Evaluation of Industrial Water Use Efficiency on an Enterprise Scale Based on Analytic Hierarchy Process, Entropy Weight Method and Self-Organizing Map: A Case Study in Zhejiang, China

Abstract

The increasingly serious imbalance between the supply and demand of water resources necessitates the establishment of a scientific and reasonable comprehensive evaluation method for industrial water use efficiency (WUE). In this study, a general method for industrial WUE evaluation on an enterprise scale was proposed by combining the analytic hierarchy process (AHP), entropy weight method (EWM), and self-organizing map (SOM), and it was tested in several areas of Zhejiang Province, China. The results show that the composite indexes generated using the AHP and EWM were different and were employed as the input of the SOM to divide enterprises into four categories. Most enterprises were classified as Class A, with a relatively high WUE, accounting for 82.5% of the total, while those in Class D, with a relatively low WUE, only accounted for 0.5% of the total. Furthermore, the differences in WUE for industry classification and spatial distribution were also analyzed. The classification results of several industries were more diverse, especially for those industries in which water plays an important role in production. Moreover, the spatial distribution of WUE classifications also implied that the clustering of enterprises has a positive effect on the improvement in WUE. In other words, it is feasible to improve WUE through industry clustering and sub-industry management. In summary, a comprehensive, detailed evaluation of industrial WUE was conducted on an enterprise scale, which can also be applied to other areas and used as a reference for local water resource managers for formulating targeted policies.

1. Introduction

With rapid socio-economic development and population growth, as well as changes in water consumption patterns, global water demand has increased annually in the past 40 years, especially in urban areas [1,2]. Unfortunately, as one of the world’s most important natural resources, the amount of water is limited globally, and the increasing demand for water has led to crises in many countries and areas [3,4,5]. Furthermore, due to uneven development between regions and industries, water use efficiency (WUE) varies among different regions and industries, and the low WUE in several regions and industries is aggravating the imbalance between the supply of and demand for water resources [6,7,8]. Therefore, it is urgently necessary to analyze and evaluate the WUE between regions and industries and formulate corresponding water resource management policies based on this evaluation and analysis.

Although the WUE in different regions and industries has been studied over recent decades, there is still no unified and comprehensive method to evaluate and analyze it [9,10]. The water use per capita, gross domestic product (GDP), and industrial added value (IAV) are the most commonly used indicators to evaluate WUE [11,12]. However, all of these indicators only reflect one aspect of WUE. The water use per capita ignores the impact of economic development, and the water use per GDP or IAV is affected by the prices of different products, and therefore does not truly reflect the WUE in different regions and industries. An available alternative is to establish a comprehensive index based on multiple indicators, which has been employed in several studies [13,14,15]. For example, Bai et al. (2017) evaluated the WUE of Bohai Bay, which is a typical water-deficient city agglomeration in China [16]. Pan et al. (2020) analyzed the changes in WUE from 2006 to 2015 in Shandong Province [17]. Moncaleano et al. (2021) assessed the influence of contextual and behavioral factors on WUE [18]. Most of the above studies used subjective or objective weighting methods to establish a comprehensive index. However, both subjective and objective weighting methods have disadvantages. Subjective weighting methods rely entirely on expert experience and subjective judgment, while objective weighting methods solely focus on the data, which may lead to biased results in WUE evaluation [19,20]. Therefore, it is necessary to balance expert opinions and data for a comprehensive assessment of WUE. Another issue worth discussing is the results of WUE evaluation research. Most of the former studies were conducted based on water use data and socio-economic data from administrative divisions, ignoring the differences in WUE across industries in different areas. A detailed evaluation of the WUE of enterprises is still lacking. In addition, traditional WUE classifications are based on specific threshold values set by experts, which highly rely on the experts’ experience and subjective preferences [21]. Accordingly, greater effort is still needed to conduct more detailed WUE investigations and develop a new general method that does not rely solely on expert opinions for evaluating and classifying the WUE of enterprises across various industries and regions, which is crucial for formulating targeted and effective water resource management strategies.

The development of machine learning, especially unsupervised learning methods, provides a new opportunity for evaluating and analyzing WUE in different regions and industries [22,23]. Unsupervised learning methods can mine valuable information and patterns from data without any pre-labeling or labeling, which makes them an efficient technique in various research fields, such as classification, dimensionality reduction, and association rule learning. The self-organizing map (SOM), as a special form of self-organizing neural network, uses competition mechanisms to map complex data to a two-dimensional or three-dimensional grid structure, which can intuitively show the topological relationships and clustering structure between input data [24]. In many previous studies, authors have pointed out that the SOM has great performance in both high-dimensional and low-dimensional classification problems, providing a new and reliable approach for WUE evaluation and classification [25].

