A massive data processing method for DSGE analysis and decision-making on the relationship between marine industry and ecology.

Abstract

This invention provides a massive data processing method for DSGE analysis and decision-making related to marine industries and ecology. Belonging to the technical field of massive data processing for DSGE models, this invention collects multi-source heterogeneous data streams and performs spatiotemporal alignment preprocessing. It then performs threshold cointegration tests on the preprocessed data to identify nonlinear threshold points and establish a state transition matrix. The data is input into a heterogeneous computing scheduler for resource allocation and grid partitioning. A multi-objective decomposition framework combined with particle swarm optimization and simulated annealing algorithms is used to complete parameter calibration. A sliding window mechanism is established, and a recursive least squares algorithm is used for incremental updates. A low-rank approximate decomposition is performed on the high-dimensional weight matrix, and the optimized low-rank submatrix is ​​used as the weight parameters of the DSGE model for decision analysis. This solves the technical problem that massive heterogeneous data is difficult to efficiently support DSGE model parameter calibration and decision optimization under real-time processing conditions.

Description

Technical Field

[0001] This invention belongs to the field of massive data processing technology of DSGE models. Specifically, it relates to a massive data processing method for DSGE analysis and decision-making on the relationship between marine industry and ecology. Background Technology

[0002] The correlation analysis between marine industry and the ecological environment typically employs DSGE (Dynamic Stochastic General Equilibrium) models to characterize the coupling relationship between the economic system and the ecosystem. Traditional methods rely on batch data processing and offline parameter estimation for model calibration. However, in current marine industry decision support systems, due to the characteristics of multi-source monitoring data—high-frequency arrival, heterogeneous formats, and inconsistent spatiotemporal resolution—traditional batch processing methods require periodic recalculation of all historical data, resulting in enormous computational resource consumption and failing to meet real-time decision-making needs. Furthermore, traditional parameter estimation methods use a single computational architecture to process all types of data, failing to optimize resource allocation for the parallel characteristics and computational complexity differences of the data, leading to low computational efficiency. In other words, existing technologies face the technical challenge of efficiently supporting DSGE model parameter calibration and decision optimization under real-time processing conditions with massive amounts of heterogeneous data. Summary of the Invention

[0003] In view of this, the present invention provides a massive data processing method for DSGE analysis and decision-making related to marine industry and ecology, which can solve the technical problem in the prior art that massive heterogeneous data is difficult to efficiently support DSGE model parameter calibration and decision optimization under real-time processing conditions.

[0004] This invention is implemented as follows: It provides a massive data processing method for DSGE analysis and decision-making related to marine industries and ecology. This method collects multi-source heterogeneous data streams and performs spatiotemporal alignment preprocessing. Threshold cointegration tests are performed on the preprocessed data to identify nonlinear threshold points and establish a state transition matrix. The data is input into a heterogeneous computing scheduler for resource allocation. A multi-objective decomposition framework is used to decompose the parameter optimization problem into single-objective sub-problems, and global and local searches are performed. A sliding window mechanism is established to retain data from the most recent time period, and a recursive least squares algorithm is used for incremental updates. A low-rank approximate decomposition is performed on the high-dimensional weight matrix, and the optimized low-rank submatrix is ​​used as the weight parameters of the DSGE model for decision analysis. The heterogeneous computing scheduler enables adaptive allocation of computing resources driven by data characteristics. Combined with low-rank matrix decomposition, the model parameter size is reduced. A recursive least squares algorithm is used to achieve incremental updates of the DSGE model parameters.

[0005] Specifically, the step of collecting multi-source heterogeneous data streams and performing spatiotemporal alignment preprocessing involves collecting raw data streams from industrial monitoring systems, ecological environment monitoring stations, and policy databases; performing missing value imputation on the raw data streams; identifying outlier data points exceeding three standard deviations and replacing them with moving averages; unifying the time scale of data collected at different frequencies; extracting the latitude and longitude coordinates of each monitoring station to achieve matching of data with different spatial resolutions; and performing minimum-maximum normalization transformation on all numerical data.

[0006] The raw data stream includes industrial economic indicators, ecological and environmental parameters, policy text information, and technical specification documents. For unstructured policy texts, a large language model is used to perform semantic understanding and information extraction. For semi-structured technical documents, a large language model is used to parse the technical indicator tables and specification paragraphs.

[0007] Specifically, the use of a large language model to perform semantic understanding and information extraction on unstructured policy texts involves guiding the large language model to identify numerical parameters, entity information and their relationships in policy documents through structured prompt word templates, using mind chain prompting technology to guide the large language model to analyze the logical relationship between policy measures and their targets, and using a few-shot learning method to enable the large language model to identify policy intensity descriptors and convert them into numerical weights.

[0008] Specifically, the step of performing a threshold cointegration test on the preprocessed data to identify nonlinear threshold points involves selecting pollution concentration as the threshold variable, setting a candidate set of threshold values ​​to cover the range of sample quantiles ∈ [0.15, 0.85], dividing the sample into two subsamples for each threshold value in the candidate set of threshold values, estimating the cointegration vector and calculating the sum of squared residuals in each subsample, and identifying the threshold value that minimizes the sum of squared residuals as the optimal threshold point.

[0009] Specifically, establishing the state transition matrix involves dividing the data into a low-concentration zone with pollution concentrations below a first threshold, a medium-concentration zone between the first and second thresholds, and a high-concentration zone with pollution concentrations above the second threshold. Within each zone, the long-term cointegration relationship between industrial output and ecological indicators is estimated, and a state transition matrix is ​​constructed to characterize the switching probability of the system between different zones.

[0010] Specifically, the process of inputting data into a heterogeneous computing scheduler for resource allocation involves constructing a computing demand assessment function to quantify the parallelism and computational complexity characteristics of the data. When the parallelism characteristic value is greater than 0.7 and the computational complexity characteristic value is greater than 0.6, the data is determined to be highly parallel and allocated to the GPU; otherwise, it is determined to be sequentially dependent data and allocated to the CPU.

[0011] The first hidden layer of the heterogeneous computing scheduler adopts a representation learning framework based on the information bottleneck principle. It minimizes the mutual information constraint of redundant information between input features and intermediate representations, while maximizing the mutual information between intermediate representations and computing resource allocation decisions to maintain key information. It introduces variational approximation techniques to estimate mutual information terms and uses the Lagrange multiplier method to dynamically balance information compression rate and task completion fidelity.

[0012] Specifically, the establishment of the training dataset for the heterogeneous computing scheduler involves collecting data features and actual execution performance data of historical computing tasks, calculating the CPU execution efficiency and GPU execution efficiency for each historical computing task, and marking it as a GPU task when the GPU execution efficiency exceeds the CPU execution efficiency by more than 1.5 times; otherwise, it is marked as a CPU task.

[0013] Specifically, the multi-objective decomposition framework is used to decompose the parameter optimization problem into single-objective sub-problems. This involves defining a set of optimization objective functions that include economic efficiency objective functions, ecological benefit objective functions, and policy synergy objective functions, using a uniform design method to generate weight vectors, decomposing the multi-objective optimization problem into single-objective sub-problems, and using Chebyshev aggregation functions to measure the approximation of individual solutions to reference points.

[0014] Specifically, the execution of global and local search involves exploring the global parameter space using the particle swarm optimization algorithm. When the fitness function fails to improve for 20 consecutive generations, it is considered to be in a stagnant state. The algorithm is then switched to simulated annealing to perform a local fine-grained search. An adaptive neighborhood adjustment strategy is used to dynamically adjust the neighborhood size based on the convergence speed of the single objective subproblem.

[0015] Specifically, the sliding window mechanism retains data from the most recent time period. The sliding window time length is set to 720 hours. When new data arrives, the earliest data in the sliding window is deleted and the new data is added to the end of the sliding window. Key statistics, including mean, variance, and covariance matrix, are calculated for the data in the sliding window.

[0016] Specifically, the incremental update using the recursive least squares algorithm involves calculating the current parameter estimate using a recursive formula with the new observation and the parameter estimate from the previous time step. The forgetting factor is set to 0.98, and the parameter change rate of the DSGE model is calculated. When the parameter change rate exceeds a preset threshold of 5%, a local recalibration process is triggered.

[0017] Specifically, the low-rank approximation decomposition of the high-dimensional weight matrix involves identifying high-dimensional weight matrices with dimensions > 1000 in the heterogeneous computing scheduler as decomposition objects, performing singular value decomposition on the high-dimensional weight matrix, calculating the cumulative energy percentage of singular values, selecting the top k largest singular values ​​that make the cumulative energy percentage reach 95%, and constructing a low-rank approximation matrix.

[0018] Specifically, the construction of the low-rank approximation matrix involves retaining the top k largest singular values ​​and their corresponding singular vectors, representing the low-rank approximation matrix as the product of two low-rank sub-matrices, introducing a nuclear norm regularization term to constrain the rank of the low-rank sub-matrices, employing an alternating minimization strategy to optimize the two low-rank sub-matrices, and setting the convergence criterion as the change in the objective function value of the low-rank decomposition in two consecutive iterations being <0.0001.

[0019] The low-rank decomposition objective function is defined as the square of the Frobenius norm plus the nuclear norm regularization term between the high-dimensional weight matrix and the low-rank approximation matrix. The parameters of the low-rank submatrix are updated using the gradient descent method. The original high-dimensional weight matrix is ​​replaced by the optimized low-rank submatrix. The optimized low-rank submatrix is ​​used as the weight parameter of the DSGE model for the trade-off analysis and decision support of the economic efficiency objective function, the ecological benefit objective function, and the policy synergy objective function.

