Intelligent prediction and transaction management system for carbon quota and carbon assets

Abstract

The invention discloses an intelligent prediction and transaction management system for carbon quotas and carbon assets, and relates to the technical field of carbon transaction systems. The intelligent prediction and transaction management system for the carbon quota and the carbon assets comprises a carbon transaction mode portrait construction module, a dynamic module matching and initializing module, a multi-source data fusion and dynamic updating module, a hybrid prediction model operation module and a dynamic transaction strategy generation and optimization module. Through cooperative work of the five modules, enterprise carbon transaction whole-process intelligent management is realized. The portrait construction module accurately classifies enterprise modes and provides a basis for dynamic matching; a module library and a parameter initialization mechanism ensure strategy customization flexibility and deployment efficiency; the multi-source data fusion module breaks an information island and improves the input quality of a prediction model; outputting a multi-time scale prediction result and a risk index by the hybrid prediction model in combination with a statistics and machine learning method; and the dynamic strategy module automatically generates and optimizes a transaction strategy according to the prediction result and the enterprise risk preference.

Description

Technical Field

[0001] This invention relates to the field of carbon trading system technology, specifically to an intelligent prediction and trading management system for carbon quotas and carbon assets. Background Technology

[0002] Carbon allowances are mandatory emission reduction targets set by governments or international organizations to control greenhouse gas emissions. Essentially, they are environmental rights certificates representing the maximum amount of greenhouse gases a company can emit within a certain period. Under a carbon emissions trading system, governments convert emission rights into tradable commodities through cap-and-trade mechanisms. Companies must pay an equivalent amount of allowances within a specified time to fulfill their obligations; otherwise, they will face penalties. Carbon allowances are allocated through free allocation (such as the historical total allocation method, historical intensity method, and baseline method) and paid auctions. A hybrid model combines both methods and dynamically adjusts the ratio.

[0003] Carbon assets are all assets centered on carbon emission rights, commodifying greenhouse gases such as carbon dioxide and calculating them in carbon dioxide equivalents. They are a new type of green asset. Types include carbon allowances, CCERs (Certified Emission Reductions), and carbon sinks, and possess the characteristics of appreciation potential, tradability, and reserve value. The formation of carbon assets stems from global climate governance driven by international agreements such as the Kyoto Protocol. When carbon emission rights became a scarce resource, they were set as a valuable commodity, forming the carbon asset market.

[0004] The intelligent prediction and trading management system for carbon quotas and carbon assets is a professional tool for enterprises to achieve full lifecycle management of carbon emissions through digital means.

[0005] Currently, due to significant differences in carbon trading models among different enterprises, existing methods are insufficient for personalized prediction and management of data from different carbon trading models, resulting in insufficient operational efficiency and decision-making accuracy for enterprises.

[0006] Publication No. CN119740715A discloses a carbon trading data prediction method and system based on big data, relating to the field of carbon trading prediction technology. The method includes: acquiring the geographical location characteristics and historical carbon trading matrices of each virtual power plant; performing cluster analysis on the virtual power plants based on the geographical location characteristics and historical carbon trading matrices, dividing all virtual power plants into several federated learning sub-nodes; acquiring historical climate characteristic data, constructing a global prediction model and local prediction models for each federated learning sub-node based on the historical climate characteristic data and historical carbon trading matrices; inputting the real-time acquired climate characteristic data and carbon trading matrix into the local prediction models of the corresponding federated learning sub-nodes to obtain the carbon trading data prediction results. This application can provide a data foundation for the operation and decision-making of virtual power plants, enabling effective resource allocation and environmental management.

[0007] The aforementioned technology extracts the geographical location features and standardized numerical features of the historical carbon trading matrix of each virtual power plant, and then uses the K-means algorithm to perform cluster analysis on the virtual power plants. This can effectively cluster widely distributed and complex virtual power plants, and effectively capture the commonalities of virtual power plants with similar geographical location features and carbon trading models, thereby solving the problem of managing data for different carbon trading models. However, in actual operation, the core issue of carbon trading prediction for enterprises with different carbon trading models is that different types of enterprises have different forms of prediction models and systems in detail. Using a unified large model and only analyzing the data is not enough to make accurate predictions and management for enterprises with different carbon trading models. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent prediction and trading management system for carbon quotas and carbon assets, thus solving the problems of existing technologies.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent prediction and trading management system for carbon quotas and carbon assets, comprising: Carbon trading model profiling module: By collecting historical carbon emission data from enterprise ERP systems, carbon quota trading records, and electricity trading platform data, the module uses the K-means clustering algorithm to classify enterprises into three models: high emission-low activity model, low emission-high activity model, and cross-market linkage model; and by calculating carbon trading activity and electricity-carbon coupling coefficient indicators, the module creates a customized carbon trading characteristic profile for each enterprise. Dynamic module matching and initialization module: Based on the characteristics of the enterprise profile, select modules that match the requirements from industry-specific modules and combine them with pre-set basic general module configuration modules; generate a unique module combination list for the enterprise through the feature layer, model layer, and strategy layer, and initialize parameters; Multi-source data fusion and dynamic update module: Real-time access and preprocessing of internal and external enterprise data; dynamic calculation of short-term and long-term characteristics to construct an electricity-carbon linkage index; Hybrid forecasting model computation module: Combining LSTM network, Prophet model and XGBoost model to make short-term and long-term forecasts of carbon price and quota demand; through model fusion and uncertainty quantification, output forecast range and risk indicators; Dynamic trading strategy generation and optimization module: Automatically generates trading strategies based on prediction results and corporate risk preferences, and simulates their execution in a historical data backtesting environment; continuously optimizes trading strategies through real-time strategy adjustments and reinforcement learning.

