Carbon asset value prediction and dynamic classification management system and method

Abstract

The application discloses a carbon asset value prediction and dynamic classification management system and method, relates to the technical field of carbon emission management and artificial intelligence, and comprises the following steps: obtaining historical carbon price time series, carbon transaction microstructure indexes, public opinion heat and sentiment scores, search index change rates, enterprise energy consumption load data and macro energy price data corresponding to carbon assets, performing normalization and time alignment processing on various data, constructing an input data sequence, obtaining market expectation factors by performing feature coding on the carbon transaction microstructure indexes, public opinion heat and sentiment scores and search index change rates, inputting the input data sequence and the market expectation factors into a hybrid neural network model for calculation, outputting quantile prediction results and corresponding confidence intervals of carbon asset prices in a future time window, and calculating a supply-demand gap distribution based on enterprise emission prediction results and carbon asset holding amounts.

Description

Technical Field

[0001] This invention relates to the fields of carbon emission management and artificial intelligence technology, specifically to a carbon asset value prediction and dynamic classification management system and method. Background Technology

[0002] With the continuous advancement of dual-carbon goals and the increasing demand from enterprises for carbon compliance, carbon achievement, and carbon asset management, carbon assets such as carbon quotas, CCERs, and green certificates have gradually evolved from simple compliance tools into comprehensive resources with compliance, trading, and financial attributes. Against this backdrop, industrial parks and enterprises not only need to account for and record carbon flows, but also need to further conduct value assessments, status identification, and disposal management of existing carbon assets to support compliance arrangements, holding decisions, trading timing selection, and financing utilization. Especially given that carbon market price fluctuations are influenced by multiple factors such as compliance cycles, policy changes, market sentiment, trading behavior, and energy prices, relying solely on static ledgers or single price observations is insufficient to meet the practical needs of enterprises for refined, dynamic, and intelligent management of carbon assets. Therefore, predictive analysis, classification, identification, and strategic linkage based on future value changes of carbon assets have become an important direction for the development of carbon emission management and intelligent decision-making technologies.

[0003] The existing technology has the following shortcomings: In existing technologies, most carbon management platforms still focus on real-time carbon flow accounting, emission monitoring, and data storage. For existing carbon assets such as carbon allowances, CCERs, and green certificates, a relatively static and extensive management approach is generally adopted, which has the following main shortcomings: First, existing value prediction methods are usually based solely on historical carbon price time series analysis, making it difficult to effectively integrate multi-source heterogeneous data such as transaction microstructure, public opinion heat, search index, enterprise energy load, and macro energy prices. In particular, it is difficult to reflect the impact of policy events and market expectations on the value of carbon assets in an engineerable way, resulting in insufficient stability and reliability of prediction results. Second, existing asset classification methods are mostly divided according to a single dimension such as asset type and maturity time, lacking dynamic adaptive classification capabilities that consider multi-dimensional characteristics such as compliance urgency, value potential, and liquidity, and cannot adjust classification boundaries according to market changes and enterprise preferences. Third, existing strategy matching mechanisms mostly rely on fixed rules and lack the ability to perceive the confidence and uncertainty of prediction results. When market volatility increases or prediction deviations are high, they cannot switch to conservative strategies or trigger manual review in a timely manner, thus making it difficult to balance the efficiency of automated management and the security of decision-making.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a carbon asset value prediction and dynamic classification management system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for carbon asset value prediction and dynamic classification management, comprising the following steps: This study acquires historical carbon price time series, carbon trading microstructure indicators, public opinion heat and sentiment scores, search index change rate, enterprise energy load data, and macro energy price data corresponding to carbon assets. It normalizes and aligns these data types to construct an input data sequence. Market expectation factors are obtained by feature encoding the carbon trading microstructure indicators, public opinion heat and sentiment scores, and search index change rate. The input data sequence and market expectation factors are then fed into a hybrid neural network model for calculation. The model outputs the quantile prediction results and corresponding confidence intervals for carbon asset prices within future time windows. Finally, the supply-demand gap distribution is calculated based on enterprise emission predictions and carbon asset holdings. Based on quantile prediction results, confidence intervals, and supply-demand gap distribution, a feature vector containing compliance urgency, value potential, and liquidity is constructed for each carbon asset. After normalization of the feature vector, a weighted distance is used for clustering calculation. The weights are obtained by constructing samples from historical strategy adoption results and return feedback and solving them through linear discriminant analysis, thereby obtaining dynamic classification results of carbon assets that change with quantile prediction results. The prediction confidence level is determined by using the dynamic classification results and confidence intervals of carbon assets. When the prediction confidence level is lower than the preset threshold, a manual review instruction is output. When the prediction confidence level is not lower than the preset threshold, the disposal rules are matched according to the dynamic classification results of carbon assets and the disposal parameters are adjusted, and the carbon asset disposal results are output.

[0007] Preferably, by introducing market behavior characteristics to enhance the informational capacity of price forecast inputs and constructing a unified form for representing market expectations, the steps are as follows: Collect micro-structure indicators of carbon trading, public opinion heat, sentiment scores and search index change rates, perform timestamp alignment, outlier removal and missing value imputation on various types of data, and form an original feature set according to a unified time granularity; The data from different sources in the original feature set are standardized, each feature is mapped to a unified numerical range, and vector concatenation is performed according to a preset dimension order to generate the market expectation feature vector. The market expectation feature vector is input into an encoding network containing a multi-layer linear transformation structure, and matrix multiplication, bias superposition and nonlinear activation are performed sequentially to obtain a low-dimensional compressed representation vector. The low-dimensional representation vector is combined with historical price time series data in terms of feature dimensions to construct a unified input tensor, which then participates in the subsequent price quantile prediction calculation process.

