Carbon market policy impact causal deduction platform based on double difference method

Abstract

The invention belongs to the field of data processing, and discloses a carbon market policy impact causal deduction platform based on a double difference method. Comprising the following steps: collecting multi-dimensional feature data of pilot regions and non-pilot regions of the carbon market, carrying out standardization processing on the data, designing a multiple interpolation algorithm based on machine learning, and filling missing values in combination with auxiliary variables of regional economic levels and energy structures; and if the deviation between the filling value and the actual observation value exceeds 10%, marking the filling value as low-credibility data and reducing the weight of the filling value in subsequent analysis. Generating a sample weight matrix according to the characteristic difference between the pilot region and the non-pilot region; through weight distribution, the non-pilot area is more matched with the pilot area in the feature space. A sliding window mechanism is adopted, if the trend difference exceeds a preset threshold value, weight distribution is automatically adjusted, a parallel trend test report is output, and carbon market policy impact causal deduction based on the double difference method is achieved.

Description

Technical Field

[0001] In the field of data processing, and more specifically, this invention relates to a causal extrapolation platform for carbon market policy impacts based on the difference-in-differences method. Background Technology

[0002] Traditional difference-in-differences (DID) methods rely on the assumption of parallel trends, but significant differences exist between pilot and non-pilot regions in the actual carbon market before policy implementation. For example, research shows that the carbon emission reduction effect in the economically developed eastern region (11.5%) far exceeds that in the western region (4.2%), and traditional DID models struggle to capture this regional heterogeneity. Furthermore, carbon emission data suffers from non-random missing data, particularly in the western region where the missing data rate reaches as high as 30% in some years, leading to significant estimation bias.

[0003] The carbon market is influenced by multiple factors, including energy price fluctuations (such as a 2.5% increase in global natural gas consumption in 2024), abnormal climate events (such as frequent extreme heat events), and technological innovations (such as new hydrogen production technologies). However, existing models lack the ability to quantitatively analyze the dynamic coupling mechanism of policy, market, and technology.

[0004] The radiation effect of carbon trading pilot areas on surrounding areas (such as joint emission reduction caused by technology diffusion) needs to rely on spatial econometric models. However, existing technologies have not achieved deep coupling between DID and SDM, and cannot simultaneously identify the net policy effect and spatial spillover effect.

[0005] To address the aforementioned issues, this patent proposes a carbon market policy inference platform that integrates a multimodal causal inference engine. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and achieve the aforementioned objectives, the present invention provides the following technical solution: a causal extrapolation platform for the impact of carbon market policies based on the difference-in-differences method, comprising: a data preprocessing and feature extraction submodule: collecting multidimensional feature data from carbon market pilot areas and non-pilot areas, standardizing the data, and extracting the main feature vectors through principal component analysis to reduce the data dimensionality.

[0007] Non-random data missing data imputation submodule: Design a machine learning-based multiple imputation algorithm, combined with auxiliary variables such as regional economic level and energy structure, to fill in missing values; perform consistency checks on the imputed data, and if the deviation between the imputed value and the actual observed value exceeds 10%, it is marked as low confidence data and its weight in subsequent analysis is reduced.

[0008] The adaptive weight allocation submodule constructs a weight generation algorithm based on gradient boosting trees, generates a sample weight matrix according to the feature differences between pilot and non-pilot areas, and makes the non-pilot areas more closely match the pilot areas in the feature space through weight allocation.

[0009] The dynamic parallel trend test submodule uses a sliding window mechanism to calculate the trend difference in carbon emission intensity between pilot areas and weighted non-pilot areas periodically within the time window before policy implementation. If the trend difference exceeds a preset threshold, the weight allocation is automatically adjusted until the trend difference converges. The parallel trend test report is output, including the trend difference curve, confidence interval, and significance test results.

[0010] Policy effect estimation submodule: Based on the adjusted weight matrix, a difference-in-differences model is constructed to estimate the net effect of policy implementation on carbon emission intensity in pilot areas.

[0011] Preferably, the double difference method includes: A dynamic spatial weight matrix is ​​constructed based on three dimensions: inter-provincial technology flow, energy transmission topology, and geographical proximity. The spatial Durbin difference-in-difference joint dynamic spatial weight matrix is ​​used to embed the spatial weight matrix into the DID model to estimate the direct effects of the policy on the pilot area and the indirect effects on the surrounding areas. The total emission reduction effect is decomposed into direct and indirect effects, and an effect decomposition report is generated.