In this study, examples of the socio-economic and water use data of industrial enterprises from various industries were collected. A new classification model was developed based on the analytic hierarchy process (AHP), entropy weight method (EWM), and SOM to divide enterprises’ WUE into several categories. Additionally, the impact of industries and regions on WUE was also analyzed at the enterprise scale. The results of this study contribute to providing a new method to evaluate the WUE of enterprises and can be used as a reference for local governments to formulate water resource management policies and refined guidance for enterprises with low WUE. An overview of the study area and data sources is presented in the next section. The methodology of this study is detailed in Section 3. Section 4 and Section 5 set out the results and discussion, respectively, and finally, the conclusions are provided in Section 6.

2. Materials

2.1. Study Area

Zhejiang Province, located in northeastern China, is one of the most developed areas in the country, with a large population and a high concentration of industrial areas. Due to the influence of topographic factors, socio-economic development in Zhejiang Province is quite uneven. Therefore, five of the most developed county-level administrative areas in Zhejiang Province, which can provide enough enterprise samples for the analysis of WUE, were selected as the study area, namely Wenling, Pinghu, Jiashan, Changxing, and Linhai. According to the statistical data from the Chinese government, all 5 were among the top 100 counties in China in 2023, ranking 22nd, 36th, 40th, 48th, and 65th, respectively. The total population, GDP, and IAV of the five counties in 2023 were 45.8 million people, CNY 505.4 billion, and CNY 218.4 billion, respectively. The location and digital elevation model (DEM) of the study area is shown in Figure 1. Detailed socio-economic data for the counties are presented in

Table 1. Basic information about the study areas.

Area Name	Population (Million Persons)	GDP (Billion Yuan)	IAV (Billion Yuan)
Wenling	14.4	135.1	43.3
Pinghu	6.9	100.9	55.6
Jiashan	6.6	90.8	47.9
Changxing	6.8	89.4	39.8
Linhai	11.1	89.2	31.8
Total	45.8	505.4	218.4

.

As typical coastal plain areas affected by a monsoon climate, the five counties receive a large amount of annual precipitation, leading to dense rivers and abundant water resources. However, due to the uneven temporal distribution of water resources and the lack of large-scale water storage engineering, these areas have experienced several water resource crises in the past. With the rapid economic and population growth in the region, the increasing water consumption, especially industrial water consumption, has further exacerbated the problem of insufficient water resources, posing a huge challenge to water resource management. Therefore, it is vital to establish a comprehensive industrial enterprise WUE evaluation model to provide a reference for improving the WUE of industrial enterprises and ultimately solving the problem of the disproportionate supply of and demand for water resources.

2.2. Data Sources

In this study, all the data of enterprises were collected from local government reports, including the names, annual GDP, annual IAV, annual total water use (TWU), land areas (LA), and industry classification and address in 2024. An explanation of each variable used in this study is presented in

Table 2. Overview of variables used in this paper.

Variable Name	Abbreviation	Explanation	Unit
Gross domestic product	GDP	The total value of goods or services produced by an enterprise in 2024.	Yuan
Industrial added value	IAV	The market value added in the production of an enterprise in 2024.	Yuan
Total water use	TWS	The total amount of water used by an enterprise in 2024, including the water used for production and employees’ living.	m3
Land areas	LA	The land area of an enterprise.	km2
Industry classification	/	The industry classification of an enterprise in the standards of the Chinese government.	/

. One of the main advantages of the data used in this study is that they were collected by local governments from enterprises, which ensured the authenticity of the data and that the dataset fully covered the local enterprises. In addition, a series of methods were employed to preprocess the data. Firstly, enterprises with an annual TWU lower than 10,000 cubic meters were excluded, as these may be affected by observational or statistical errors due to the low value. Then, the industry of each enterprise was determined based on the industrial classification for national economic activities from the national standards of China, which divides economic activities into 20 broad sections and 97 detailed divisions. More detailed information about these sections and divisions can be found in

Table A1. The industrial classification of national economic activities.