[0020] This invention achieves adaptive allocation of computing resources driven by data features by establishing a heterogeneous computing scheduler. Highly parallel data is allocated to GPUs for grid-parallel computing, while sequentially dependent data is allocated to CPUs for recursive computing, fully leveraging the performance advantages of different computing architectures. By introducing a representation learning framework based on the information bottleneck principle, the heterogeneous computing scheduler automatically learns the mapping relationship between data features and optimal computing resources during training. Simultaneously, it employs low-rank matrix factorization to reduce the size of model parameters and combines this with recursive least squares algorithm to incrementally update DSGE model parameters, avoiding redundant computation of all historical data and significantly improving computational throughput. In summary, this invention solves the technical problem mentioned in the background art of efficiently supporting DSGE model parameter calibration and decision optimization under real-time processing conditions with massive heterogeneous data. Attached Figure Description

[0021] Figure 1 This is a pulse diagram illustrating the greening of petrochemical seawater desalination and aquaculture enterprises under the carbon tax policy in the example.

[0022] Figure 2 This is a multi-combination pulse diagram of petrochemical seawater desalination and greening of aquaculture enterprises under the carbon tax policy in the example.

[0023] Figure 3 This is a pulse diagram of the greening of petrochemical seawater desalination and aquaculture enterprises under the subsidy policy in the example.

[0024] Figure 4 This is a multi-combination pulse diagram of petrochemical seawater desalination and greening of aquaculture enterprises under the subsidy policy in the example.

[0025] Figure 5The pulse diagram illustrates the technological innovations in the petrochemical seawater desalination and greening of aquaculture enterprises in the embodiments.

[0026] Figure 6 The example illustrates a multi-combination pulse diagram of petrochemical seawater desalination and greening of aquaculture enterprises under the technological innovations described.

[0027] Figure 7 This is a diagram showing the changes in social welfare under the carbon tax policy in the example.

[0028] Figure 8 This is a diagram showing the changes in social welfare under the subsidy policy in the example.

[0029] Figure 9 This is a flowchart illustrating the resource allocation decision-making process of the heterogeneous computing scheduler in this embodiment.

[0030] Figure 10 This is the singular value energy distribution diagram of the high-dimensional weight matrix in the embodiment. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0032] This invention provides a method for processing massive amounts of data for DSGE analysis and decision-making regarding the relationship between marine industries and ecology, comprising the following steps:

[0033] A method for processing massive amounts of data for DSGE analysis and decision-making regarding the relationship between marine industry and ecology, comprising the following steps:

[0034] S1. Collect multi-source heterogeneous data streams and perform spatiotemporal alignment preprocessing. Correct data points that exceed the statistical range through outlier identification algorithms. Achieve spatial matching based on geographic information systems and eliminate dimensional differences by normalization transformation.

[0035] S2. Perform threshold cointegration test on the preprocessed data to identify nonlinear threshold points. Divide the data into different concentration zones according to the threshold, estimate the cointegration relationship in each zone, and establish a state transition matrix to characterize the zone switching law.

[0036] S3. Input the data into the heterogeneous computing scheduler to determine resource allocation. Based on the parallelism and computational complexity characteristics, allocate sequentially dependent data to the CPU and highly parallel data to the GPU and perform grid partitioning.

[0037] S4. The parameter optimization problem is decomposed into single-objective sub-problems using a multi-objective decomposition framework. A global search is performed using the particle swarm algorithm. When the fitness stagnates, the simulated annealing algorithm is switched to perform a local search. Parameter calibration is completed by combining adaptive neighborhood adjustment.

[0038] S5. Establish a sliding window mechanism to retain the data of the most recent time period and use a recursive least squares algorithm to perform incremental updates. When the parameter change rate exceeds the threshold, a local recalibration process is triggered. Combined with GPU acceleration, the DSGE model can be refreshed quickly.

[0039] S6. Perform low-rank approximate decomposition on the high-dimensional weight matrix, determine the optimal decomposition dimension through singular value decomposition and introduce nuclear norm regularization constraints, and optimize the parameters of the low-rank submatrix using an alternating minimization strategy. Use the optimized low-rank submatrix as the weight parameters of the DSGE model for decision analysis.

[0040] The specific steps of collecting multi-source heterogeneous data streams and performing spatiotemporal alignment preprocessing include: collecting raw data streams from industry monitoring systems, ecological environment monitoring stations, and policy databases, wherein the raw data streams contain industry economic indicators, ecological environment parameters, policy text information, and technical specifications; performing missing value imputation on the raw data streams, using a combination of forward and backward imputation to handle time-series breakpoints; calculating the mean and standard deviation of each monitoring indicator, identifying outlier data points exceeding three times the standard deviation range, and replacing them with moving averages; unifying the time scale of data from different collection frequencies, and generating intermediate time values ​​using linear interpolation; extracting the latitude and longitude coordinates of each monitoring station, and using the nearest neighbor interpolation algorithm to match data with different spatial resolutions; performing minimum-maximum normalization transformation on all numerical data, mapping the data to a standard interval of 0 to 1; and performing semantic understanding and information processing on unstructured policy texts using a large language model. Information extraction involves using structured prompt templates to guide the large language model in identifying numerical parameters, entity information, and their relationships in policy documents. Mind chain prompting technology guides the large language model in analyzing the logical connections between policy measures and their targets. A few-shot learning method enables the large language model to identify policy intensity descriptors and convert them into numerical weights. The output of the large language model is organized into structured records containing policy type, policy parameters, scope of application, effective date, and policy intensity according to a predefined pattern. For semi-structured technical documents, the large language model parses technical indicator tables and specification paragraphs. Prompt engineering guides the large language model to extract the numerical range, unit information, and test conditions of technical parameters. The multimodal understanding capabilities of the large language model are used to process technical documents containing charts and extract key data points from the charts. The extracted technical parameters are organized into structured data tables according to technical category, parameter name, parameter value, parameter unit, and test conditions.

[0041] The specific steps for performing threshold cointegration tests on preprocessed data to identify nonlinear threshold points include: selecting pollution concentration as the threshold variable, setting the candidate threshold value set to cover the range of sample quantiles from 0.15 to 0.85; dividing the sample into two subsamples for each threshold value in the candidate threshold value set, estimating the cointegration vector and calculating the residual sum of squares in each subsample; generating a simulated distribution and calculating the critical value using the bootstrap method, and testing the significance of the threshold effect using the likelihood ratio statistic; identifying the threshold value that minimizes the residual sum of squares as the optimal threshold point, dividing the data into a low-concentration zone with pollution concentration below the first threshold value, a medium-concentration zone between the first and second threshold values, and a high-concentration zone with pollution concentration above the second threshold value; estimating the long-term cointegration relationship between industrial output and ecological indicators in each zone, and extracting the error correction term for each zone; constructing a state transition matrix to characterize the switching probability of the system between different zones, determining the transition probability parameters using the maximum likelihood estimation method, and embedding the zone transition rule into the state-space equation of the DSGE model.

[0042] The threshold cointegration test is a cointegration relationship testing method based on a threshold regression model. It divides the sample into different regimes by setting a threshold variable, allowing for differences in cointegration vectors within each regime, thereby capturing the nonlinear long-term equilibrium relationship between variables. The test statistic uses the likelihood ratio to measure the significance of the threshold effect. The state transition matrix is ​​a square matrix describing the probability of transitions between different states of a stochastic process. Matrix elements represent the probability of the system transitioning from one regime to another, and the sum of elements in each row of the matrix is ​​1. The state transition matrix characterizes the dynamic switching patterns of the industrial ecosystem between regimes with different pollution concentrations.

[0043] The specific steps for inputting data into the heterogeneous computing scheduler for resource allocation include: constructing a computing demand assessment function to quantify the parallelism and computational complexity characteristics of the data. The parallelism characteristic is calculated by comprehensively considering data dimension, data independence, and iterative dependencies. The computational complexity characteristic is obtained by comprehensively evaluating floating-point operations, memory access patterns, and branch prediction difficulty. The output value of the computing demand assessment function is compared with a preset threshold. When the parallelism characteristic value is greater than 0.7 and the computational complexity characteristic value is greater than 0.6, the data is determined to be highly parallel and allocated to the GPU; otherwise, it is determined to be sequentially dependent and allocated to the CPU. The highly parallel data allocated to the GPU is divided into grids, and the optimal grid size is dynamically adjusted according to the data size, the number of GPU cores, and the cache capacity. The divided data blocks are mapped to the GPU's streaming multiprocessor array. Each streaming multiprocessor is responsible for processing one or more data blocks, and fast access to data within the same data block is achieved through shared memory. The sequentially dependent data allocated to the CPU is recursively computed according to the time sequence, utilizing the CPU's high clock speed to handle control logic and branch prediction.

[0044] The heterogeneous computing scheduler has the following structure: the input layer receives parallelism feature vectors and computational complexity feature vectors from the data; the first hidden layer employs a representation learning framework based on the information bottleneck principle, minimizing the mutual information between the input features and the intermediate representation to constrain redundant information, while maximizing the mutual information between the intermediate representation and the computing resource allocation decision to preserve key information. The first hidden layer contains 64 neuron units, and the activation function uses parameterized rectified linear units; the second hidden layer performs a nonlinear transformation on the output of the first hidden layer, using a sparse connection mechanism to reduce the number of parameters. The second hidden layer contains 32 neuron units; the output layer outputs CPU allocation probabilities and GPU allocation probabilities, selecting the computing resource with the higher probability as the final allocation result; variational approximation techniques are introduced in the first hidden layer to estimate mutual trust. The system employs a reparameterization technique to achieve end-to-end training. It uses the Lagrange multiplier method to dynamically balance information compression rate and task completion fidelity, with the Lagrange multiplier adaptively adjusted based on the average processing time of the current batch of data, the target processing time, and the parameter change rate. Low-rank approximation optimization based on matrix factorization is performed on the high-dimensional weight matrix. The weight matrix of the first hidden layer is decomposed into the product of two low-rank sub-matrices. Singular value decomposition is used to determine the singular vectors corresponding to the k largest singular values. A nuclear norm regularization term is introduced to constrain the rank of the low-rank sub-matrices. An alternating minimization strategy is used to optimize the parameters of the two low-rank sub-matrices. The rank parameter k in the low-rank approximation optimization is determined based on the cumulative singular value energy ratio. The rank is stopped when the sum of squares of the current k singular values ​​accounts for 95% of the total sum of squares of singular values.

