A water resource carrying capacity prediction method based on a water resource supply and demand prediction model

Abstract

The application discloses to belong to the technical field of water resources management, and particularly relates to a water resources carrying capacity prediction method based on a water resources supply and demand prediction model, which comprises the following steps: collecting economic, water supply and hydro-meteorological data of a research area; constructing an input-output model, combining with structural decomposition analysis to identify virtual water core influencing factors; constructing a water resources supply and demand deep learning prediction model, inputting influencing factor data under different future scenarios, and outputting water resources supply and demand prediction values; using the supply and demand prediction values and future social and economic data to calculate water resources supply and demand indexes and concentrated carrying indexes, and integrating the two indexes into a water resources carrying capacity index through an entropy weight method. The application realizes accurate prediction of regional water resources supply and demand and dynamic quantification of water resources carrying capacity under multiple social economy-climate scenarios, can accurately identify core driving factors of water resources carrying capacity, and provides technical support for sustainable allocation of water resources in arid areas and coordinated development of social economy and ecological environment.

Description

Technical Field

[0001] This invention belongs to the field of water resource management technology, and specifically relates to a method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model. Background Technology

[0002] Water resources are a core foundation for achieving the UN Sustainable Development Goals and ensuring regional water security. With rising water demand driven by socio-economic development and shrinking water resources due to climate change and human activities, the contradiction in water resource allocation is becoming increasingly prominent. Many regions, especially arid areas, have long faced the challenge of synergistic development between socio-economic development and water resources under the constraint of water scarcity. Water resource carrying capacity assessment is a key means to ensure the sustainable use of water resources and support long-term regional development. It can be used to determine whether the pressure of socio-economic development exceeds the local water resource carrying capacity, identify core constraints, and provide pathways for water-scarce regions to solve resource allocation problems. Therefore, there is an urgent need for new methods adapted to socio-economic and climate change to accurately diagnose water supply and demand imbalances and quantify water resource carrying capacity. Existing research has proposed various water resource carrying capacity assessment methods, such as supply and demand balance analysis, entropy weight TOPSIS, and fuzzy comprehensive evaluation. Some studies have also combined scenario analysis, deep learning, and hydrological models to conduct related assessments. While virtual water analysis is significant for water resource demand management, it is rarely incorporated into a comprehensive economic dimension water resource carrying capacity assessment framework.

[0003] Input-output analysis is a common method in virtual water analysis, capable of accurately tracking direct and indirect water flows in the industrial chain, revealing virtual water transfer and spatial spillover effects. Combined with structural decomposition analysis, it can also isolate the influence of various driving factors. However, it lacks the ability to predict future water resource carrying capacity and is difficult to adapt to dynamic scenario simulations. In recent years, deep learning frameworks integrating convolutional neural networks, long short-term memory networks, and attention mechanisms have demonstrated strong spatiotemporal prediction advantages, yet they are rarely applied to multi-regional dynamic water resource carrying capacity simulations. Existing water resource carrying capacity research mostly focuses only on direct water use and employs static assessment frameworks, failing to comprehensively characterize the impact of virtual water and the temporal evolution of carrying capacity, and also struggling to accurately identify core driving factors. There is a significant gap in research combining input-output-structural decomposition analysis with deep learning models.

[0004] Therefore, there is an urgent need for a water resource carrying capacity prediction method based on a water resource supply and demand forecasting model, to carry out dynamic quantitative research on water resource carrying capacity that integrates whole-economy virtual water analysis and deep learning technology, so as to provide technical support for improving the reliability, accuracy and applicability of the assessment and supporting the sustainable allocation of regional water resources. Summary of the Invention

[0005] The purpose of this invention is to propose a method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model, comprising the following steps:

[0006] Step S1: Obtain economic and hydrological / meteorological data for the study area; calculate the virtual water transfer characteristics of each industry sector along the supply chain based on input-output analysis; decompose the changes in virtual water transfer characteristics into multiple driving factors using structural decomposition analysis, and select the core driving factor set.

[0007] Step S2: Process the core driving factor set, hydrological and meteorological data, and socio-economic time series data, align them according to the time dimension, and construct structured input data containing time dimension, regional dimension, and driving factor dimension; based on the structured input data, form a multi-driving factor multi-regional spatial matrix under each time slice to characterize the spatial distribution relationship between multiple regions and the structural correlation relationship between driving factors;

[0008] Step S3: Input the spatial matrix into a convolutional neural network to extract features from the regional spatial distribution characteristics and the correlation structure between driving factors represented in the spatial matrix; input the extracted spatial features into a long short-term memory network in chronological order for time series modeling; introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series to obtain a water resource supply and demand prediction model that integrates regional spatial heterogeneity, the correlation structure of driving factors, and the dynamic features of the time series.

[0009] Step S4: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios, output water resource supply and demand forecasting results, and calculate water resource carrying capacity index based on the forecasting results to achieve dynamic assessment of water resource carrying capacity in multiple regions and scenarios.

[0010] The economic data include: GDP, regional annual input-output table, consumer price index, consumption, import and export and investment data; the hydrological and meteorological data include: water consumption and total water resources in the regional water resources bulletin, regional historical runoff, precipitation, wind speed, pressure, humidity, solar radiation and temperature data.

[0011] The virtual water transfer characteristics of each industry sector along the supply chain, calculated based on input-output analysis, include:

[0012] Data standardization processing: unify the classification of industrial sectors in the input-output table within the research area, map the water use data and economic data to the industrial sectors one by one, and construct an industrial economic water use matching dataset;

[0013] Construction of basic coefficient matrix: Based on the input-output table, calculate the direct consumption coefficient matrix and the total consumption coefficient matrix to quantify the economic interdependence between upstream and downstream sectors of the industrial chain.

