Method for measuring water budget benchmark quota of water user based on multi-factor fusion evaluation

Abstract

The application discloses a kind of based on multi-factor fusion evaluation's water budget benchmark quota of water user measure method, comprising: obtaining the multi-source data of water user and constructing multidimensional dynamic feature tensor;Heterogeneous graph structure is constructed, and implicit state vector of user of fusion space-time association is generated using graph attention network and time series convolution network;Real-time meteorological factor and market quotation factor are obtained, and after splicing with implicit state vector, input bayesian deep neural network to obtain the probability density function sequence of future water demand;With the sequence as random constraint, combine water permit and industry index to build multi-objective optimization model to obtain candidate benchmark quota, determine preliminary benchmark quota after decision preference screening;Market sensitivity index is generated using historical water right transaction data, and dynamic adjustment coefficient is generated by inputting fuzzy reasoning system combined with meteorological stress index, and the final water budget benchmark quota is output after calibration. To improve the scientificity, accuracy and adaptability of water resources budget management.

Description

Technical Field

[0001] This invention relates to the field of water resources management technology, and in particular to a method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment. Background Technology

[0002] In the water budget management system, scientifically setting the budget benchmark for water users is a core element, directly affecting the accurate allocation of budget targets and the effective implementation of water-saving incentives. Existing technologies typically employ a simple weighted average method based on historical water consumption, or calculate quotas according to national, industry, and local water quota standards, combining this with the upper limit of water withdrawal permits to determine the annual water quota. However, existing methods have the following technical shortcomings: First, the forecasting methods are relatively singular, usually using simple weighted averages to predict output, failing to fully consider the impact of dynamic factors such as production fluctuations, weather changes, and market supply and demand on water demand, resulting in limited forecast accuracy. Second, the quota determination process lacks effective integration of quota information from different sources, often using minimum values ​​or simple averages, making it difficult to reflect individual user differences and the potential for water efficiency improvement. Third, the adjustment mechanism is relatively rigid, using fixed rules and preset coefficients for correction, unable to respond in real-time to fluctuations in the water rights market and changes in weather conditions, making it difficult to achieve dynamic optimization of budget quotas. Fourth, there is a lack of quantitative processing of water use uncertainty; the determined quota is a single value, unable to characterize water use risks at different confidence levels, and difficult to support risk-controlled budget management decisions.

[0003] To address the aforementioned issues, there is an urgent need for a method to calculate the water budget benchmark for water users that can integrate multi-source heterogeneous data, dynamically capture spatiotemporal correlations, quantify water use uncertainties, and achieve multi-objective collaborative optimization, in order to improve the scientific nature, accuracy, and adaptability of water resource budget management. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment. Its important purpose is to improve the scientificity, accuracy and adaptability of water resource budget management.

[0005] To achieve the above objectives, the first aspect of this invention provides a method for calculating the water budget baseline for water users based on multi-factor fusion assessment, comprising: We acquire historical water usage data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data from water users. After multi-source data fusion and spatiotemporal alignment processing, we construct a multidimensional dynamic feature tensor. Based on the multidimensional dynamic feature tensor, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industrial association as edges. Graph attention network and temporal convolutional network are used to generate user latent state vectors that integrate spatiotemporal associations. Real-time meteorological factors and market conditions are obtained, concatenated with the user's latent state vector, and then input into a Bayesian deep neural network. Probability distribution modeling is performed through variational inference, and a probability density function sequence of water demand in the future budget period is output. Using the probability density function sequence of water demand as stochastic constraints, a multi-objective optimization model is constructed by combining the upper limit of water abstraction permits and industry budget control indicators. Candidate benchmark quotas are obtained by solving the Pareto front, and preliminary benchmark quotas are determined based on decision preferences. A market situation predictor is constructed and a market sensitivity index is generated using historical water rights transaction data. A weather stress index is calculated using weather forecasts and input into a fuzzy inference system to generate dynamic adjustment coefficients. The preliminary benchmark quota is then calibrated with flexible boundaries, and the final water budget benchmark quota for water users is output.

[0006] In this solution, the acquisition of historical water consumption data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data from water users, followed by multi-source data fusion and spatiotemporal alignment processing to construct a multidimensional dynamic feature tensor, specifically includes: Historical water usage data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data of water users are obtained. Outliers are identified and removed using the 3σ principle based on a sliding window. For the identified outliers, linear interpolation of three time points before and after the point is used for replacement. For missing data, the interpolation method is selected according to the data type. Cubic spline interpolation is used to fill in continuous data, and forward imputation is used for discrete data. After the data cleaning is completed, Z-score standardization is performed on all continuous numerical features to obtain the cleaned time series data. The cleaned time series data are uniformly aligned to the daily scale, and based on the geographical coordinates of the water users, the inverse distance weighted interpolation method is used to map the meteorological observation data of the surrounding meteorological stations to each user's location, generating user-specific daily meteorological features. The daily aligned data for each user is associated with the user identifier and date to generate a wide table with user-time as the primary key. Derived features are constructed to enhance time dynamics and interactivity. At the same time, industry water quota standards, user industry category and water-saving level are attached as static attribute features to each time slice. The derived features include the lag features and sliding window statistical features of water consumption series, the year-on-year and month-on-month change rates of production and operation data, the cumulative effect features of meteorological data, and the lag features of market data. Based on the water supply network topology data, the upstream and downstream node information and network connection relationship of each user are extracted and encoded into a static adjacency matrix. A multidimensional dynamic feature tensor is constructed with time step as the first dimension, individual water user as the second dimension, and feature channel as the third dimension. The feature channel contains all original features, constructed derived features, and user-specific daily meteorological features.