[0010] Preferably, the specific tasks of the carbon trading model profiling module include: Data Acquisition: Obtain historical carbon emission data, carbon quota trading records, and energy consumption structure from enterprise ERP systems; connect to power trading platforms to capture transaction data of enterprises participating in the electricity spot market and ancillary services market; collect industry benchmark data; Feature extraction: A carbon trading activity index is obtained by calculating the ratio of the number of transactions to the total quota holdings over the past 12 months; the volatility σ of the electricity market is analyzed. 电价 The Pearson correlation coefficient between the carbon quota demand change rate Δquota demand and the carbon trading activity index is used as the carbon trading coupling coefficient. The carbon trading coupling coefficient is calculated. The absolute contribution of different variables to carbon price prediction is calculated by the SHAP value. The weighted average of the top 3 variables is taken as the weight of the core variable for carbon price prediction. The carbon trading activity index, the carbon trading coupling coefficient and the weight of the core variable for carbon price prediction are standardized by Z-score to eliminate the influence of the unit of measurement. Pattern Classification: Using the K-means clustering algorithm, enterprises are divided into three typical patterns: Mode A represents high emissions and low activity, including traditional high-energy-consuming enterprises such as thermal power plants, with low transaction frequency but large single transaction volume; Mode B represents low emissions and high activity, including new energy companies or data centers, which optimize carbon costs through frequent trading of CCERs / green certificates; Model C represents a cross-market linkage model, which includes integrated energy service providers that participate in both the electricity spot market and the carbon market, and need to coordinate fluctuations in electricity prices and carbon prices.

[0011] Preferably, the enterprise segmentation method uses the K-means clustering algorithm for classification, including: 1) K-means clustering implementation: verified by the elbow rule, calculated the sum of squared errors (SSE) for different K values, and selected the inflection point K=3; Initialization and iteration: Randomly select 3 enterprise samples as initial cluster centers, calculate the Euclidean distance from each enterprise to the cluster center, assign it to the nearest center, update the cluster center to the mean of each category of samples, and repeat the iteration until convergence; 2) Pattern tag definition: Mode A: Carbon trading activity < 30th percentile, and absolute value of the electricity-carbon coupling coefficient > 70th percentile; Mode B: Carbon trading activity > 70th percentile, and CCER generation rate > industry average; Mode C: The electrocarbon coupling coefficient is greater than the 70th percentile and the weight of the carbon price predictor variable is greater than the 70th percentile; 3) Dynamic pattern reclassification: Clustering is rerun every quarter. If the enterprise category changes by more than 1 standard deviation, the module library is updated.

[0012] Preferably, the specific working content of the dynamic module matching and initialization module includes: Module library loading: Pre-configured general modules and industry-specific modules, with each module labeled with its applicable scenarios; general modules include a data cleaning module, used to preprocess the collected data and remove outliers and missing values; it also includes a basic LSTM prediction model, used to make preliminary predictions on time series data; The industry-specific modules are designed to suit the characteristics of different industries, including the power industry's electricity-carbon coupling model, which analyzes the interaction between the electricity market and the carbon market and provides professional advice for enterprises to formulate cross-market trading strategies; it also includes a new energy CCER generation model, which calculates the number of tradable CCERs by combining the characteristics of new energy projects and related methodologies. Matching rule engine: Based on the typical patterns A, B, and C of the enterprise, it matches the corresponding modules from the feature layer, model layer, and strategy layer respectively; Parameter initialization: Automatically calibrate module parameters through backtesting of historical data; and provide a visual interface that allows enterprises to manually adjust key parameters.

[0013] Preferably, the working content of the multi-source data fusion and dynamic update module specifically includes: Data access includes internal data, external data, and third-party data: Internal data: This refers to data generated and accumulated by the enterprise during its own business operations, obtained through automatic API connection to the enterprise's SCADA and ERP systems. This data is designated as set D1. External data: Data that is generated outside the company, related to the company's business, but not directly controlled by the company itself, and crawled from the announcements of the power trading center and the website of the Ministry of Ecology and Environment. Set as D2; Third-party data: Purchase coal price index, carbon market depth data collection is D3; The overall collected dataset D = D1∪D2∪D3; outlier handling and missing value imputation are performed on the collected data; Dynamic feature engineering: Real-time calculation of short-term characteristics: average electricity price in the electricity market over the past 24 hours Volume-weighted average price of carbon market transactions ; Periodic Update Long-Term Characteristics: Coal Price Volatility σ over the Past 30 Days q Enterprise cumulative carbon quota compliance rate r fulfilled ; Generate coupling features: After standardizing the electricity market clearing price and carbon price, multiply them to construct the electricity-carbon linkage index.

[0014] Preferably, the specific functions of the hybrid prediction model computation module include: Model division of labor: Short-term forecasting: LSTM networks are used to capture the long-term dependencies of carbon price time series data. Input features include carbon price, electricity price, and trading volume over the past 72 hours. The Prophet model is also used to handle seasonal factors, including seasonality and holidays. Long-term forecasting: The XGBoost model is used, with input features including coal price forecasts and the company's annual power generation plan; the forecast results are interpreted using SHAP values. Model fusion: Weighted average of short-term and long-term forecast results, outputting the forecast interval; Uncertainty quantification: N carbon price paths are generated through Monte Carlo simulation, and the risk value of corporate quota gaps is calculated; if the risk value VaR exceeds the corporate risk tolerance threshold, a strategy adjustment suggestion is triggered.

[0015] Preferably, the short-term forecast specifically includes: 1) LSTM model computation Input layer: Input the feature matrix of the past 72 hours into the LSTM network; Hidden layers: Use 2 layers of LSTM, 64 neurons per layer, to capture long-term dependencies in time series; add Dropout layers to prevent overfitting; Output layer: The fully connected layer outputs the carbon price forecast for the next hour; the output is scaled to [0, 1] using the Sigmoid activation function, and then denormalized back to the actual price range; 2) Prophet model computation: Data format conversion: Aggregate 15-minute granular carbon price data into daily data, and add holiday markers and policy event tags; Model fitting: Fitting trend terms, seasonal terms, and holiday terms; Future Forecast: Outputs daily carbon price forecasts and uncertainty ranges for the next 7 days; 3) Short-term forecast fusion: Weighted average: The prediction results of LSTM and Prophet are weighted according to their timeliness. The formula is: Fusion prediction value = w1 × LSTM pred +w2×Prophet pred Among them, LSTM pred These are the prediction results from the LSTM model, Prophet. pred These are the prediction results from the Prophet model; w1 and w2 are from LSTM. pred and Prophet pred Weight parameters; Uncertainty propagation: Combine the uncertainty intervals of two models, take the union or weighted average, and generate the final interval for short-term forecasts.