[0012] Preferably, the spatial weight matrix includes: Real-time acquisition of multi-dimensional high-frequency data streams, including carbon price fluctuations, climate anomalies, and energy prices; construction of a long short-term memory neural network to predict future carbon price sequences, and embedding them as time-varying covariates into the DID model; The DID model is extended by adding an interaction term for carbon price fluctuations, and a dynamic risk report is output, including the changing trends and confidence intervals of policy effects under different carbon price scenarios.

[0013] Preferably, the DID model includes: It accesses four types of data sources, including macro data, micro data, real-time data, and spatial data; it designs a blockchain anchor verification channel, triggers data consistency verification through smart contracts, and automatically initiates manual review if the deviation between provincial data and satellite inversion values ​​exceeds 15%. A multimodal data alignment algorithm is used to standardize data with different spatiotemporal resolutions into a unified format and store them in a distributed data lake.

[0014] Preferably, the distributed data lake includes: The policy scenario library contains 12 policy combinations, including carbon tax, quota auctions, and green certificate trading; Based on the DID model, we simulate the carbon emission path when the policy is not implemented, and calculate the net effect and confidence interval of the policy. Generate a three-dimensional assessment report, including the net effect of the policy, simulation of the cost-benefit of corporate emission reduction, and the path of regional industrial structure transformation.

[0015] Preferably, the policy scenario library includes: Construct a multi-stakeholder model involving power companies, government regulatory agencies, and carbon traders, and train a Q-learning strategy based on historical data; The emission reduction effect output by the DID model is used as a reward function and input into the MAS model to optimize the subject's decision-making. Based on the simulation results, policy optimization suggestions are automatically generated.

[0016] Preferably, the multi-agent model includes: The architecture uses a layered parallel approach. The first layer uses a Spark cluster to compute K_O provincial DID coefficients in parallel. The second layer uses GPUs to accelerate matrix operations. The third layer uses the D3.js engine to visualize the difference results in real time. It supports dynamic resource allocation and automatically expands computing nodes when the data scale exceeds a preset threshold.

[0017] Preferably, the hierarchical parallel architecture includes: The thermal layer uses color gradients to show the rate of change in carbon intensity for each province; The network layer uses node size to represent the direct effect value and edge width to represent the spillover effect intensity, supporting the retrospective spatiotemporal evolution from 2011 to 2025; Users can view detailed effect analysis reports for specific time points or regions by dragging the timeline or clicking on nodes.

[0018] The technical effects and advantages of this invention's carbon market policy impact causal extrapolation platform based on the difference-in-differences method are as follows: It pioneered an interdisciplinary framework combining econometrics, spatial geography, and deep learning to overcome the limitations of traditional policy evaluation. A three-in-one verification system integrating blockchain, satellite remote sensing, and ground monitoring has been established to reduce the risk of data falsification by 80%. It supports counterfactual simulation and comparison of multiple policy combinations, providing intelligent support for decision-making; Real-time processing of hundreds of millions of data points is achieved through a distributed architecture, resulting in a significant improvement in computing efficiency. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of a carbon market policy impact causal inference platform based on the difference-in-differences method according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1

[0022] Please see Figure 1 As shown in this embodiment, a carbon market policy impact causal extrapolation platform based on the difference-in-differences method includes: Collect multidimensional characteristic data of carbon market pilot areas and non-pilot areas, including but not limited to carbon emission intensity, energy consumption structure, industrial added value ratio, population density, etc. The data is standardized and principal component analysis (PCA) is used to extract the main eigenvectors, reducing the data dimensionality and ensuring the stability of subsequent calculations.

[0023] Example: Taking pilot regions (such as Beijing, Shanghai, Guangdong, and 8 other provinces and municipalities) and non-pilot regions (such as Shanxi and Gansu) of China's carbon market in 2023 as examples, data on carbon emission intensity (tons of CO2 / 10,000 yuan of GDP), energy consumption structure (coal share, natural gas share, etc.), and industrial added value share were collected from 2018 to 2023. After PCA analysis, the first three principal components were extracted, explaining 85% of the variance of the original data, and used for subsequent analysis.