[0007] In this scheme, based on the multidimensional dynamic feature tensor, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industry associations as edges. A user latent state vector that integrates spatiotemporal associations is generated using graph attention networks and temporal convolutional networks. Specifically, this includes: Obtain a multidimensional dynamic feature tensor, and extract the node feature matrix corresponding to each time step from the multidimensional dynamic feature tensor. The rows of the node feature matrix correspond to each water user, and the columns correspond to all features in the feature channel, including historical water consumption, production and operation indicators, meteorological factors, water rights market statistics, and static attribute features. The static adjacency matrix encoded according to the topology data of the water supply network is acquired synchronously, and an industry association adjacency matrix is ​​constructed. The industry association adjacency matrix is ​​calculated based on the user's industry classification, the upstream and downstream relationship of the industrial chain, and the Pearson correlation coefficient of the historical water use sequence. Based on the node feature matrix, static adjacency matrix, and industry association adjacency matrix at each time step, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industry association relationships as edges. The heterogeneous graph structure is then input into a graph attention network, and independent attention mechanisms are set for network topology relationships and industry association relationships respectively. The attention coefficients between each target user node and its two types of neighbor nodes are calculated through a set independent attention mechanism. The attention coefficients are obtained by concatenating the feature vectors of the target node and its neighbor nodes after linear transformation and inputting them into a single-layer feedforward neural network. The final attention weights are obtained after LeakyReLU activation and Softmax normalization. Based on the calculated final attention weight, the feature vectors of industry-related neighbor nodes and pipeline topology neighbor nodes are weighted and summed respectively to obtain two aggregated feature vectors of the target node under pipeline topology relationship and industry-related relationship. These two aggregated feature vectors are then concatenated to serve as the spatial aggregated feature vector of the target node at the current time step. The above operation is repeated for all user nodes to obtain the spatial aggregated feature vectors of all users at the current time step. A multi-head attention mechanism is used to execute multiple independent graph attention calculations in parallel. Each group uses different linear transformation parameters to concatenate multiple spatial aggregated feature vectors to obtain an enhanced spatial aggregated feature vector. The forward propagation of the graph attention network is executed sequentially for all time steps to generate a three-dimensional spatial feature sequence with dimensions of time step number, number of users, and hidden feature dimension. The three-dimensional spatial feature sequence is input into a temporal convolutional network, which is composed of multiple residual blocks containing dilated causal convolutional layers. Through layer-by-layer convolutional operations, the short-term fluctuations, periodic patterns and long-term trends of each user's water use behavior are extracted along the time dimension. The output of the last time step is taken as the latent state vector of each user.

[0008] In this solution, the acquisition of real-time meteorological factors and market conditions, concatenated with the user's latent state vector, is input into a Bayesian deep neural network. Probability distribution modeling is performed through variational inference, outputting a probability density function sequence of water demand within the future budget period. Specifically, this includes: It acquires daily precipitation, maximum temperature, minimum temperature, relative humidity, and evaporation data, and combines this with the geographical coordinates of each water user to generate user-specific real-time meteorological factors; at the same time, it obtains real-time transaction records from the water rights trading platform to generate the weighted average transaction price, total transaction volume, and price volatility of the most recent trading day, and organizes and generates market trend factors on a daily basis. Derivative features are constructed for real-time meteorological factors and market conditions factors. The forecast mean and upper and lower limits of the forecast interval of each meteorological element in the real-time meteorological factors are extracted as the first derivative features and correlated. The first-order price difference and relative strength index of the market conditions factors are used as the second derivative features and correlated. After completing the feature construction, the real-time meteorological factors and market conditions factors are aligned according to the time step to form a real-time dynamic feature matrix. The user's hidden state vector is obtained and its dimensions are expanded. Then, it is concatenated with the real-time dynamic feature matrix in the feature dimension to obtain a joint feature matrix that integrates static user features and dynamic environment features. The joint feature matrix is ​​input into a pre-built Bayesian deep neural network, which is composed of multiple fully connected layers stacked together. Its weight parameters are random variables that follow a prior distribution. The variational inference method is used to approximate the posterior distribution of the weights, and the variational parameters are optimized by minimizing the variational distribution and KL divergence. The log-normal distribution is chosen as the output distribution type. Two neurons are set in the output layer to predict the log-normal mean and log-normal standard deviation, respectively. For the input at each time step, the network outputs a pair of distribution parameters through forward propagation, which represent the probability distribution of water demand for that day. Finally, the output parameters of all time steps are combined along the time dimension to generate a sequence of probability density functions for water demand in future budget cycles.

[0009] In this scheme, a multi-objective optimization model is constructed using the probability density function sequence of water demand as a stochastic constraint, combined with the upper limit of water withdrawal permits and industry budget control indicators. Candidate benchmark quotas are obtained by solving the Pareto front, and a preliminary benchmark quota is determined based on decision preferences. Specifically, this includes: Obtain the probability density function sequence, extract the log-normal distribution parameters for each time step from the probability density function sequence, calculate the water consumption quantiles for each time step under the preset confidence level, sum the quantiles for all time steps to obtain the daily quantile sequence as the risk boundary, sum the daily quantiles to obtain the annual total risk, and use it as the annual equivalent form of the stochastic constraint. The first constraint is to obtain the upper limit of water withdrawal permits for water users, and the second constraint is to obtain industry budget control indicators and convert them into an annual total upper limit. The annual budget amount is used as the decision variable, and two optimization objectives are set: minimizing the annual budget amount and maximizing the probability of water demand satisfaction. A multi-objective optimization model is constructed. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm. An initial population is randomly generated and the constraints of each candidate solution are checked. The probability of water demand satisfaction for each candidate solution is calculated by Monte Carlo simulation based on the probability density function sequence of water demand. The total annual budget and the probability of water demand satisfaction are used as objective function values ​​for population evaluation. The population is divided into Pareto fronts at different levels by non-dominated sorting. Through selection, crossover and mutation operations, the evolution is iterated until convergence, resulting in a set of Pareto front candidate solutions. Each candidate solution corresponds to an annual budget and the probability of meeting water demand, generating a candidate baseline quota set. The entropy weight TOPSIS method is introduced for multi-attribute decision-making. Each candidate solution in the candidate benchmark quota set is taken as a scheme, and the total annual budget and the probability of water demand satisfaction are taken as evaluation indicators. A decision matrix is ​​constructed and the indicators are homogenized. The entropy value of each indicator is calculated as the indicator weight. Calculate the Euclidean distance between each candidate solution and the ideal solution and the negative ideal solution. Calculate the relative closeness between the candidate solution and the ideal solution based on the Euclidean distance. Sort the candidate solutions from high to low relative closeness and select the candidate solution with the highest relative closeness as the initial benchmark quota.