[0008] Preferably, a multi-quantile output structure is constructed based on the expression of the future price distribution of carbon assets to form interval-based prediction results. The steps are as follows: Construct a prediction structure containing multiple parallel output nodes, each node corresponding to a preset quantile, and assign an independent set of parameters to each node; A linear transformation is performed on the fused features, and weight matrix multiplication and bias superposition are performed for different quantiles to obtain the corresponding quantile prediction values; The output results of multiple quantiles are organized and arranged in the order of time steps to form a multidimensional prediction sequence covering future time windows; The median forecast is extracted as a price reference sequence, and the upper and lower quantile results are combined to generate a continuous interval representation for subsequent processing.

[0009] Preferably, the supply and demand relationship is characterized by combining changes in emissions with asset holdings, and then extended to express the uncertainty characteristics in an interval form. The steps are as follows: Obtain emission forecasts for future timeframes and current carbon asset holdings data, and perform unified and time-aligned processing on data from different sources; For each time step, the difference between emissions and holdings is calculated to form a continuous supply and demand gap sequence, which is then numerically standardized. A random perturbation set is constructed for the emission prediction residuals, multiple emission paths are generated through repeated sampling, and the corresponding gap result is calculated for each path. Statistical analysis was performed on all gap samples to extract the median and quantile intervals, forming a distribution representation of the supply and demand gap.

[0010] Preferably, a carbon asset urgency index is constructed based on compliance pressure and time factors to characterize the role of different assets in the compliance process. The steps are as follows: Read the median of the supply-demand gap, the quantity of individual assets, and the offset ratio parameters to calculate the maximum scale of assets that can be used to fulfill obligations; The calculation results are subjected to non-negativity constraints and proportional normalization to obtain the gap ratio and limit its range. Obtain information on the total duration and remaining time of the contract fulfillment cycle, and perform proportional conversion on the time variable to obtain the time progress indicator; The gap ratio, time schedule, and penalty items are linearly weighted and combined to form a unified numerical expression of compliance urgency.

[0011] The preferred approach is to construct a revenue potential expression method that combines price change trends with asset-added attributes to characterize the future value changes of different carbon assets. The steps are as follows: Obtain the median price at the end of the forecast time window and the current market price data, and perform alignment and consistency checks on the price series; The difference between the predicted price and the current price is calculated and the ratio is normalized to obtain a series of price change rate indicators. Obtain green certificate prices and benchmark electricity price data and calculate the premium rate, while also obtaining information on the discount rate for pledged financing corresponding to the assets; A weighted combination calculation is performed on the price change rate, premium rate, and financing discount rate to form a unified value potential indicator output.

[0012] Preferably, the feature weights in the classification process are dynamically adjusted based on historical return feedback information to reflect the influence of different dimensions on the results. The steps are as follows: Collect historical strategy execution records and corresponding profit results, filter and process the data, and form a labeled sample set; The samples are classified according to the profit results, and positive and negative sample sets are generated and stored in a structured manner. The intra-class scatter matrix and inter-class scatter matrix are calculated based on the sample set, and the distribution of samples of different categories is expressed in a matrix form. The corresponding eigenvectors are obtained by solving the eigenvalue problem, and the updated set of weight parameters is obtained by performing normalization.

[0013] Preferably, the weights of each dimension are normalized proportionally by combining the feature vector results in the weight parameter update process, and the weight parameters are assigned to the corresponding dimensions of compliance urgency, value potential and liquidity according to the preset mapping relationship. At the same time, the updated weight parameters are written into the clustering calculation process to participate in the subsequent feature distance calculation and classification processing.

[0014] Preferably, a disposal process is constructed based on the reliability of the prediction results and asset classification information to complete strategy selection and output processing, with the following steps: Build a set of strategies that include multiple handling methods, and configure corresponding parameters and execution rules for each strategy; Calculate the prediction confidence level and compare it with a preset threshold to form the judgment conditions for different decision paths; Under different judgment conditions, perform manual review process or perform mapping and matching between classification labels and strategies; The matched strategy results are adjusted in terms of parameters and structured processing information is generated for output.