[0045] The steps for establishing the training dataset for the heterogeneous computing scheduler include: collecting data features and actual execution performance data of historical computing tasks. The data features include data size, data dimension, data sparsity, number of nested loops, and number of conditional branches. The actual execution performance data includes CPU execution time, GPU execution time, and memory bandwidth utilization. For each historical computing task, the CPU execution efficiency and GPU execution efficiency are calculated, defining execution efficiency as the number of floating-point operations completed per unit time. Based on the execution efficiency comparison results, each historical computing task is labeled with its optimal computing resource category. When the GPU execution efficiency exceeds the CPU execution efficiency by more than 1.5 times, it is labeled as a GPU task; otherwise, it is labeled as a CPU task. The labeled dataset is divided into a training set and a validation set in an 8:2 ratio. The training set is augmented by adding Gaussian noise, random scaling, and feature combination to generate expanded samples, increasing the size of the training set to three times its original size. The training steps of the heterogeneous computing scheduler include: initializing the weight parameters of the heterogeneous computing scheduler, using the Xavier initialization method to set the initial weight values ​​to follow a uniform distribution with a mean of 0 and a variance related to the number of neuron units; defining the loss function as a weighted sum of the classification cross-entropy loss and the mutual information regularization term, wherein the mutual information regularization term is approximated by a variational lower bound; updating the weight parameters using the Adam optimization algorithm, setting the learning rate to 0.001, the momentum parameter to 0.9, and the second-order moment estimation parameter to 0.999; introducing a learning rate decay strategy during training, multiplying the learning rate by 0.5 every 20 training epochs; and using early stopping to prevent overfitting, stopping the validation set loss when it has not decreased for 10 consecutive training epochs. Training is stopped when the average processing time of the current batch of data is exceeded. In each training batch, the Lagrange multiplier is dynamically adjusted based on the average processing time deviation. When the actual average processing time exceeds the target processing time, the Lagrange multiplier value is increased to enhance information compression; conversely, the Lagrange multiplier value is decreased to retain more information. After training, a performance evaluation is performed on the validation set, calculating classification accuracy, precision, recall, and F1 score. A classification accuracy of 90% or higher is required for deployment. The optimization of low-rank sub-matrices is achieved using the alternating direction multiplier method, fixing one low-rank sub-matrix while optimizing another, iterating until the objective function converges. The convergence criterion for the objective function is that the change in the objective function value between two consecutive iterations is less than 0.0001.

[0046] The implementation of the representation learning framework based on the information bottleneck principle in the heterogeneous computing scheduler is as follows: The input feature vector is considered the original information source, the intermediate hidden layer representation is considered the compressed representation, and the output decision is considered the final task objective. Information compression is achieved by minimizing the mutual information between the input feature vector and the intermediate hidden layer representation, removing redundant features irrelevant to the task. Key task-related information is preserved by maximizing the mutual information between the intermediate hidden layer representation and the output decision. The information bottleneck objective function is defined as the mutual information of the output decision minus the product of the Lagrange multiplier and the mutual information of the input feature vector. Since mutual information is difficult to calculate directly, a variational approximation method is used to introduce a variational distribution to approximate the true conditional distribution. The random sampling process is represented as a deterministic function with added noise using a reparameterization technique, allowing the gradient to be propagated back through... The process involves several steps: During forward propagation, the input feature vector is mapped by the encoder to mean and variance parameters of the intermediate hidden layer representation. The intermediate hidden layer representation is sampled from a parameterized Gaussian distribution and then output by the decoder to make the final decision. During backpropagation, the gradient of the information bottleneck objective function with respect to the network parameters is calculated, and the network parameters are updated through gradient descent. The Lagrange multiplier, as a hyperparameter controlling the degree of information compression, is dynamically adjusted according to task performance. When task accuracy decreases, the Lagrange multiplier is decreased to retain more information; when computation time exceeds the limit, the Lagrange multiplier is increased to enhance compression. The representation learning framework based on the information bottleneck principle enables the heterogeneous computing scheduler to learn the optimal representation that compresses redundant features while maintaining key task information, thereby reducing the computational burden while ensuring allocation accuracy.

[0047] The technical principle of CPU and GPU collaborative data processing is as follows: The heterogeneous computing scheduler quantifies the parallelism and computational complexity characteristics of data through a computational demand evaluation function, determines the data type based on a preset threshold, and allocates it to corresponding computing resources. This fully leverages the advantages of the CPU in handling control logic and sequential dependency computation, as well as the advantages of the GPU in large-scale parallel computing, avoiding resource waste and performance degradation caused by forcibly allocating unsuitable parallelization tasks to the GPU. A representation learning framework based on the information bottleneck principle enables the heterogeneous computing scheduler to automatically learn the mapping relationship between data features and optimal computing resources during training. Through the dual constraints of minimizing redundant information and maximizing task-related information, it ensures that scheduling decisions are both accurate and efficient. The introduction of variational approximation techniques and the Lagrange multiplier method solves the problem of mutual information difficulty. By directly addressing computational issues, a dynamic balance between information compression rate and task completion fidelity is achieved, enabling the heterogeneous computing scheduler to maintain stable performance under different computational loads. Low-rank matrix factorization significantly reduces the number of model parameters and storage space requirements by decomposing high-dimensional weight matrices into the product of low-rank submatrices, while kernel norm regularization ensures that the decomposed model maintains the approximate accuracy of the original model. The dynamic mesh partitioning strategy adaptively adjusts the optimal mesh size based on the data scale and GPU hardware characteristics, ensuring that each streaming multiprocessor receives an appropriate workload and avoiding idle computing resources caused by load imbalance. The entire collaborative processing mechanism significantly improves computational throughput when processing massive heterogeneous data, providing solid computational support for subsequent parameter optimization and model updates, enabling the system to meet the real-time processing needs of high-frequency data streams.

[0048] The steps of decomposing the parameter optimization problem into single-objective sub-problems using a multi-objective decomposition framework specifically include: defining a set of optimization objective functions, including economic efficiency objective functions, ecological benefit objective functions, and policy synergy objective functions; generating 100 weight vectors using a uniform design method, where each weight vector defines a search direction from the origin to different positions on the Pareto front; decomposing the multi-objective optimization problem into 100 single-objective sub-problems, each single-objective sub-problem corresponding to a weight vector; using a Chebyshev aggregation function to measure the approximation of individual solutions to a reference point, where the Chebyshev aggregation function calculates the weighted maximum value of the distance between the individual solution and the reference point in each objective dimension; initializing the population individuals, using Latin hypercube sampling to generate initial solutions in the parameter space, ensuring that the initial solutions are uniformly distributed in the parameter space; defining the neighborhood of each single-objective sub-problem as the nearest single-objective sub-problems in Euclidean space to the weight vector; and performing global parameter space exploration using a particle swarm optimization algorithm, where each particle maintains its current position, velocity, and historical best position. The population maintains the globally optimal position; when updating particle velocity, inertia, cognition, and social terms are comprehensively considered; when the fitness function has not improved for 20 consecutive generations, it is determined to be in a stagnant state, triggering an algorithm switch; switching to simulated annealing algorithm to perform local fine search, generating perturbation solutions in the neighborhood of the current optimal solution; calculating the objective function value of the perturbation solution, if it is better than the current solution, the perturbation solution is accepted, if it is worse than the current solution, it is accepted with a certain probability, the acceptance probability decreases as the temperature parameter decreases; an adaptive neighborhood adjustment strategy is adopted, dynamically adjusting the neighborhood size according to the convergence speed of the single objective subproblem, expanding the neighborhood for single objective subproblems with slower convergence to enhance information exchange, and shrinking the neighborhood for single objective subproblems with faster convergence to strengthen local search; maintaining an elite archive to store all non-dominated solutions, each time a new solution is generated, comparing the non-dominated relationship with the solutions in the elite archive, adding the non-dominated solution to the elite archive and deleting elite archive solutions dominated by the new solution; after 500 iterations, outputting the Pareto solution set in the elite archive, the Pareto solution set is the set of optimal solutions that weigh multiple objective functions.

[0049] The steps of establishing a sliding window mechanism to retain data from the most recent time period and using a recursive least squares algorithm for incremental updates specifically include: setting the sliding window time length to 720 hours, retaining monitoring data and calculation results from the most recent 720 hours within the sliding window; when new data arrives, deleting the earliest data within the sliding window and adding the new data to the end of the sliding window to maintain a constant amount of data within the sliding window; calculating key statistics, including mean, variance, and covariance matrix, for the data within the sliding window, which are used for incremental updates of DSGE model parameters; using a recursive least squares algorithm to update the DSGE model parameters online, whereby the recursive least squares algorithm uses a recursive formula to calculate the current parameter estimate using the new observation value and the parameter estimate from the previous time step, avoiding recalculating all historical data; the gain matrix of the recursive least squares algorithm is calculated based on the covariance matrix from the previous time step and the current observation vector, and the covariance matrix is ​​recursively updated according to a forgetting factor; the forgetting factor is set to 0.9. 8. The recursive least squares algorithm assigns greater weight to recent data, while its influence on long-term data gradually diminishes; the parameter change rate of the DSGE model is calculated as the norm of the difference between the current parameter estimate and the previous parameter estimate, divided by the norm of the previous parameter estimate; when the parameter change rate exceeds a preset threshold of 5%, the DSGE model structure is deemed to have changed significantly, triggering a local recalibration process; the local recalibration process performs complete parameter estimation on the significantly changed parameter subset, keeping other parameter subsets unchanged; the computational tasks requiring recalibration are allocated to GPUs for execution, utilizing the parallel computing capabilities of GPUs to accelerate sparse matrix operations; the conjugate gradient method is used to solve a large-scale sparse linear equation system, which is derived from the first-order optimality condition of the DSGE model; the condition number of the large-scale sparse linear equation system is improved through preconditioning techniques to accelerate the convergence speed of the conjugate gradient method; after the online incremental update is completed, the updated DSGE model parameters and state variables are output.