[0010] In this scheme, the construction of a market situation predictor and the generation of a market sensitivity index using historical transaction data, the calculation of a weather stress index using weather forecasts, and the input of these into a fuzzy inference system to generate dynamic adjustment coefficients, the calibration of the preliminary benchmark quota using flexible boundaries, and the output of the final water budget benchmark quota for water users, specifically include: Historical water rights trading data is obtained, including the weighted average trading price, total trading volume, and price volatility for each trading day over the past three years. These data are then organized into a fixed-length sliding window sample in chronological order and input into a market trend predictor based on a long short-term memory network to obtain daily predicted trading prices and volumes for a preset period. The price change rate is calculated based on the fluctuation range of the daily predicted transaction price and the average transaction price within a preset period. The transaction activity index is calculated based on the ratio of the predicted transaction volume to the historical average transaction volume for the same period. The market sensitivity index is obtained by weighted summing of the price change rate and the transaction activity index, which represents the degree of supply and demand tension in the water rights market within the preset period. Obtain daily precipitation, evaporation, maximum temperature and minimum temperature forecast data for each user's location within a preset period, calculate precipitation anomaly index, evaporation potential index and high temperature stress index, and obtain the meteorological stress index by weighted summation of the three indices, which characterizes the degree of meteorological risk within the preset period; The market sensitivity index and the weather stress index are input variables into the fuzzy inference system. The fuzzy inference system includes a fuzzification module, a rule base, an inference engine, and a defuzzification module. The rule base is pre-established based on expert experience and contains fuzzy subsets corresponding to the market sensitivity index and the weather stress index. The membership degree of the input variable to each fuzzy subset is calculated by the membership function. Fuzzy logic reasoning is performed according to the rule base. The activation intensity of each rule is calculated by the minimum value method and applied to the output fuzzy subset to obtain the truncated output fuzzy set corresponding to each rule. The truncated output fuzzy sets of all rules are superimposed to obtain the total output fuzzy set. The centroid method is used to defuzzify the total output fuzzy set to obtain the dynamic adjustment coefficient, and the preliminary benchmark quota is calibrated with elastic boundaries. The system presets a flexible calibration limit range and determines whether the calibration value exceeds the upper or lower limit of the flexible calibration limit range. If it exceeds the limit, the boundary value of the range is used as the final limit. If it is within the range, the calibration value is used directly. Finally, the system outputs the final water budget benchmark limit for water users. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0012] Figure 1 A flowchart of the first method for calculating the water budget benchmark amount for water users based on multi-factor fusion evaluation, provided in an embodiment of the present invention; Figure 2 A flowchart of the second method for calculating the water budget benchmark amount for water users based on multi-factor fusion evaluation, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0013] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0014] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0015] Figure 1 A flowchart of the first method for calculating the water budget benchmark amount for water users based on multi-factor fusion evaluation, provided in an embodiment of the present invention; like Figure 1 As shown, the present invention provides a first method flowchart for calculating the water budget benchmark quota for water users based on multi-factor fusion assessment, including: S102: Obtain historical water usage data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data from water users, and construct a multi-dimensional dynamic feature tensor after multi-source data fusion and spatiotemporal alignment processing. S104. Based on the multidimensional dynamic feature tensor, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industrial association as edges. The graph attention network and temporal convolutional network are used to generate user latent state vectors that integrate spatiotemporal associations. S106, Obtain real-time meteorological factors and market conditions factors, concatenate them with the user's hidden state vector, and input them into a Bayesian deep neural network. Through variational inference, perform probability distribution modeling and output the probability density function sequence of water demand in the future budget period. S108 uses the probability density function sequence of water demand as a stochastic constraint, combines the upper limit of water withdrawal permits and industry budget control indicators to construct a multi-objective optimization model, obtains candidate benchmark quotas by solving the Pareto front, and determines the preliminary benchmark quotas based on decision preferences. S110 constructs a market situation predictor and generates a market sensitivity index using historical water rights transaction data. It calculates a weather stress index through weather forecasts and inputs it into a fuzzy inference system to generate dynamic adjustment coefficients. It then performs elastic boundary calibration on the preliminary benchmark quota and outputs the final water budget benchmark quota for water users.

[0016] Furthermore, in a preferred embodiment of the present invention, the step of acquiring historical water consumption data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data from water users, and constructing a multidimensional dynamic feature tensor after multi-source data fusion and spatiotemporal alignment processing, specifically includes: Historical water usage data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data of water users are obtained. Outliers are identified and removed using the 3σ principle based on a sliding window. For the identified outliers, linear interpolation of three time points before and after the point is used for replacement. For missing data, the interpolation method is selected according to the data type. Cubic spline interpolation is used to fill in continuous data, and forward imputation is used for discrete data. After the data cleaning is completed, Z-score standardization is performed on all continuous numerical features to obtain the cleaned time series data. The cleaned time series data are uniformly aligned to the daily scale, and based on the geographical coordinates of the water users, the inverse distance weighted interpolation method is used to map the meteorological observation data of the surrounding meteorological stations to each user's location, generating user-specific daily meteorological features. The daily aligned data for each user is associated with the user identifier and date to generate a wide table with user-time as the primary key. Derived features are constructed to enhance time dynamics and interactivity. At the same time, industry water quota standards, user industry category and water-saving level are attached as static attribute features to each time slice. The derived features include the lag features and sliding window statistical features of water consumption series, the year-on-year and month-on-month change rates of production and operation data, the cumulative effect features of meteorological data, and the lag features of market data. Based on the water supply network topology data, the upstream and downstream node information and network connection relationship of each user are extracted and encoded into a static adjacency matrix. A multidimensional dynamic feature tensor is constructed with time step as the first dimension, individual water user as the second dimension, and feature channel as the third dimension. The feature channel contains all original features, constructed derived features, and user-specific daily meteorological features.

[0017] It should be noted that the first step is to acquire multi-source basic data from water users. This includes production and operation data, such as indicators reflecting production activities like product output, output value, and operating rate. These data may be recorded on a daily, weekly, or monthly basis. Regional meteorological and hydrological data includes observations from meteorological or hydrological stations, such as precipitation, evaporation, temperature, and humidity, typically sampled daily or hourly. Water rights trading market data includes event-based data such as transaction time, transaction volume, and transaction price. Spatial geographic information data includes static spatial attributes such as the geographical coordinates of water users, their administrative region, water supply network connections, and the watershed they are located in. In addition, static information such as industry water consumption quota standards (national, provincial, and municipal levels), user industry category, water-saving enterprise designation, water balance test results, and historical records of exceeding planned water usage needs to be collected to provide a foundation for subsequent feature construction.

[0018] After data collection, the data is first cleaned: A sliding window 3σ principle is used to identify outliers in water consumption, and these outliers are replaced with linear interpolations from three preceding and following time points. For missing data, continuous data such as water consumption and temperature are filled using cubic spline interpolation, while discrete data such as operating status are filled using forward imputation. After cleaning, all continuous features are standardized using Z-scores. Next, all data is aligned to a daily scale: water consumption monitoring data is summed daily to obtain daily water consumption; production operation data, if monthly, is linearly interpolated to daily levels; meteorological data is taken as daily averages; and water rights trading data is statistically analyzed as the total number of transactions, total amount, and weighted average price per day. Based on the user's geographic coordinates, inverse distance weighted interpolation is used to map observations from surrounding meteorological stations to the user's location, generating unique daily meteorological features. The aligned data is then associated with the user identifier and date to generate a wide table with "user-time" as the primary key. To enhance the ability to capture temporal dynamics, a series of derived features are constructed: for water consumption sequences, features with lags of 1, 3, and 7 days are constructed, along with the mean and standard deviation of the past 7-day sliding window; for production data, the year-on-year change rate compared to the same day last year and the month-on-month change rate compared to the previous day are calculated; for meteorological data, cumulative effect features such as the cumulative precipitation over the past 7 days are constructed; and for market data, lagging features such as the average transaction price over the past 3 days are constructed. Furthermore, static attributes that do not change over time, such as industry water quota standards, user industry categories, and water-saving levels, are added as constant features to each time slice. Finally, the upstream and downstream node relationships of users are extracted based on the water supply network topology data and encoded into a static adjacency matrix. A multidimensional dynamic feature tensor is constructed with the time step as the first dimension, the individual water user as the second dimension, and the feature channel as the third dimension. The feature channel includes all original features, constructed derived features, user-specific daily meteorological features, and static attributes, providing a standardized input data format for subsequent graph neural networks and temporal convolutional networks.