[0015] The carbon asset value prediction and dynamic classification management system includes a multi-source data fusion prediction module, a dynamic adaptive clustering module, and a risk perception strategy matching module. The multi-source data fusion prediction module acquires historical carbon price time series, carbon trading microstructure indicators, public opinion heat and sentiment scores, search index change rate, enterprise energy load data, and macro energy price data corresponding to carbon assets. It normalizes and aligns various types of data to construct input data sequences. By feature encoding carbon trading microstructure indicators, public opinion heat and sentiment scores, and search index change rate, it obtains market expectation factors. The input data sequences and market expectation factors are input into a hybrid neural network model for calculation, outputting the quantile prediction results and corresponding confidence intervals of carbon asset prices within the future time window. Based on the enterprise emission prediction results and carbon asset holdings, it calculates the supply and demand gap distribution. The dynamic adaptive clustering module constructs a feature vector for each carbon asset based on quantile prediction results, confidence intervals, and supply-demand gap distribution. This vector includes compliance urgency, value potential, and liquidity. After normalizing the feature vector, a weighted distance is used for clustering calculation. The weights are obtained by constructing samples from historical strategy adoption results and return feedback and solving them through linear discriminant analysis. This results in a dynamic classification of carbon assets that changes with the quantile prediction results. The risk perception strategy matching module uses the dynamic classification results and confidence intervals of carbon assets to determine the prediction confidence level. When the prediction confidence level is lower than the preset threshold, it outputs a manual review instruction. When the prediction confidence level is not lower than the preset threshold, it matches the disposal rules according to the dynamic classification results of carbon assets and adjusts the disposal parameters, and outputs the carbon asset disposal results.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a multi-dimensional input feature system by introducing multi-source market expectation factors such as high-frequency market microstructure data, public opinion intensity, and search index change rates. This system can indirectly and promptly capture the impact of policy releases and emergencies on carbon asset prices. Compared to methods that directly rely on text semantic modeling, it effectively avoids the problems of semantic parsing lag and sample sparsity. Furthermore, by combining quantile regression output with a probability integral transformation calibration mechanism, the prediction results not only possess point estimation capabilities but also stable and reliable interval prediction capabilities. This makes the confidence interval coverage closer to the theoretical nominal value and reduces prediction errors to a certain extent.

[0017] This invention introduces a weight adaptive update mechanism based on linear discriminant analysis in the asset classification process. This allows the weights of each dimension to be dynamically adjusted based on historical return feedback during clustering, thereby making the classification results more closely aligned with actual return targets. By uniformly normalizing compliance urgency, value potential, and liquidity, and combining this with dynamic quantile thresholds to generate classification labels, the influence of different units of measurement is effectively eliminated. This ensures that feature weights can truly reflect changes in corporate preferences at different stages, thereby improving the clustering purity and identification accuracy of high-value assets in the classification process.

[0018] This invention introduces prediction confidence as a decision control variable to construct a risk-aware strategy matching mechanism. When prediction uncertainty is high, it automatically triggers a manual review process, effectively avoiding misjudgments in high-risk situations in automated decision-making. Simultaneously, it dynamically adjusts strategy parameters when the confidence level meets the criteria, enabling flexible control of the decision-making process based on prediction reliability. This balances automation efficiency with decision security, reducing potential losses due to prediction bias.

[0019] This invention employs a modular design with a high degree of decoupling between functional units, allowing for independent optimization and upgrades of different components. Furthermore, the system relies on stable data sources with relatively low computational complexity, eliminating the need for high-cost natural language processing models and facilitating deployment in enterprise private cloud environments. In addition, an automated closed-loop mechanism, formed through model probability calibration and cluster weight updates, enables continuous optimization of system performance without frequent manual intervention, thereby reducing operational costs and enhancing long-term system stability. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a schematic diagram of the process of the present invention.

[0022] Figure 2 This is a flowchart of the carbon asset value prediction and dynamic classification management method of the present invention.

[0023] Figure 3 This is a schematic diagram of the carbon asset value prediction and dynamic classification management module of the present invention. Detailed Implementation

[0024] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0025] This invention provides, for example Figure 1 and Figure 2 The carbon asset value prediction and dynamic classification management method shown below includes the following specific steps: In the process of carbon asset value prediction, data is collected at fixed time intervals to ensure the consistency and comparability of input data over time. Preferably, the sampling point is the end of each day, for example, within one hour after the exchange closes, collecting market data that has fully stabilized up to that day to construct the prediction input dataset for the next trading day. This collection strategy effectively avoids future information leakage while ensuring that the data reflects the true market state, thereby improving the reliability and engineering feasibility of subsequent modeling results.

[0026] Among the collected data, one of the fundamental data sets is the historical carbon price time series, which is expressed as follows: Among them, parameters Indicates the category of carbon assets, with a value range of 100%. These correspond to different types of carbon assets; parameters Indicates a time index. ,in This represents the total number of trading days observed historically. This time series is used to describe the price changes of a specific carbon asset over historical trading days and is the core input data for subsequent time series modeling and trend analysis.

[0027] Furthermore, Indicates the first On the [number]th trading day, the [number]th The market transaction price or closing price of carbon-like assets, typically derived from publicly available trading data from carbon exchanges, reflects the combined effects of market supply and demand, policy influences, and trading behavior. By constructing a complete historical carbon price series, it is possible to capture the cyclical fluctuations, trend changes, and abnormal fluctuations in prices, providing a foundational time series characteristic for subsequent predictive models.

[0028] Because different types of carbon assets differ in price formation mechanisms, liquidity, and policy constraints, parameters are introduced. This method is used to distinguish different asset classes, which helps to achieve unified processing of multiple assets within the same modeling framework, while preserving information about class differences. In practical applications, time series can be constructed separately for each asset class, or time series from different classes can be used as multi-channel inputs to improve the model's ability to characterize cross-asset relationships.

[0029] In addition, time index The continuity of the sequence ensures the order of the data over time, which is crucial for subsequent use of sequence models (such as sliding window-based feature extraction methods). By setting a fixed-length history window, the sequence can be... This is transformed into structured input to characterize the evolution path of carbon asset prices over time.

[0030] In summary, historical carbon price time series collected at fixed time intervals... In order to clarify asset class parameters With time parameters Based on this, a data foundation is formed for carbon asset value prediction, providing key support for subsequent integration of multi-source data, construction of prediction models, and realization of dynamic classification management.