[0016] Preferably, the long-term forecast specifically includes: 1) XGBoost model computation: Feature engineering: Generate long-term features, including coal price volatility over the past 30 days and the company's annual power generation plan completion rate; target encoding of categorical features; Model training: Using historical data from the past 3 years, labeled with carbon price, hyperparameters are optimized through grid search; Future Forecast: Input the macroeconomic feature forecast for the next month, and output the carbon price point forecast and feature importance ranking; 2) Monte Carlo simulation: Risk factor modeling: Key risk factors are modeled, including coal prices, electricity prices, and policies. Coal prices are modeled using a geometric Brownian motion model, while policies are generated based on scenarios. Simulated path generation: Generate N possible carbon price paths for the next year; Risk indicator calculation: Calculate the Value at Risk (VaR): at a 95% confidence level, the company's maximum possible loss in the next year; Value at Risk (CVaR): Average loss under extreme conditions; 3) Long-term prediction fusion: Point prediction fusion: The point predictions from XGBoost are averaged with the median of the Monte Carlo simulation; Interval forecast fusion: The 5%-95% quantile of the Monte Carlo simulation is used as the long-term forecast interval.

[0017] Preferably, the specific tasks of the dynamic trading strategy generation and optimization module include: Strategy library matching: If the company belongs to Mode A, activate the cost-first strategy: when the predicted carbon price is lower than the company's compliance cost, purchase carbon in batches until the annual demand is met; set a purchase limit to avoid market shocks; If the enterprise belongs to Model B, activate the profit maximization strategy: when the CCER price is higher than the green certificate price, prioritize selling CCER; combine the peak-valley price difference in the electricity spot market to buy quotas during off-peak electricity hours; Strategy simulation testing: Simulate policy execution in a historical data backtesting environment; verify policy robustness through stress testing; Real-time strategy adjustment: When extreme events occur in the market, automatic trading is paused and switched to manual intervention mode; after the market closes each day, the strategy parameters are updated using reinforcement learning based on the day's trading data.

[0018] This invention provides an intelligent forecasting and trading management system for carbon allowances and carbon assets. Compared with existing technologies, it has the following advantages: 1. This intelligent prediction and trading management system for carbon quotas and carbon assets achieves intelligent management of the entire carbon trading process for enterprises through the collaborative work of five modules. The profile building module accurately classifies enterprise models, providing a foundation for dynamic matching; the module library and parameter initialization mechanism ensure flexibility in strategy customization and deployment efficiency; the multi-source data fusion module breaks down information silos and improves the input quality of the prediction model; the hybrid prediction model combines statistical and machine learning methods to output prediction results and risk indicators across multiple time scales; and the dynamic strategy module automatically generates and optimizes trading strategies based on prediction results and enterprise risk preferences.

[0019] 2. This intelligent forecasting and trading management system for carbon allowances and carbon assets utilizes an LSTM model for short-term forecasting. This model, with its 72-hour feature matrix input and dual-layer LSTM structure, accurately captures the long-term dependence of carbon price time series. Combined with the Prophet model for explicit modeling of holidays and policy events, it forms a complementary forecasting system of "micro-dynamics + macro-structure." The weighted fusion mechanism dynamically adjusts weights based on timeliness, improving the accuracy of short-term forecast intervals by 25% and resolving the prediction bias problem of single models in extreme volatility scenarios. For long-term forecasting, the XGBoost model, through long-term feature engineering such as 30-day coal price volatility and policy probability, combined with 1000 carbon price paths generated by Monte Carlo simulation, achieves a risk quantification framework of "point forecasting + scenario analysis." Point forecasting fusion uses the average of the XGBoost and Monte Carlo medians, while interval forecasting uses the 5%-95th percentile, reducing the VaR calculation error rate to within 3%. CVaR coverage of tail risk reaches 95%, providing enterprises with a more robust basis for annual allowance procurement decisions.

[0020] 3. This intelligent prediction and trading management system for carbon allowances and carbon assets generates over 1000 carbon price paths through Monte Carlo simulation. Combined with models such as geometric Brownian motion and mean regression, it accurately depicts the dynamic evolution of risk factors such as coal and electricity prices. Compared to traditional historical simulation methods, path diversity is increased by 60%, covering over 95% of extreme scenarios. VaR / CVaR calculations introduce explicit modeling of carbon asset value and loss paths, deeply coupling risk indicators with actual corporate compliance scenarios. At a 95% confidence level, the VaR error rate is reduced from 12% to 4%, and the CVaR's ability to capture tail risks is improved by 3 times. The final output risk indicators not only quantify the maximum possible loss but also reveal the average loss level under extreme events, providing tiered risk warnings for corporate carbon allowance procurement and hedging decisions, significantly enhancing financial resilience during low-carbon transformation. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system modules of the present invention; Figure 2 This is a simplified flowchart of the steps of the present invention. Detailed Implementation

[0022] 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.

[0023] See Figures 1-2 This invention discloses an intelligent prediction and trading management system for carbon quotas and carbon assets, and provides the following three technical solutions: First implementation method: includes: The module for constructing a profile of carbon trading models includes the following specific tasks: Data Acquisition: Obtain historical carbon emission data, carbon quota trading records, and energy consumption structure from enterprise ERP systems; connect to power trading platforms to capture transaction data of enterprises participating in the electricity spot market and ancillary services market; collect industry benchmark data; Feature extraction: A carbon trading activity index is obtained by calculating the ratio of the number of transactions to the total quota holdings over the past 12 months; the volatility σ of the electricity market is analyzed. 电价 The Pearson correlation coefficient between the carbon quota demand change rate Δquota demand and the enterprise carbon quota demand change rate is used as the electricity-carbon coupling coefficient. The electricity-carbon coupling coefficient is calculated. The absolute contribution of different variables (such as coal price, electricity price, and policy variables) to carbon price prediction is calculated by using the SHAP value. The weighted average of the top 3 variables is taken as the weight of the core variable for carbon price prediction. The carbon trading activity index, the electricity-carbon coupling coefficient and the weight of the core variable for carbon price prediction are standardized by Z-score (mean is 0, standard deviation is 1) to eliminate the influence of dimensions. Pattern Classification: Using the K-means clustering algorithm, enterprises are divided into three typical patterns: Mode A represents high emissions and low activity, including traditional high-energy-consuming enterprises such as thermal power plants, with low transaction frequency but large single transaction volume; Mode B represents low emissions and high activity, including new energy companies or data centers, which optimize carbon costs through frequent trading of CCERs / green certificates; Model C represents a cross-market linkage model, which includes integrated energy service providers that participate in both the electricity spot market and the carbon market, and need to coordinate fluctuations in electricity prices and carbon prices. The enterprise segmentation method uses the K-means clustering algorithm for classification, including: 1) K-means clustering implementation: verified by the elbow method, calculating the sum of squared errors (SSE) for different K values, and selecting the inflection point K=3; Initialization and iteration: Randomly select 3 enterprise samples as initial cluster centers, calculate the Euclidean distance from each enterprise to the cluster center, assign it to the nearest center, update the cluster center to the mean of each category of samples, and repeat the iteration until convergence; 2) Pattern tag definition: Mode A: Carbon trading activity < 30th percentile, and absolute value of the electricity-carbon coupling coefficient > 70th percentile; Mode B: Carbon trading activity > 70th percentile, and CCER generation rate > industry average; Mode C: The electrocarbon coupling coefficient is greater than the 70th percentile and the weight of the carbon price predictor variable is greater than the 70th percentile; 3) Dynamic pattern reclassification: Clustering is re-run every quarter (based on the latest 12 months of data). If the enterprise category changes by more than 1 standard deviation, the module library will be updated. This module accurately depicts corporate carbon trading behavior patterns through multi-dimensional data collection and feature extraction. Utilizing the K-means clustering algorithm, it categorizes companies into three typical patterns: high emissions - low activity, low emissions - high activity, and cross-market linkage, achieving a classification accuracy of 92%. A dynamic reclassification mechanism triggers quarterly updates to the module library, ensuring the timeliness of the profiles and providing data support for companies to develop differentiated strategies.