Section		Division
Code	Name	Code	Name
A	Agriculture, forestry, husbandry, and fishing	01	Agriculture
		02	Forestry
		03	Husbandry
		04	Fishing
		05	Agriculture, forestry, husbandry, and fishery professional and supporting activities
B	Mining	06	Mining and washing of coal
		07	Extraction of oil and gas
		08	Mining and processing of ferrous metal
		09	Mining and processing of non-ferrous metal
		10	Non-metallic mineral mining and processing
		11	Mining professional and supporting activities
		12	Other mining
C	Manufacturing	13	Processing of agricultural sideline food
		14	Manufacture of food
		15	Manufacture of wine, beverages, and tea
		16	Manufacture of tobacco products
		17	Manufacture of textiles
		18	Manufacture of wearing apparel
		19	Manufacture of leather, fur, feathers, and related products; manufacture of footwear
		20	Processing of wood; manufacture of wood, bamboo, rattan, palm, and straw products
		21	Manufacture of furniture
		22	Manufacture of paper and paper products
		23	Printing and reproduction of recorded media
		24	Manufacture of culture, education, arts, crafts, sports, and entertainment products
		25	Processing of petroleum, coal, and other fuel
		26	Manufacture of chemical raw materials and chemical products
		27	Manufacture of pharmaceuticals
		28	Manufacture of chemical fiber
		29	Manufacture of rubber and plastic products
		30	Manufacture of other non-metallic mineral products
		31	Smelting and calendaring of ferrous metal
		32	Smelting and calendaring of nonferrous metal
		33	Manufacture of metal products
		34	Manufacture of general equipment
		35	Manufacture of special equipment
		36	Manufacture of automobile
		37	Manufacture of railway, shipbuilding, aerospace, and other transportation equipment
		38	Manufacture of electrical machinery and equipment
		39	Manufacture of computers, communications, and other electronic equipment
		40	Manufacture of instrumentation
		41	Other manufacturing
		42	Comprehensive utilization of waste resources
		43	Repair of metal products, machinery, and equipment
D	Electricity, heat, gas and water production and supply	44	Electricity, heat power generation and supply
		45	Manufacture of gas and gas supply
		46	Water collection, treatment, and supply
E	Construction	47	Construction of buildings
		48	Civil engineering
		49	Construction installation
		50	Building decoration, renovation, and other construction
F	Wholesale and retail trade	51	Wholesale trade
		52	Retail trade
G	Transportation, warehousing and postal	53	Railway transport
		54	Road transport
		55	Water transport
		56	Air transport
		57	Pipeline transportation
		58	Multimodal transport and transportation agency
		59	Loading, unloading, handling, and warehousing
		60	Postal
H	Accommodation and food service activities	61	Accommodation
		62	Food and beverage service activities
I	Information transmission, software and information technology service	63	Telecommunications, radio and television and satellite transmission activities
		64	Internet and related activities
		65	Software and information technology activities
J	Financial activities	66	Monetary financial service activities
		67	Capital Market service activities
		68	Insurance
		69	Other financial activities
K	Real estate activities	70	Real estate activities
L	Leasing and business activities	71	Leasing activities
		72	Business service activities
M	Scientific research and technical activities	73	Research and experimental development
		74	Professional and technical activities
		75	Technology promotion and application activities
N	Water conservancy, environment and public facilities management	76	Water conservancy management
		77	Ecological protection and environmental management
		78	Public facilities management
		79	Land management
O	Residential services, repairs and other services	80	Residential service activities
		81	Repair of motor vehicles, electronic products, and daily products
		82	Other service activities
P	Education	83	Education
Q	Human health and social work activities	84	Human health activities
		85	Social work activities
R	Culture, sports and entertainment	86	News and publishing activities
		87	Production of radio, television, film, and sound recordings
		88	Culture and art activities
		89	Sports activities
		90	Entertainment activities
S	Public administration, social security and social organizations	91	Organs of the Communist Party of China activities
		92	National Agencies activities
		93	People’s Political Consultative Conference, democratic party activities
		94	Social security activities
		95	Mass organizations, social groups, and other member organizations activities
		96	Grassroots mass autonomous organizations and other organizations activities
T	International organizations	97	International organizations activities

in Appendix A. Finally, the spatial locations of the enterprises were obtained by using the geographic coding method based on their addresses.

3. Methods

3.1. Technical Process

In this study, a new general model based on the AHP, EWM, and SOM was developed to comprehensively evaluate the WUE of industrial enterprises in different industries and areas. The model has three main components: the construction of WUE indicators, comprehensive weight calculation, and WUE classification. The technical process of this model is shown in Figure 2.