[0031] In the process of carbon asset value prediction, market expectation data is introduced as an important supplementary information to more accurately depict the dynamic changes in market sentiment and trading behavior. This type of data does not originate directly from prices themselves, but rather reflects the behavioral characteristics and expected changes of market participants within a specific timeframe by extracting the micro-structural features of transactions. By structuring these indicators, the shortcomings of a single price series in expressing market conditions can be effectively compensated for, thus providing richer input features for subsequent prediction models.

[0032] In market expectation data, microstructure indicators of carbon exchanges constitute the core content, including the intraday best bid-ask spread, the ratio of bid to ask volume at the highest bid level, and intraday price volatility, etc. These indicators together form a vector. , used to describe the first The market condition within a trading day. Among them, the intraday best bid-ask spread is used to measure the immediate cost of transactions and market liquidity, and is an important indicator reflecting the market tension. When the spread is large, it usually indicates that the market liquidity is weak or that there is a large disagreement between buyers and sellers; when the spread is small, it indicates that the market is active and has good liquidity.

[0033] The calculation method for the best intraday bid-ask spread is as follows: In the above formula, the parameters have the following meanings: Indicates the first The intraday best bid-ask spread indicator for a trading day is an averaged measure of the bid-ask spread across all points within that trading day. Indicates the first The total number of snapshots within a trading day, i.e. the number of market quote data points collected on that trading day, is usually generated by the trading system at fixed time intervals or by event triggering. Indicates the first The first trading day Each snapshot index has a value range of [number]. This is used to iterate through all market status records within the trading day; Indicates the first The first trading day At any given snapshot time, the best selling price in the market (i.e., the best bid price) represents the lowest price at which the market is willing to sell the carbon asset. Indicates the first The first trading day At any given snapshot time, the best bid price in the market (i.e., the highest bid price) represents the highest price at which the market is willing to buy the carbon asset.

[0034] From the perspective of the formula structure, the molecular part This represents the absolute price difference at a given moment, while the denominator is... The price difference is then normalized to transform it into a relative price difference, thereby eliminating the dimensional impact of different price ranges and making price differences between different time periods or different assets comparable. By averaging all snapshot data, the overall market liquidity level and transaction cost characteristics for that trading day can be obtained.

[0035] This indicator plays a crucial role in market expectation modeling. On one hand, changes in the spread reflect market participants' uncertainty about future price movements; when market expectations are inconsistent, the divergence between buyer and seller quotes widens, leading to a widening spread. On the other hand, the spread is also closely related to market depth; in situations of insufficient liquidity, even small-scale transactions can trigger significant price fluctuations. Therefore, [the following text is incomplete and requires further context: "…"] Incorporating feature vectors It can provide important information about market micro-behavior for predictive models.

[0036] By including By unifying the modeling of multiple microstructural indicators, including those mentioned above, a feature vector reflecting the market's expected state can be constructed. This vector maintains consistency with other data along the time dimension, thus achieving effective fusion of multi-source data. This indirect representation method based on microstructure has advantages in stability and engineerability compared to direct semantic analysis of policy texts. At the same time, it can capture changes in market sentiment and trading behavior characteristics in a timely manner without introducing future information, providing key support for subsequent carbon asset value prediction.

[0037] In the process of carbon asset value prediction, public opinion intensity and sentiment scores are used to characterize the level of attention and emotional inclination of market participants towards carbon assets from the perspective of information dissemination and market perception. This type of information can reflect the impact of external factors such as policy releases and unexpected events on market expectations. Compared with directly performing semantic analysis of policy texts, it has advantages such as stable data acquisition, timely response, and lower engineering difficulty. By performing structured modeling on public opinion data, key features reflecting market sentiment fluctuations can be constructed, providing an important supplement for subsequent predictions.

[0038] Public opinion heat ( The calculation method for ) is as follows: In this formula, the parameters have the following meanings: Indicates the first The public opinion heat of a trading day is used to measure the strength of market attention on that day and its deviation from historical trends. Indicates the first This collection comprises relevant articles gathered within a single trading day, including all news, commentary, and analytical articles related to carbon assets. Represents a set A specific article index within; Article The authority of the media source is assigned a weight, with different weights assigned to different sources. For example, official channels are assigned a higher weight (e.g., 1.5), industry media are assigned a medium weight (e.g., 1.0), and self-media are assigned a lower weight (e.g., 0.5) to distinguish the credibility of information sources. Article The weighted sum of interaction metrics, including metrics such as number of reads and number of shares, which reflect the scope of dissemination and user engagement; This represents the sum of weighted interactions across all relevant articles on that day, used to characterize the overall intensity of information dissemination. The logarithmic function is used to scale the interaction quantity to avoid extreme values ​​having too much influence on the results; This indicates that a 20-day moving average is performed on the variable within the parentheses to extract the long-term trend component.

[0039] The formula consists of two parts: the first part is the logarithmically transformed weighted total daily interaction, reflecting the immediate intensity of public opinion; the second part is the 20-day moving average of the indicator, used to depict the long-term trend. The difference between the two is used to highlight short-term abnormal changes, i.e., the pulse characteristics of public opinion intensity. This approach can effectively filter out the influence of long-term stable attention, making the abnormal increase in attention caused by sudden policy changes or major events more obvious, thus serving as an important proxy signal for changes in market expectations.