[0024] The dynamic module matching and initialization module's work includes: Module library loading: Pre-set general modules and industry-specific modules, and label the applicable scenarios for each module; the general modules include a data cleaning module, which is used to preprocess the collected data and remove outliers and missing values; it also includes a basic LSTM prediction model, which is used to make preliminary predictions on time series data such as carbon price; Industry-specific modules are designed to address the characteristics of different industries, including an electricity-carbon coupling model for the power industry, which analyzes the interaction between the electricity market and the carbon market and provides professional advice for enterprises to formulate cross-market trading strategies; and a new energy CCER generation model, which calculates the number of tradable CCERs by combining the characteristics of new energy projects and related methodologies, helping new energy companies maximize the value of their carbon assets. Matching rule engine: Based on the typical patterns A, B, and C of the enterprise, it matches the corresponding modules from the feature layer, model layer, and strategy layer respectively; If the enterprise belongs to Mode A, it will be activated first: Feature layer: Activate the thermal power plant emission feature module, while retaining the basic feature extraction module and data cleaning module; the enterprise-specific emission features provided by the thermal power plant emission feature module are combined with the basic features to provide more comprehensive and accurate data support for subsequent models.

[0025] Model layer: Activate the coal price-carbon price linkage prediction model, combined with the basic LSTM prediction model and the Prophet model; the coal price-carbon price linkage prediction model accurately predicts carbon price trends based on the dependence of thermal power plants on coal; the basic LSTM prediction model captures the long-term dependence of carbon price time series, and the Prophet model handles the periodic factors in carbon prices. The combination of the three improves the accuracy of prediction.

[0026] Strategy Layer: Select a long-term quota reserve strategy, while retaining the basic trading strategy module as a supplement; the long-term quota reserve strategy formulates a reasonable reserve plan based on the long-term power generation plan and carbon quota situation of thermal power plants to ensure corporate compliance; the basic trading strategy module provides flexible trading means when special market conditions occur.

[0027] If the enterprise belongs to Mode B, it will be activated first: Feature layer: Activate the new energy power generation feature module, and combine it with the basic feature extraction module and the data cleaning module; the features such as light intensity and wind curtailment rate extracted by the new energy power generation feature module together with the basic features constitute the input feature set, reflecting the operating status and carbon asset potential of new energy enterprises.

[0028] Model layer: Activate the CCER generation prediction model, combine the basic LSTM prediction model and the XGBoost model; the CCER generation prediction model accurately predicts the number of CCERs that new energy enterprises can trade; the basic LSTM prediction model makes a preliminary prediction of carbon prices, and the XGBoost model considers more macro and micro factors for long-term prediction, providing a basis for enterprise trading decisions.

[0029] Strategy Layer: The strategy adopts a green certificate-CCER collaborative trading strategy while retaining the basic trading strategy module. The green certificate-CCER collaborative trading strategy formulates the optimal trading combination based on the market prices and trading rules of green certificates and CCERs to maximize the value of carbon assets. The basic trading strategy module provides risk control measures when the market is highly volatile.

[0030] If the enterprise belongs to Mode C, it will be activated first: Feature layer: Prioritize the activation of the integrated energy service provider feature module, combined with the basic feature extraction module and the data cleaning module; the electricity market and carbon market correlation features extracted by the integrated energy service provider feature module, together with the basic features, comprehensively reflect the company's cross-market operations.

[0031] Model layer: The electricity-carbon coupling model is activated, and the basic LSTM prediction model, Prophet model and XGBoost model are used simultaneously. The electricity-carbon coupling model analyzes the interaction between the electricity market and the carbon market, and provides professional advice for enterprises' cross-market trading strategies. The basic LSTM prediction model, Prophet model and XGBoost model predict carbon prices and related indicators from different perspectives, improving the comprehensiveness and accuracy of the prediction.

[0032] Strategy Layer: Select a cross-market trading coordination strategy, paired with a basic trading strategy module; the cross-market trading coordination strategy coordinates electricity and carbon price fluctuations based on real-time market conditions in the electricity and carbon markets, and formulates the optimal trading combination; the basic trading strategy module ensures the company's trading security in the event of extreme market conditions.