[0014] Water consumption calculation by dimension: direct water consumption and total water consumption are calculated for each department; the direct water consumption is the original water consumption in the industrial production process; the total water consumption is the direct water consumption plus the implicit water consumption indirectly transmitted from upstream and downstream of the industrial chain;

[0015] Virtual water transfer quantitative modeling: Taking total water consumption as the core and combining inter-sectoral product trade flow, a multi-regional and multi-industry virtual water flow matrix is ​​constructed to accurately characterize the input, output, and net transfer direction and scale of virtual water along the supply chain between regions and industries, forming a standardized time series dataset of virtual water transfer characteristics.

[0016] Time-series feature anchoring: Output virtual water transfer total amount, structure and spatial distribution characteristics by annual time slices, providing quantitative base data for subsequent structural decomposition analysis.

[0017] The structural decomposition analysis was used to decompose the changes in virtual water transfer characteristics into multiple driving factors, and the core driving factor set was obtained by screening:

[0018] Decomposition Model Construction: Using the time-series changes in virtual water transfer characteristics output by the IOA as the target decomposition variable, an additive structural decomposition analysis model is constructed to establish a quantitative mapping relationship between virtual water changes and driving factors;

[0019] Multi-dimensional driving factor full decomposition: The temporal fluctuations of virtual water transfer characteristics are precisely decomposed into three independent and quantifiable driving factors, including:

[0020] Water intensity drivers: the impact of changes in water efficiency per unit of industrial output on virtual water transfer;

[0021] Industrial structure drivers: characterizing the impact of regional industrial proportion adjustments and industrial chain structure upgrades on virtual water transfer;

[0022] Final demand drivers: representing the impact of changes in the scale and structure of end-user demand, including consumption, investment, and exports, on virtual water transfer;

[0023] Quantitative calculation of driving factor contribution: Calculate the absolute contribution value and relative contribution rate of each driving factor to virtual water change on a time slice and region-by-region basis, and quantitatively distinguish between dominant and secondary factors;

[0024] Targeted screening of core driving factors: Set a contribution rate threshold, remove weakly influential factors with a contribution rate below the threshold, retain factors that play a decisive role in virtual water transfer and changes in water supply and demand, and solidify a set of core driving factors with fixed dimensions and clear physical meaning;

[0025] Model input dimension binding: The core driving factor set is directly defined as the exclusive input dimension of the subsequent deep learning model, replacing the undifferentiated full data input, and completing the rigid connection from the IOA-SDA screening results to the model input structure.

[0026] Step S2 specifically includes:

[0027] Step S21: Obtain the core driving factor set, hydrological and meteorological data, and socio-economic time series data, and uniformly represent the core driving factor set, hydrological and meteorological data, and socio-economic time series data as driving factor data with time and regional identifiers;

[0028] Step S22: Standardize the driving factor data to eliminate dimensional differences between different data.

[0029] Step S23: Align the standardized driving factor data according to the time dimension to construct multi-period time series data, so that each time point corresponds to multiple driving factor data in each region;

[0030] Step S24: Based on the multi-period time series data, construct three-dimensional structured input data including time dimension, regional dimension and driving factor dimension;

[0031] Step S25: Slice the three-dimensional structured input data according to the time dimension, extract the data of the corresponding region dimension and driving factor dimension under each time slice, and form a multi-driving factor multi-region spatial matrix.

[0032] Step S3 specifically includes:

[0033] Step S31: Input the spatial matrix into a convolutional neural network for spatial feature extraction;

[0034] Step S32: Input the extracted spatial features into the Long Short-Term Memory network in chronological order for temporal modeling;

[0035] Step S33: Introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series.

[0036] The convolutional neural network is used to extract spatial heterogeneity features between different regions in the spatial matrix and spatial correlation features between driving factors.

[0037] The Long Short-Term Memory (LSTM) network is used to model the long-term dependence and dynamic evolution characteristics of water resource supply and demand over time.

[0038] The weights in the attention mechanism are initialized and constrained based on the contribution of driving factors obtained from structural decomposition analysis.

[0039] Step S4 specifically includes:

[0040] Step S41: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios to obtain the water resource supply forecast and water resource demand forecast for each region within the corresponding time range.

[0041] Step S42: Based on the predicted water supply and water demand, calculate the water supply and demand index for each region in different time periods;

[0042] Step S43: Based on the predicted water supply and water demand, calculate the concentrated carrying capacity index of each region to characterize whether the water resource utilization exceeds a preset threshold.

[0043] Step S44: Based on the water resource supply and demand index and the concentrated carrying capacity index, calculate the water resource carrying capacity index and realize dynamic assessment of water resource carrying capacity under multiple regions, multiple scenarios and long-term series.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention, based on a water resource supply and demand forecasting model, enables more accurate calculation of virtual water flow across the entire economic chain and identification of key drivers of virtual water. Simultaneously, it efficiently captures the nonlinear spatiotemporal dynamic characteristics of water resource supply and demand, thereby achieving accurate prediction of regional water resource supply and demand under multiple socioeconomic and climate scenarios and dynamic quantification of water resource carrying capacity. It effectively explores the differentiated impacts of different scenarios on water resource carrying capacity in arid regions, providing a powerful tool for solving the problem of water resource supply and demand imbalance in arid regions, quantifying water resource carrying capacity, and supporting the sustainable socioeconomic and ecological development of arid regions. Specifically, it includes:

[0046] (1) By using input-output analysis to calculate the virtual water transfer along the supply chain of each industrial sector and combining it with structural decomposition analysis to decompose the virtual water change, water resource consumption can be upgraded from a single water consumption statistic to a structured representation across the industrial chain, thereby realizing the characterization of the transmission path of water resources in the economic system. At the same time, by decomposing the virtual water change into driving factors such as water intensity, industrial structure and final demand, the key factors affecting water resource change can be identified from the mechanism level, avoiding the problem of relying solely on historical data for empirical analysis, thereby improving the interpretability and reliability of the model input data.