[0050] The forgetting factor is a parameter in the recursive least squares algorithm that controls the decay rate of the influence of historical data. Its value ranges from 0 to 1. The closer the value is to 1, the slower the decay of the influence of historical data. The closer the value is to 0, the more the recursive least squares algorithm focuses on recent data. By adjusting the forgetting factor, the DSGE model can balance its tracking of long-term trends and its response to short-term fluctuations.

[0051] The specific steps for performing low-rank approximate decomposition on high-dimensional weight matrices include: identifying high-dimensional weight matrices with dimensions exceeding 1000 in the heterogeneous computing scheduler as decomposition objects; performing singular value decomposition on the high-dimensional weight matrices, decomposing them into the product of a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix; calculating the cumulative energy percentage of the singular values, selecting the top k largest singular values ​​that achieve a cumulative energy percentage of 95%; and constructing a low-rank approximate matrix, retaining the top k largest singular values ​​and their corresponding singular vectors. Discard smaller singular values; represent the low-rank approximation matrix as the product of two low-rank submatrices, where the first low-rank submatrix is ​​the product of the left singular vector and the square root of the singular value, and the second low-rank submatrix is ​​the product of the square root of the singular value and the right singular vector; introduce a nuclear norm regularization term to constrain the rank of the low-rank submatrix, where the nuclear norm is defined as the sum of the singular values ​​of the matrix; employ an alternating minimization strategy to optimize the two low-rank submatrices, fixing the first low-rank submatrix to optimize the second low-rank submatrix, and fixing the second low-rank submatrix to optimize the first low-rank submatrix. The rank submatrix is ​​used alternately until the low-rank decomposition objective function converges. The low-rank decomposition objective function is defined as the square of the Frobenius norm between the high-dimensional weight matrix and the low-rank approximation matrix, plus the nuclear norm regularization term. The low-rank submatrix parameters are updated using gradient descent, and the partial derivatives of the low-rank decomposition objective function with respect to the elements of the low-rank submatrix are calculated. The low-rank submatrix parameters are updated along the negative gradient direction. The convergence criterion is set as the change in the low-rank decomposition objective function value between two consecutive iterations being less than 0.0001. The iteration stops after convergence. The optimized low-rank submatrix replaces the original high-dimensional weight matrix, significantly reducing the number of parameters and storage space of the heterogeneous computing scheduler. The approximation error of the low-rank approximation matrix is ​​verified, and the relative error between the low-rank approximation matrix and the high-dimensional weight matrix is ​​calculated, requiring the relative error to be less than 2%. The structure of the heterogeneous computing scheduler after low-rank decomposition and the low-rank submatrix parameters are output. The optimized low-rank submatrix is ​​used as the weight parameters of the DSGE model for trade-off analysis and decision support of the economic efficiency objective function, ecological benefit objective function, and policy synergy objective function.

[0052] The nuclear norm is the sum of all singular values ​​of a matrix, representing a convex relaxation form of the matrix's rank function. In optimization problems, it serves as a regularization term constraining the matrix's rank, encouraging optimization results to tend towards a low-rank structure. It is widely used in matrix completion and sparse representation learning. The Frobenius norm is the square root of the sum of the squares of matrix elements, measuring the size of the matrix. In low-rank approximation, it measures the difference between the low-rank approximation matrix and the high-dimensional weight matrix. Minimizing the square of the Frobenius norm is equivalent to least squares estimation. The alternating minimization strategy is an iterative method for solving multivariate optimization problems. It divides variables into several groups, fixing one group of variables in each iteration and optimizing another group, repeating this process until convergence. The alternating minimization strategy decomposes joint optimization into multiple simpler sub-optimization problems, reducing the difficulty of solving them.

[0053] The multi-source heterogeneous data stream encompasses comprehensive information on the marine industry and ecological environment, typically including industrial economic data, ecological environment data, policy data, and technical data. Industrial economic data includes petrochemical industry output value, seawater desalination capacity, aquaculture output, and port operation data. Ecological environment data includes petrochemical wastewater discharge data, carbon emissions, and blue carbon absorption by aquaculture. Concentration and water quality monitoring indicators; policy data including government carbon tax rate setting documents, blue carbon subsidy policy documents, technical subsidy intensity regulations, and environmental protection regulations; technical data including seawater desalination membrane technical parameters, petrochemical emission reduction technical indicators, and aquaculture technical efficiency data; the multi-source heterogeneous data stream includes monthly collected industry monitoring data and quarterly collected ecological environment monitoring data in the time dimension; it is distributed across monitoring stations in different geographical locations in the spatial dimension; and the data format includes structured numerical data, semi-structured document data, and unstructured policy text data. The multi-source heterogeneous data stream originates from the real-time sensor network of the industry monitoring system, the periodic sampling and analysis of ecological environment monitoring stations, the document management system of the policy database, and the industry standard records of the technical database. The unstructured policy text data is semantically understood and information extracted using a large language model. Structured prompt templates guide the large language model to identify carbon tax rates, subsidy amounts, and implementation dates in policy documents. The large language model's contextual understanding capabilities are used to locate company names, industry types, geographical regions, and their interrelationships within the policy documents. Mind chain prompting technology guides the large language model to analyze the logical connections between policy measures and their target audience. A few-shot learning method enables the large language model to identify policy intensity descriptors in policy documents and convert them into numerical weights according to predefined quantification rules. The structured information output by the large language model is organized according to a predefined JSON format into structured records containing policy type, policy parameters, scope of application, effective date, and policy intensity. The semi-structured technical document data is parsed using a large language model to analyze technical indicator tables and technical specification paragraphs. The large language model extracts the numerical range, unit information, and test conditions of technical parameters through prompts for engineering guidance. It utilizes the multimodal understanding capabilities of the large language model to process technical documents containing charts and extract key data points from them. A confirmatory prompting strategy ensures the accuracy and consistency of the large language model's extraction results. The extracted technical parameters are organized into structured data tables according to technical category, parameter name, parameter value, parameter unit, and test conditions. After preprocessing and structure transformation, the multi-source heterogeneous data streams constitute the foundational data resources for marine industry and ecological correlation analysis, providing data support for DSGE model construction, multi-sectoral coupled simulation, policy scenario analysis, and decision support.

[0054] The DSGE model is a mathematical model describing the dynamic evolution of a macroeconomic system. By establishing the optimal behavior of representative economic agents, market clearing conditions, and stochastic shock processes, it characterizes the fluctuation patterns of economic variables near the long-term equilibrium state. The DSGE model includes household utility maximization, firm profit maximization, government budget constraints, and market equilibrium conditions. The state-space representation of economic variables is obtained by solving the first-order optimality conditions of the DSGE model. This state-space representation consists of state equations and observation equations. The state equations describe the dynamic relationship between endogenous variables and exogenous shocks, while the observation equations establish the correspondence between endogenous variables and actual observed data. The parameters of the DSGE model are calibrated using Bayesian estimation or maximum likelihood estimation methods. The calibrated DSGE model is then used for... This study analyzes the dynamic impacts of industrial and environmental policies on economic growth and ecological protection. By simulating economic trajectories under different policy scenarios, it provides quantitative evidence for policy formulation. The economic efficiency objective function, ecological benefit objective function, and policy synergy objective function are constructed by weighted combination of endogenous variables from the DSGE model. The low-rank sub-matrix parameters are embedded as weight coefficients in the weighted calculation of the objective functions. By adjusting the low-rank sub-matrix parameters, a trade-off between different policy objectives is achieved. The Pareto solution set provides multiple low-rank sub-matrix parameter configuration schemes. Decision-makers select appropriate low-rank sub-matrix parameter configurations according to their actual policy preferences. The selected low-rank sub-matrix parameters are substituted into the DSGE model to perform policy simulation and effect evaluation. The dynamic prediction results of industrial output, pollution emissions, and residents' welfare are output for decision support.

[0055] Furthermore, this invention addresses the technical problem that structural differences in cointegration relationships within industrial ecosystems under different pollution concentration regimes make it difficult for traditional linear models to accurately characterize the dynamic features of the system. This invention identifies nonlinear threshold points by performing a threshold cointegration test. Based on these thresholds, the data is divided into low-concentration, medium-concentration, and high-concentration regimes. Within each regime, the long-term cointegration relationship between industrial output and ecological indicators is estimated, and error correction terms are extracted. A state transition matrix is ​​constructed to characterize the switching probability of the system between different regimes, and the regime transition rules are embedded into the state-space equations of the DSGE model. This method allows for differences in cointegration vectors under different pollution concentration regimes, capturing the nonlinear long-term equilibrium relationship of the industrial ecosystem. This enables the model to accurately reflect the dynamic response characteristics of the system under different environmental pressure levels, providing policymakers with more precise scenario analysis tools and improving the reliability of policy simulation and effect evaluation.