[0024] To address the issue of non-random missing values ​​in carbon emission data, a machine learning-based multiple imputation algorithm (MICE) is designed to fill in missing values ​​by incorporating auxiliary variables such as regional economic level and energy structure. The imputed data is subjected to a consistency test. If the deviation between the imputed value and the actual observed value exceeds 10%, it is marked as low confidence data and its weight in subsequent analysis is reduced.

[0025] Example: In the case of missing carbon emission data for Gansu Province in 2019, the MICE algorithm was used, combined with variables such as Gansu's GDP, coal consumption, and wind power installed capacity, to fill in the missing values. The data after filling had an 8% deviation from the satellite inversion value (NASA OCO-2), meeting the consistency requirement, and the weight remained at 1.0; however, if the deviation were 12%, the weight would be reduced to 0.7.

[0026] A weight generation algorithm based on gradient boosting is constructed to generate a sample weight matrix according to the feature differences between pilot and non-pilot areas; By assigning weights, non-pilot areas are better matched with pilot areas in the feature space, satisfying the parallel trend assumption.

[0027] Example: Taking Shanxi Province (a non-pilot region) and Guangdong Province (a pilot region) as examples, the characteristic differences between the two provinces from 2018 to 2022 were calculated. It was found that Shanxi Province deviated from the average characteristics of the pilot regions due to its high dependence on coal (coal accounting for 65%). Through weight calculation, the weight of Shanxi Province was reduced to 0.3, while the weight of Jiangsu Province, which is closer to the characteristics of the pilot regions, was increased to 1.7.

[0028] Within the time window before policy implementation (e.g., the first 5 years), a sliding window mechanism is used to calculate the difference in carbon emission intensity trends between pilot areas and weighted non-pilot areas period by period; If the trend difference exceeds a preset threshold (e.g., ±0.5%), the weight allocation will be automatically adjusted until the trend difference converges. Output a parallel trend test report, including the trend difference curve, confidence interval, and significance test results.

[0029] Example: Before the policy was implemented from 2020 to 2022, the average annual growth rate of carbon emission intensity in the pilot areas (8 provinces and municipalities) and the weighted non-pilot areas was calculated. It was found that the growth rate in Shanxi Province was too high (1.2%) due to the expansion of the coal industry, exceeding the threshold ±0.5%. The system automatically lowered the weight of Shanxi to 0.2 and increased the weight of Jiangsu to 1.8, so that the trend difference converged to 0.4%, satisfying the parallel trend assumption.

[0030] Based on the adjusted weight matrix, a difference-in-differences (DID) model is constructed to estimate the net effect of policy implementation on carbon emission intensity in pilot areas; Example: Taking the 2023 carbon market expansion policy as an example, the estimation results show that the carbon emission intensity in pilot areas decreased by an average of 11.5%, while there was no significant change in non-pilot areas. The report shows a confidence interval of [-0.128, -0.102], indicating that the policy effect is robust.

[0031] Spatial Weight Matrix Generator: Based on three dimensions—inter-provincial technology flow (e.g., patent cooperation networks), energy transmission topology (e.g., power grid architecture), and geographical proximity—a dynamic spatial weight matrix is ​​constructed; the weight calculation formula is as follows: Spatial Durbin Difference-Difference (DID-SDM) Joint Estimator: Embeds the spatial weight matrix into the DID model to estimate the direct effects of policies on pilot areas and the indirect effects on surrounding areas; Effect decomposer: Decomposes the total emission reduction effect into direct effects (the emission reduction rate of the pilot area itself) and indirect effects (the driving rate on the surrounding 200km area), and outputs an effect decomposition report.

[0032] Example: Taking the Yangtze River Delta region as an example, when constructing the spatial weight matrix, the weight for technology flow between Shanghai and Jiangsu (based on patent cooperation) is 0.6, the weight for power grid connection is 0.5, the weight for geographical distance is 0.4, and the total weight is 0.53. DID-SDM estimation results show that Shanghai's carbon trading pilot policy reduced local carbon emission intensity by 0.167%, while the spillover effects on Jiangsu and Zhejiang were 0.053% and 0.048%, respectively, mainly attributed to the diffusion of green technologies and power transmission.