3.2. Construction of WUE Indicators

Three different WUE indicators based on GDP, IAV, LA, and TWU were employed to build a comprehensive index and evaluate the WUE of different enterprises, namely water consumption per output value (WCO), water consumption to generate a certain unit of added value (WCAV), and water consumption per square kilometer of land (WCL). The three indicators are introduced below:

(1)

WCO

The WCO refers to the water consumption required to generate a certain unit of output value. In this study, the ratio between TWU and GDP of the enterprises was used to estimate the WCO:

$$WCO={\displaystyle \frac{TWU}{GDP}}$$

(1)

(2)

WCAV

The WCAV refers to the water consumption required to generate a certain unit of added value. The ratio between TWU and IAV of the enterprises was used to estimate the WCAV:

$$WCAV={\displaystyle \frac{TWU}{IAV}}$$

(2)

(3)

WCL

The WCL refers to the water consumption per square kilometer of land occupied by water-using enterprises:

$$WCL={\displaystyle \frac{TWU}{LA}}$$

(3)

To eliminate the influence of dimensionality, all three indicators were standardized before building the WUE evaluation model.

$$y_{i,j}={\displaystyle \frac{x_{i,j}-{\overline{x}}_j}{{\sigma }_j}}$$

(4)

where $x_{i,j}$ is the value of the $i$th sample for the $j$th variable, $\overline{x_j}$ is the mean of the $j$th variable, ${\sigma }_j$ is the standard deviation of the $j$th variable, and $y_{i,j}$ is the standardized result for $x_{i,j}$.

3.3. Construction of Comprehensive Index of WUE

3.3.1. Analytic Hierarchy Process

The AHP is one of the most commonly used subjective weighting methods for constructing a comprehensive index [26,27]. Rather than prescribing a “perfect” set of weights, the AHP provides a comprehensive and rational framework to help researchers find the weight of each element that best suits their purpose and allows them to understand the problem. The central aim of the AHP is to decompose an abstract problem into a hierarchy of more easily comprehended sub-problems, each of which can be analyzed independently. According to the hierarchical structure, the relative importance of each variable in each hierarchy was determined using a pairwise comparison based on expert opinions. A detailed guide to building the AHP model can be found in Appendix B. However, as a subjective weighting method, the reliability of the AHP model is influenced by the rationality of expert opinions; in particular, when the number of experts is small, some unreasonable judgment matrices may lead to wrong weights.

3.3.2. Entropy Weight Method

The EWM is an objective weighting method based on information entropy [28,29]. In information theory, entropy refers to the degree of chaos or disorder in a system, i.e., entropy is a measure of uncertainty. The information entropy of a certain random variable $X$ can be derived as

$$H(x)=-{\displaystyle \sum _{x\in K}{P}_X}(x)\log P_X(x)$$

(5)

where $H\left(x\right)$ is the information entropy for a specific value $x$ of random variable, i.e., X = x, $K$ is the value domain of $X$, and $P_{X }\left(x\right)$ is the probability mass of $X$.

According to the definition of information theory, an indicator with a high level of entropy has a large degree of information confusion and uncertainty, which means that the indicator can only provide very limited information, and its weight should be small. Therefore, EWM is widely used as an objective weight estimation method that does not involve any expert experience. However, there are still some defects that are difficult to overcome. Firstly, each indicator is independent in EWM, which may ignore the mutual influence between the indicators. Additionally, the weights generated with EWM are completely dependent on data, which may lead to weight distortion; that is, the calculated weight from EWM cannot truly reflect the importance of the indicators. A feasible way to solve the above problems is to combine EWM with other subjective weight methods. More detailed information on building an EWM model can be found in Appendix C.

3.3.3. Construction of Comprehensive Index

Once the weights for the indicators are determined with AHP or EWM, the comprehensive index of WUE can be obtained with

$$S_j={\displaystyle \sum _{i=1}^n{D}_{i,j}{w}_i}$$

(6)

where $S_j$ is the comprehensive index for the $j$th sample, $D_{i,j}$ is the value of the $i$th indicator from the $j$th sample, $w_i$ is the weight for the $i$th indicator from AHP or EWM, and $n$ is the number of indicators.

3.4. Self-Organizing Map

The SOM is an unsupervised learning artificial neural network algorithm that maps high-dimensional data with complex and nonlinear relationships to a low-dimensional space with simple geometric structures and relationships through competitive learning, displaying the topological and clustering relationships between data [30].$X$ is the input samples, $M$ is the number of samples, $N$ is the dimension of the input sample data, and $\mathrm{P}$ is the number of neurons in the neural network. The main steps in the SOM are as follows:

Step 1: Network initialization.

Randomly initialize each neuron in the network. The weight vector of each neuron $j$ is $W_j$, which contains $N$ weight values, as seen in Equation (7):

$$W_j={\left[w_{j1},w_{j2},\dots \dots ,w_{jN}\right]}^T$$

(7)

Step 2: Select the best matching unit.