[0056] Specifically, the principle of this invention is as follows: The solution to this technical problem lies in using a heterogeneous computing scheduler to allocate data processing tasks to the most suitable computing architecture based on their parallelism and computational complexity characteristics. This avoids the resource waste caused by forcibly assigning unsuitable sequentially dependent tasks to the graphics processor (GPU), while simultaneously leveraging the GPU's massively parallel computing capabilities to accelerate the execution of highly parallel tasks. The representation learning framework based on the information bottleneck principle removes redundant features by minimizing the mutual information between input features and intermediate representations, while maximizing the mutual information between intermediate representations and output decisions to preserve key task information. This allows the scheduler to learn the optimal representation that both compresses redundancy and retains key information. Combined with the Lagrange multiplier method, it dynamically balances information compression rate and task completion fidelity, ensuring that scheduling decisions remain stable and accurate under different computational loads. Low-rank matrix factorization identifies the main energy distribution direction of the high-dimensional weight matrix through singular value decomposition and retains the singular vectors corresponding to the first few largest singular values. Kernel norm regularization is used to constrain the rank of the decomposed matrix, significantly reducing the number of model parameters and storage space requirements. The recursive least squares algorithm uses a forgetting factor to assign greater weight to recent data and implements online incremental updates of parameters through a recursive formula. When the parameter change rate exceeds a threshold, a local recalibration process is triggered to perform a complete estimation on the significantly changed subset of parameters. The sliding window mechanism ensures that the model is always updated based on the latest time period data. The entire technical solution achieves efficient real-time processing of massive heterogeneous data streams and rapid calibration of DSGE models through the organic combination of optimized allocation of computing resources, dimensionality reduction of model parameters, and incremental update mechanism, thereby supporting the needs of real-time decision optimization.

[0057] The following provides a specific embodiment 1 of the present invention. The specific implementation of step S1 in this embodiment 1 is the same as that described above, and will not be repeated in detail here. The specific implementation of each other step is described in detail below.

[0058] The specific implementation of step S2 is to perform a threshold cointegration test on the preprocessed data to identify nonlinear threshold points, and select pollution concentration as the threshold variable. The unit is Set the candidate set of threshold values ​​as In the formula For the first One candidate threshold value, in units of , This represents the total number of candidate threshold values, expressed in units of quantiles. The candidate threshold values ​​cover the range of sample quantiles from 0.15 to 0.85. For each threshold value in the candidate set... The sample is divided into two sub-sample sets. and subsample set The cointegration vectors are estimated separately, where... This represents the set of times when the pollution concentration does not exceed the threshold value. Let represent the set of times when the pollution concentration exceeds the threshold value, and calculate the sum of squared residuals for each subsample. In the formula Threshold value The corresponding residual sum of squares, in units of 100 million yuan squared. For subsamples Inner Time The cointegrated regression residuals, in units of 100 million yuan. For subsamples Inner Time The cointegrated regression residuals, in units of 100 million yuan, are used to generate a simulated distribution and calculate critical values ​​using the bootstrap method. The likelihood ratio statistic is expressed as... In the formula This is the likelihood ratio statistic, which is dimensionless. This refers to the sample size, expressed in units of samples. Given the sum of squared residuals without considering the threshold effect (in billions of yuan), identify the threshold value that minimizes the sum of squared residuals as the optimal threshold point. The data is divided into low-concentration regions. medium concentration zone collection and high concentration zone collection In the formula , , These represent the sets of times under low, medium, and high concentration conditions, respectively. and These are the first and second optimal threshold values, respectively, in units of 1. Estimate industrial output separately within each district. With ecological indicators The long-term cointegration relationship between them is In the formula The area index takes values ​​from 1 to 3. For a moment The industrial output, expressed in 100 million yuan. For reference purposes, industry output is expressed in 100 million yuan, with an empirical value of 10 billion yuan. For a moment The ecological indicators, which are dimensionless, have been standardized. For district system The intercept term, in units of 100 million yuan, For district system The cointegration coefficient is dimensionless. The error correction term is expressed in hundreds of millions of yuan. The state transition matrix is ​​constructed as follows: , where matrix elements Indicates that the system is based on the regional system. Transfer to district system The probability, dimensionless, satisfies The formula for estimating the transition probability is: In the formula To be from the district system Transfer to district system The number of observations, in units of times. For the district system The total number of observations, expressed in times, is used to determine the transition probability parameters through maximum likelihood estimation. The likelihood function is: In the formula The value is the likelihood function, which is dimensionless. For a moment The zone status takes values ​​from 1 to 3.

[0059] The specific implementation of step S3 is as follows: Data is input into a heterogeneous computing scheduler for resource allocation determination; a computing demand assessment function is constructed to quantify the parallelism and computational complexity characteristics of the data; the parallelism characteristic is calculated using the following formula: In the formula This is a parallelism eigenvalue, dimensionless, ranging from 0 to 1. Data dimension, unit is dimension. The maximum data dimension, in units of dimensions, is an empirical value. dimension, This is a data independence metric, dimensionless, calculated by the proportion of data blocks with no dependencies on each other to the total number of data blocks. Its value ranges from 0 to 1. The maximum independence index is dimensionless, with an empirical value of 1. This represents the number of iterative dependency levels, in layers. This represents the maximum number of iteration layers, expressed in layers. An empirical value is 100 layers. , , The weighting coefficients are dimensionless and satisfy the following conditions: The default values ​​are 0.4, 0.3, and 0.3, respectively. The formula for calculating the computational complexity feature is as follows: In the formula The eigenvalues ​​for calculating complexity are dimensionless and range from 0 to positive infinity. This represents the number of floating-point operations, expressed in units of times. The reference number of floating-point operations is expressed in units of times, and the empirical value is... Second-rate, This refers to the number of memory accesses, measured in times. For reference, the number of memory accesses is in times, and the empirical value is... Second-rate, The number of branch prediction failures, expressed in times. The number of branch predictions is for reference, in units of times, and the empirical value is... Second-rate, , , The weighting coefficients are dimensionless and satisfy the following conditions: The default values ​​are 0.5, 0.3, and 0.2. and Data is identified as highly parallel and allocated to the GPU. The data allocated to the GPU is then meshed, and the optimal mesh size is calculated using the following formula: In the formula The optimal grid size is expressed in blocks. This represents the maximum grid size supported by the GPU, in blocks; an empirical value is 65535 blocks. The data size is expressed in units of data. This refers to the number of GPU cores, expressed in units. The amount of data processed per thread block, in units of numbers; an empirical value is 256 numbers. (symbol) This represents the function for rounding up.

[0060] The specific implementation of step S4 is to decompose the parameter optimization problem into single-objective sub-problems using a multi-objective decomposition framework, and define a set of optimization objective functions that includes an economic efficiency objective function. Ecological benefit objective function and policy coordination objective function In the formula This is the parameter vector of the DSGE model. A standardized economic efficiency index, dimensionless. The standardized ecological benefit index is dimensionless. To standardize the policy synergy index, which is dimensionless, a uniform design method was used to generate 100 weight vectors. In the formula The weight vector index takes values ​​from 1 to 100. For the first The weight vector of the th weight vector Each component, dimensionless, satisfies and Chebyshev aggregation function is expressed as In the formula The value of the Chebyshev aggregation function is dimensionless. The reference point vector has all components that are dimensionless. For the objective function Ideal value, dimensionless. and The objective functions are respectively The maximum and minimum values ​​in the population are dimensionless. The particle swarm optimization algorithm's update rate formula is: In the formula For particles In the The first generation Velocity components in one parameter dimension For particles In the The first generation Velocity components in one parameter dimension The inertial weight is dimensionless and has an empirical value of 0.7. and The acceleration factor is dimensionless, and its empirical value is 2.0. and A random number between 0 and 1, dimensionless. For particles The The individual historical best position of a particle in each parameter dimension. In the historical iterations, the first The position where each parameter obtains the optimal objective function value. For particles In the The first generation The position of each parameter dimension For the population in the first The globally optimal position in the nth parameter dimension, i.e., the position of all particles in the historical iteration. The position when each parameter obtains the globally optimal objective function value is given by the position update formula. The normalization formula for the change in the objective function value of the simulated annealing algorithm is: In the formula This is the change in the normalized objective function value, which is dimensionless. The new solution after the perturbation. Given the current solution, the acceptance probability is... In the formula Let be the acceptance probability of the perturbation solution, dimensionless, and ranging from 0 to 1. The parameter is a dimensionless analogous parameter to the Boltzmann constant, with an empirical value of 1.0. The current temperature parameter is dimensionless, and the temperature update formula is: In the formula For the first The temperature parameters of the time are dimensionless. The cooling coefficient is dimensionless, ranging from 0.8 to 0.95, with an empirical value of 0.9. The average convergence rate is calculated using the following formula: In the formula For the first The average convergence rate of the algebra is dimensionless. For the first The population average objective function value of generation is dimensionless, and the adaptive adjustment formula for neighborhood size is as follows: In the formula For the first The neighborhood size of a generation, in units of 1. For the first The neighborhood size of a generation, in units of 1. Let be the target convergence rate, dimensionless, with an empirical value of 0.05. The adjustment coefficient is dimensionless and has an empirical value of 0.3.

[0061] The specific implementation of step S5 is as follows: a sliding window mechanism is established to retain data from the most recent time period. The sliding window time length is set to 720 hours, and the parameter update formula for the recursive least squares algorithm is as follows: In the formula For a moment The parameter estimation vector has the same dimension as the number of parameters. For a moment The parameter estimation vector, Here is the gain matrix. For a moment The observed values, For a moment The regression vector has the same dimension as the number of parameters, and the gain matrix is ​​calculated using the following formula: In the formula For a moment The covariance matrix, whose dimension is the product of the number of parameters. Let be the forgetting factor, dimensionless, with an empirical value of 0.98. The formula for updating the covariance matrix is: In the formula For a moment The covariance matrix and the formula for calculating the rate of change of parameters are as follows: In the formula The rate of change of the parameter is dimensionless and has the following symbol. Denotes the Euclidean norm of a vector, when The local recalibration process is triggered at this time, and the iterative formula of the conjugate gradient method is as follows: In the formula For the first The solution vector of the nth iteration. For the first The step size of the next iteration is dimensionless and is determined through line search. For the first The search direction for the next iteration is updated using the following formula: In the formula For the first The residual vector of the next iteration The coefficient matrix, A constant vector, The conjugate coefficient is dimensionless and is calculated using the following formula: .