[0040] In sentiment analysis, the sentiment tendency of each article is obtained by classifying its content, and then aggregated to form an overall sentiment score. The sentiment score of a single article is calculated as follows: in: Article The emotional score, with a value range of , where 1 represents extreme optimism and -1 represents extreme pessimism; This indicates the output of the sentiment analysis model. The probability of belonging to the positive / positive category, which is usually calculated by a pre-trained model.

[0041] Through the above linear transformation, the original location The probability of an interval is mapped to The intervals allow for symmetry in emotional expression, facilitating subsequent weighted averaging and fusion.

[0042] After obtaining the sentiment score for each individual article, all related articles are weighted and aggregated to obtain the overall sentiment trend. The calculation method is as follows: The parameters are defined as follows: Indicates the first The overall sentiment score for each trading day is used to characterize the overall market sentiment. Article The correlation coefficient between the article content and the carbon asset price trend can be determined by thematic models or other methods to measure the actual relevance of the article content to the carbon asset market, or a default value can be used when precise calculation is lacking. A very small positive number is introduced to prevent the denominator from being zero, in order to ensure the numerical stability of the calculation; the numerator part This represents a relevance-weighted sum of the sentiment scores for all articles; the denominator is... This indicates the normalization process for the sum of weights.

[0043] This aggregation method introduces relevance weights to differentiate the importance of different articles, giving greater weight to articles highly correlated with carbon asset price changes in the sentiment calculation, thereby reducing the interference of irrelevant or weakly relevant information on the overall sentiment results. The final result... It can accurately reflect the overall direction and intensity of market sentiment on that day.

[0044] By controlling public opinion heat With Emotional Score By jointly introducing the model, we can characterize changes in market expectations from two dimensions: attention and sentiment, providing more comprehensive information support for carbon asset value prediction.

[0045] In the process of carbon asset value prediction, to characterize changes in public and market participants' attention to carbon assets, the search index change rate is introduced as an important market expectation feature. Search behavior typically responds quickly to policy announcements or market fluctuations; therefore, modeling changes in search volume can reflect marginal changes in market sentiment from a side perspective, exhibiting stronger foresight and sensitivity compared to price data.

[0046] First, we define the search index time series, which is expressed as follows: The meanings of each parameter are as follows: Indicates the first The search index value at each point in time is used to reflect the public's level of attention to topics related to carbon assets on that day. Indicates the first Search volume statistics for a given set of keywords at a specific time point; carbon trading, carbon quotas, CCER, compliance: a set of keywords used to extract search data. These keywords cover the core trading objects and compliance behaviors in the carbon market, and can comprehensively reflect the market's focus; subscript This indicates a time index, corresponding to a specific transaction day or statistical period.

[0047] Based on the above definition, a search index sequence that changes over time can be obtained, which can be used to describe the dynamic changes in market attention.

[0048] Based on this, in order to depict the changing trend of search volume, the search index change rate is introduced. The calculation method is as follows: The parameters in this formula have the following meanings: Indicates the first The rate of change of search index at each point in time is used to measure the degree of change in current search volume relative to the previous point in time. This indicates the search index at the current point in time. Indicates the search index at the previous point in time; molecule The denominator represents the absolute change in search volume. This represents the average search volume between the current and previous time points, used to normalize the changes. It is a very small positive number, used to prevent numerical instability caused by the denominator being zero or a very small value.

[0049] The rate of change is presented in a symmetrical form, using the average of the current value and the value at the previous time step as the normalization benchmark. This approach is more efficient than directly using... The traditional growth rate calculation method with the denominator has higher stability. When the search volume fluctuates drastically or the search volume was close to zero in the previous moment, the symmetric rate of change can effectively avoid the problem of excessively large or distorted values, thereby improving the robustness of the feature.

[0050] Furthermore, this rate of change directly reflects the marginal change in market attention, i.e., whether attention has increased or decreased compared to the previous moment, and the magnitude of the change. When policies are introduced, implementation deadlines approach, or abnormal market fluctuations occur, the public and market participants tend to increase the frequency of searching for relevant information, thereby leading to... A significant increase occurred, making It shows a positive surge; conversely, when market attention decreases, the indicator tends to decline or become negative.

[0051] By modeling the year-on-year change rate of the search index and combining it with symmetric normalization, we can highlight abnormal fluctuations in search behavior while ensuring numerical stability. These abnormal changes often correspond to significant turning points in market expectations, making this indicator of significant value in carbon asset valuation.

[0052] It should be noted that the search index and its rate of change are calculated based on data that has stabilized by the end of the day and do not include future information, thus ensuring that no information leakage is introduced during the prediction process. This data processing method ensures the temporal consistency of the model input and the feasibility of the prediction results, and also provides a foundation for unified modeling with other market expectation data (such as microstructure indicators and public opinion indicators).

[0053] In the process of carbon asset value prediction, to accurately depict the intrinsic relationship between corporate carbon emission behavior and energy consumption, corporate energy load data is introduced as an important input feature. The consumption of different types of energy (such as coal, natural gas, and electricity) by enterprises during their production and operation directly affects their carbon emission levels, and consequently, the demand and value changes of carbon assets. Therefore, by presenting corporate energy load data in a structured manner, key support can be provided for the analysis of carbon asset supply and demand and value prediction.