The Euclidean distance between the $i$th sample and the weight vector of the $j$th neuron is calculated, and the neuron with the smallest Euclidean distance is selected as the best matching unit.

$$d_{i,j}={{\displaystyle \sum _{k=1}^N(x_{i,k}-w_{j,k})}}^2$$

(8)

where $d_{i,v}$ is the Euclidean distance between the $i$th sample and the weight vector of the $j$th neuron, $x_{i,k}$ is the $k$th variable of the $i$th sample, and $w_{j,k}$ is the $k$th weight value of the $j$th neuron.

Step 3: Determine the neighborhood range.

Taking the best matching unit as the center, the neurons in the circle with radius $r$ are determined as the best matching neighborhood.

Step 4: Update and iteration.

Assuming that there are $Q(Q\le P)$ neurons in the circle, the weight vectors of the neurons in the neighborhood range are updated by using the learning rate.

$$W_q(s+1)=W_q(s)+\alpha \times h(d_{q,u})\times (X_i-W_q(s)),q=1,2,\cdots Q$$

(9)

where $W_q\left(s\right)$ is the weight vector of the $q$th neuron in the neighborhood range at the $s$th iteration, $\alpha$ is the learning rate, $d_{q,u}$ is the Euclidean distance between the $W_q\left(s\right)$ and the weight vector of the best matching unit, and $h(·)$ is the kernel function, which can be written as seen in Equation (10):

$$h(x)=e^{-x^2/2r^2}$$

(10)

where $r$ is the radius of the best matching neighborhood range.

Furthermore, it is worth noting that the input data of SOM were preprocessed. The input of the SOM was normalized to ensure that the maximum of each input variable was 1 and the minimum was 0. The input data for SOM were the comprehensive WUE indexes of the enterprises based on AHP and EWM, which were both non-beneficial indicators. The normalization equation can be found in Equation (A6) in Appendix C.

4. Results

4.1. Overview of Data

Since a variety of data from different industrial enterprises and areas were employed, a comprehensive analysis of the data employed in this study was conducted as the first step. After the preprocessing introduced in Section 2.2, the data from 629 enterprises in 30 different industries were selected. Figure 3 shows the distribution of industries for the selected enterprises. It is worth noting that the identification code (ID) of each industry is based on the industrial classification for national economic activities from the national standard of China, which can be found in Table A1 (Appendix A). As shown in Figure 3, the industry distribution of enterprise samples from various areas was not uniform. It is obvious that there were several dominant industries with significantly larger numbers in Wenling and Changxing, while the industry distribution of other areas was more even. Furthermore, the number of enterprise samples in most industries exceeds 10 (22 out of a total of 30 industries), which provided a solid data foundation for subsequent study.

Figure 4 shows the empirical cumulative distribution function (ECDF) of the four variables, i.e., TWU, LA, GDP, and IAV. For convenience of comparison, all the data used are standardized. The ranges for the four variables are between −2 and 14, and most of them are concentrated within the range of [−1, 1]. The values of TWU are the most concentrated, while LA is the most discrete.

The correlation coefficients (CCs) between the variables were also analyzed, and the results are presented in Figure 5. The CC between GDP and IAV was higher than 0.8, while the others were between 0.4 and 0.6. Since the samples used in this study were all from industrial enterprises, there were similarities in the definitions of GDP and IAV, leading to the high CC between the two variables. However, according to the definition of the two variables, their connotations were not similar; the IAV refers to the newly added value in industrial production activities, while GDP refers to the value of all products in industrial production. Therefore, the two variables were both employed in this study as input indicators for WUE classification.

In summary, the data used in this study cover a large number of enterprises from different industries and scales, offering great representation and providing a solid foundation for subsequent WUE research.

4.2. Construction of Comprehensive WUE Index

The three indicators, i.e., WCO, WCAV, and WCL, for each enterprise were calculated using the above data. Figure 6 presents a histogram of the three indicators. Although all three indicators reflect WUE from different perspectives, their value ranges and distributions are different. Therefore, it is necessary to build a comprehensive index based on these indicators.

To construct the comprehensive index of WUE, the first step was to calculate the weights based on the AHP and EWM. For the AHP, due to the limited number of indicators in this study, a simple two-layer structure was employed, as shown in Figure 7.

For the AHP, expert opinions were collected through a survey questionnaire distributed among local government managers and water resource experts. The pairwise comparison between the three indicators was determined using Saaty’s 1–9 scale, which can be found in

Table A2. Saaty’s 1–9 scale of pairwise comparison.