[0062] The specific implementation of step S6 is to process the high-dimensional weight matrix... Perform singular value decomposition In the formula is the number of rows in the weight matrix, in dimensionality. The column number of the weight matrix, in dimensionality. It is a left singular vector matrix, and each column vector is an orthogonal unit vector. This is a singular value diagonal matrix, where the diagonal elements are non-negative singular values ​​arranged in descending order. Given a right singular vector matrix, where each column vector is an orthogonal unit vector, calculate the cumulative energy percentage of the singular values. In the formula For the front The cumulative energy percentage of each singular value, dimensionless. The number of singular values ​​retained, expressed in units of 1. For the first A singular value, dimensionless. Let be the rank of the matrix, in units of , and choose such that The smallest The value is used to construct a low-rank approximation matrix. In the formula For the front Left singular vector matrix, For the front A diagonal matrix composed of singular values. For the front The right singular vector matrix represents the low-rank approximation matrix as the product of two low-rank submatrices. In the formula This is the first low-rank submatrix. Represents a diagonal matrix The diagonal matrix obtained by taking the square root of each diagonal element. Let be the second low-rank submatrix, and let the objective function of the low-rank decomposition be... In the formula The objective function value is a low-rank decomposition and is dimensionless. Denotes the Frobenius norm. The nuclear norm is the sum of all singular values ​​of a matrix. is the regularization coefficient, dimensionless, with an empirical value of 0.01. The subgradient of the nuclear norm is calculated as follows: In the formula and Each is a matrix The left and right singular vector matrices are fixed in the alternating minimization strategy. optimization The gradient is ,fixed optimization The gradient is In the formula and Each is a matrix The formula for updating the parameters of the left singular vector matrix, the right singular vector matrix, and the low-rank submatrix is ​​as follows: and In the formula This is the index for the number of iterations, in units of iterations. Let be the learning rate, dimensionless, with an empirical value of 0.001, and the convergence criterion be... The formula for calculating relative error is: In the formula For relative error, dimensionless, and requiring .

[0063] To better understand and implement this invention, the following is a specific application scenario example 2: The technical team, addressing the core characteristics of a certain region—"heavy chemical industry dominance - intensified ecological constraints - urgent green transformation"—faced a real-time analysis need for multi-source heterogeneous data, including a petrochemical industry output value of 200 billion yuan, a seawater desalination capacity of 182,000 tons / day, blue carbon sequestration from seaweed farming in Haizhou Bay, and shore power from 29 berths in the port. Simultaneously, they needed to process... Daily average 24 Ecological challenges include exceeding standards, high COD concentrations in nearshore waters, and insufficient utilization of blue carbon resources. The technical team has built a system capable of processing massive amounts of heterogeneous data and executing multi-objective optimization decisions. Based on the theory of multi-sectoral dynamic stochastic general equilibrium, it achieves industry-ecology-policy coupled analysis of five main entities: petrochemicals, seawater desalination, aquaculture, ports, and government, providing quantitative decision support for the region's environmental planning and carbon neutrality goals.

[0064] The technical team first performed data acquisition and preprocessing. Raw data streams were collected from the region's industry monitoring system, including data on the petrochemical industry's output value of 200 billion yuan. Daily average 24 Data included: daily seawater desalination capacity of 182,000 tons and daily reclaimed water data; blue carbon sequestration data from laver aquaculture in Haizhou Bay, including 50,000 tons of carbon sequestration annually by *Porphyra tenera*; and carbon reduction data from shore power at 29 berths in the port and the Tianwan Nuclear Power Plant, totaling 1.07 million tons per year. Nearshore water quality monitoring data, including COD concentration, dissolved oxygen, nitrogen, phosphorus, and other ecological parameters, was collected from ecological and environmental monitoring stations. Policy documents, such as government carbon tax documents and documents related to the central government's 203 million yuan pollution prevention and control fund, were collected from policy databases. Technical specifications, including seawater desalination membrane technology parameters, petrochemical emission reduction technology indicators, and specifications for intelligent aquaculture equipment, were extracted from technical databases. The collected raw data streams presented challenges such as missing value imputation, outlier identification, inconsistent time scales, and spatial resolution differences. The technical team performed missing value imputation on the raw data streams, using a combination of forward and backward imputation to handle time-series breakpoints. The mean and standard deviation of each monitoring indicator were calculated, and outliers exceeding three times the standard deviation were identified and replaced with moving averages. Data from different collection frequencies were standardized in terms of time scale, and linear interpolation was used to generate intermediate values, unifying monthly industry data and quarterly ecological monitoring data to a quarterly scale. The latitude and longitude coordinates of each monitoring station were extracted, and a nearest neighbor interpolation algorithm was used to match data with different spatial resolutions. Based on GIS technology, monthly industry data and quarterly nearshore water quality data were spatiotemporally aligned. A minimum-maximum normalization transformation was performed on all numerical data, mapping different units of measurement, such as output value in billions of yuan and pollution discharge in tons, to a standard range of 0 to 1.

[0065] For unstructured policy texts, the technical team employed a large language model (MLM) for semantic understanding and information extraction. Structured prompt templates guided the MLM to identify numerical parameters, entity information, and their relationships within policy documents, including key parameters such as carbon tax rates, subsidy amounts, and implementation dates. Mind chain prompting technology guided the MLM to analyze the logical connections between policy measures and their target audiences. A few-shot learning method enabled the MLM to identify policy intensity descriptors and convert them into numerical weights. The MLM output was organized into structured records containing policy type, policy parameters, scope of application, effective date, and policy intensity, according to a predefined pattern. For semi-structured technical documents, the team used the MLM to parse technical indicator tables and specification paragraphs. Prompt engineering guided the MLM to extract the numerical range, unit information, and testing conditions of technical parameters. The MLM's multimodal understanding capabilities were utilized to process technical documents containing charts and graphs and extract key data points from them. The extracted technical parameters were organized into structured data tables according to technology category, parameter name, parameter value, parameter unit, and testing conditions. Based on industry types including petrochemicals, seawater desalination, aquaculture, and ports, and data attributes including economic and ecological categories, a marine industry-ecology related database for the region was constructed, data access permissions were configured, and data integrity was improved to over 95%.

[0066] The technical team performed threshold cointegration tests on the preprocessed data to identify nonlinear threshold points. Pollution concentration was selected as the threshold variable, and the candidate threshold values ​​were set to cover the range of sample quantiles from 0.15 to 0.85. For each threshold value in the candidate set, the sample was divided into two subsamples, and the cointegration vector was estimated and the sum of squared residuals was calculated within each subsample. A bootstrap method was used to generate a simulated distribution and calculate the critical value. The significance of the threshold effect was tested using the likelihood ratio statistic. The threshold value that minimized the sum of squared residuals was identified as the optimal threshold point, dividing the data into a low-concentration zone (below the first threshold), a medium-concentration zone (between the first and second thresholds), and a high-concentration zone (above the second threshold). Within each zone, the long-term cointegration relationship between industrial output and ecological indicators was estimated, and error correction terms for each zone were extracted. A state transition matrix was constructed to characterize the switching probability of the system between different zones. The transition probability parameters were determined using maximum likelihood estimation, and the zone transition rules were embedded into the state-space equations of the DSGE model. The state transition matrix indicates that low-concentration regimes have a strong self-sustaining capacity, medium-concentration regimes are at risk of transitioning to high-concentration regimes, and high-concentration regimes require policy intervention to transition to low-concentration regimes.

[0067] The technical team constructed five principal optimization equations, including a petrochemical enterprise production equation with a capital output elasticity of 0.6; a blue carbon absorption equation for aquaculture enterprises with a unit output blue carbon coefficient of 0.1; and a government carbon tax-subsidy budget equation to establish a balance between carbon tax revenue and subsidy expenditure. Based on macroeconomic benchmarks, a time discount factor of 0.99 corresponds to an annual real interest rate of 4%, and a capital depreciation rate of 0.025 corresponds to an annual depreciation of 2.5%. Based on industry data from the new district, a petrochemical unit pollution coefficient of 0.5 and a seawater desalination wastewater input elasticity of 0.3 were used. Based on literature values, a pollution negative utility weight of 0.1 and a pollution natural decay rate of 0.1 were used. 32 parameters of the model were assigned values. By simultaneously solving state-space equations such as the household consumption equation and the petrochemical enterprise capital marginal return equation, the steady-state values ​​of each variable, including the steady-state output of petrochemicals and the steady-state blue carbon content of aquaculture, were solved to determine the benchmark point for dynamic analysis of the model. The five principal equations achieved a goodness of fit of 92%, providing an accurate theoretical basis for policy simulation.

[0068] The technical team inputs the data into a heterogeneous computing scheduler for resource allocation. A computing demand assessment function is constructed to quantify the parallelism and computational complexity characteristics of the data. The parallelism characteristic is calculated comprehensively based on data dimensionality, data independence, and iterative dependencies, while the computational complexity characteristic is obtained through a comprehensive evaluation of floating-point operations, memory access patterns, and branch prediction difficulty. Figure 9As shown, the heterogeneous computing scheduler consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives the parallelism feature vector and the computational complexity feature vector of the data. The first hidden layer employs a representation learning framework based on the information bottleneck principle. It minimizes the mutual information between the input features and the intermediate representation to constrain redundant information, while maximizing the mutual information between the intermediate representation and the computing resource allocation decision to preserve key information. The first hidden layer contains 64 neurons, and the activation function uses parameterized rectified linear units. The second hidden layer performs a nonlinear transformation on the output of the first hidden layer, using a sparse connection mechanism to reduce the number of parameters. The second hidden layer contains 32 neurons. The output layer outputs the CPU allocation probability and the GPU allocation probability, selecting the computing resource with the higher probability as the final allocation result. The output value of the computing demand evaluation function is compared with a preset threshold. When the parallelism feature value is greater than 0.7 and the computational complexity feature value is greater than 0.6, the data is determined to be highly parallel and allocated to the GPU; otherwise, it is determined to be sequence-dependent data and allocated to the CPU.