[0019] Furthermore, in a preferred embodiment of the present invention, the step of constructing a heterogeneous graph structure based on the multidimensional dynamic feature tensor, with water users as nodes and water supply network topology and industry association relationships as edges, and generating user latent state vectors that integrate spatiotemporal association relationships using graph attention networks and temporal convolutional networks, specifically includes: Obtain a multidimensional dynamic feature tensor, and extract the node feature matrix corresponding to each time step from the multidimensional dynamic feature tensor. The rows of the node feature matrix correspond to each water user, and the columns correspond to all features in the feature channel, including historical water consumption, production and operation indicators, meteorological factors, water rights market statistics, and static attribute features. The static adjacency matrix encoded according to the topology data of the water supply network is acquired synchronously, and an industry association adjacency matrix is ​​constructed. The industry association adjacency matrix is ​​calculated based on the user's industry classification, the upstream and downstream relationship of the industrial chain, and the Pearson correlation coefficient of the historical water use sequence. Based on the node feature matrix, static adjacency matrix, and industry association adjacency matrix at each time step, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industry association relationships as edges. The heterogeneous graph structure is then input into a graph attention network, and independent attention mechanisms are set for network topology relationships and industry association relationships respectively. The attention coefficients between each target user node and its two types of neighbor nodes are calculated through a set independent attention mechanism. The attention coefficients are obtained by concatenating the feature vectors of the target node and its neighbor nodes after linear transformation and inputting them into a single-layer feedforward neural network. The final attention weights are obtained after LeakyReLU activation and Softmax normalization. Based on the calculated final attention weight, the feature vectors of industry-related neighbor nodes and pipeline topology neighbor nodes are weighted and summed respectively to obtain two aggregated feature vectors of the target node under pipeline topology relationship and industry-related relationship. These two aggregated feature vectors are then concatenated to serve as the spatial aggregated feature vector of the target node at the current time step. The above operation is repeated for all user nodes to obtain the spatial aggregated feature vectors of all users at the current time step. A multi-head attention mechanism is used to execute multiple independent graph attention calculations in parallel. Each group uses different linear transformation parameters to concatenate multiple spatial aggregated feature vectors to obtain an enhanced spatial aggregated feature vector. The forward propagation of the graph attention network is executed sequentially for all time steps to generate a three-dimensional spatial feature sequence with dimensions of time step number, number of users, and hidden feature dimension. The three-dimensional spatial feature sequence is input into a temporal convolutional network, which is composed of multiple residual blocks containing dilated causal convolutional layers. Through layer-by-layer convolutional operations, the short-term fluctuations, periodic patterns and long-term trends of each user's water use behavior are extracted along the time dimension. The output of the last time step is taken as the latent state vector of each user.

[0020] It should be noted that, firstly, the node feature matrix for each time step is obtained from the multidimensional dynamic feature tensor. The rows of this matrix correspond to each water user, while the columns cover all feature dimensions reflecting the user's status, specifically including historical water consumption, production and operation indicators, meteorological factors, water rights market statistics, and static attribute features such as industry water quota standards, user's industry category, and water-saving level. Simultaneously, a static adjacency matrix based on the water supply network topology data encoding is obtained. Based on the user's industry classification, upstream and downstream relationships in the industrial chain, and the Pearson correlation coefficient of historical water consumption sequences, an industry association adjacency matrix is ​​constructed. For example, dyeing and printing factories and weaving factories belonging to the same textile industry chain, or users in the same industry with highly similar water consumption patterns, will be assigned higher industry association weights. After the node feature matrix and the two adjacency matrices together form a heterogeneous graph structure, it is input into a graph attention network, with independent attention mechanisms set for network topology relationships and industry association relationships.

[0021] In the attention calculation process, for each target user node, its feature vector and the feature vectors of its neighboring nodes are first linearly transformed. The transformed vectors are then concatenated and input into a single-layer feedforward neural network. After LeakyReLU activation and Softmax normalization, attention weights are obtained. These weights reflect the degree of influence of different neighbors on the target node's water usage behavior. For example, water usage fluctuations of upstream enterprises in a pipeline network may have a transmission effect on downstream areas, while users with strong industrial linkages may experience water usage synchronization due to production collaboration. Based on the attention weights, the feature vectors of the two types of neighboring nodes are weighted and summed to obtain the aggregated feature vector of the target node under the two relationships. These two aggregated feature vectors are then concatenated as the spatial aggregated feature of the current time step. To avoid the randomness of a single attention calculation, a multi-head attention mechanism is used to execute multiple independent calculations in parallel, each using different linear transformation parameters. The outputs from these multiple sets are concatenated to obtain the enhanced spatial aggregated feature vector. The above operations are performed sequentially for all time steps to generate a three-dimensional spatial feature sequence.

[0022] The sequence is input into a temporal convolutional network composed of multiple stacked dilated causal convolutional residual blocks. The use of dilated convolutions allows the network to obtain a large receptive field with fewer layers, effectively capturing short-term fluctuations, periodic patterns (such as water usage troughs on weekends and holidays), and long-term trends (such as seasonal variations) in water use behavior. After multiple convolutional layers, the output of the last time step is taken as the latent state vector for each user. This vector integrates the spatiotemporal coupling information of the user's own historical water use dynamics, the spatial influence of surrounding users, and environmental factors, providing a high-dimensional feature representation for subsequent water demand probability prediction.