Intensity of Importance	Verbal Judgment of Preference	Explanation
1	Equal importance	Two activities contribute equally to the objective.
3	Weak importance	Experience and judgment slightly favor one activity over another.
5	Strong importance	Experience and judgment strongly favor one activity over another.
7	Very strong importance	An activity is favored very strongly over another.
9	Extreme importance	The judgment favoring one activity over another is of the highest possible order of affirmation.
2,4,6,8	Intermediate values between the two adjacent judgements	/

in Appendix B. According to the results of the survey questionnaire, the judgment matrix can be determined as

Table 3. Judgment matrix of indicator layer on target layer.

Judgment Matrix	WCO	WCAV	WCL
WCO	1	2	7
WCAV	1/2	1	5
WCL	1/7	1/5	1

.

The consistency index (CI) of the matrix was 0.0071, and the consistency ratio (CR) was calculated as 0.0122, which was much smaller than the threshold of 0.1, meaning that the judgment matrix passed the consistency test. Finally, the weights for WCO, WCAV, and WCL based on the AHP were 0.5917, 0.3332, and 0.0751, respectively.

The weights of the three indicators calculated using EWM were based entirely on the input data. The weights of WCO, WCAV, and WCL were 0.2252, 0.3623, and 0.4125, respectively. The weights generated using the AHP and EWM are presented in

Table 4. Weights of variables from AHP and EWM.

Method	Weights of Variables
	WCO	WCAV	WCL
AHP	0.5917	0.3332	0.0751
EWM	0.2252	0.3623	0.4125

. The weights obtained with the two methods were different, especially for WCO and WCL. The main reason for this phenomenon is that the weights from the AHP were based on expert opinions, which may be biased according to the experts’ experience and preferences, while the weights from the EWM fully rely on the distribution of the data, which ignores the actual meaning of the variables. In addition, the comprehensive WUE indexes based on the AHP and EWM were calculated. They are illustrated in Figure 8, which shows that the distribution of the comprehensive WUE index from the AHP was more discrete, while that of the index from the EWM was more concentrated between 0 and 0.5. Furthermore, the different distributions of the two comprehensive WUE indexes also indicated that the WUE evaluation results for enterprises were different based on the subjective and objective weighting methods. Therefore, it is necessary to develop a classification model for comprehensively evaluating the WUE of different enterprises.

4.3. WUE Classification of Enterprises

The WUE classification model for enterprises was established using the SOM model based on the comprehensive indexes from the AHP and EWM. In this study, the number of nodes in the SOM was set to four, and the learning rate was set to 0.01. The results of the SOM model are shown in Figure 9.

According to the classification results from the SOM, the WUE values of the enterprises were divided into four different categories. As an unsupervised classification model, the classification results directly generated using the SOM only indicate that these enterprises should be classified into one class based on WUE rather than directly reflecting the level of WUE. In this study, the distribution of the two comprehensive WUE indexes from the AHP and EWM for each class was used to indicate the WUE of that category. As shown in Figure 10, the classification of enterprise samples located in the upper-right corner had the worst comprehensive WUE indexes from both the AHP and EWM, which was named Class D. With the decreases in comprehensive WUE indexes, the middle part of the enterprise samples was divided into two categories, Class B and Class C. Finally, the enterprise samples with the lowest comprehensive WUE indexes in the lower-left corner were named Class A. In other words, the WUE decreases from Class A to Class D sequentially. It can also be found in Figure 10 that the majority of the enterprises selected in this study had reasonably high WUE. More specifically, the number of enterprises classified as Class A was the highest, at 520, accounting for 82.7% of the total, while only three enterprises were classified as Class D, and there was a big gap between Class C and Class D. Although the classification results indicated that most enterprises in the study areas had good performance with regard to WUE, there were still some enterprises with reasonably poor WUE, especially the three enterprises in Class D. Local water resource managers can develop targeted policies based on the above classification results, such as tiered water prices and water saving rewards. For example, local government can give a low water price or rewards to the enterprises belonging to Class A, while the enterprises in Class D may need to be regulated or even penalized.

5. Discussion

5.1. WUE in Different Industries

The detailed data collected from industrial enterprises in this study provide a strong basis for analyzing WUE in terms of industries. Due to the uneven distribution of industries in the enterprise samples, only 14 industries with more than 20 samples were selected for analysis. The statistical results of WUE classification in different industries are presented in Figure 11 and Table 5. As shown in this figure and table, more than 85% of enterprises in most of the selected industries were classified as Class A, which exceeds the overall average (82.3%). However, there are four industries that merit special attention, namely ID 17 Manufacture of textiles, ID 18 Manufacture of wearing apparel, ID 22 Manufacture of paper and paper products, and ID 26 Manufacture of chemical raw materials and chemical products. Although no enterprises in these four industries were classified as Class D, the proportion of enterprises belonging to Class A was significantly lower, while the proportion belonging to Class B and Class C was higher than for other industries. According to the specific definition of these four industries in terms of national economic activities, the main reasons for their notability can be summarized as follows:

(1)

For ID 17 Manufacture of textiles and ID 18 Manufacture of wearing apparel, water is an essential raw material throughout the entire production process, used for rinsing, printing, and dyeing, etc.

