteorological Factors (L/Climate Factor Step Size)
			Personal Cleaning	Showering	Culinary Activities	Environmental Cleaning	Laundry	Flushing	Total
WIND	0 to 5 m/s	+1 m/s	−0.19 (±0.06)	+0.08 (±0.01)	−0.05 (±0.01)	-	−0.07 (±0.01)	-	−0.23 (±0.09)
TASMIN	−20 °C to 20 °C	+10 °C	−4.27 (±1.17)	+5.76 (±4.22)	-	-	−0.1 (±0.01)	+0.03 (±0.01)	+0.92 (±5.41)
TAS	−10 °C to 30 °C	+10 °C	−2.95 (±4.38)	+3.46 (±5.85)	+0.68 (±0.27)	-	−0.07 (±0.01)	+0.02 (±0.01)	+1.14 (±10.52)
PR	0 to 40 mm/day	+10 mm/day	−1.19 (±0.67)	+0.65 (±0.15)	-	-	−0.07 (±0.01)	+0.01 (±0.01)	−0.6 (±0.84)
HURS	10% to 90%	+10%	+0.25 (±0.04)	+0.65 (±0.21)	-	−0.65 (±0.23)	-	-	+0.25 (±0.48)
TASMAX	−10 °C to 40 °C	+10 °C	−0.77 (±2.26)	+1.32 (±5.49)	+0.27 (±0.25)	-	+0.05 (±0.01)	+0.33 (±0.41)	+1.2 (±8.42)
GT	−20 °C to 40 °C	+10 °C	−0.01 (±0.01)	+0.03 (±0.03)	-	-	-	-	+0.02 (±0.04)

). Using the display equation developed by GP fitting, the mechanism of the influence of a single meteorological factor on water use was determined by isolating individual meteorological factors while considering the range and probability of influence of other variables from the questionnaire survey. By incorporating meteorological factors into the equation fitted with the display GP, changes in the use trends of water use types influenced by individual meteorological factors were determined.

The results showed that meteorological factors have different effects on water use behavior. Of these factors, temperature exerted the greatest impact on total household water use, with a 10 °C increase in TASMAX resulting in a 1.2 (±8.42) L/d rise in water use nonlinearly. However, this influence was not uniform across all water use types. For instance, water use trends for culinary activities, environmental cleaning, laundry, and flushing exhibited minimal variation with changes in meteorological factors, whereas personal cleaning and showering revealed more substantial impacts. In particular, increases in temperature and precipitation led to greater water use for showering, increasing by 5.76 (±4.22) L/10 °C for TASMIN and 0.65 (±0.15) L/10 mm for PR. Interestingly, the water use trends for personal cleaning and showering in response to climate variables were largely opposite, likely owing to the interchangeable nature of these two water use types. Given the high variability in individual water use trends and differing responses to climate change, the dispersion of water use relative to the mean is substantial.

Based on the magnitude of water use variation in Table 2, meteorological factors with significant impacts on water use were identified to create trend charts of water use types with respect to meteorological factors (Figure 6). The shaded area in Figure 6 represents the 25–75% probability interval for meteorological factors in Beijing. Figure 6A–H represent the impact of individual meteorological factors on water use behavior.

The variation in mean values in Figure 6 closely reflects the data presented in Table 2. While all graphs display nonlinear trends, Figure 6G uniquely presents nonmonotonic changes. Additionally, all curves show amplitude variations, particularly when comparing ranges within and outside the 25–75% probability interval (Figure 6). The study further identified the range in which individual meteorological factors have the most severe impacts on water use behavior. When PR varies between 0 and 10 mm, TAS between 10 °C and 20 °C, and TASMIN between −10 °C and 0 °C, water use for personal cleaning is most severely affected. Similarly, when TAS varies between 0 °C and 10 °C and TASMIN between −20 °C and 10 °C, showering water use is the most severely affected.

In Figure 6E, variations are more pronounced within the 25–75% range than outside it, whereas in Figure 6H, the reverse is true. Figure 6G highlights monotonic changes in the use trends of water use types, with an observed extreme value. Notably, personal cleaning reaches its minimum use when TASMIN falls within the 0 °C–10 °C range. In Figure 6H, when TASMIN ranges from −20 °C to −10 °C, the increase in water use for showering is much greater than it is for TASMIN of −10 °C to 0 °C. As depicted in Figure 6F, when TAS reaches 30 °C, the lower extreme of the box is located within the box, suggesting that at least 5% of the data share the same value. This overlap indicates a potential shift in the mechanism driving showering behavior at this temperature.