[0033] Real-time acquisition of multi-dimensional high-frequency data streams, including carbon price fluctuations (such as EU ETS futures prices), climate anomalies (such as NASA's global temperature anomaly index), and energy prices (such as Brent crude oil futures); A long short-term memory neural network (LSTM) is constructed to predict the future carbon price sequence (P_t), and it is embedded as a time-varying covariate into the DID model; The DID model is extended by adding an interaction term for carbon price fluctuations, and a dynamic risk report is output, including the changing trends and confidence intervals of policy effects under different carbon price scenarios.

[0034] Example: Taking the EU ETS carbon price fluctuation in 2024 as an example, LSTM forecast results show that if the carbon price rises from €50 / ton to €60 / ton, the policy effect elasticity coefficient (\beta_3) will increase from 0.65 to 0.78, indicating that the policy's emission reduction effect will be enhanced. The report shows that under the high carbon price scenario, emission reductions in pilot areas can be increased to 12.8%, with a confidence interval of [11.5%, 14.1%].

[0035] It accesses four types of data sources, including macro data (such as carbon emission databases for 30 out of 34 provinces in China from 2003 to 2023), micro data (such as enterprise-level carbon accounting systems), real-time data (such as carbon quota auction data from the Shanghai Environment Exchange), and spatial data (such as CO2 concentration grids from NASA's carbon satellite). The design incorporates a blockchain-based verification channel, which triggers data consistency verification via smart contracts. If the deviation between provincial data and satellite inversion values ​​exceeds 15%, manual review will be automatically initiated. A multimodal data alignment algorithm is used to standardize data with different spatiotemporal resolutions into a unified format and store them in a distributed data lake.

[0036] Example: Taking Hebei Province's Q4 2023 data as an example, the provincial carbon emission inventory showed emissions of 120 million tons, while the NASA OCO-2 satellite inversion value was 138 million tons, a discrepancy of 15%. The system triggered a smart contract, initiated manual review, and ultimately confirmed that the satellite data was more accurate, updating the provincial inventory. The merged data is stored in a data lake in a 1km×1km raster format.

[0037] The policy scenario library contains 12 policy combinations, including carbon tax, quota auctions, and green certificate trading; Based on the DID model of claim 1, the carbon emission path is simulated when the policy is not implemented, and the net effect and confidence interval of the policy are calculated. A three-dimensional assessment report is generated, including the net effect of the policy, simulation of the cost-benefit of corporate emission reduction (such as the prediction of quota gaps for coal-fired power companies), and the path of regional industrial structure transformation (such as the empirical evidence of promoting industrial agglomeration by referring to the Chongqing land ticket policy).

[0038] Example: Taking the inclusion of the steel industry in the carbon market as an example, counterfactual analysis shows that without the policy, the carbon emission intensity in the eastern region would increase by 7.8%. The decision support report shows that after the policy was implemented, emission reductions in the eastern region increased by 14.2%, but costs for steel companies in the west increased by 23%. The system recommends the optimized solution of "implementing a 3-year transitional quota subsidy in the west".

[0039] A multi-stakeholder model involving power companies, government regulatory agencies, and carbon traders is constructed, and a Q-learning strategy is trained based on historical data. The emission reduction effect output by the DID model in claim 1 is used as a reward function and input into the MAS model to optimize the decision-making of the stakeholders. Based on the simulation results, policy optimization suggestions are automatically generated, such as suggesting that the quota auction floor price be increased when the emission reduction exceeds the expected 10%.

[0040] Example: Taking power companies as an example, MAS simulation shows that if the government raises the reserve price for quota auctions by 10%, power companies will proactively deploy hydrogen energy technology, increasing emissions reduction by 12%. The system recommends raising the reserve price from 50 yuan / ton to 55 yuan / ton in Q1 2025 to achieve the emissions reduction target.

[0041] Layered parallel architecture: The first layer uses a Spark cluster to compute K_O provincial DID coefficients in parallel; the second layer uses GPU to accelerate matrix operations (such as inverting the spatial weight matrix); the third layer uses the D3.js engine to realize real-time visualization of the difference results. It supports dynamic resource allocation and automatically expands computing nodes when the data size exceeds a preset threshold (such as 100GB).