[0047] (2) By using a convolutional neural network to extract features from a spatial matrix composed of multiple regions and multiple driving factors, it is possible to effectively capture the spatial heterogeneity between different regions and the correlation structure between driving factors. Furthermore, by using a long short-term memory network to perform time-series modeling of the above spatial features, it is possible to depict the dynamic evolution of water resource supply and demand over a long time scale. On this basis, an attention mechanism is introduced to weight the key time nodes and key driving factors in the time series, so that the model can focus on the key features that have a greater impact on the prediction results, thereby improving the prediction accuracy and model stability.

[0048] (3) By introducing the driving factors and their contribution relationships obtained from structural decomposition analysis into the attention mechanism, the attention weights are guided or constrained, so that when the model performs feature weighting, it no longer completely relies on the data-driven learning process, but distinguishes the degree of influence of different driving factors in combination with economic mechanism information, so that the attention allocation is more in line with the actual water resource change pattern. This method can effectively reduce the model's sensitivity to noisy data, improve the identification ability of key driving factors, and enhance the stability and interpretability of prediction results. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to the present invention.

[0050] Figure 2 A graph showing the contribution of factors to virtual water changes in an embodiment of the present invention;

[0051] Figure 3 The graph shows the changes in training loss and Nash coefficient of the degree learning model in this embodiment of the invention with the number of training rounds; where (a) is the model training in Inner Mongolia, (b) is the model training in Shaanxi, (c) is the model training in Ningxia, and (d) is the model training in the three provinces of Inner Mongolia, Shaanxi and Ningxia.

[0052] Figure 4 This is a graph showing the predicted changes in water consumption and water intensity under different scenarios according to an embodiment of the present invention;

[0053] Figure 5 This is a graph showing the predicted changes in available water resources under different scenarios according to an embodiment of the present invention.

[0054] Figure 6 This is a map showing the predicted changes and distribution of water resource carrying capacity under different scenarios according to an embodiment of the present invention;

[0055] Figure 7 This is a data flow diagram of the water resource carrying capacity prediction method based on a water resource supply and demand prediction model, according to an embodiment of the present invention. Detailed Implementation

[0056] This invention provides a method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model. The invention will be further described in detail below with reference to the accompanying drawings.

[0057] like Figure 1 As shown, an embodiment of the present invention discloses a method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model, comprising the following steps:

[0058] Step S1: Obtain economic and hydrological / meteorological data for the study area; calculate the virtual water transfer characteristics of each industry sector along the supply chain based on input-output analysis; decompose the changes in virtual water transfer characteristics into multiple driving factors using structural decomposition analysis, and select the core driving factor set.

[0059] Step S2: Process the core driving factor set, hydrological and meteorological data, and socio-economic time series data, align them according to the time dimension, and construct structured input data containing time dimension, regional dimension, and driving factor dimension; based on the structured input data, form a multi-driving factor multi-regional spatial matrix under each time slice to characterize the spatial distribution relationship between multiple regions and the structural correlation relationship between driving factors;

[0060] Step S3: Input the spatial matrix into a convolutional neural network to extract features from the regional spatial distribution characteristics and the correlation structure between driving factors represented in the spatial matrix; input the extracted spatial features into a long short-term memory network in chronological order for time series modeling; introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series to obtain a water resource supply and demand prediction model that integrates regional spatial heterogeneity, the correlation structure of driving factors, and the dynamic features of the time series.

[0061] Step S4: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios, output water resource supply and demand forecasting results, and calculate water resource carrying capacity index based on the forecasting results to achieve dynamic assessment of water resource carrying capacity in multiple regions and scenarios.

[0062] This invention discloses a method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model. By coupling input-output analysis with deep learning to construct a water resource supply and demand forecasting model, it can more accurately calculate virtual water flow across the entire economic chain, identify key driving factors of virtual water, and efficiently capture the nonlinear spatiotemporal dynamic characteristics of water resource supply and demand. This enables accurate prediction of regional water resource supply and demand and dynamic quantification of water resource carrying capacity under multiple socioeconomic and climate scenarios. It effectively explores the differentiated impacts of different scenarios on water resource carrying capacity in arid regions and is a powerful tool for solving the problem of water resource supply and demand imbalance in arid regions, quantifying water resource carrying capacity, and supporting the sustainable development of socioeconomic and ecological environment in arid regions.

[0063] The water resource supply and demand forecasting model includes: input-output analysis (IOA), structural decomposition analysis (SDA), and a deep learning forecasting model for water resource supply and demand (CNN-LSTM-Attention, CLA). By combining the economic mechanism information provided by IOA and SDA with the spatiotemporal modeling capabilities of CLA based on convolutional neural networks, long short-term memory networks, and attention mechanisms, an integrated modeling process is achieved, encompassing "virtual water structure analysis—driving factor extraction—spatial feature modeling—temporal dynamic prediction—carrying capacity assessment." This significantly improves the interpretability and engineering applicability of the results while ensuring the model's predictive accuracy.

[0064] The following is an explanation of each sub-step.

[0065] Step S1: Obtain economic and hydrological / meteorological data for the study area; calculate the virtual water transfer characteristics of each industry sector along the supply chain based on input-output analysis; decompose the changes in virtual water transfer characteristics into multiple driving factors using structural decomposition analysis, and select the core driving factor set.

[0066] The economic data include: GDP, regional annual input-output table, consumer price index, consumption, import and export and investment data; the hydrological and meteorological data include: water consumption and total water resources in the regional water resources bulletin, regional historical runoff, precipitation, wind speed, pressure, humidity, solar radiation and temperature data.