[0023] Furthermore, in a preferred embodiment of the present invention, the acquisition of real-time meteorological factors and market conditions, concatenating them with the user's latent state vector, and inputting them into a Bayesian deep neural network, performs probability distribution modeling through variational inference, and outputs a probability density function sequence of water demand within the future budget period, specifically including: It acquires daily precipitation, maximum temperature, minimum temperature, relative humidity, and evaporation data, and combines this with the geographical coordinates of each water user to generate user-specific real-time meteorological factors; at the same time, it obtains real-time transaction records from the water rights trading platform to generate the weighted average transaction price, total transaction volume, and price volatility of the most recent trading day, and organizes and generates market trend factors on a daily basis. Derivative features are constructed for real-time meteorological factors and market conditions factors. The forecast mean and upper and lower limits of the forecast interval of each meteorological element in the real-time meteorological factors are extracted as the first derivative features and correlated. The first-order price difference and relative strength index of the market conditions factors are used as the second derivative features and correlated. After completing the feature construction, the real-time meteorological factors and market conditions factors are aligned according to the time step to form a real-time dynamic feature matrix. The user's hidden state vector is obtained and its dimensions are expanded. Then, it is concatenated with the real-time dynamic feature matrix in the feature dimension to obtain a joint feature matrix that integrates static user features and dynamic environment features. The joint feature matrix is ​​input into a pre-built Bayesian deep neural network, which is composed of multiple fully connected layers stacked together. Its weight parameters are random variables that follow a prior distribution. The variational inference method is used to approximate the posterior distribution of the weights, and the variational parameters are optimized by minimizing the variational distribution and KL divergence. The log-normal distribution is chosen as the output distribution type. Two neurons are set in the output layer to predict the log-normal mean and log-normal standard deviation, respectively. For the input at each time step, the network outputs a pair of distribution parameters through forward propagation, which represent the probability distribution of water demand for that day. Finally, the output parameters of all time steps are combined along the time dimension to generate a sequence of probability density functions for water demand in future budget cycles.

[0024] It should be noted that when integrating real-time dynamic information for probabilistic prediction, it is first necessary to obtain real-time meteorological factors for the future budget period. Specifically, daily precipitation, maximum temperature, minimum temperature, relative humidity, and evaporation data are extracted from numerical weather prediction. Based on the geographical coordinates of each water user, a user-specific daily weather forecast sequence is generated using bilinear interpolation. Simultaneously, real-time transaction records are obtained from the water rights trading platform, and the weighted average transaction price and total transaction volume of the most recent trading day are calculated. The price volatility over the past seven days is also calculated, and this data is organized daily to form market trend factors. To fully exploit the information content in the real-time factors, derivative features are constructed for these factors. For meteorological factors, the forecast mean and upper and lower limits of the forecast intervals for each element are extracted as the first derivative feature to characterize the uncertainty range of future weather conditions. For market trend factors, the first difference of prices is calculated to reflect price change trends, and a relative strength index is constructed to measure the comparison of market buying and selling forces. These indicators are used as the second derivative feature. After completing the feature construction, the meteorological factors and market conditions are aligned according to the time step to form a real-time dynamic feature matrix. The rows of the matrix correspond to each future day, and the columns correspond to each meteorological element and market indicator.

[0025] Since the hidden state vector is a fixed-length vector that does not change over time, it needs to be copied and expanded along the time dimension to make its number of rows the same as the real-time dynamic feature matrix. The expanded hidden state matrix and the real-time dynamic feature matrix are concatenated along the feature dimension to obtain a joint feature matrix, which simultaneously contains user static spatiotemporal features and future dynamic environmental information. This joint feature matrix is ​​then input into a Bayesian deep neural network. This network consists of multiple fully connected layers stacked together. Unlike traditional neural networks, its weight parameters are treated as random variables following a prior distribution, rather than fixed values. A variational inference method is used to approximate the posterior distribution of the weights, optimizing the variational parameters by minimizing the KL divergence between the variational distribution and the true posterior distribution. This probabilistic modeling approach allows the network to quantify prediction fluctuations caused by data noise and model uncertainty. The network output layer selects a log-normal distribution as the distribution type for water demand, setting two neurons to predict the log mean and log standard deviation, respectively. For each future day's input, the network outputs a pair of distribution parameters through forward propagation, fully describing the probability density of each possible value of water consumption for that day. By combining the output parameters of all time steps along the time dimension, a probability density function sequence of water demand in the future budget cycle is finally generated. This sequence not only gives the expected value of water consumption, but also quantifies the uncertainty of the prediction, providing the necessary probabilistic information for subsequent baseline quota optimization based on chance-constrained planning.

[0026] Furthermore, in a preferred embodiment of the present invention, the step of constructing a multi-objective optimization model using the probability density function sequence of water demand as a stochastic constraint, combined with the upper limit of water withdrawal permits and industry budget control indicators, obtaining candidate benchmark quotas by solving the Pareto front, and determining preliminary benchmark quotas based on decision preferences, specifically includes: Obtain the probability density function sequence, extract the log-normal distribution parameters for each time step from the probability density function sequence, calculate the water consumption quantiles for each time step under the preset confidence level, sum the quantiles for all time steps to obtain the daily quantile sequence as the risk boundary, sum the daily quantiles to obtain the annual total risk, and use it as the annual equivalent form of the stochastic constraint. The first constraint is to obtain the upper limit of water withdrawal permits for water users, and the second constraint is to obtain industry budget control indicators and convert them into an annual total upper limit. The annual budget amount is used as the decision variable, and two optimization objectives are set: minimizing the annual budget amount and maximizing the probability of water demand satisfaction. A multi-objective optimization model is constructed. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm. An initial population is randomly generated and the constraints of each candidate solution are checked. The probability of water demand satisfaction for each candidate solution is calculated by Monte Carlo simulation based on the probability density function sequence of water demand. The total annual budget and the probability of water demand satisfaction are used as objective function values ​​for population evaluation. The population is divided into Pareto fronts at different levels by non-dominated sorting. Through selection, crossover and mutation operations, the evolution is iterated until convergence, resulting in a set of Pareto front candidate solutions. Each candidate solution corresponds to an annual budget and the probability of meeting water demand, generating a candidate baseline quota set. The entropy weight TOPSIS method is introduced for multi-attribute decision-making. Each candidate solution in the candidate benchmark quota set is taken as a scheme, and the total annual budget and the probability of water demand satisfaction are taken as evaluation indicators. A decision matrix is ​​constructed and the indicators are homogenized. The entropy value of each indicator is calculated as the indicator weight. Calculate the Euclidean distance between each candidate solution and the ideal solution and the negative ideal solution. Calculate the relative closeness between the candidate solution and the ideal solution based on the Euclidean distance. Sort the candidate solutions from high to low relative closeness and select the candidate solution with the highest relative closeness as the initial benchmark quota.