[0042] Example: Taking panel data from 30 provinces from 2003 to 2023 as an example, traditional single-machine calculation of DID coefficients takes 2 hours, while this system, through parallel calculation using a Spark cluster, only takes 8 minutes. The spatial weight matrix inversion operation is accelerated by GPU, reducing the time from 1 hour to 10 minutes.

[0043] In the visual interactive interface, the heat map uses color gradients to display the rate of change in carbon intensity for each province; for example, dark red indicates a decrease of more than 10%, and light pink indicates a decrease of less than 2%. The network layer uses node size to represent the direct effect value and edge width to represent the spillover effect intensity, supporting the retrospective spatiotemporal evolution from 2011 to 2025; Users can view detailed effect analysis reports for specific time points or regions by dragging the timeline or clicking on nodes.

[0044] Example: Taking 2023 as an example, the heat map shows that Shandong reduced emissions by 12.3% (dark red) and Gansu reduced emissions by 1.8% (light pink). The network diagram shows that the Shanghai node radiates to Jiangsu, Zhejiang, and Anhui (side width is 5), while the Shanxi node is isolated and unconnected. Users can click on the Shanghai node to view the spillover effect decomposition report.

[0045] The following section uses the assessment of China's carbon market expansion policy in 2025 (including the steel industry in the carbon market) as an example to describe in detail the specific implementation of this patent, covering the system architecture, workflow and key steps.

[0046] The platform adopts a four-layer logical architecture, which includes, from bottom to top: Data layer: blockchain evidence repository, multi-source data lake (macro / micro / real-time / spatial data); Computation engine layer: Spark cluster, GPU-accelerated matrix operation library; The core layer of the model consists of: DID evaluation module, LSTM risk engine, and SDM spatial analysis module. Application layer: Policy dashboard (visual interface), enterprise decision support terminal.

[0047] Workflow Step 1: Data Fusion and Verification Data input: Collect industrial carbon emission data from pilot provinces (Beijing, Shanghai, Guangdong and other 8 provinces and cities) and non-pilot provinces (Shanxi, Gansu and other provinces) from 2020 to 2025, and simultaneously access global variables such as EU carbon price, photovoltaic power generation and NASA OCO-2 satellite CO2 concentration during the same period.

[0048] Data verification: The blockchain verification module detected an anomaly in Hebei Province's Q4 2023 data (18% deviation from the satellite inversion value), triggering a smart contract to initiate manual review. Ultimately, the satellite data was confirmed to be more accurate, and the provincial list was updated. The merged data is stored in a data lake in a 1km×1km raster format.

[0049] Step 2: Dynamic Parallel Trend Test Feature extraction: The main feature vectors of pilot and non-pilot areas are extracted by PCA, such as carbon emission intensity, coal share, and industrial added value share.

[0050] Data imputation: For missing data in Gansu Province in 2023, the MICE algorithm was used for imputation, with an imputation value deviation of 8% and a weight of 1.0.

[0051] Weighting: The characteristic differences between Shanxi Province and Guangdong Province were calculated, with Shanxi's weight adjusted to 0.3 and Jiangsu's weight adjusted to 1.7.

[0052] Example 2

[0053] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the carbon market policy impact causal inference platform based on the dual difference method described above.

[0054] Since the electronic device described in this embodiment is the electronic device used to implement the carbon market policy impact causal inference platform based on the difference-in-differences method described in this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the difference-in-differences method described in this application embodiment. Therefore, how the electronic device implements the method in this application embodiment will not be described in detail here. Any electronic device used by those skilled in the art to implement the carbon market policy impact causal inference platform based on the difference-in-differences method described in this application embodiment falls within the scope of protection of this application.

[0055] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0056] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A causal extrapolation platform for carbon market policy shocks based on the difference-in-differences method, characterized in that, include: The data preprocessing and feature extraction submodule collects multidimensional feature data from carbon market pilot areas and non-pilot areas, standardizes the data, and extracts the main feature vectors through principal component analysis to reduce the data dimensionality. Non-random data missing data imputation submodule: Design a machine learning-based multiple imputation algorithm, combined with auxiliary variables such as regional economic level and energy structure, to fill in missing values; perform consistency checks on the imputed data, and if the deviation between the imputed value and the actual observed value exceeds 10%, it is marked as low confidence data and its weight in subsequent analysis is reduced; The adaptive weight allocation submodule constructs a weight generation algorithm based on gradient boosting trees, generates a sample weight matrix according to the feature differences between pilot and non-pilot areas, and makes the non-pilot areas more closely match the pilot areas in the feature space through weight allocation. Dynamic parallel trend verification submodule: Within the time window before policy implementation, a sliding window mechanism is used to calculate the difference in carbon emission intensity trends between pilot areas and weighted non-pilot areas period by period; If the trend difference exceeds a preset threshold, the weight allocation will be automatically adjusted until the trend difference converges. Output a parallel trend test report, including the trend difference curve, confidence interval, and significance test results; Policy effect estimation submodule: Based on the adjusted weight matrix, a difference-in-differences model is constructed to estimate the net effect of policy implementation on carbon emission intensity in pilot areas.