(2)

The production of paper and paper products requires combining pulp fibers suspended in water through various processes to produce paper sheets. Therefore, water, as a production medium, is vital for ID 22 Manufacture of paper and paper products.

(3)

Water is mainly used as a coolant material in ID 26 Manufacture of chemical raw materials and chemical products.

In summary, the common characteristic of these four industries is that a large amount of water is necessary for their production. In other words, the production of these four industries is closely related to water consumption. Due to the large differences in WUE classifications among these four industries, it is essential to formulate corresponding WUE management policies for them. These results also suggest that industries in which water plays an important role in production are more sensitive to changes in WUE.

Figure 11. WUE classifications of enterprises for the selected industries.

Figure 11. WUE classifications of enterprises for the selected industries.

Table 5. The number and proportion of enterprises with different WUE classifications in various industries.

Industry ID	Total	Class A		Class B		Class C		Class D
		Count	Proportion	Count	Proportion	Count	Proportion	Count	Proportion
ID13	26	22	84.6%	4	15.4%	0	0.0%	0	0.0%
ID17	44	20	45.5%	16	36.4%	8	18.2%	0	0.0%
ID18	23	15	65.2%	5	21.7%	3	13.0%	0	0.0%
ID21	23	22	95.7%	1	4.3%	0	0.0%	0	0.0%
ID22	23	17	73.9%	3	13.0%	3	13.0%	0	0.0%
ID26	29	18	62.1%	7	24.1%	4	13.8%	0	0.0%
ID29	24	20	83.3%	4	16.7%	0	0.0%	0	0.0%
ID30	45	40	88.9%	5	11.1%	0	0.0%	0	0.0%
ID33	51	46	90.2%	5	9.8%	0	0.0%	0	0.0%
ID34	55	55	100.0%	0	0.0%	0	0.0%	0	0.0%
ID35	26	21	80.8%	4	15.4%	1	3.8%	0	0.0%
ID36	36	35	97.2%	1	2.8%	0	0.0%	0	0.0%
ID38	44	41	93.2%	3	6.8%	0	0.0%	0	0.0%
ID39	35	31	88.6%	4	11.4%	0	0.0%	0	0.0%

Table 5. The number and proportion of enterprises with different WUE classifications in various industries.

Industry ID	Total	Class A		Class B		Class C		Class D
		Count	Proportion	Count	Proportion	Count	Proportion	Count	Proportion
ID13	26	22	84.6%	4	15.4%	0	0.0%	0	0.0%
ID17	44	20	45.5%	16	36.4%	8	18.2%	0	0.0%
ID18	23	15	65.2%	5	21.7%	3	13.0%	0	0.0%
ID21	23	22	95.7%	1	4.3%	0	0.0%	0	0.0%
ID22	23	17	73.9%	3	13.0%	3	13.0%	0	0.0%
ID26	29	18	62.1%	7	24.1%	4	13.8%	0	0.0%
ID29	24	20	83.3%	4	16.7%	0	0.0%	0	0.0%
ID30	45	40	88.9%	5	11.1%	0	0.0%	0	0.0%
ID33	51	46	90.2%	5	9.8%	0	0.0%	0	0.0%
ID34	55	55	100.0%	0	0.0%	0	0.0%	0	0.0%
ID35	26	21	80.8%	4	15.4%	1	3.8%	0	0.0%
ID36	36	35	97.2%	1	2.8%	0	0.0%	0	0.0%
ID38	44	41	93.2%	3	6.8%	0	0.0%	0	0.0%
ID39	35	31	88.6%	4	11.4%	0	0.0%	0	0.0%

5.2. WUE in Different Areas

Since the enterprise samples of this study were collected from different areas of Zhejiang Province, another interesting topic is the relationship between WUE classification and spatial distribution. In this study, two industries, namely ID 17 Manufacture of textiles and ID 18 Manufacture of wearing apparel, were selected as examples to analyze the differences in WUE between areas. The main reasons for selecting these two industries are as follows:

(1)

According to the above results, the production of these two industries has a close relationship with water consumption, and their WUE classifications are also somewhat diverse.