3.4. Future Trends in Household Water Use Under CMIP6 Scenarios

The GP model was used to simulate future household water use trends in Beijing under typical IPCC SSP scenarios (Figure 7). As stated, climate parameters were derived from the following three CanESM5 scenarios: SSP1-2.6, SSP3-7.0, and SSP5-8.5. The shaded areas in Figure 7 indicate 95% CIs. Household water use is anticipated to increase across all scenarios due to global warming. The SSP1-2.6 scenario, characterized by lower greenhouse gas emissions and minimal climate change, predicts a slight increase in household water use. In this scenario, temperatures are expected to rise at a rate of 0.9 °C per decade until the 2050s. The mean annual household water use reported in the questionnaire was 141 L/d, slightly higher than the 139 L/d reported in the 2020 statistical bulletin. Forecasts of water use changes are based on the bulletin values, with household water use projected to increase by 0.2% per decade, reaching approximately 140 L/d. Showering water use is expected to drive 94% of this increase, rising by 1.8 L per decade. Beyond the 2050s, household water use is expected to stabilize at approximately 137.7 L/d.

Under SSP3-7.0 and SSP5-8.5, water use increase follows similar patterns, although with differing magnitudes. Notably, temperatures are projected to increase at rates of 0.6 °C and 0.7 °C per decade under SSP3-7.0 and SSP5-8.5, respectively. The frequency of extreme temperature events is expected to rise, with extreme hot days and low temperatures increasing by 415% and 459% under SSP3-7.0 and SSP5-8.5, respectively. Future increases in extreme temperature and abnormal weather events will lead to higher water use for personal cleaning and showering. Under SSP3-7.0 and SSP5-8.5, personal cleaning water use is projected to increase by 4.4 L and 4.5 L per decade, respectively, reaching approximately 58 L/d and 59 L/d by 2100. Showering water use is expected to increase by 2.6 L and 3.8 L per decade, respectively, reaching approximately 72 L/d and 81 L/d by 2100. Total water use under SSP3-7.0 and SSP5-8.5 is projected to increase by 2.7 L and 3.9 L per decade, respectively, reaching 160.3 L/d and 170.3 L/d by 2100. Based on Beijing’s projected population of 23 million by 2035, total household water use is expected to rise by 21–33%, equivalent to an additional 1.16–1.27 billion cubic meters, compared with 2020.

4. Discussion

4.1. Influence of Meteorological Factors on Personal Water Use Trends

Water is essential for maintaining physiological functions and ensuring hygiene, cleanliness, and temperature regulation. Daily behaviors are influenced by individual preferences and random factors. Tracking surveys and climate-related data [6] confirm a global increase in household water use. From a physiological perspective, rising temperatures affect drinking and cooking water use. For every 10 °C increase in TAS, culinary water use increases by 0.68 (±0.27) L/d. The effects of climate change on temperature regulation and cleaning water use are more pronounced. As temperatures rise, showering water use increases by 3.46 (±5.85) L/d, whereas personal cleaning water use decreases by 2.95 (±4.38) L/d for every 10 °C increase in TAS. Temperature significantly influences human comfort. A 10 °C increase in TASMAX results in a 0.05 (±0.01) L/d increase in laundry water use. HURS strongly affects the living environment, with a 10% increase reducing environmental cleaning water use by 0.65 (±0.23) L/d. The correlation between flushing water use and climate change is minimal, likely because behavior data collected on a daily scale do not reveal significant changes.