[0054] Enterprise energy load data is collected on a daily basis, and is recorded on the [day / month] [day]. The consumption of different energy types at each point in time is expressed as follows: The meanings of each parameter are as follows: Indicates the first At the 1st time point, the 1st The consumption of this type of energy, which is usually derived from statistics from the company's energy management system or production system; This represents an energy type index used to distinguish different categories of energy, such as coal, gas, and electricity. Different energy sources differ in terms of carbon emission coefficients and cost structures. This indicates a time index, corresponding to a specific statistical period, generally in calendar days, used to reflect the time-series changes in a company's energy consumption behavior.

[0055] By uniformly collecting data on the consumption of various energy sources, an energy consumption characteristic vector of an enterprise at a specific point in time can be constructed, and its expression is as follows: in, Indicates time At any moment, by multiple This constitutes a multi-dimensional vector that comprehensively reflects the enterprise's overall energy consumption structure and load level at a given point in time. Each dimension of the vector corresponds to the consumption of different energy types, thus preserving information on the differences in energy structure.

[0056] Enterprise energy consumption feature vector Forming a continuous sequence over time can be used to characterize the changing trends of a company's production activities. For example, when a company's production load increases, multiple energy dimensions... Energy consumption often increases in tandem with carbon assets, reflecting a potential rise in future carbon emissions; conversely, when businesses enter a low-load operation phase and energy consumption data decreases, it may correspond to a reduction in carbon emission demand. This trend provides an important basis for subsequent analysis of the carbon asset supply and demand gap.

[0057] Furthermore, different energy types correspond to different carbon emission coefficients, therefore, in subsequent analysis, this can be addressed by... By weighting energy consumption data and mapping it to potential carbon emission levels, a modeling of the relationship between energy consumption and carbon asset demand is achieved. Even without explicitly introducing emission calculation formulas, the data retains... The multidimensional structure also helps the model learn autonomously the impact of different energy sources on carbon emissions.

[0058] By building Based on the energy consumption feature vector It can introduce real production behavior data from the enterprise side, complementing market data, so that carbon asset value prediction not only relies on market prices and sentiment information, but also takes into account the operating status of the real economy, thereby improving the accuracy and stability of the overall prediction results.

[0059] In the process of carbon asset value prediction, macroeconomic and energy price data are introduced as key input variables to comprehensively characterize the impact of the external macroeconomic environment on carbon asset prices. Energy prices, as a significant component of industrial production costs, are closely linked to carbon emission behavior. Changes in different energy prices affect companies' energy structure choices, thereby indirectly influencing carbon emission levels and the demand for carbon assets. Therefore, incorporating coal, gas, and electricity prices into a unified modeling framework helps enhance the ability to characterize the drivers of carbon asset price fluctuations.

[0060] Macroeconomic energy price data is collected over time and is expressed in the following format: The meanings of each parameter are as follows: Indicates the first Coal prices at a specific point in time; Indicates the first Natural gas prices at a specific point in time; Indicates the first Electricity prices at a specific point in time; This indicates a time index, typically in units of trading days.

[0061] The above-mentioned energy prices can be further combined and represented as a macroeconomic energy price vector, denoted as: in, Indicates time A vector composed of multiple energy prices at any given time is used to uniformly express the state of macroeconomic energy prices.

[0062] To ensure consistency in numerical scale across data from different sources, all numerical sequences are standardized according to their respective historical intervals. The Z-score normalization method is used to map data of different dimensions to a unified scale range. At the same time, data of different frequencies are resampled to unify them into daily data in the time dimension, thereby achieving alignment and fusion of multi-source data.

[0063] After data preprocessing, multiple features are integrated to construct the first... The input tensor at each time point is expressed in the following form: The meanings of each parameter are as follows: Indicates time The input tensor at time step 1 is used for subsequent model calculations; These represent different types of carbon assets over time. The price; This represents the feature vector composed of carbon trading microstructure indicators; Indicates the level of public opinion intensity; Indicates the emotional score; Indicates the rate of change in search index; Represents the energy consumption characteristic vector of an enterprise; This represents the macroeconomic energy price vector.

[0064] Building upon this, the market expectation-related features are further integrated. First, market microstructure indicators, public opinion heat, sentiment scores, and search index change rates are concatenated to construct a market expectation feature vector, expressed as follows: in, Indicates time The set of market expectation characteristics at any given moment.

[0065] Subsequently, a two-layer fully connected network is used to compress the feature vector to obtain a low-dimensional market expectation factor, which is calculated as follows: In this formula, the parameters have the following meanings: Indicates time The market expectation factor at any given time is represented by a low-dimensional vector. This represents the characteristic vector of market expectations; These represent the weight matrices of the first and second fully connected layers, respectively. These represent the bias terms of the corresponding layers; This represents the linear rectification activation function, used to introduce nonlinear characteristics; in, This indicates that the vector is a 16-dimensional low-dimensional representation, used to characterize the market's expected direction and strength regarding the future price of carbon assets at the current point in time.

[0066] This encoding process compresses high-dimensional market expectation features into a low-dimensional representation through nonlinear mapping, which helps reduce redundant information and improve feature expression efficiency. Simultaneously, this low-dimensional representation can serve as an important input to subsequent prediction models, participating in model training along with price time series and other features. The entire encoding network is jointly optimized with the downstream prediction process, using price prediction error as the optimization objective, thereby achieving effective learning and expression of market expectation information without the need for manual annotation.