[0069] Perform grid partitioning on the highly parallel data allocated to the GPU, based on the data size, the number of GPU cores (2560), and the cache capacity (48). The optimal grid size is dynamically adjusted to 32×32. The partitioned data blocks are mapped to the GPU's streaming multiprocessor array, with each streaming multiprocessor responsible for processing one or more data blocks. Fast access to data within the same block is achieved through shared memory. For sequentially dependent data allocated to the CPU, recursive computation is performed in time-series order, utilizing the CPU's high clock speed of 3.6GHz. Advantages include handling control logic and branching decisions. A variational approximation technique is introduced in the first hidden layer to estimate mutual information terms, and end-to-end training is achieved through reparameterization techniques. The Lagrange multiplier method is used to dynamically balance information compression rate and task completion fidelity; the Lagrange multiplier is adaptively adjusted based on the average processing time of the current batch of data, the target processing time, and the parameter change rate.

[0070] The technical team established a training dataset for the heterogeneous computing scheduler. Data features and actual execution performance data of historical computing tasks were collected. These data features included data size, data dimensionality, data sparsity, number of nested loops, and number of conditional branches. The actual execution performance data included CPU execution time, GPU execution time, and memory bandwidth utilization. For each historical computing task, CPU and GPU execution efficiency were calculated, with execution efficiency defined as the number of floating-point operations performed per unit time. Based on the execution efficiency comparison results, each historical computing task was labeled with its optimal computing resource category: a GPU task was labeled if its GPU execution efficiency exceeded that of the CPU by more than 1.5 times; otherwise, it was labeled as a CPU task. The labeled dataset was divided into training and validation sets in an 8:2 ratio. Data augmentation was performed on the training set by adding Gaussian noise, random scaling, and feature combination to generate expanded samples, increasing the training set size to three times its original size. The weight parameters of the heterogeneous computing scheduler were initialized using the Xavier initialization method, setting the initial weight values ​​to follow a uniform distribution with a mean of 0 and a variance related to the number of neurons. The loss function is defined as the weighted sum of the classification cross-entropy loss and the mutual information regularization term, which is approximated by a variational lower bound. The Adam optimization algorithm is used to update the weight parameters, with a learning rate of 0.001, a momentum parameter of 0.9, and a second-moment estimation parameter of 0.999. A learning rate decay strategy is introduced during training, multiplying the learning rate by 0.5 every 20 training epochs. Early stopping is used to prevent overfitting; training stops when the validation set loss does not decrease for 10 consecutive training epochs. In each training batch, the Lagrange multiplier is dynamically adjusted based on the average processing time deviation of the current batch of data. When the actual average processing time exceeds the target processing time, the Lagrange multiplier value is increased to enhance information compression; conversely, the Lagrange multiplier value is decreased to retain more information. After training, performance is evaluated on the validation set, calculating classification accuracy, precision, recall, and F1 score. A classification accuracy of 90% or higher is required for deployment.

[0071] The technical team performed low-rank approximate decomposition on high-dimensional weight matrices. High-dimensional weight matrices with dimensions exceeding 1000 in heterogeneous computing schedulers were identified as decomposition objects. Singular value decomposition was performed on the high-dimensional weight matrices, decomposing them into the product of a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. For example... Figure 10As shown, the cumulative energy percentage of singular values ​​is calculated, and the top k largest singular values ​​that achieve a cumulative energy percentage of 95% are selected. A low-rank approximation matrix is ​​constructed, retaining the top k largest singular values ​​and their corresponding singular vectors, while discarding smaller singular values. The low-rank approximation matrix is ​​represented as the product of two low-rank submatrices. The first low-rank submatrix is ​​composed of the product of the left singular vector and the square root of the singular value, and the second low-rank submatrix is ​​composed of the product of the square root of the singular value and the right singular vector. A nuclear norm regularization term is introduced to constrain the rank of the low-rank submatrix, where the nuclear norm is defined as the sum of the singular values ​​of the matrix. The optimization of the low-rank submatrix is ​​achieved using the alternating direction multiplier method, fixing one low-rank submatrix and optimizing the other, iterating until the objective function converges. The convergence criterion for the objective function is that the change in the objective function value between two consecutive iterations is less than 0.0001. The parameters of the low-rank submatrix are updated using the gradient descent method. The partial derivatives of the low-rank decomposition objective function with respect to the elements of the low-rank submatrix are calculated, and the parameters of the low-rank submatrix are updated along the negative gradient direction. Replacing the original high-dimensional weight matrix with the optimized low-rank submatrix significantly reduces the number of parameters and storage space in the heterogeneous computing scheduler. The approximation error of the low-rank approximation matrix is ​​verified by calculating the relative error between the low-rank approximation matrix and the high-dimensional weight matrix, requiring the relative error to be less than 2%. The structure of the heterogeneous computing scheduler after low-rank decomposition and the parameters of the low-rank submatrix are output.

[0072] The technical team employed a multi-objective decomposition framework to break down the parameter optimization problem into single-objective subproblems. A set of optimization objective functions was defined, including economic efficiency, ecological benefit, and policy synergy objectives. A uniform design method was used to generate 100 weight vectors, each defining a search direction from the origin to different positions on the Pareto front. The multi-objective optimization problem was further decomposed into 100 single-objective subproblems, each corresponding to a weight vector. A Chebyshev aggregation function was used to measure the approximation of individual solutions to a reference point, calculating the weighted maximum distance between the individual solution and the reference point across all objective dimensions. The population was initialized, and initial solutions were generated in the parameter space using Latin hypercube sampling, ensuring a uniform distribution of initial solutions within the parameter space. The neighborhood of each single-objective subproblem was defined as the nearest single-objective subproblems in Euclidean space to the weight vector. A global parameter space exploration was performed using a particle swarm optimization algorithm. Each particle maintained its current position, velocity, and historical best position, while the population maintained the global best position. When updating particle velocities, inertia, cognition, and social terms were comprehensively considered. When the fitness function fails to improve for 20 consecutive generations, it is considered stagnant, triggering an algorithm switch. The algorithm switches to simulated annealing to perform a local fine-grained search, generating perturbation solutions within the neighborhood of the current optimal solution. The objective function value of the perturbation solution is calculated; if it is better than the current solution, it is accepted; otherwise, it is accepted with a certain probability, which decreases as the temperature parameter decreases. An adaptive neighborhood adjustment strategy is employed, dynamically adjusting the neighborhood size based on the convergence speed of the single-objective subproblem. For slow-converging subproblems, the neighborhood is expanded to enhance information exchange; for fast-converging subproblems, the neighborhood is shrunk to strengthen local search. An elite archive is maintained, storing all non-dominated solutions. Each time a new solution is generated, its non-dominated relationship is compared with the solutions in the elite archive. Non-dominated solutions are added to the elite archive, and solutions dominated by the new solution are deleted. After 500 iterations, the Pareto solution set in the elite archive is output; this Pareto solution set represents the optimal set of solutions that balance multiple objective functions.

[0073] The technical team established a sliding window mechanism to retain data from the most recent time period and used a recursive least squares algorithm to perform incremental updates. The sliding window time length was set to 720. The sliding window retains the most recent 720 The monitoring data and calculation results are used. When new data arrives, the earliest data within the sliding window is deleted, and the new data is added to the end of the sliding window to keep the amount of data in the sliding window constant. Key statistics, including mean, variance, and covariance matrix, are calculated for the data within the sliding window. These key statistics are used for incremental updates of DSGE model parameters. A recursive least squares algorithm is used to update the DSGE model parameters online. This algorithm uses a recursive formula to calculate the current parameter estimate using the new observation and the parameter estimate from the previous time step, avoiding recalculation of all historical data. The gain matrix of the recursive least squares algorithm is calculated based on the covariance matrix from the previous time step and the current observation vector. The covariance matrix is ​​recursively updated using a forgetting factor. The forgetting factor is set to 0.98, giving the recursive least squares algorithm a larger weight for recent data and gradually decreasing its influence on older data. The parameter change rate of the DSGE model parameters is calculated as the norm of the difference between the current parameter estimate and the parameter estimate from the previous time step, divided by the norm of the parameter estimate from the previous time step. When the rate of parameter change exceeds a preset threshold of 5%, a significant change in the DSGE model structure is determined, triggering a local recalibration process. This local recalibration process performs complete parameter estimation on the significantly changed subset of parameters, while keeping other parameter subsets unchanged. The computational tasks requiring recalibration are allocated to the GPU for execution, leveraging the GPU's parallel computing capabilities to accelerate sparse matrix operations. A large-scale sparse linear equation system, derived from the first-order optimality conditions of the DSGE model, is solved using the conjugate gradient method. Preconditioning techniques are used to improve the condition number of the large-scale sparse linear equation system, accelerating the convergence speed of the conjugate gradient method. After the online incremental update is completed, the updated DSGE model parameters and state variables are output.