[0027] It should be noted that when constructing the multi-objective optimization model, the log-normal distribution parameters for each time step are first extracted from the probability density function sequence of water demand. Based on a preset confidence level, such as 90%, the water consumption quantiles for each time step at that confidence level are calculated, representing the upper limit of daily water consumption that may be reached in the future, ensuring that the probability of actual water consumption falling below this value is exactly 90%. The quantiles for all time steps are summed to obtain a daily quantile sequence as the risk boundary. Further summation yields the annual total risk, which serves as the annual equivalent form of the stochastic constraint, representing the maximum possible value of annual water demand at a given confidence level. The upper limit of water withdrawal permits for water users is obtained as the first constraint; this value is usually approved by the water administration department and represents the maximum amount of water a user can withdraw throughout the year. Simultaneously, industry budget control indicators, such as unit product water consumption limits or reuse rate requirements, are obtained and converted into an annual total upper limit as the second constraint. Using the total annual budget as the decision variable, two conflicting optimization objectives—minimizing the total annual budget and maximizing the probability of water demand satisfaction—are set to construct the multi-objective optimization model. The probability of meeting water demand is defined as the percentage of samples whose total annual water consumption does not exceed the budget, obtained by sampling a large number of daily probability distributions through Monte Carlo simulation, given a total annual budget.

[0028] A non-dominated sorting genetic algorithm is used to solve the model. A set of candidate annual budget totals is randomly generated as the initial population. Each candidate solution needs to pass the water withdrawal permit and industry indicator constraints. Based on the probability density function sequence of water demand, the probability of water demand satisfaction for each candidate solution is calculated. The annual budget total and the probability of water demand satisfaction are used as the objective function values ​​to evaluate the individuals in the population. Individuals are divided into different levels of Pareto fronts through non-dominated sorting, with individuals at the first front being non-dominated. Through selection, crossover, and mutation operations, iterative evolution is performed until convergence, resulting in a set of candidate solutions distributed at the Pareto fronts. Each solution corresponds to an annual budget total and the probability of water demand satisfaction, forming a candidate baseline quota set. The entropy-weighted TOPSIS method is introduced for multi-attribute decision-making. Each candidate solution is used as a decision scheme, and the annual budget total and the probability of water demand satisfaction are used as evaluation indicators to construct a decision matrix. The indicators are homogenized by taking the reciprocal of the annual budget total to make it a benefit-type indicator, similar to the probability of water demand satisfaction. The entropy value of each indicator is calculated. The smaller the entropy value, the greater the discriminative power of the indicator for decision-making, and the indicator weights are determined accordingly. The Euclidean distance between each candidate solution and the ideal solution is calculated, thus obtaining the relative closeness between each candidate solution and the ideal solution. The candidate solutions with the highest closeness are selected as the initial baseline allowance, which achieves an optimal balance between water conservation and ensuring water supply security.

[0029] Figure 2A flowchart of the second method for calculating the water budget benchmark amount for water users based on multi-factor fusion evaluation, provided in an embodiment of the present invention; like Figure 2 As shown, the present invention provides a second method flowchart for calculating the water budget benchmark quota for water users based on multi-factor fusion assessment, including: S202: Obtain historical water rights trading data, including the weighted average trading price, total trading volume, and price volatility for each trading day over the past three years. Organize these data into a fixed-length sliding window sample in chronological order and input it into a market situation predictor based on a long short-term memory network to obtain daily predicted trading prices and predicted trading volumes for a preset period. S204: The price change rate is calculated based on the fluctuation range of the daily predicted transaction price and the average transaction price within the preset period. The transaction activity index is calculated based on the ratio of the predicted transaction volume to the historical average transaction volume for the same period. The market sensitivity index is obtained by weighted summing of the price change rate and the transaction activity index, which represents the supply and demand tension of the water rights market within the preset period. S206: Obtain daily precipitation, evaporation, maximum temperature and minimum temperature forecast data for each user's location within a preset period, calculate precipitation anomaly index, evaporation potential index and high temperature stress index, and obtain the meteorological stress index by weighted summation of the three indices, which characterizes the degree of meteorological risk within the preset period. S208, the market sensitivity index and the weather stress index are input as input variables into the fuzzy inference system. The fuzzy inference system includes a fuzzification module, a rule base, an inference engine, and a defuzzification module. The rule base is pre-established based on expert experience and contains fuzzy subsets corresponding to the market sensitivity index and the weather stress index. S210: Calculate the membership degree of the input variable to each fuzzy subset through the membership function, perform fuzzy logic reasoning based on the rule base, calculate the activation intensity of each rule using the minimum value method and apply it to the output fuzzy subset to obtain the truncated output fuzzy set corresponding to each rule; S212, superimpose the truncated output fuzzy sets of all rules to obtain the total output fuzzy set, use the centroid method to defuzzify the total output fuzzy set to obtain the dynamic adjustment coefficient, and perform elastic boundary calibration on the preliminary benchmark quota. S214, preset flexible calibration limit range, determine whether the calibration value exceeds the upper or lower limit of the flexible calibration limit range. If it exceeds, the boundary value of the range is used as the final amount. If it is within the range, the calibration value is used directly. Finally, the final water budget benchmark amount for the water user is output.

[0030] It should be noted that, firstly, historical water rights trading data is acquired, including the weighted average trading price, total trading volume, and price volatility for each trading day over the past three years. This data comes from the water rights trading platform and reflects the market supply and demand situation for water rights indicators. This data is organized chronologically into a fixed-length sliding window sample; for example, using data from the past 60 days to predict the market situation for the next 7 days. This sample is then input into a market trend predictor built on a Long Short-Term Memory (LSTM) network. This network automatically learns the long-term dependencies in the trading sequence through its gating mechanism. After nonlinear transformation through multiple hidden layers, it outputs the daily predicted trading price and predicted trading volume for a preset period. Based on the daily predicted trading price, its volatility relative to the average trading price of the past 30 days is calculated to obtain the price change rate, which reflects the short-term trend of water rights market prices. Based on the ratio of the predicted trading volume to the historical average trading volume for the same period, a trading activity index is calculated, reflecting the level of market activity. The price change rate and the trading activity index are weighted and summed to obtain the market sensitivity index. The greater the upward pressure on water rights indicators, the higher the scarcity of regional water resource use rights.

[0031] Simultaneously, daily precipitation, evaporation, and forecast data for maximum and minimum temperatures for each user's location within a preset period are acquired. The precipitation anomaly index is calculated, which is the ratio of cumulative precipitation within the preset period to the historical average precipitation for the same period. A ratio less than 0.8 indicates drought, and a ratio greater than 1.2 indicates flooding. The evaporation potential index is calculated, which is the ratio of average evaporation within the preset period to the historical average evaporation for the same period. The high-temperature stress index is calculated, which is the percentage of days within the preset period where the maximum temperature exceeds a threshold. The three indices are weighted and summed to obtain the meteorological stress index. The market sensitivity index and the meteorological stress index are used as input variables and fed into the fuzzy inference system. This system includes a fuzzification module, a rule base, an inference engine, and a defuzzification module. The rule base is pre-established based on the experience of water resource management experts. For example, one rule states: if market sensitivity is high and meteorological stress is severe, the dynamic adjustment coefficient is positive, indicating that the quota needs to be increased to cope with the dual pressures of market scarcity and meteorological risk. The membership degree of the input variable to each fuzzy subset is calculated using a membership function. The activation intensity of each rule is calculated using the minimum value method and applied to the output fuzzy subset, resulting in a truncated output fuzzy set for each rule. The truncated output fuzzy sets of all rules are superimposed to obtain the total output fuzzy set. After defuzzification using the centroid method, the precise value of the dynamic adjustment coefficient is obtained.