[0067] The virtual water transfer characteristics of each industry sector along the supply chain, calculated based on input-output analysis, include:

[0068] Data standardization processing: unify the classification of industrial sectors in the input-output table within the research area, map the water use data and economic data to the industrial sectors one by one, and construct an industrial economic water use matching dataset;

[0069] Construction of basic coefficient matrix: Based on the input-output table, calculate the direct consumption coefficient matrix and the total consumption coefficient matrix to quantify the economic interdependence between upstream and downstream sectors of the industrial chain.

[0070] Water consumption calculation by dimension: direct water consumption and total water consumption are calculated for each department; the direct water consumption is the original water consumption in the industrial production process; the total water consumption is the direct water consumption plus the implicit water consumption indirectly transmitted from upstream and downstream of the industrial chain;

[0071] Virtual water transfer quantitative modeling: Taking total water consumption as the core and combining inter-sectoral product trade flow, a multi-regional and multi-industry virtual water flow matrix is ​​constructed to accurately characterize the input, output, and net transfer direction and scale of virtual water along the supply chain between regions and industries, forming a standardized time series dataset of virtual water transfer characteristics.

[0072] Time-series feature anchoring: Output virtual water transfer total amount, structure and spatial distribution characteristics by annual time slices, providing quantitative base data for subsequent structural decomposition analysis.

[0073] The structural decomposition analysis was used to decompose the changes in virtual water transfer characteristics into multiple driving factors, and the core driving factor set was obtained by screening:

[0074] Decomposition Model Construction: Using the time-series changes in virtual water transfer characteristics output by the IOA as the target decomposition variable, an additive structural decomposition analysis model is constructed to establish a quantitative mapping relationship between virtual water changes and driving factors;

[0075] Multi-dimensional driving factor full decomposition: The temporal fluctuations of virtual water transfer characteristics are precisely decomposed into three independent and quantifiable driving factors, including:

[0076] Water intensity drivers: the impact of changes in water efficiency per unit of industrial output on virtual water transfer;

[0077] Industrial structure drivers: characterizing the impact of regional industrial proportion adjustments and industrial chain structure upgrades on virtual water transfer;

[0078] Final demand drivers: representing the impact of changes in the scale and structure of end-user demand, including consumption, investment, and exports, on virtual water transfer;

[0079] Quantitative calculation of driving factor contribution: Calculate the absolute contribution value and relative contribution rate of each driving factor to virtual water change on a time slice and region-by-region basis, and quantitatively distinguish between dominant and secondary factors;

[0080] Targeted screening of core driving factors: Set a contribution rate threshold, remove weakly influential factors with a contribution rate below the threshold, retain factors that play a decisive role in virtual water transfer and changes in water supply and demand, and solidify a set of core driving factors with fixed dimensions and clear physical meaning;

[0081] Model input dimension binding: The core driving factor set is directly defined as the exclusive input dimension of the subsequent deep learning model, replacing the undifferentiated full data input, and completing the rigid connection from the IOA-SDA screening results to the model input structure.

[0082] Step S2: Process the core driving factor set, hydrological and meteorological data, and socio-economic time series data, align them according to the time dimension, and construct structured input data containing time dimension, regional dimension, and driving factor dimension; based on the structured input data, form a multi-driving factor multi-regional spatial matrix under each time slice to characterize the spatial distribution relationship between multiple regions and the structural correlation relationship between driving factors;

[0083] Step S2 involves constructing the Input-Output Model (IOA).

[0084] Step S2 specifically includes:

[0085] Step S21: Obtain the core driving factor set, hydrological and meteorological data, and socio-economic time series data, and uniformly represent the core driving factor set, hydrological and meteorological data, and socio-economic time series data as driving factor data with time and regional identifiers;

[0086] Step S22: Standardize the driving factor data to eliminate dimensional differences between different data.

[0087] Step S23: Align the standardized driving factor data according to the time dimension to construct multi-period time series data, so that each time point corresponds to multiple driving factor data in each region;

[0088] Step S24: Based on the multi-period time series data, construct three-dimensional structured input data including time dimension, regional dimension and driving factor dimension;

[0089] Step S25: Slice the three-dimensional structured input data according to the time dimension, extract the data of the corresponding region dimension and driving factor dimension under each time slice, and form a multi-driving factor multi-region spatial matrix.

[0090] In this embodiment, based on the collected economic and hydrological and meteorological data of the study area, an input-output model of the study area is established, and the departmental virtual water is calculated by combining water use data. The core influencing factors of virtual water in each region are determined by combining structural decomposition analysis.

[0091] Step S3: Input the spatial matrix into a convolutional neural network to extract features from the regional spatial distribution characteristics and the correlation structure between driving factors represented in the spatial matrix; input the extracted spatial features into a long short-term memory network in chronological order for time series modeling; introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series to obtain a water resource supply and demand prediction model that integrates regional spatial heterogeneity, the correlation structure of driving factors, and the dynamic features of the time series.

[0092] Step S3 specifically includes:

[0093] Step S31: Input the spatial matrix into a convolutional neural network for spatial feature extraction;

[0094] Step S32: Input the extracted spatial features into the Long Short-Term Memory network in chronological order for temporal modeling;

[0095] Step S33: Introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series.

[0096] The convolutional neural network is used to extract spatial heterogeneity features between different regions in the spatial matrix and spatial correlation features between driving factors.

[0097] The Long Short-Term Memory (LSTM) network is used to model the long-term dependence and dynamic evolution characteristics of water resource supply and demand over time.

[0098] The weights in the attention mechanism are initialized and constrained based on the contribution of driving factors obtained from structural decomposition analysis.

[0099] In this embodiment, a deep learning prediction model for water resource supply and demand is established based on the core economic factors of virtual water and the core meteorological factors of water resource supply in each region, as well as the corresponding historical data. For the prediction of dynamic water resource supply and demand in multiple regions, the CNN-LSTM-Attention model, the CNN-LSTM model, and the LSTM model will all be used as candidate models for regional prediction. The performance of the models is compared by comparing the R-squared, RMSE, NSE, and MAE indicators to select the optimal prediction model. The water resource supply and demand factor data under different future scenarios are then input, and the predicted values ​​of water resource supply and demand are finally output.