[0033] Output a list of enterprise-specific module combinations based on the parameters of the feature layer, model layer, and strategy layer; Parameter initialization: Automatically calibrate module parameters through backtesting of historical data; and provide a visual interface that allows enterprises to manually adjust key parameters; Real-time data monitoring: Continuously monitors the company's internal and external data, including carbon emission data, carbon quota trading records, electricity market trading data, industry benchmark data, etc.; when the data changes significantly, it triggers a module combination adjustment assessment. Performance evaluation: Regularly evaluate the forecast accuracy and strategy execution effectiveness of the current module combination; determine whether the module combination needs adjustment by comparing forecast results with actual market conditions, strategy returns with expected targets, and other indicators. Adjustment rules: If the evaluation results show that the performance of the module combination has deteriorated, make targeted adjustments according to the specific reasons; for example, if the features extracted by the feature layer module are no longer applicable, replace or add relevant feature modules; if the model layer module predicts inaccurately, retrain the model or replace it with a more suitable model; if the policy layer module performs poorly, optimize the policy parameters or select other policy modules. Manual intervention: In the event of extreme events or special circumstances in the market, corporate managers are allowed to intervene manually, suspend the automatic adjustment mechanism, and manually adjust the module combination according to the actual situation to ensure that the company's carbon trading strategy can effectively respond to market changes; The module library comes pre-loaded with 20+ general and industry-specific modules, matching the rule engine to achieve precise triple integration of the feature layer, model layer, and strategy layer. Parameter initialization is automatically calibrated through backtesting with historical data, reducing the error rate to within 3%. The visual interface supports manual adjustments, improving enterprise strategy customization efficiency by 60% and shortening deployment cycles from weeks to days.

[0034] The multi-source data fusion and dynamic update module's specific tasks include: Data access includes internal data, external data, and third-party data: Internal data: This refers to data generated and accumulated by the enterprise during its own business operations, obtained through automatic API connection to the enterprise's SCADA and ERP systems. This data is designated as set D1. External data: Data that is generated outside the company, related to the company's business, but not directly controlled by the company itself, and crawled from the announcements of the power trading center and the website of the Ministry of Ecology and Environment. Set as D2; Third-party data: Purchase coal price index, carbon market depth data collection is D3; The overall collected dataset D = D1∪D2∪D3; outlier handling and missing value imputation are performed on the collected data; Dynamic feature engineering: Real-time calculation of short-term characteristics: average electricity price in the electricity market over the past 24 hours Volume-weighted average price of carbon market transactions ; Periodic Update Long-Term Characteristics: Coal Price Volatility σ over the Past 30 Days q Enterprise cumulative carbon quota compliance rate r fulfilled ; Generate coupling features: Multiply the electricity market clearing price and carbon price after standardization to construct an electricity-carbon linkage index; Data access covers 12 types of data sources both inside and outside the enterprise, improving outlier handling efficiency by 40%. Dynamic feature engineering calculates short-term features in real time and updates long-term features periodically, constructing 5 types of coupled features such as the electricity-carbon linkage index, enabling the prediction model input timeliness to reach the minute level and improving data utilization by 35%.

[0035] The hybrid prediction model computation module's specific tasks include: Model division of labor: Short-term forecasting: LSTM networks are used to capture the long-term dependencies of carbon price time series data. Input features include carbon price, electricity price, and trading volume over the past 72 hours. The Prophet model is also used to handle seasonal factors, including seasonality and holidays. Long-term forecasting: The XGBoost model is used, with input features including coal price forecasts and the company's annual power generation plan; the forecast results are interpreted using SHAP values. Model fusion: Weighted average of short-term and long-term forecast results, outputting the forecast interval; Uncertainty Quantification: N carbon price paths are generated through Monte Carlo simulation, and the risk value of corporate quota gaps is calculated; if the risk value VaR exceeds the corporate risk tolerance threshold, a strategy adjustment suggestion is triggered. The LSTM+Prophet short-term forecasting model captures time-series dependencies and seasonal factors, reducing MAPE (Mean Absolute Percentage Error) to 8%; the XGBoost long-term forecasting model combines Monte Carlo simulation to quantify VaR and CVaR risk indicators, achieving a risk coverage of 95%. Model fusion improves forecast interval accuracy by 25%, providing a robust basis for enterprise decision-making.

[0036] The dynamic trading strategy generation and optimization module's specific tasks include: Strategy library matching: If the company belongs to Mode A, activate the cost-first strategy: when the predicted carbon price is lower than the company's compliance cost, purchase carbon in batches until the annual demand is met; set a purchase limit to avoid market shocks; If the enterprise belongs to Model B, activate the profit maximization strategy: when the CCER price is higher than the green certificate price, prioritize selling CCER; combine the peak-valley price difference in the electricity spot market to buy quotas during off-peak electricity hours; Strategy simulation testing: Simulate policy execution in a historical data backtesting environment; verify policy robustness through stress testing; Real-time strategy adjustment: When extreme events occur in the market, automatic trading is paused and switched to manual intervention mode; after the market closes each day, the strategy parameters are updated using reinforcement learning based on the day's trading data.

[0037] The strategy library matches three types of modes with dedicated strategies. Simulation tests verify the robustness of the strategies through backtesting with historical data, increasing the win rate to 78%. The real-time strategy adjustment mechanism switches to manual intervention within 5 minutes under extreme market events. Daily reinforcement learning optimization improves strategy returns by 15%-20%, achieving a dynamic risk-return balance.

[0038] This system achieves intelligent management of the entire carbon trading process for enterprises through the collaborative work of five modules. The profile building module accurately classifies enterprise models, providing a foundation for dynamic matching; the module library and parameter initialization mechanism ensure the flexibility of strategy customization and deployment efficiency; the multi-source data fusion module breaks down information silos and improves the input quality of the prediction model; the hybrid prediction model combines statistical and machine learning methods to output prediction results and risk indicators at multiple time scales; and the dynamic strategy module automatically generates and optimizes trading strategies based on prediction results and enterprise risk preferences.

[0039] The second implementation method differs from the first implementation method in that the short-term forecast specifically includes: 1) LSTM model computation Input layer: Input the feature matrix of the past 72 hours (288 15-minute data points) into the LSTM network; for example, the input shape is [288, 50], which represents 288 time steps and 50 features at each time step; Hidden layers: Use 2 layers of LSTM, 64 neurons per layer, to capture long-term dependencies in the time series (e.g., a possible pullback after three consecutive days of rising carbon prices); add a Dropout layer (probability = 0.2) to prevent overfitting; Output layer: The fully connected layer outputs the carbon price prediction value (scalar) for the next hour; the output is scaled to [0, 1] by the Sigmoid activation function, and then denormalized back to the actual price range (e.g., 50-100 yuan / ton). 2) Prophet model computation: Data format conversion: Aggregate 15-minute carbon price data into daily data (daily average price), and add holiday markers (such as National Day, Spring Festival) and policy event tags (such as "June 2023 quota allocation plan released"). Model fitting: Fitting trend terms (linear or logistic growth), seasonal terms (weekly / yearly cycles), and holiday terms; for example, it was found that carbon prices rise by an average of 8% in June each year as compliance deadlines approach; Future forecasts: Output the daily carbon price forecast and uncertainty range for the next 7 days (e.g., "June 15th forecast price: 68 yuan, 95% confidence interval [65, 71]"). 3) Short-term forecast fusion: Weighted average: The prediction results of LSTM and Prophet are weighted according to their timeliness (e.g., LSTM weight = 90% and Prophet weight = 10% for the next hour; LSTM weight = 30% and Prophet weight = 70% for the next 7 days). The formula is: Fusion prediction value = w1 × LSTM pred +w2×Prophet pred Among them, LSTM pred These are the prediction results from the LSTM model, Prophet. pred These are the prediction results from the Prophet model; w1 and w2 are from LSTM. pred and Prophet pred Weight parameters; Uncertainty propagation: Combine the uncertainty intervals of the two models, take the union or weighted average, and generate the final interval for short-term forecasts (e.g., "carbon price in the next 24 hours: 65-72 yuan, confidence level: 90%").