[0032] The initial benchmark quota is multiplied by a dynamic adjustment coefficient to obtain the initial calibration value. A preset flexible calibration limit range is set at ±10% of the initial benchmark quota to prevent drastic fluctuations from impacting user production plans. The calibration value is then checked against the upper or lower limit of the range. If it exceeds this limit, the range boundary value is used as the final quota; otherwise, the calibration value is directly adopted. The final output is the water budget benchmark quota for each water user. This quota comprehensively considers the user's historical water usage patterns, industry constraints, real-time water rights market signals, and weather forecast information, achieving dynamic adaptability and refined control of budget management.

[0033] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment, characterized in that, include: We acquire historical water usage data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data from water users. After multi-source data fusion and spatiotemporal alignment processing, we construct a multidimensional dynamic feature tensor. Based on the multidimensional dynamic feature tensor, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industrial association as edges. Graph attention network and temporal convolutional network are used to generate user latent state vectors that integrate spatiotemporal associations. Real-time meteorological factors and market conditions are obtained, concatenated with the user's latent state vector, and then input into a Bayesian deep neural network. Probability distribution modeling is performed through variational inference, and a probability density function sequence of water demand in the future budget period is output. Using the probability density function sequence of water demand as stochastic constraints, a multi-objective optimization model is constructed by combining the upper limit of water abstraction permits and industry budget control indicators. Candidate benchmark quotas are obtained by solving the Pareto front, and preliminary benchmark quotas are determined based on decision preferences. A market situation predictor is constructed and a market sensitivity index is generated using historical water rights transaction data. A weather stress index is calculated using weather forecasts and input into a fuzzy inference system to generate dynamic adjustment coefficients. The preliminary benchmark quota is then calibrated with flexible boundaries, and the final water budget benchmark quota for water users is output.

2. The method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment as described in claim 1, characterized in that, The process involves acquiring historical water usage data from water users, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data. After multi-source data fusion and spatiotemporal alignment processing, a multidimensional dynamic feature tensor is constructed, specifically including: Historical water usage data, production and operation data, regional meteorological and hydrological data, water rights trading market data, and spatial geographic information data of water users are obtained. Outliers are identified and removed using the 3σ principle based on a sliding window. For the identified outliers, linear interpolation of three time points before and after the point is used for replacement. For missing data, the interpolation method is selected according to the data type. Cubic spline interpolation is used to fill in continuous data, and forward imputation is used for discrete data. After the data cleaning is completed, Z-score standardization is performed on all continuous numerical features to obtain the cleaned time series data. The cleaned time series data are uniformly aligned to the daily scale, and based on the geographical coordinates of the water users, the inverse distance weighted interpolation method is used to map the meteorological observation data of the surrounding meteorological stations to each user's location, generating user-specific daily meteorological features. The daily aligned data for each user is associated with the user identifier and date to generate a wide table with user-time as the primary key. Derived features are constructed to enhance time dynamics and interactivity. At the same time, industry water quota standards, user industry category and water-saving level are attached as static attribute features to each time slice. The derived features include the lag features and sliding window statistical features of water consumption series, the year-on-year and month-on-month change rates of production and operation data, the cumulative effect features of meteorological data, and the lag features of market data. Based on the water supply network topology data, the upstream and downstream node information and network connection relationship of each user are extracted and encoded into a static adjacency matrix. A multidimensional dynamic feature tensor is constructed with time step as the first dimension, individual water user as the second dimension, and feature channel as the third dimension. The feature channel contains all original features, constructed derived features, and user-specific daily meteorological features.

3. The method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment according to claim 1, characterized in that, The process involves constructing a heterogeneous graph structure based on the multidimensional dynamic feature tensor, using water users as nodes and water supply network topology and industry relationships as edges. A graph attention network and a temporal convolutional network are then used to generate a user latent state vector that integrates spatiotemporal relationships. Specifically, this includes: Obtain a multidimensional dynamic feature tensor, and extract the node feature matrix corresponding to each time step from the multidimensional dynamic feature tensor. The rows of the node feature matrix correspond to each water user, and the columns correspond to all features in the feature channel, including historical water consumption, production and operation indicators, meteorological factors, water rights market statistics, and static attribute features. The static adjacency matrix encoded according to the topology data of the water supply network is acquired synchronously, and an industry association adjacency matrix is ​​constructed. The industry association adjacency matrix is ​​calculated based on the user's industry classification, the upstream and downstream relationship of the industrial chain, and the Pearson correlation coefficient of the historical water use sequence. Based on the node feature matrix, static adjacency matrix, and industry association adjacency matrix at each time step, a heterogeneous graph structure is constructed with water users as nodes and water supply network topology and industry association relationships as edges. The heterogeneous graph structure is then input into a graph attention network, and independent attention mechanisms are set for network topology relationships and industry association relationships respectively. The attention coefficients between each target user node and its two types of neighbor nodes are calculated through a set independent attention mechanism. The attention coefficients are obtained by concatenating the feature vectors of the target node and its neighbor nodes after linear transformation and inputting them into a single-layer feedforward neural network. The final attention weights are obtained after LeakyReLU activation and Softmax normalization. Based on the calculated final attention weight, the feature vectors of industry-related neighbor nodes and pipeline topology neighbor nodes are weighted and summed respectively to obtain two aggregated feature vectors of the target node under pipeline topology relationship and industry-related relationship. These two aggregated feature vectors are then concatenated to serve as the spatial aggregated feature vector of the target node at the current time step. The above operation is repeated for all user nodes to obtain the spatial aggregated feature vectors of all users at the current time step. A multi-head attention mechanism is used to execute multiple independent graph attention calculations in parallel. Each group uses different linear transformation parameters to concatenate multiple spatial aggregated feature vectors to obtain an enhanced spatial aggregated feature vector. The forward propagation of the graph attention network is executed sequentially for all time steps to generate a three-dimensional spatial feature sequence with dimensions of time step number, number of users, and hidden feature dimension. The three-dimensional spatial feature sequence is input into a temporal convolutional network, which is composed of multiple residual blocks containing dilated causal convolutional layers. Through layer-by-layer convolutional operations, the short-term fluctuations, periodic patterns and long-term trends of each user's water use behavior are extracted along the time dimension. The output of the last time step is taken as the latent state vector of each user.