[0100] The specific process is as follows:

[0101] Convolutional Neural Networks (CNNs), as the core model of deep learning, utilize local perception and weight sharing mechanisms to automatically extract high-dimensional data features through convolutional kernels, avoiding manual feature selection. They employ a dense front-end and sparse back-end structure, effectively mitigating the gradient decay problem during backpropagation. The CNN model structure consists of a series of components: convolutional layers, pooling layers, fully connected layers, and an output layer. Its computation process is as follows:

[0102] ;

[0103] ;

[0104] In the formula, o(i,j) is the feature value output by the convolution operation at position (i,j); L(m,n) is the weight parameter of the convolution kernel at position (m,n), used to capture the local correlation of the input data; Z(i+m, j+n) is the original value of the input data at position (i+m,j+n); i is the row index of the output feature value; j is the column index of the output feature value; m is the row offset of the convolution kernel; n is the column offset of the convolution kernel; o k These are the cell values ​​corresponding to the output matrix; x i It is a pooling region.

[0105] LSTM is a type of neural network with powerful nonlinear capabilities. The LSTM structure consists of three key components: a memory adjustment unit, a real-time response unit, and an output generation unit. This process can be represented as:

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] In the formula, i t , f t O t , g t C t and h t These are the input gate, forget gate, output gate, candidate cell state, and cell state, respectively; x t is the input data at time t; W is the weight matrix, whose subscripts indicate the connection method; b is the bias vector, whose subscripts correspond to the gating type; It is the Sigmoid activation function; tanh is the hyperbolic tangent activation function.

[0113] Attention is a widely used technique in deep learning that enables deep learning models to weight important parts of the input data, focusing on the information most relevant to the current goal. The calculation process is as follows:

[0114] ;

[0115] ;

[0116] In the formula, Q represents the query vector; K represents the key vector; V represents the value vector; d k Let Q, K, and V represent the dimensions. The similarity function F(Q,K) calculates a sample-specific attention score, which is then normalized using softmax to derive a time weighting factor.

[0117] Step S4: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios, output water resource supply and demand forecasting results, and calculate water resource carrying capacity index based on the forecasting results to achieve dynamic assessment of water resource carrying capacity in multiple regions and scenarios.

[0118] Step S4 specifically includes:

[0119] Step S41: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios to obtain the water resource supply forecast and water resource demand forecast for each region within the corresponding time range.

[0120] Step S42: Based on the predicted water supply and water demand, calculate the water supply and demand index for each region in different time periods;

[0121] Step S43: Based on the predicted water supply and water demand, calculate the concentrated carrying capacity index of each region to characterize whether the water resource utilization exceeds a preset threshold.

[0122] Step S44: Based on the water resource supply and demand index and the concentrated carrying capacity index, calculate the water resource carrying capacity index and realize dynamic assessment of water resource carrying capacity under multiple regions, multiple scenarios and long-term series.

[0123] In this embodiment, the output water supply and demand forecasts and socio-economic data under future scenarios are used to calculate two sub-indicators of water resource carrying capacity, namely the water resource supply and demand index and the concentrated carrying capacity index. Finally, the two sub-indicators are unified into the water resource carrying capacity index according to the entropy weight method.

[0124] The specific process is as follows:

[0125] The forecast results and socioeconomic data are input into the water resource carrying capacity assessment method to determine the future water resource carrying capacity (CCWR). Specifically, this involves combining future water supply and demand with population and economic development to calculate and assess the capacity, using the following formula:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] In the formula, IWSD and CCI represent the water resource surplus / deficit situation and whether water resource utilization has exceeded the critical point, respectively. w1 and w2 are weights, calculated using the entropy weight method. Specifically, if IWSD is greater than zero, it indicates that water resources are in a surplus state; conversely, it indicates a shortage state. Furthermore, if CCI is greater than zero, it indicates that CI is at a high level; conversely, it indicates that CI is at a low level.

[0131] In this embodiment, a water resource carrying capacity prediction method based on a water resource supply and demand forecasting model is applied. Input-output tables, economic data, water use data, water supply data, and hydrological and meteorological data for the study area are collected. An input-output model is constructed based on this data to calculate departmental virtual water and identify core influencing factors of virtual water through structural decomposition analysis. Based on these core influencing factors and historical data, CNN-LSTM-Attention, CNN-LSTM, and LSTM candidate deep learning models are constructed. The optimal model is selected through performance index comparison. The model is then input with influencing factor data under different future scenarios and outputs predicted water resource supply and demand values. Using the predicted supply and demand values ​​and future socioeconomic data, a water resource supply and demand index and a concentrated carrying capacity index are calculated. These are then integrated into a water resource carrying capacity index using the entropy weight method. This method achieves accurate prediction of regional water resource supply and demand and dynamic quantification of water resource carrying capacity under multiple socioeconomic and climate scenarios. It can accurately identify core driving factors of water resource carrying capacity and has high reliability, accuracy, and applicability, providing technical support for the sustainable allocation of water resources and the coordinated development of socioeconomic and ecological environments in arid areas.

[0132] To verify the water resource carrying capacity prediction method based on a water resource supply and demand forecasting model disclosed in this invention, the following specific implementation method is disclosed. In this specific embodiment, the data flow of the water resource carrying capacity prediction method based on the water resource supply and demand forecasting model is as follows: Figure 7As shown, a method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model is disclosed, comprising the following steps:

[0133] Step S1; Data Collection: Collect economic and social data for a historical period in the region, including input-output tables, trade, GDP and population data, as well as hydrological and meteorological data such as wind speed, air pressure, rainfall, and radiation.