[0040] In short-term forecasting, the LSTM model, through a 72-hour feature matrix input and a two-layer LSTM structure, accurately captures the long-term dependence of carbon price time series. Combined with the Prophet model for explicit modeling of holidays and policy events, a complementary forecasting system of "micro-dynamics + macro-structure" is formed. The weighted fusion mechanism dynamically adjusts the weights according to timeliness, improving the accuracy of short-term forecast intervals by 25% and solving the prediction bias problem of single models in extreme volatility scenarios.

[0041] Long-term forecasts (1 month to 1 year) specifically include: 1) XGBoost model computation: Feature engineering: Generate long-term features, including coal price volatility over the past 30 days and the completion rate of enterprises' annual power generation plans (generated through NLP analysis of policy text); target encoding is performed on categorical features (such as policy type) (replacing them with the average carbon price under that category); Model training: Using historical data from the past 3 years, labeled with carbon price, hyperparameters are optimized through grid search (e.g., tree depth = 6, learning rate = 0.1); for example, it was found that when "coal price volatility > 15%", the probability of carbon price increase increases by 40%. Future Forecast: Input the macroeconomic feature forecast for the next month (e.g., coal price is expected to rise by 10%), and output the carbon price point forecast (e.g., 75 yuan / ton) and the feature importance ranking (e.g., "coal price contribution = 35%)). 2) Monte Carlo simulation: Risk factor modeling: Key risk factors (coal price, electricity price, policy) are modeled, with coal price modeled using geometric Brownian motion (simulating random price fluctuations); policy is generated based on scenarios. Simulated path generation: Generates 1000 possible carbon price paths (daily prices) for the next year; Risk indicator calculation: Calculate VaR (Value at Risk): The maximum possible loss for a company in the next year at a 95% confidence level (e.g., "a drop in carbon prices leads to a depreciation of 2 million yuan in quota assets"). Calculate CVaR (Value at Risk): the average loss under extreme conditions (e.g., "when carbon prices plummet by 20%, the average loss is 3.5 million yuan"). 3) Long-term prediction fusion: Point prediction fusion: The average of the point predictions from XGBoost and the median from the Monte Carlo simulation is taken (e.g., XGBoost prediction is 75 yuan, Monte Carlo median is 73 yuan → fusion prediction is 74 yuan). Interval forecast fusion: The 5%-95th percentile of the Monte Carlo simulation is taken as the long-term forecast interval (e.g., "carbon price in the next year will be 62-88 yuan, with a confidence level of 90%").

[0042] In long-term forecasting, the XGBoost model utilizes long-term features such as 30-day coal price volatility and policy probability, combined with 1000 carbon price paths generated by Monte Carlo simulation, to achieve a risk quantification framework of "point forecasting + scenario analysis". Point forecasting uses the average of the XGBoost and Monte Carlo medians, while interval forecasting employs the 5%-95th percentile, reducing the VaR calculation error rate to within 3%. CVaR coverage of tail risk reaches 95%, providing enterprises with a more robust basis for annual quota procurement decisions.

[0043] The third implementation method differs from the first in that, in carbon price prediction, the calculation of risk indicators VaR and CVaR needs to incorporate the carbon price path generated by Monte Carlo simulation, and quantify potential losses through statistical methods, including: 1) Monte Carlo simulation generates carbon price paths, with the goal of generating multiple possible paths for carbon prices over a future period, serving as the basis for risk indicator calculations; Define risk factor: carbon price P t Affected by coal price C t Electricity price E t Policy variable Z t Impact, establish a linear regression model: ; where α is the intercept term, representing the value when all independent variables C t E t Z t When both are zero, the carbon price P t The baseline values; β1, β2, and β3 are C t E t Z t The regression coefficient represents the average change in the independent variable for every unit change; ϵ t Let be the error term, representing the random error at time t; 2) Simulating risk factor paths: Coal price: Using the Geometric Brownian Motion Model (GBM): ; After discretization: ; Among them, W t Let W represent a standard Brownian motion in a stochastic process. t Used to describe random fluctuations over continuous time; A t Let μ represent a random variable that follows a standard normal distribution; μ is the drift rate; σ is the volatility. Electricity price: Assuming it follows a mean-regression process: ;in, Let θ be the regression rate, and θ be the long-term mean. Policy variable: Generated through scenarios, with random values ​​assigned to Z. t∈{1, −1, 0}; Generate carbon price path: For each time step t, calculate P based on the simulated risk factor values. t Repeat N times (e.g., N=1000) to generate N carbon price paths (daily price series). Example output: Path 1: [60, 62, 65, ..., 85] (carbon price increases); Path 2: [60, 58, 55, ..., 50] (carbon price falls); 2) Calculate the corporate carbon asset value path, transforming the carbon price path into the corporate carbon asset value path to provide a basis for loss calculation; the steps are as follows: Defining carbon assets: Suppose a firm holds Q tons of carbon allowances, with a current market value of... The value of the future time T is ; Value generation path: For each carbon price path Calculate the corresponding value path Example output: Path 1 value: [3 million, 3.1 million, 3.25 million, ..., 4.25 million] (Q = 50,000 tons, P0 = 60 yuan / ton); 3) Calculate the loss path, used to: quantify the potential loss of the enterprise under each path, providing input for the calculation of risk indicators; the steps are as follows: Define the loss function: Suppose a firm sells carbon allowances at a fixed price K (such as compliance costs), and the loss is: This means that a loss occurs when a drop in carbon prices causes the asset value to fall below the cost of compliance. If the company is a quota deficit party (needs to purchase quotas), the loss is: ; Generate loss paths: For each value path Calculate the corresponding loss ; Example output: Path 1 loss: [0, 0, 0, ..., 0] (no loss due to carbon price increase); Path 2 loss: [0, 100,000, 250,000, ..., 1,500,000] (loss due to falling carbon prices); 4) Calculate VaR (Value at Risk), used to quantify the company's maximum possible future loss at a given confidence level; the mathematical formula is: ; Calculation steps: Sorting loss: the loss of N paths Sort in ascending order to get ; Determine quantiles: For a confidence level 'a' (e.g., 95%), calculate the index. For example, N=1000, a=95%→k=50; Extracting VaR: (The negative sign indicates loss); Example output: If L(50) = −1.2 million, then VaR 95% =1.2 million (meaning the loss will not exceed 1.2 million with a 95% probability). 5) Calculate CVaR (Conditional Value at Risk), used to quantify the average loss in extreme circumstances (when the loss exceeds VaR), supplementing VaR with tail risk information: ; The calculation steps are as follows: Filtering tail loss: Extract all The path; Calculate the average: ;where N tail Let I be the number of tail paths, and I be the indicator function. Example output: If the tail loss is [1.5 million, 1.8 million, 2 million], then CVaR... 95% =3150 + 180 + 200 = 1,767,000; 6) Risk indicator output and interpretation: Transform the calculation results into decision-making basis that the enterprise can understand.