4. The method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment as described in claim 1, characterized in that, The process of acquiring real-time meteorological factors and market conditions, concatenating them with the user's latent state vector, and inputting the result into a Bayesian deep neural network, uses variational inference to model the probability distribution and outputs a sequence of probability density functions for water demand within the future budget period. Specifically, this includes: It acquires daily precipitation, maximum temperature, minimum temperature, relative humidity, and evaporation data, and combines this with the geographical coordinates of each water user to generate user-specific real-time meteorological factors; at the same time, it obtains real-time transaction records from the water rights trading platform to generate the weighted average transaction price, total transaction volume, and price volatility of the most recent trading day, and organizes and generates market trend factors on a daily basis. Derivative features are constructed for real-time meteorological factors and market conditions factors. The forecast mean and upper and lower limits of the forecast interval of each meteorological element in the real-time meteorological factors are extracted as the first derivative features and correlated. The first-order price difference and relative strength index of the market conditions factors are used as the second derivative features and correlated. After completing the feature construction, the real-time meteorological factors and market conditions factors are aligned according to the time step to form a real-time dynamic feature matrix. The user's hidden state vector is obtained and its dimensions are expanded. Then, it is concatenated with the real-time dynamic feature matrix in the feature dimension to obtain a joint feature matrix that integrates static user features and dynamic environment features. The joint feature matrix is ​​input into a pre-built Bayesian deep neural network, which is composed of multiple fully connected layers stacked together. Its weight parameters are random variables that follow a prior distribution. The variational inference method is used to approximate the posterior distribution of the weights, and the variational parameters are optimized by minimizing the variational distribution and KL divergence. The log-normal distribution is chosen as the output distribution type. Two neurons are set in the output layer to predict the log-normal mean and log-normal standard deviation, respectively. For the input at each time step, the network outputs a pair of distribution parameters through forward propagation, which represent the probability distribution of water demand for that day. Finally, the output parameters of all time steps are combined along the time dimension to generate a sequence of probability density functions for water demand in future budget cycles.

5. The method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment according to claim 1, characterized in that, The process involves using the probability density function sequence of water demand as a stochastic constraint, combining the upper limit of water abstraction permits and industry budget control indicators to construct a multi-objective optimization model. Candidate benchmark quotas are obtained by solving the Pareto front, and a preliminary benchmark quota is determined based on decision preferences. Specifically, this includes: Obtain the probability density function sequence, extract the log-normal distribution parameters for each time step from the probability density function sequence, calculate the water consumption quantiles for each time step under the preset confidence level, sum the quantiles for all time steps to obtain the daily quantile sequence as the risk boundary, sum the daily quantiles to obtain the annual total risk, and use it as the annual equivalent form of the stochastic constraint. The first constraint is to obtain the upper limit of water withdrawal permits for water users, and the second constraint is to obtain industry budget control indicators and convert them into an annual total upper limit. The annual budget amount is used as the decision variable, and two optimization objectives are set: minimizing the annual budget amount and maximizing the probability of water demand satisfaction. A multi-objective optimization model is constructed. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm. An initial population is randomly generated and the constraints of each candidate solution are checked. The probability of water demand satisfaction for each candidate solution is calculated by Monte Carlo simulation based on the probability density function sequence of water demand. The total annual budget and the probability of water demand satisfaction are used as objective function values ​​for population evaluation. The population is divided into Pareto fronts at different levels by non-dominated sorting. Through selection, crossover and mutation operations, the evolution is iterated until convergence, resulting in a set of Pareto front candidate solutions. Each candidate solution corresponds to an annual budget and the probability of meeting water demand, generating a candidate baseline quota set. The entropy weight TOPSIS method is introduced for multi-attribute decision-making. Each candidate solution in the candidate benchmark quota set is taken as a scheme, and the total annual budget and the probability of water demand satisfaction are taken as evaluation indicators. A decision matrix is ​​constructed and the indicators are homogenized. The entropy value of each indicator is calculated as the indicator weight. Calculate the Euclidean distance between each candidate solution and the ideal solution and the negative ideal solution. Calculate the relative closeness between the candidate solution and the ideal solution based on the Euclidean distance. Sort the candidate solutions from high to low relative closeness and select the candidate solution with the highest relative closeness as the initial benchmark quota.

6. The method for calculating the water budget benchmark amount for water users based on multi-factor fusion assessment according to claim 1, characterized in that, The process of constructing a market situation predictor and generating a market sensitivity index using historical transaction data, calculating a weather stress index using weather forecasts, and inputting these into a fuzzy inference system to generate dynamic adjustment coefficients, performing elastic boundary calibration on the preliminary benchmark quota, and outputting the final water budget benchmark quota for water users, specifically includes: Historical water rights trading data is obtained, including the weighted average trading price, total trading volume, and price volatility for each trading day over the past three years. These data are then organized into a fixed-length sliding window sample in chronological order and input into a market trend predictor based on a long short-term memory network to obtain daily predicted trading prices and volumes for a preset period. The price change rate is calculated based on the fluctuation range of the daily predicted transaction price and the average transaction price within a preset period. The transaction activity index is calculated based on the ratio of the predicted transaction volume to the historical average transaction volume for the same period. The market sensitivity index is obtained by weighted summing of the price change rate and the transaction activity index, which represents the degree of supply and demand tension in the water rights market within the preset period. Obtain daily precipitation, evaporation, maximum temperature and minimum temperature forecast data for each user's location within a preset period, calculate precipitation anomaly index, evaporation potential index and high temperature stress index, and obtain the meteorological stress index by weighted summation of the three indices, which characterizes the degree of meteorological risk within the preset period; The market sensitivity index and the weather stress index are input variables into the fuzzy inference system. The fuzzy inference system includes a fuzzification module, a rule base, an inference engine, and a defuzzification module. The rule base is pre-established based on expert experience and contains fuzzy subsets corresponding to the market sensitivity index and the weather stress index. The membership degree of the input variable to each fuzzy subset is calculated by the membership function. Fuzzy logic reasoning is performed according to the rule base. The activation intensity of each rule is calculated by the minimum value method and applied to the output fuzzy subset to obtain the truncated output fuzzy set corresponding to each rule. The truncated output fuzzy sets of all rules are superimposed to obtain the total output fuzzy set. The centroid method is used to defuzzify the total output fuzzy set to obtain the dynamic adjustment coefficient, and the preliminary benchmark quota is calibrated with elastic boundaries. The system presets a flexible calibration limit range and determines whether the calibration value exceeds the upper or lower limit of the flexible calibration limit range. If it exceeds the limit, the boundary value of the range is used as the final limit. If it is within the range, the calibration value is used directly. Finally, the system outputs the final water budget benchmark limit for water users.

Images