[0067] In the process of carbon asset value prediction, to achieve an accurate depiction of future price distribution, a hybrid neural network structure is used to model multi-source features. By fusing convolutional neural networks and long short-term memory networks, joint extraction of local features and temporal dependencies is achieved. In this process, features with a length of... Using historical time windows as input, historical data is mapped to future time windows with a length of [missing information]. The prediction interval is obtained, thus yielding the probability distribution expression of carbon asset prices.

[0068] First, in the local feature extraction stage, a one-dimensional convolution operation is performed on the input window data, and its calculation form is as follows: The meanings of each parameter are as follows: The feature representation of the output of the convolutional layer is used to describe the change pattern of the input sequence within a local time range; This represents the convolution kernel weight parameters, used for feature extraction from the input data; This indicates that the input data sequence is derived from the history window. It contains multidimensional features across multiple time steps; Indicates the convolutional layer bias term; This represents the convolution operation; This represents the linear rectified activation function, used to introduce nonlinear expressive power.

[0069] Through the above convolution operations, short-term fluctuation features in time series can be extracted, such as local trend changes and short-cycle fluctuation structures, thus providing high-quality input for subsequent time series modeling.

[0070] In the temporal dependency modeling stage, the feature sequence output by the convolution is input into the Long Short-Term Memory network, and its state update form is as follows: in: Indicates at time step The hidden state vector is used to record sequence history information; Indicates at time step The memory cell state is used to store long-term dependency information; Indicates the time step of the convolutional layer The output characteristics; These represent the hidden state and the memory state of the previous time step, respectively. Represents a long short-term memory network unit; Represents the time step index, with a range of to .

[0071] By recursively calculating, the state sequence within the entire time window can be obtained, and the hidden state of the last time step can be selected. As a global representation of this time window, this vector comprehensively reflects the long-term and short-term dependencies in the historical sequence.

[0072] During the integration phase, the global representation is combined with market expectation factors, and its expression is as follows: in: This represents the fused feature vector; This represents the global time-series features corresponding to the historical window; This represents a low-dimensional market expectation factor, used to characterize the direction and strength of market expectations for future prices.

[0073] This splicing operation allows for a unified expression of time-series information extracted from historical data and features constructed based on market expectations, thereby enhancing the model's ability to characterize the drivers of price changes.

[0074] In the output phase, quantile regression is used to predict the probability of future prices, and its calculation form is as follows: in: Indicates the first Carbon assets in the future The first moment Quantile predictions; Indicates the corresponding quantile The output layer weight parameters; Indicates the corresponding quantile The bias term; This represents the fused feature vector; Indicates the quantile level, with values ​​ranging from 1 to 2. These correspond to the 10th, 50th, and 90th percentiles, respectively. Indicates the prediction step size, ranging from... to .

[0075] Using the above method, multiple quantile prediction values ​​can be obtained simultaneously, among which... The corresponding median forecast reflects the most likely price level, while and These respectively constitute the boundaries of the prediction interval.

[0076] During model training, the quantile loss function is used for optimization, and its expression is as follows: in: Represents the overall loss function; Indicates the length of the prediction period; Represents the quantile loss function; Indicates the true price; This indicates the predicted quantile value.

[0077] The quantile loss function is defined as follows: in: Indicates prediction error; Indicates the quantile level.

[0078] This loss function, by weighting different quantile errors, enables the model to learn complete price distribution information, rather than just a single predicted value. In this way, predictions incorporating uncertainty information can be obtained, thus providing more reliable data support for subsequent risk assessment and decision-making.

[0079] In the process of carbon asset valuation, in addition to price forecasting, it is also necessary to characterize the future supply and demand relationship to further support risk assessment and resource allocation in the decision-making process. Therefore, based on the price forecast results, supply and demand gap forecasting and uncertainty propagation mechanisms are introduced to describe the carbon asset supply and demand balance of enterprises in the future time frame.

[0080] Based on a company's historical emission trends and future production plans, it is possible to predict future emissions. The daily emissions are estimated and expressed as follows: , indicating the future The projected emissions for the day are calculated by taking into account historical emissions data and future production plans, and are used to reflect the company's future carbon emission needs. This indicates the predicted time step, corresponding to a specific future time point.

[0081] Based on this, we introduce the current carbon asset holdings, which are expressed as follows: This represents the total amount of carbon assets currently held by a company, including carbon allowances, CCERs, and green certificates. This indicator is used to measure the carbon asset reserves that a company currently has available for compliance or trading.

[0082] The supply-demand gap can be obtained by calculating the difference between emission demand and current holdings, and it can be expressed as follows: in, Indicating the future The carbon asset supply and demand gap is calculated daily. When it is positive, it indicates that future emission demand exceeds current holdings, resulting in a carbon asset shortage. When it is negative, it indicates that holdings exceed emission demand, resulting in surplus assets that can be traded or stored.

[0083] To account for the inherent errors in emission prediction models and characterize prediction uncertainty, the emission prediction residuals are subjected to multiple random sampling processes. The Monte Carlo method is then used to propagate the uncertainty of the supply-demand gap. Specifically, the residuals of the emission prediction model are sampled 1000 times to construct multiple possible emission scenarios, and the corresponding calculations are performed for each scenario. Thus, the empirical distribution of the supply and demand gap is obtained.

[0084] Based on this empirical distribution, key statistics can be extracted, including the median and its range. The median of the supply-demand gap is expressed as: in, It represents the 50th percentile of the supply-demand gap distribution, used to reflect the level of the most likely supply-demand gap.