[0074] The technical team used the optimized low-rank submatrix as weight parameters in the DSGE model for decision analysis. The policy variables were defined to range from a carbon tax rate of 0.1 to 0.5 yuan / ton and a blue carbon subsidy rate of 0.3 to 0.8 yuan / ton. The tiered subsidy for seawater desalination ranges from 3.6 to 4.2 yuan / ton, supporting both single and continuous policy shock settings. A logarithmic linearization model is used to generate policy shock response paths. For example... Figure 1 As shown, under the carbon tax policy, the green transformation of petrochemical enterprises is gradually improving. In the short term (0-10 periods), the analysis shows that petrochemical output will decrease by 3% to 5% under the impact of the carbon tax. In the long term (30 periods), the analysis shows that the output value of aquaculture converges to a steady-state growth rate of 4% to 6%. Figure 2 As shown, under the carbon tax policy, the combined impulse responses of petrochemical, seawater desalination, and aquaculture enterprises exhibit synergistic evolution characteristics. Subsidy elasticity calculations indicate that a 1% increase in blue carbon subsidies leads to a 0.8% increase in aquaculture blue carbon. Baseline scenarios include no carbon tax subsidies, a single scenario including only carbon tax, and a combined scenario including carbon tax plus membrane technology subsidies. Comparison of output scenarios shows that the combined scenario improves petrochemical emission reduction efficiency by 25% compared to the single scenario. For example... Figure 3 As shown, under the subsidy policy, petrochemical, seawater desalination, and aquaculture enterprises exhibit differentiated response models to their green transformation. For example... Figure 4 As shown, the multi-combination pulse diagram of greening in petrochemical, seawater desalination, and aquaculture enterprises under the subsidy policy reveals the time lag effect of policy transmission. For example... Figure 5 As shown, technological innovation is accelerating the greening process for petrochemical, seawater desalination, and aquaculture enterprises. Figure 6 As shown in the diagram, the multi-combination pulse diagram of greening of petrochemical, seawater desalination and aquaculture enterprises under technological innovation shows the technology spillover effect.

[0075] The technical team calculated the region's GDP growth rate under various policy scenarios. The combined policy scenario showed an annual growth rate increase of 0.8%. Changes in petrochemical enterprise profits indicated a short-term decrease of 2% due to the carbon tax impact, followed by a long-term recovery of 1.5% due to technological innovation. The return on investment for seawater desalination reached 8% when subsidies were at 4.2 yuan / ton. Quantitative ecological indicators showed that the carbon tax impact resulted in a COD reduction of 2056.78 tons / year and a port carbon emission reduction of 1.07 million tons / year. The increase in blue carbon from aquaculture, achieved through the expansion of *Porphyra yezoensis* production, resulted in annual carbon sequestration of 54,000 tons. Concentration dropped to 22 A coupled coordination model was used to calculate the coordination coefficient between carbon tax and subsidies. The combined scenario coordination coefficient was 0.75, an improvement of 0.3 compared to the single scenario. The optimal policy combination was identified as a carbon tax of 0.3 yuan / ton plus a blue carbon subsidy of 0.6 yuan / ton. Figure 7 As shown, changes in social welfare under carbon tax policies exhibit a pattern of short-term decline followed by long-term recovery. For example... Figure 8 As shown, social welfare changes under subsidy policies exhibit a continuous upward trend. Changes in social welfare under technological innovation demonstrate a significant incremental effect of welfare brought about by technological progress. Based on the optimal scenario, policy recommendations are output: carbon tax to be increased in a 5% annual gradient, with subsidies prioritized for seawater desalination membrane technology. An early warning is triggered when the probability of excessively high carbon taxes leading to the relocation of the petrochemical industry exceeds 15%, and a phased implementation plan is provided. The risk of a 6% to 8% decrease in treatment capacity due to the reduction of seawater desalination subsidies is mitigated by a phased subsidy reduction. A web-based interactive interface is provided to allow users to adjust parameters such as pollution negative utility weights and the emission reduction efficiency of environmental protection technologies, updating simulation results in real time. Different data access and parameter adjustment permissions are assigned to local government and enterprise users to ensure data security.

[0076] This invention, based on the information bottleneck principle, implements an intelligent mapping between data features and computing resources in a heterogeneous computing scheduler. By minimizing redundant information and maximizing task-related information, it ensures both accurate and efficient scheduling decisions. It fully leverages the advantages of CPUs in handling control logic and sequence dependency computations, as well as the advantages of GPUs in large-scale parallel computing, avoiding resource waste and performance degradation caused by forcibly allocating unsuitable parallelization tasks to GPUs. The introduction of variational approximation techniques and the Lagrange multiplier method solves the problem of directly calculating mutual information, achieving a dynamic balance between information compression rate and task completion fidelity, enabling the heterogeneous computing scheduler to maintain stable performance under different computing loads. Low-rank matrix factorization significantly reduces the number of model parameters and storage space requirements by decomposing high-dimensional weight matrices into products of low-rank submatrices. Simultaneously, kernel norm regularization ensures that the decomposed model maintains the approximate accuracy of the original model, providing technical support for real-time updates of large-scale DSGE models. Multiple historical data reviews have shown that this invention improves the update efficiency of DSGE models by 27% to 32%, greatly enhancing the real-time update capability of DSGE models.

[0077] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing massive amounts of data for DSGE analysis and decision-making regarding the relationship between marine industry and ecology, characterized in that, Multi-source heterogeneous data streams are collected and spatiotemporally aligned preprocessing is performed. Threshold cointegration tests are performed on the preprocessed data to identify nonlinear threshold points and establish a state transition matrix. The data is then input into a heterogeneous computing scheduler for resource allocation. A multi-objective decomposition framework is used to decompose the parameter optimization problem into single-objective sub-problems and perform global and local searches. A sliding window mechanism is established to retain data from the most recent time period, and a recursive least squares algorithm is used for incremental updates. Low-rank approximate decomposition is performed on the high-dimensional weight matrix, and the optimized low-rank submatrix is ​​used as the weight parameters of the DSGE model for decision analysis. The heterogeneous computing scheduler enables adaptive allocation of computing resources driven by data features. Combined with low-rank matrix decomposition, the model parameter size is reduced, and a recursive least squares algorithm is used to implement incremental updates of the DSGE model parameters.

2. The method according to claim 1, characterized in that, The steps of collecting multi-source heterogeneous data streams and performing spatiotemporal alignment preprocessing specifically involve collecting raw data streams from industrial monitoring systems, ecological environment monitoring stations, and policy databases; performing missing value imputation on the raw data streams; identifying outlier data points exceeding three standard deviations and replacing them with moving averages; unifying the time scale of data collected at different frequencies; extracting the latitude and longitude coordinates of each monitoring station to achieve matching of data with different spatial resolutions; and performing minimum-maximum normalization transformation on all numerical data.

3. The method according to claim 2, characterized in that, The raw data stream includes industrial economic indicators, ecological and environmental parameters, policy text information, and technical specification documents. For unstructured policy texts, a large language model is used to perform semantic understanding and information extraction. For semi-structured technical documents, a large language model is used to parse the technical indicator tables and specification paragraphs.

4. The method according to claim 3, characterized in that, The process of using a large language model to perform semantic understanding and information extraction on unstructured policy texts involves guiding the large language model to identify numerical parameters, entity information and their relationships in policy documents through structured prompt word templates, using mind chain prompting technology to guide the large language model to analyze the logical relationship between policy measures and their targets, and using a few-shot learning method to enable the large language model to identify policy intensity descriptors and convert them into numerical weights.

5. The method according to claim 4, characterized in that, The step of performing a threshold cointegration test on the preprocessed data to identify nonlinear threshold points specifically involves selecting pollution concentration as the threshold variable, setting a candidate set of threshold values ​​to cover the range of sample quantiles ∈ [0.15, 0.85], dividing the sample into two subsamples for each threshold value in the candidate set of threshold values, estimating the cointegration vector and calculating the sum of squared residuals in each subsample, and identifying the threshold value that minimizes the sum of squared residuals as the optimal threshold point.

6. The method according to claim 5, characterized in that, The establishment of the state transition matrix specifically involves dividing the data into a low-concentration zone with pollution concentrations below the first threshold, a medium-concentration zone between the first and second thresholds, and a high-concentration zone with pollution concentrations above the second threshold. Within each zone, the long-term cointegration relationship between industrial output and ecological indicators is estimated, and a state transition matrix is ​​constructed to characterize the switching probability of the system between different zones.

7. The method according to claim 6, characterized in that, The process of inputting data into a heterogeneous computing scheduler for resource allocation involves constructing a computing demand assessment function to quantify the parallelism and computational complexity characteristics of the data. When the parallelism characteristic value is greater than 0.7 and the computational complexity characteristic value is greater than 0.6, the data is determined to be highly parallel and allocated to the GPU; otherwise, it is determined to be sequentially dependent data and allocated to the CPU.

8. The method according to claim 7, characterized in that, The first hidden layer of the heterogeneous computing scheduler adopts a representation learning framework based on the information bottleneck principle. It minimizes the mutual information between the input features and the intermediate representation to constrain redundant information, while maximizing the mutual information between the intermediate representation and the computing resource allocation decision to preserve key information. It introduces variational approximation techniques to estimate mutual information terms and uses the Lagrange multiplier method to dynamically balance information compression rate and task completion fidelity.

9. The method according to claim 8, characterized in that, The establishment of the training dataset for the heterogeneous computing scheduler specifically involves collecting data features and actual execution performance data of historical computing tasks, calculating the CPU execution efficiency and GPU execution efficiency for each historical computing task, and marking it as a GPU task when the GPU execution efficiency exceeds the CPU execution efficiency by more than 1.5 times; otherwise, it is marked as a CPU task.

10. The method according to claim 9, characterized in that, The multi-objective decomposition framework decomposes the parameter optimization problem into single-objective sub-problems. Specifically, it defines a set of optimization objective functions including economic efficiency objective function, ecological benefit objective function, and policy synergy objective function, uses a uniform design method to generate weight vectors, decomposes the multi-objective optimization problem into single-objective sub-problems, and uses Chebyshev aggregation function to measure the approximation of individual solutions to reference points.

Images