[0134] Step S2; Input-Output Model Construction and Virtual Water Analysis: This embodiment uses an input-output model combined with structural decomposition analysis to analyze virtual water flow and its core driving factors. The input-output model can quantify the correlation between production and consumption across industries, reveal economic interdependence, and track the flow of resources and pollutants along the supply chain. The specific calculation process is as follows:

[0135] ;

[0136] ;

[0137] In the formula, ε represents the physical water consumption intensity matrix, ε=[ε j ] 1×n ; P=[p j ] 1×n p j Represents the physical water consumption of the j-th department during the production process; U is the total input matrix of the departments, U=[u j ] 1×n H is the intermediate input matrix, H = [h ij ] n×n ;U diagonalize to U diag Similarly, ε is diagonalized to ε diag Y represents the department's final demand, Y = [y i ] n×1 I represents the identity matrix, A represents the technical coefficient matrix, (IA) -1 It is the Leontief inverse matrix; W d This refers to the virtual water supply for the final needs of the departments. In this example, the 42 economic sectors of the provincial-level administrative unit are summarized and merged into 9 major categories of economic sectors. The research area is the three provinces and autonomous regions of Inner Mongolia, Shaanxi and Ningxia. The final consumption types include import and export, investment, government consumption, urban residents' consumption and rural residents' consumption.

[0138] Next, Structural Decomposition Analysis (SDA) is used to calculate the contribution of various factors to the virtual water change. SDA focuses on factor decomposition of the dependent variable's changes and has significant advantages in quantifying the key drivers of observed environmental and economic phenomena. Therefore, combining the above virtual water calculation formula with the SDA method can decompose the contribution of the independent variable within a certain time period. The specific decomposition process is as follows:

[0139] ;

[0140] ;

[0141] ;

[0142] In the formula, Q represents e, i.e., actual water consumption intensity; L is the Leontief inverse matrix; Y (final demand) is decomposed into YC (household consumption), YG (government consumption), C (investment), E (exports), and I (imports); t and 0 represent the end and start periods of the variable, respectively; W t-0 This reflects the changes in virtual water between two periods; This indicates the contribution of water intensity to the virtual water change between two periods; This indicates the contribution of industrial structure to the virtual water change between two periods; other variables are interpreted similarly. G represents the total contribution of all factors; G represents the combination of all other factors multiplied by the current factor. The final decomposition result is as follows: Figure 2 As shown, Q represents water intensity, L represents industrial structure, YC represents household consumption, YG represents government consumption, C represents investment, E represents exports, and I represents imports. W2007, W2012, W2017, and W2022 represent the total virtual water resources in 2007, 2012, 2017, and 2022, respectively. In the Shaanxi-Gansu-Ningxia region, water intensity has a significant impact on the reduction of virtual water resources, contributing an average of -38.1% in Inner Mongolia, -42.1% in Shaanxi, and -45.5% in Ningxia. Investment and trade also have a significant impact: on average, investment contributes 16.7%, imports 41.8%, and exports 45.3%. Industrial structure plays an important role in Inner Mongolia (31%), but its impact is smaller in Ningxia (8.2%) and Shaanxi (2.1%). Household consumption is a significant driver of water consumption in Ningxia (36.9%) and Shaanxi (15.6%), but its contribution is relatively small in Inner Mongolia (8.7%). In Inner Mongolia, the five main factors influencing water consumption are water intensity, industrial structure, investment, exports, and imports. Since agriculture accounts for the largest share of virtual water consumption in the province, industrial structure here specifically refers to the proportion of the agricultural sector. In contrast, in Shaanxi and Ningxia, the five key factors are water intensity, household consumption, investment, exports, and imports.

[0143] Step S3: After identifying the key influencing factors of water supply and demand, the model is trained using Python software based on collected historical data (1995-2023) and the constructed deep learning model. In the CLA model used to predict water consumption, 80% of the data is allocated to the training set, while 20% is reserved for the test set. Economic factor data for 2026-2050 are based on predictions using SSPs-RCPs. Total water resource data comes from provincial water resource bulletins from 1995 to 2023, and meteorological data from 1995 to 2023 comes from ERA5. Specific factors include relative humidity, air temperature, vertical wind speed, horizontal wind speed, surface temperature, 2-meter air temperature, sea level pressure, net shortwave radiation, evaporation, and precipitation. In the CLA model for predicting available water resources, 70% of the data is used for the training set, and 30% is used for the test set. Meteorological data for the next few years (2026 to 2050) are projected from CMIP6 data under SSPs-RCPs. Figure 3 The training average loss and Nash coefficient of the CLA method are shown, where (a) represents the model training in Inner Mongolia, (b) in Shaanxi, (c) in Ningxia, and (d) represents the model training in the three provinces of Inner Mongolia, Shaanxi, and Ningxia. These results show that the performance of CLA remains stable and there is no obvious overfitting phenomenon. Finally, the relevant parameters were continuously adjusted to lock in the optimal deep learning model for subsequent prediction. The specific model performance comparison results are shown in Tables 1 and 2 below.

[0144] Table 1. Comparison of Water Resource Demand Model Performance

[0145]

[0146] Note: CLA stands for CNN-LSTM-Attention; CL stands for CNN-LSTM; L stands for LSTM.