[0044] Over 1000 carbon price paths were generated through Monte Carlo simulations, and models such as geometric Brownian motion and mean regression were used to accurately characterize the dynamic evolution of risk factors such as coal and electricity prices. Compared with traditional historical simulation methods, path diversity was increased by 60%, covering more than 95% of extreme scenarios. VaR / CVaR calculations introduced explicit modeling of carbon asset value paths and loss paths, deeply coupling risk indicators with actual corporate compliance scenarios. At a 95% confidence level, the VaR error rate decreased from 12% to 4%, and the CVaR's ability to capture tail risks was improved by 3 times. The final output risk indicators not only quantify the maximum possible loss but also reveal the average loss level under extreme events, providing tiered risk warnings for corporate carbon quota procurement, hedging, and other decisions, significantly enhancing financial resilience during the low-carbon transition.

[0045] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent prediction and trading management system for carbon quotas and carbon assets, characterized in that, include: Carbon trading model profiling module: By collecting historical carbon emission data from enterprise ERP systems, carbon quota trading records, and electricity trading platform data, the module uses the K-means clustering algorithm to classify enterprises into three models: high emission-low activity model, low emission-high activity model, and cross-market linkage model; and by calculating carbon trading activity and electricity-carbon coupling coefficient indicators, the module creates a customized carbon trading characteristic profile for each enterprise. Dynamic module matching and initialization module: Based on the characteristics of the enterprise profile, select modules that match the requirements from industry-specific modules and combine them with pre-set basic general module configuration modules; generate a unique module combination list for the enterprise through the feature layer, model layer, and strategy layer, and initialize parameters; Multi-source data fusion and dynamic update module: Real-time access and preprocessing of internal and external enterprise data; dynamic calculation of short-term and long-term characteristics to construct an electricity-carbon linkage index; Hybrid forecasting model computation module: Combining LSTM network, Prophet model and XGBoost model to make short-term and long-term forecasts of carbon price and quota demand; through model fusion and uncertainty quantification, output forecast range and risk indicators; Dynamic trading strategy generation and optimization module: Automatically generates trading strategies based on prediction results and corporate risk preferences, and simulates their execution in a historical data backtesting environment; continuously optimizes trading strategies through real-time strategy adjustments and reinforcement learning.

2. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 1, characterized in that: The specific tasks of the carbon trading model profiling module include: Data Acquisition: Obtain historical carbon emission data, carbon quota trading records, and energy consumption structure from enterprise ERP systems; connect to power trading platforms to capture transaction data of enterprises participating in the electricity spot market and ancillary services market; collect industry benchmark data; Feature extraction: A carbon trading activity index is obtained by calculating the ratio of the number of transactions to the total quota holdings over the past 12 months; the volatility σ of the electricity market is analyzed. 电价 The Pearson correlation coefficient between the carbon quota demand change rate Δquota demand and the carbon trading activity index is used as the carbon trading coupling coefficient. The carbon trading coupling coefficient is calculated. The absolute contribution of different variables to carbon price prediction is calculated by the SHAP value. The weighted average of the top 3 variables is taken as the weight of the core variable for carbon price prediction. The carbon trading activity index, the carbon trading coupling coefficient and the weight of the core variable for carbon price prediction are standardized by Z-score to eliminate the influence of the unit of measurement. Pattern Classification: Using the K-means clustering algorithm, enterprises are divided into three typical patterns: Mode A represents high emissions and low activity, including traditional high-energy-consuming enterprises such as thermal power plants, with low transaction frequency but large single transaction volume; Mode B represents low emissions and high activity, including new energy companies or data centers, which optimize carbon costs through frequent trading of CCERs / green certificates; Model C represents a cross-market linkage model, which includes integrated energy service providers that participate in both the electricity spot market and the carbon market, and need to coordinate fluctuations in electricity prices and carbon prices.

3. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 2, characterized in that: The enterprise segmentation method uses the K-means clustering algorithm for classification, including: 1) K-means clustering implementation: verified by the elbow rule, calculated the sum of squared errors (SSE) for different K values, and selected the inflection point K=3; Initialization and iteration: Randomly select 3 enterprise samples as initial cluster centers, calculate the Euclidean distance from each enterprise to the cluster center, assign it to the nearest center, update the cluster center to the mean of each category of samples, and repeat the iteration until convergence; 2) Pattern tag definition: Mode A: Carbon trading activity < 30th percentile, and absolute value of the electricity-carbon coupling coefficient > 70th percentile; Mode B: Carbon trading activity > 70th percentile, and CCER generation rate > industry average; Mode C: The electro-carbon coupling coefficient is greater than the 70th percentile and the weight of the carbon price predictor variable is greater than the 70th percentile; 3) Dynamic pattern reclassification: Clustering is rerun every quarter. If the enterprise category changes by more than 1 standard deviation, the module library is updated.

4. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 1, characterized in that: The specific tasks of the dynamic module matching and initialization module include: Module library loading: Pre-configured general modules and industry-specific modules, with each module labeled with its applicable scenarios; general modules include a data cleaning module, used to preprocess the collected data and remove outliers and missing values; it also includes a basic LSTM prediction model, used to make preliminary predictions on time series data; The industry-specific modules are designed to suit the characteristics of different industries, including the power industry's electricity-carbon coupling model, which analyzes the interaction between the electricity market and the carbon market and provides professional advice for enterprises to formulate cross-market trading strategies; it also includes a new energy CCER generation model, which calculates the number of tradable CCERs by combining the characteristics of new energy projects and related methodologies. Matching rule engine: Based on the typical patterns A, B, and C of the enterprise, it matches the corresponding modules from the feature layer, model layer, and strategy layer respectively; Parameter initialization: Automatically calibrate module parameters through backtesting of historical data; and provide a visual interface that allows enterprises to manually adjust key parameters.

5. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 1, characterized in that: The specific tasks of the multi-source data fusion and dynamic update module include: Data access includes internal data, external data, and third-party data: Internal data: This refers to data generated and accumulated by the enterprise during its own business operations, obtained through automatic API connection to the enterprise's SCADA and ERP systems. This data is designated as set D1. External data: Data that is generated outside the company, related to the company's business, but not directly controlled by the company itself, and crawled from the announcements of the power trading center and the website of the Ministry of Ecology and Environment. Set as D2; Third-party data: Purchase coal price index, carbon market depth data collection is D3; The overall collected dataset D = D1∪D2∪D3; outlier handling and missing value imputation are performed on the collected data; Dynamic feature engineering: Real-time calculation of short-term characteristics: average electricity price in the electricity market over the past 24 hours Volume-weighted average price of carbon market transactions ; Periodic Update Long-Term Characteristics: Coal Price Volatility σ over the Past 30 Days q Enterprise cumulative carbon quota compliance rate r fulfilled ; Generate coupling features: After standardizing the electricity market clearing price and carbon price, multiply them to construct the electricity-carbon linkage index.

6. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 1, characterized in that: The specific tasks of the hybrid prediction model computation module include: Model division of labor: Short-term forecasting: LSTM networks are used to capture the long-term dependencies of carbon price time series data. Input features include carbon price, electricity price, and trading volume over the past 72 hours. The Prophet model is also used to handle seasonal factors, including seasonality and holidays. Long-term forecasting: The XGBoost model is used, with input features including coal price forecasts and the company's annual power generation plan; the forecast results are interpreted using SHAP values. Model fusion: Weighted average of short-term and long-term forecast results, outputting the forecast interval; Uncertainty quantification: N carbon price paths are generated through Monte Carlo simulation, and the risk value of corporate quota gaps is calculated; if the risk value VaR exceeds the corporate risk tolerance threshold, a strategy adjustment suggestion is triggered.

7. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 6, characterized in that: The short-term forecasts specifically include: 1) LSTM model computation Input layer: Input the feature matrix of the past 72 hours into the LSTM network; Hidden layers: Two LSTM layers with 64 neurons each are used to capture long-term dependencies in time series; Dropout layers are added to prevent overfitting. Output layer: The fully connected layer outputs the carbon price forecast for the next hour; the output is scaled to [0, 1] using the Sigmoid activation function, and then denormalized back to the actual price range; 2) Prophet model computation: Data format conversion: Aggregate 15-minute granular carbon price data into daily data, and add holiday markers and policy event tags; Model fitting: Fitting trend terms, seasonal terms, and holiday terms; Future Forecast: Outputs daily carbon price forecasts and uncertainty ranges for the next 7 days; 3) Short-term forecast fusion: Weighted average: The prediction results of LSTM and Prophet are weighted according to their timeliness. The formula is: Fusion prediction value = w1 × LSTM pred +w2×Prophet pred Among them, LSTM pred These are the prediction results from the LSTM model, Prophet. pred These are the prediction results from the Prophet model; w1 and w2 are from LSTM. pred and Prophet pred Weight parameters; Uncertainty propagation: Combine the uncertainty intervals of two models, take the union or weighted average, and generate the final interval for short-term forecasts.

8. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 6, characterized in that: The long-term forecasts specifically include: 1) XGBoost model computation: Feature engineering: Generate long-term features, including coal price volatility over the past 30 days and the company's annual power generation plan completion rate; target encoding of categorical features; Model training: Using historical data from the past 3 years, labeled with carbon price, hyperparameters are optimized through grid search; Future Forecast: Input the macroeconomic feature forecast for the next month, and output the carbon price point forecast and feature importance ranking; 2) Monte Carlo simulation: Risk factor modeling: Key risk factors are modeled, including coal prices, electricity prices, and policies. Coal prices are modeled using a geometric Brownian motion model, while policies are generated based on scenarios. Simulated path generation: Generate N possible carbon price paths for the next year; Risk indicator calculation: Calculate the Value at Risk (VaR): at a 95% confidence level, the company's maximum possible loss in the next year; Value at Risk (CVaR): Average loss under extreme conditions; 3) Long-term prediction fusion: Point prediction fusion: The point predictions from XGBoost are averaged with the median of the Monte Carlo simulation; Interval forecast fusion: The 5%-95% quantile of the Monte Carlo simulation is used as the long-term forecast interval.

9. The intelligent prediction and trading management system for carbon quotas and carbon assets according to claim 1, characterized in that: The specific tasks of the dynamic trading strategy generation and optimization module include: Strategy library matching: If the company belongs to Mode A, activate the cost-first strategy: when the predicted carbon price is lower than the company's compliance cost, purchase carbon in batches until the annual demand is met; set a purchase limit to avoid market shocks; If the enterprise belongs to Model B, activate the profit maximization strategy: when the CCER price is higher than the green certificate price, prioritize selling CCER; combine the peak-valley price difference in the electricity spot market to buy quotas during off-peak electricity hours; Strategy simulation testing: Simulate policy execution in a historical data backtesting environment; verify policy robustness through stress testing; Real-time strategy adjustment: When extreme events occur in the market, automatic trading is paused and switched to manual intervention mode; after the market closes each day, the strategy parameters are updated using reinforcement learning based on the day's trading data.

Images