Daily household water use is a complex process influenced by factors such as personal and household characteristics, water use equipment, and climate. Studies across the world have demonstrated that socioeconomic variables, such as water tariffs, play a key role in shaping water use behavior. While domestic water prices remain relatively consistent across China, notable differences exist in the lifestyles of urban populations. Further research is needed to examine how socioeconomic and meteorological factors interact to influence household water use. Moreover, extreme weather events (e.g., heatwaves) are closely associated with peak household water consumption. While this study provides a preliminary analysis, future research should investigate how extreme events affect long-term trends in household water use. According to the 2020 China Water Resources Bulletin, per capita water use was 135 L/d. As a megacity experiencing long-term water scarcity, Beijing’s per capita water use is significantly lower than that of cities with abundant water resources, such as New York City (447 L/d), Tokyo (220 L/d), Guangzhou (177 L/d), and Shanghai (160 L/d). Cities with abundant water resources may experience greater variability in water use due to climate change compared with those without abundant resources. Therefore, this study carries substantial research and practical value for water-stressed megacities facing climate challenges, particularly those experiencing water scarcity conditions comparable to those of Beijing, offering new insights into urban water management under climate change [55]. Further city- and public location-scale research is necessary to ensure sustainable living standards for residents in densely populated cities. Household water use is also significantly influenced by other systems, particularly energy systems. This study examined the household water use–climate nexus as an initial step toward investigating the interconnected water use–energy–climate system at the household level.

4.2. Uncertainties in the Research

This study determined interpretable relationships between household water use and meteorological factors using a questionnaire-based survey. A large sample size, stratified sampling, and face-to-face tracking interviews reduced random errors and provided high-precision data on water use trends. However, some uncertainty remains in data acquisition. While questionnaires provided valuable insights, they introduced potential subjectivity and variability that can affect household water use estimates. Additionally, differences between long-term and temporary residents, along with regional cultural habits, could impact data representativeness.

To accurately analyze and predict the influence of meteorological factors on household water use, this study optimized model performance and verified its accuracy. Nonetheless, some uncertainty persists. This study employed the explainable machine learning algorithm GP to establish mathematical relations between household water use and climate factors. However, the results may be influenced by the number of training iterations and the choice of target indices. Optimization strategies and training frequency may also impact the model’s accuracy. Additionally, owing to GP’s unique structure, training outcomes may vary across sessions, significantly affecting the frequency statistics and subsequent analysis of meteorological factors. Given the substantial computational cost of GP, parameter optimization was not conducted. In the sensitivity analysis, only meteorological factors with frequencies exceeding 3% in the GP training results were considered. However, the combined effects of meteorological factors on household water use, along with their interactive mechanisms, require further investigation. When projecting future water demand, accounting for the inherent uncertainty of the GP model and the structural limitations of the climate projection model is essential. Although the CanESM5 model has errors, it performs better than many CMIP models [56].

The findings of this study are region-specific: the GP training equations are shaped by regional and situational factors, which remain consistent regardless of training iterations or the fitness function used. Household water use varies with regional influences and individual differences, making it challenging to fully align it with the preset conditions of SSP scenarios. In China, household water use is a basic necessity with limited direct impact from policy interventions. Nonetheless, the results of this study offer valuable insights for future sustainable water management planning. Although our study focused on Beijing, the impacts of climate change on human behavior are similar across the globe, and the quantification of such impacts is still in its infancy. This study provides annual-scale water use predictions based solely on meteorological factors, without accounting for other influential elements, seasonal dynamics, or extreme weather events. Further research with a broader scope is essential to enhance the understanding of these interactions.

GP is robust, simple, and flexible. Moreover, it has strong capabilities for solving complex nonlinear problems [57]. Artificial neural networks have been widely employed to address nonlinear problems; however, they are often regarded as black-box models [57,58], with gradient boosting techniques (e.g., XGBoost) being primarily used for classification and regression tasks. However, their interpretability diminishes when handling complex feature interactions [59]. Multiple regression models are typically used to capture linear relationships and predict continuous numerical values [60]. The transparency of the GP model enables researchers and practitioners to better understand features and make informed decisions. Unlike traditional linear regressions, GP for symbolic regression does not require a predefined model structure and can simultaneously search for the optimal model structure and its coefficients. However, this self-evolving process is stochastic; the same parameters may yield different structures, and the associated uncertainty requires further investigation.

4.3. Policy Impact and Suggestions for Water Supply Management

Climate change presents a global challenge that demands action. Although the scientific community agrees on its catastrophic impacts, international policy responses remain diverse. For example, Europe relies on legal frameworks and financial instruments such as the Recovery and Resilience Facility; Lithuania emphasizes carbon neutrality through legislative measures (e.g., green taxation), whereas Bulgaria focuses on the structural transformation of the energy system. By contrast, China primarily adopts national strategic planning supported by various local regulations and indicators, with an emphasis on engineering solutions and a lack of coordinated governance [61,62,63]. Research on climate change responses at the household level plays a critical role in mitigation efforts [64]. This study identifies key meteorological thresholds that can inform adaptive water management strategies, including Internet of Things (IoT)-enabled dynamic pricing during heatwaves, to enhance conservation and system efficiency. Furthermore, real-time meteorological data collected via IoT systems can support behavioral interventions, such as issuing warning messages to encourage residents to conserve water, particularly during periods of extreme heat or cold.