[0147] Table 2 Comparison of Model Performance for Available Water Resources

[0148]

[0149] Figure 4The results show the projected future water consumption, presenting dynamic water consumption and intensity data across multiple regions. Under different SSPs-RCPs scenarios, water resource usage in the Inner Mongolia-Shaanxi-Ningxia region is projected to increase overall with economic development from 2026 to 2050. The increases are 6.2%, 7.3%, 3.9%, and 10.3% under SSP126, SSP245, SSP370, and SSP585 scenarios, respectively, with SSP585 showing the highest water consumption and SSP370 the lowest. SSP585, lacking sustainable management, exhibits almost unrestrained water consumption growth. SSP370, influenced by deglobalization and reduced regional competition, shows lower production and water consumption but increased water intensity, indicating weaker coordination between economic development and water resources. Overall, water intensity shows a downward trend, reflecting improved coordination between water and the economy. However, Inner Mongolia experiences a brief increase in water intensity from 2036 to 2043, presenting a window of pressure for water resource management. The low water intensity of SSP585 is merely a superficial phenomenon of rapid economic expansion, exacerbating actual ecological and climate risks; only low-emission pathways such as SSP126 can achieve coordinated economic and water resource development through structural optimization, technological innovation, and strict control.

[0150] Figure 5 The results demonstrate the future availability of water resources. From 2026 to 2050, the available water resources in the Shaanxi-Ningxia region show significant spatiotemporal differences under different SSPs-RCPs scenarios. Scenario SSP126 shows increased precipitation and runoff, leading to an overall increase in regional water resources. Scenario SSP585, however, is affected by increased evaporation and ecological degradation, resulting in lower water resources with large interannual fluctuations. By province, Inner Mongolia experiences an initial increase followed by a decrease under Scenario SSP126, while under Scenario SSP585, water resources decrease sharply and fluctuate dramatically. Shaanxi shows continuous growth under Scenario SSP126, while under Scenario SSP585, water resources remain low with extreme dry and wet conditions. Ningxia remains the most water-scarce region, with water resources at extremely low levels under both scenarios. The results indicate that emission pathways significantly impact regional water resources, and high emission scenarios will exacerbate regional water shortages and fluctuation risks, highlighting significant differences in water resource endowments among provinces within the region.

[0151] Step S4 finally inputs the predicted water supply and demand values ​​into the water resource carrying capacity calculation step to obtain the dynamic water resource carrying capacity of multiple regions. The results are as follows: Figure 6As shown, the water resource carrying capacity of Inner Mongolia, Shaanxi, and Ningxia is highest under scenario SSP126 and lowest under scenario SSP585. Inner Mongolia's SSP126 and SSP245 distributions are similar, but it has more years with high values; Shaanxi's SSP245 and SSP370 are similar, reflecting its current development trajectory being close to regional competition and a medium-to-high emission path; Ningxia has the most obvious comprehensive advantages under SSP126 and should firmly adhere to this development path. Water resource carrying capacity is directly related to water security and sustainable development; the higher the capacity, the better it can support development. Inner Mongolia faces relatively low water resource pressure; Shaanxi is approaching the pressure threshold and needs to strengthen demand-side management; Ningxia has inherently insufficient water resources, and efficiency improvements alone are insufficient, requiring strict water use control, structural adjustment, and reliance on virtual water and regional trade to alleviate pressure.

[0152] The water resource carrying capacity prediction method based on a water resource supply and demand forecasting model disclosed in this invention solves the technical challenge of combining mechanistic and statistical methods in water resource carrying capacity prediction. This is because input-output analysis (IOA-SDA) and deep learning (CNN-LSTM-Attention) are two completely heterogeneous and mechanistically incompatible technical systems. These two systems have fundamental conflicts in data structure, modeling logic, spatiotemporal scale, and computational objectives, making true coupling impossible for a long time.

[0153] Because IOA (Integrated Water Accounting) is a static, structured model based on economic accounts, it only focuses on historical accounting and lacks accurate time-series forecasting capabilities, making it unable to handle nonlinear, non-stationary, and highly fluctuating water resource time-series data. CLA (Clear Water Analysis), a pure data-driven black-box model, while capable of capturing spatiotemporal dynamics, lacks economic mechanism constraints, and its prediction results are detached from industrial water use logic, failing to connect with core elements such as virtual water and driving factors. The two also suffer from a mismatch in data dimensions: IOA uses structured economic data of sector × industry, while CLA uses high-dimensional spatiotemporal data of driving factors × region × time series, making direct fusion impossible. Therefore, existing research either uses only IOA for mechanism analysis or only CLA for prediction, failing to form a unified, end-to-end dynamic carrying capacity calculation framework.

[0154] This invention successfully couples IOA-SDA and CNN-LSTM-Attention by constructing a spatiotemporal feature mapping mechanism under mechanistic constraints. By combining the economic mechanism information provided by input-output analysis and structural decomposition analysis with the spatiotemporal modeling capabilities based on convolutional neural networks, long short-term memory networks, and attention mechanisms, an integrated modeling process is achieved, encompassing "virtual water structure analysis—driving factor extraction—spatial feature modeling—temporal dynamic prediction—bearing capacity assessment." This significantly improves the interpretability and engineering applicability of the results while ensuring the model's prediction accuracy.

[0155] Another embodiment of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the water resource carrying capacity prediction method based on the water resource supply and demand prediction model according to the present invention.

[0156] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor performs the water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to the present invention.

Claims

1. A method for predicting water resource carrying capacity based on a water resource supply and demand forecasting model, characterized in that, Includes the following steps: Step S1: Obtain economic and hydrological / meteorological data for the study area; calculate the virtual water transfer characteristics of each industry sector along the supply chain based on input-output analysis; decompose the changes in virtual water transfer characteristics into multiple driving factors using structural decomposition analysis, and select the core driving factor set. Step S2: Process the core driving factor set, hydrological and meteorological data, and socio-economic time series data, align them according to the time dimension, and construct structured input data containing time dimension, regional dimension, and driving factor dimension; based on the structured input data, form a multi-driving factor multi-regional spatial matrix under each time slice to characterize the spatial distribution relationship between multiple regions and the structural correlation relationship between driving factors; Step S3: Input the spatial matrix into a convolutional neural network to extract features from the regional spatial distribution characteristics and the correlation structure between driving factors represented in the spatial matrix; input the extracted spatial features into a long short-term memory network in chronological order for time series modeling; introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series to obtain a water resource supply and demand prediction model that integrates regional spatial heterogeneity, the correlation structure of driving factors, and the dynamic features of the time series. Step S4: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios, output water resource supply and demand forecasting results, and calculate water resource carrying capacity index based on the forecasting results to achieve dynamic assessment of water resource carrying capacity in multiple regions and scenarios.

2. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 1, characterized in that, The economic data include: GDP, regional annual input-output table, consumer price index, consumption, import and export and investment data; the hydrological and meteorological data include: water consumption and total water resources in the regional water resources bulletin, regional historical runoff, precipitation, wind speed, pressure, humidity, solar radiation and temperature data.

3. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 1, characterized in that, The virtual water transfer characteristics of each industry sector along the supply chain, calculated based on input-output analysis, include: Data standardization processing: unify the classification of industrial sectors in the input-output table within the research area, map the water use data and economic data to the industrial sectors one by one, and construct an industrial economic water use matching dataset; Construction of basic coefficient matrix: Based on the input-output table, calculate the direct consumption coefficient matrix and the total consumption coefficient matrix to quantify the economic interdependence between upstream and downstream sectors of the industrial chain. Water consumption calculation by dimension: direct water consumption and total water consumption are calculated for each department; the direct water consumption is the original water consumption in the industrial production process; the total water consumption is the direct water consumption plus the implicit water consumption indirectly transmitted from upstream and downstream of the industrial chain; Virtual water transfer quantitative modeling: Taking total water consumption as the core and combining inter-sectoral product trade flow, a multi-regional and multi-industry virtual water flow matrix is ​​constructed to accurately characterize the input, output, and net transfer direction and scale of virtual water along the supply chain between regions and industries, forming a standardized time series dataset of virtual water transfer characteristics. Time-series feature anchoring: Output virtual water transfer total amount, structure and spatial distribution characteristics by annual time slices, providing quantitative base data for subsequent structural decomposition analysis.

4. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 1, characterized in that, The structural decomposition analysis was used to decompose the changes in virtual water transfer characteristics into multiple driving factors, and the core driving factor set was obtained by screening: Decomposition Model Construction: Using the time-series changes in virtual water transfer characteristics output by the IOA as the target decomposition variable, an additive structural decomposition analysis model is constructed to establish a quantitative mapping relationship between virtual water changes and driving factors; Multi-dimensional driving factor full decomposition: The temporal fluctuations of virtual water transfer characteristics are precisely decomposed into three independent and quantifiable driving factors, including: Water intensity drivers: the impact of changes in water efficiency per unit of industrial output on virtual water transfer; Industrial structure drivers: characterizing the impact of regional industrial proportion adjustments and industrial chain structure upgrades on virtual water transfer; Final demand drivers: representing the impact of changes in the scale and structure of end-user demand, including consumption, investment, and exports, on virtual water transfer; Quantitative calculation of driving factor contribution: Calculate the absolute contribution value and relative contribution rate of each driving factor to virtual water change on a time slice and region-by-region basis, and quantitatively distinguish between dominant and secondary factors; Targeted screening of core driving factors: Set a contribution rate threshold, remove weakly influential factors with a contribution rate below the threshold, retain factors that play a decisive role in virtual water transfer and changes in water supply and demand, and solidify a set of core driving factors with fixed dimensions and clear physical meaning; Model input dimension binding: The core driving factor set is directly defined as the exclusive input dimension of the subsequent deep learning model, replacing the undifferentiated full data input, and completing the rigid connection from the IOA-SDA screening results to the model input structure.

5. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 1, characterized in that, Step S2 specifically includes: Step S21: Obtain the core driving factor set, hydrological and meteorological data, and socio-economic time series data, and uniformly represent the core driving factor set, hydrological and meteorological data, and socio-economic time series data as driving factor data with time and regional identifiers; Step S22: Standardize the driving factor data to eliminate dimensional differences between different data. Step S23: Align the standardized driving factor data according to the time dimension to construct multi-period time series data, so that each time point corresponds to multiple driving factor data in each region; Step S24: Based on the multi-period time series data, construct three-dimensional structured input data including time dimension, regional dimension and driving factor dimension; Step S25: Slice the three-dimensional structured input data according to the time dimension, extract the data of the corresponding region dimension and driving factor dimension under each time slice, and form a multi-driving factor multi-region spatial matrix.

6. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Input the spatial matrix into a convolutional neural network for spatial feature extraction; Step S32: Input the extracted spatial features into the Long Short-Term Memory network in chronological order for temporal modeling; Step S33: Introduce an attention mechanism into the long short-term memory network to weight key time nodes and key driving factors in the time series.

7. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 6, characterized in that, The convolutional neural network is used to extract spatial heterogeneity features between different regions in the spatial matrix and spatial correlation features between driving factors.

8. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 6, characterized in that, The Long Short-Term Memory (LSTM) network is used to model the long-term dependence and dynamic evolution characteristics of water resource supply and demand over time.

9. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 6, characterized in that, The weights in the attention mechanism are initialized and constrained based on the contribution of driving factors obtained from structural decomposition analysis.

10. The water resource carrying capacity prediction method based on a water resource supply and demand prediction model according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Based on the water resource supply and demand forecasting model, input data under different future socio-economic and climate scenarios to obtain the water resource supply forecast and water resource demand forecast for each region within the corresponding time range. Step S42: Based on the predicted water supply and water demand, calculate the water supply and demand index for each region in different time periods; Step S43: Based on the predicted water supply and water demand, calculate the concentrated carrying capacity index of each region to characterize whether the water resource utilization exceeds a preset threshold. Step S44: Based on the water resource supply and demand index and the concentrated carrying capacity index, calculate the water resource carrying capacity index and realize dynamic assessment of water resource carrying capacity under multiple regions, multiple scenarios and long-term series.

Images