(2)

These two industries are the dominant industries in two areas, i.e., the number of enterprises from these two industries is the largest in the local area. Specifically, ID 17 Manufacture of textiles is the dominant industry in Changxing, accounting for 27.4% of the total number of enterprise samples, and ID 18 Manufacture of wearing apparel is the dominant industry in Pinghu, accounting for 10.3% of the total number of enterprise samples.

According to the classification results, the number of enterprise samples in Class A in these two industries was significantly higher than the average, 78.6% and 77.8%, respectively (

Table 6. The number and proportion of enterprises with different WUE classifications for ID 17 in Changxing and ID 18 in Pinghu.

Industry ID	Area	Total	Class A		Class B		Class C
			Count	Proportion	Count	Proportion	Count	Proportion
ID17	Changxing	14	11	78.6%	2	14.3%	1	7.1%
	Total	44	20	45.5%	16	36.4%	8	18.2%
ID18	Pinghu	18	14	77.8%	3	16.7%	1	5.6%
	Total	23	15	65.2%	5	21.7%	3	13.0%

). These results indicate that the clustered industries have a positive effect on WUE.

5.3. Limitation and Future Work

In this study, the WUE of a large number of enterprises was analyzed and evaluated through a multi-model methodology. As with all scientific research, this study has several limitations. First, due to the limitations in data availability, the study area and sample size were relatively small, and not all the industrial enterprises were involved in this study, which means that the specific classification standard used here cannot be directly applied to other areas or industries. Secondly, the temporal variation in WUE was not analyzed in this study. Due to the limitations in the availability of data, only three WUE indicators from four variables were employed in this study. Furthermore, although the results imply that the clustering of enterprises may improve regional WUE, a thorough analysis of the influence of enterprise clustering on WUE is still lacking, which is an interesting topic that needs more attention in future. However, it is still meaningful to evaluate the WUE from the current dataset to provide a reference for local water resource managers for formulating targeted policies. In addition, the method proposed in this article is general, and it can also be used to evaluate WUE based on other datasets collected from different areas and industries. Researchers can also select other indicators of WUE as the input for this model according to their own research objectives, rather than employing WCO, WCAV, and WCL as in this paper. In other words, our method is flexible and can be used to evaluate and compare the WUE across different spatiotemporal resolutions or industries based on different input indicators, as well as provide a reference for identifying the main influencing factors of WUE. It is feasible to apply this method for the evaluation and comparison of WUE at a larger scale or even a global scale in future research.

With the advent of the big data era, increasing global attention has been paid to the collection and analysis of different data, providing an opportunity for future research on WUE in terms of a larger spatiotemporal scale. In addition, the rapid development of deep learning has also provided a series of new approaches for unsupervised classification. Although it is interesting to employ more innovative approaches to better distinguish the WUE of different enterprises, industries, or regions, a comprehensive comparison between these unsupervised classification models is not included in this article. Moreover, although an obvious gap between the weights obtained through the AHP and EWM was found in this study, no further analysis was conducted, which is another direction worth exploring in future. Finally, an attribution analysis of the differences in WUE between the enterprises from the same industry or different industries is not included in this article. More efforts are still needed in future to identify the main causes of WUE difference.

6. Conclusions

With rapid economic development and population growth, water shortages pose a major global challenge. As many previous studies have pointed out, an effective way to improve the WUE of a region or industry is by reducing water consumption while ensuring socio-economic development and solving the current water crisis. However, due to the difficulties in data collection and unified analysis methods, there is still a lack of WUE evaluation at the enterprise scale. In this article, a general evaluation method for WUE was proposed based on AHP, EWM, and SOM, and the method was tested in several areas in Zhejiang Province, China. The results show that the data collected in this study were quite comprehensive, comprising 629 enterprise samples from 17 different industrial industries in six areas. The different results from AHP and EWM indicate that there is a gap between the weights determined using subjective and objective methods. Therefore, it is necessary to develop a more comprehensive method to evaluate WUE. An SOM was employed to classify the results from AHP and EWM for the evaluation of WUE. Four different classifications were established based on WUE, named, from high to low, Class A, Class B, Class C, and Class D. Of the 629 enterprise samples, 519 were Class A enterprises, 83 were Class B enterprises, 24 were Class C enterprises, and 3 were Class D enterprises. Additionally, the differences in WUE classification in terms of industry and area also suggest that the gap in WUE was more obvious in industries where production was closely related to water consumption, and the clustering of enterprises had a positive effect on improving WUE. In conclusion, the method developed in this study has proven to be an effective tool for WUE evaluation and classification and can be flexibly applied to research objects from different areas or industries, which is helpful for formulating targeted water resource management policies to improve WUE and is feasible for widespread application in the future.