The urban water supply system is a critical component of city infrastructure. This study proposes technical solutions to support the development of innovative smart water management systems. By integrating the key thresholds identified in this study with real-time data on temperature, a dual-threshold smart management system can be designed to enable intelligent water supply from the household level to the network scale [65]. Combining household water use prediction models with spectral analysis enhances water usage evaluation in buildings [66]. Moreover, regional behavioral patterns should be incorporated into planning. Through PCA-based spatial clustering, infrastructure investments can be prioritized in areas beyond temperature thresholds, particularly by strengthening pipe cooling infrastructure in regions with high shower demand.

5. Conclusions

To quantitatively analyze the relationship between climate factors, household habits, and individual water use, this study adopted a systematic approach involving the construction of numerical models and data collection. This methodology quantifies the impact of meteorological factors on urban residential water use patterns, enabling the exploration of previously unknown influences on urban household water use behavior. In this research, the explainable machine learning algorithm GP was employed to link questionnaire responses about household water use patterns with meteorological data, identify key meteorological factors influencing water use, and predict future residential water use trends.

This study used Beijing as a representative megacity in China to quantify household water use behavior under climate change, identify the underlying nonlinear response mechanisms, determine key meteorological thresholds, and propose dynamic water tariff adjustments alongside infrastructure optimization policies. By quantifying the relationship between climate and household behavior at the microlevel, this study addresses a critical gap and provides an empirical basis for cross-jurisdictional policy harmonization. The limitations of data regionalization and GP model parameter optimization in this paper still need to be further considered in the subsequent research. The main findings are summarized as follows.

Household water use in Beijing exhibited distinct seasonal variations. Correlation analysis revealed a clear association between specific water use types and meteorological factors, although the correlation between total water use and meteorological factors was less pronounced. Temperature changes influenced water use behaviors, with a complementary relationship observed between personal cleaning and showering. Water use behaviors under meteorological influences were nonlinear and nonmonotonic. Temperature emerged as the most significant determinant of water use. TASMAX had the most pronounced effect, increasing water use for showering by 1.32 (±5.49) L/d and reducing water use for personal cleaning by 0.77 (±2.26) L/d, resulting in a net increase of 1.2 (±8.42) L/d in total water use for every 10 °C rise. In terms of water usage behavior, showering and personal cleaning accounted for the largest proportion of water use and were most affected by meteorological factors. Furthermore, the study identified the range in which individual meteorological factors had the most severe impacts on water use behavior. When PR varied between 0 and 10 mm, TAS between 10 °C and 20 °C, and TASMIN between −10 °C and 0 °C, a significant change in personal cleaning water use was recorded. Similarly, when TAS varied between 0 °C and 10 °C and TASMIN between −20 °C and 10 °C, showering water use exhibited the most substantial change.

CanESM5 (CMIP6) was utilized to predict changes in personal water use in Beijing. Under scenarios SSP1-2.6, SSP3-7.0, and SSP5-8.5, water use demonstrated an overall increasing trend. SSP1-2.6 exhibited the slowest growth, stabilizing at 137.7 L/d, whereas SSP5-8.5 showed the sharpest increase, averaging 3.9 L/d per decade and reaching 170.3 L/d by 2100. By 2035, considering Beijing’s projected population of 23 million, total household water use is expected to rise by 21–33%, equivalent to an additional 1.16–1.27 billion cubic meters, compared with 2020. These findings highlight significant challenges for ensuring urban water security, emphasizing the need for sustainable urban water management strategies to account for the impacts of climate change. The ranges of meteorological factors that affected household water use most severely can provide insights into peak urban water use and regional energy consumption. This study serves as a reference for assessing, evaluating, and predicting household water use under varying climate conditions worldwide. The proposed methodology can be adapted to other regions, although regional differences must be carefully considered when selecting household water use influencing factors and designing GP models. By acquiring a clearer understanding of how meteorological factors influence household water use, water resource managers and policymakers can make more informed decisions about urban sustainable water resource allocation and conservation strategies.