2. The carbon market policy impact causal extrapolation platform based on the difference-in-differences method according to claim 1, characterized in that, The difference-in-differences method includes: A dynamic spatial weight matrix is ​​constructed based on three dimensions: inter-provincial technology flow, energy transmission topology, and geographical proximity. The spatial Durbin difference-in-difference joint dynamic spatial weight matrix is ​​used to embed the spatial weight matrix into the DID model to estimate the direct effects of the policy on the pilot area and the indirect effects on the surrounding areas. The total emission reduction effect is decomposed into direct and indirect effects, and an effect decomposition report is generated.

3. The carbon market policy impact causal extrapolation platform based on the difference-in-differences method according to claim 2, characterized in that, The spatial weight matrix includes: Real-time acquisition of multi-dimensional high-frequency data streams, including carbon price fluctuations, climate anomalies, and energy prices; construction of a long short-term memory neural network to predict future carbon price sequences, and embedding them as time-varying covariates into the DID model; The DID model is extended by adding an interaction term for carbon price fluctuations, and a dynamic risk report is output, including the changing trends and confidence intervals of policy effects under different carbon price scenarios.

4. The carbon market policy impact causal extrapolation platform based on the difference-in-differences method according to claim 3, characterized in that, The DID model includes: It accesses four types of data sources, including macro data, micro data, real-time data, and spatial data; it designs a blockchain anchor verification channel, triggers data consistency verification through smart contracts, and automatically initiates manual review if the deviation between provincial data and satellite inversion values ​​exceeds 15%. A multimodal data alignment algorithm is used to standardize data with different spatiotemporal resolutions into a unified format and store them in a distributed data lake.

5. The carbon market policy impact causal extrapolation platform based on the difference-in-differences method according to claim 4, characterized in that, The distributed data lake includes: The policy scenario library contains 12 policy combinations, including carbon tax, quota auctions, and green certificate trading; Based on the DID model, we simulate the carbon emission path when the policy is not implemented, and calculate the net effect and confidence interval of the policy. Generate a three-dimensional assessment report, including the net effect of the policy, simulation of the cost-benefit of corporate emission reduction, and the path of regional industrial structure transformation.

6. The carbon market policy impact causal inference platform based on the difference-in-differences method according to claim 5, characterized in that, The policy scenario library includes: Construct a multi-stakeholder model involving power companies, government regulatory agencies, and carbon traders, and train a Q-learning strategy based on historical data; The emission reduction effect output by the DID model is used as a reward function and input into the MAS model to optimize the subject's decision-making. Based on the simulation results, policy optimization suggestions are automatically generated.

7. The carbon market policy impact causal extrapolation platform based on the difference-in-differences method according to claim 6, characterized in that, The multi-agent model includes: The architecture uses a layered parallel approach. The first layer uses a Spark cluster to compute K_O provincial DID coefficients in parallel. The second layer uses GPUs to accelerate matrix operations. The third layer uses the D3.js engine to visualize the difference results in real time. It supports dynamic resource allocation and automatically expands computing nodes when the data scale exceeds a preset threshold.

8. The carbon market policy impact causal inference platform based on the difference-in-differences method according to claim 7, characterized in that, The hierarchical parallel architecture includes: The thermal layer uses color gradients to show the rate of change in carbon intensity for each province; The network layer uses node size to represent the direct effect value and edge width to represent the spillover effect intensity, supporting the retrospective spatiotemporal evolution from 2011 to 2025; Users can view detailed effect analysis reports for specific time points or regions by dragging the timeline or clicking on nodes.

Images