[0085] Meanwhile, to characterize the range of uncertainty surrounding the supply-demand gap, an 80% confidence interval is introduced, expressed as follows: in, and These represent the 10th and 90th percentiles of the supply-demand gap distribution, respectively, and are used to define the fluctuation range of the supply-demand gap within a relatively high probability range.

[0086] After obtaining the supply-demand gap forecast, a joint forecast of carbon asset prices and supply-demand conditions can be further generated. Regarding price forecasting, the focus is on the median and range within the quantile forecast results. The median price is expressed as: in, Indicates the first Carbon assets in the future The 50th percentile forecast at time 50 is used to represent the most likely price level. This indicates an index for carbon asset types.

[0087] The corresponding price range is represented as follows: in, and They represent the first The 10% and 90% quantile forecasts of carbon-like asset prices are used to construct an 80% confidence interval, thereby reflecting the range of uncertainty in price forecasts.

[0088] By simultaneously outputting the price forecast range and the supply-demand gap range, a multi-dimensional description of the future state of carbon assets can be formed. On the one hand, the price distribution reflects the market-level value change trend; on the other hand, the supply-demand gap distribution reflects the internal compliance pressure and resource adequacy of enterprises. Combining the two can provide more comprehensive information support for subsequent decision-making, such as determining whether it is necessary to purchase carbon assets in advance, and whether the conditions for sale or pledging are met.

[0089] In summary, by introducing supply and demand gap forecasting and uncertainty propagation mechanisms, and combining them with quantile price forecasting results, a unified expression of multidimensional information on carbon asset prices, demand, and risks can be achieved within the same framework, thereby significantly improving the practicality and decision support capabilities of the forecasting results.

[0090] In the dynamic classification of carbon assets, to achieve a refined characterization of different asset states, a three-dimensional feature vector based on predicted output is introduced. Compliance urgency is one of the core dimensions, used to measure the ability of an individual carbon asset to mitigate the overall supply gap under compliance constraints. This feature not only reflects the relationship between the number of assets and the overall supply-demand gap of enterprises, but also comprehensively considers policy restrictions and time factors, thus forming a quantitative description of the asset urgency.

[0091] For any carbon asset In time At any given moment, the urgency of compliance is defined as follows: The meanings of the parameters in the above formula are as follows: Indicates the first The compliance urgency of a carbon asset is used to measure the urgency of the asset's compliance requirements at the current moment. , which is a weighting coefficient used to adjust the contribution ratio of different factors in the calculation of compliance urgency; This represents the median absorption gap corresponding to the asset, defined as follows: This represents the median of the overall supply-demand gap for enterprises, which is the 50th percentile of the supply-demand gap distribution. Indicates the first The number of carbon assets; This indicates the offset ratio limit for this type of carbon asset, which is used to constrain the maximum proportion of such assets used in compliance. For example, certain assets (such as CCERs) are not allowed to exceed a certain proportion of emissions under policy. This indicates the maximum amount of the shortfall that the asset can be used to cover, within the limits permitted by policy. Furthermore: This indicates that the absorbable gap is subject to non-negative constraints, and the value is 0 when the gap is negative, thereby avoiding negative urgency in the case of oversupply. This indicates the current holdings of this type of carbon asset, used to normalize the gap; : is a very small positive number, used to prevent numerical instability caused by a denominator of zero or too small; This indicates the ratio of the gap to the holdings, used to characterize the relative importance of the asset in mitigating the gap; In terms of time: Indicates the remaining time until the performance deadline; This represents the total time from the current moment to the start of the performance period; This indicates the percentage of progress in fulfilling the contract. The smaller the remaining time, the larger this value, reflecting the increasing urgency over time. also: This indicates a penalty clause, used to depict the additional costs or risks that may arise when performance requirements are not met; This is used to incorporate potential risk factors into the urgency calculation, so that the urgency level increases accordingly when there is a risk of violation.

[0092] The above structure allows supply and demand gaps, asset size, policy constraints, and time pressure to be uniformly mapped to a single indicator. This allows for a quantitative description of the urgency of carbon asset compliance. This indicator reflects the actual strength of the role of different assets in the compliance system. A higher value indicates that the asset has a higher priority in meeting performance requirements at the current stage; a lower value indicates that the asset has limited effect on alleviating the current shortfall or has less performance pressure.

[0093] Overall, this feature construction method, without introducing additional assumptions, transforms existing prediction results into asset-level decision indicators by reprocessing them, thus providing a foundation for subsequent multidimensional feature modeling and dynamic classification.

[0094] In the dynamic assessment of carbon assets, value potential measures the ability of an individual asset to generate returns over a future timeframe and its associated financial attributes. It is a crucial indicator connecting price forecasts with asset decisions. This characteristic comprehensively considers expected price changes, the premium capacity of a specific asset, and financial leverage attributes, thereby quantitatively characterizing the potential for future value appreciation of an asset.

[0095] For any carbon asset Its value potential is defined as follows: The meanings of the parameters in the above formula are as follows: Indicates the first A value potential indicator for carbon assets, used to characterize the future price growth potential and added value of the asset; These are weighting coefficients used to adjust the contribution ratios of price expectations, premium attributes, and financial attributes to the overall value potential; In the price expectation section: The 50th percentile forecast, representing the median price of the asset at the end of a future time window, is the most likely future price level to be achieved, as output by the forecasting model. This indicates the current market price of the asset; This represents the relative rate of change in prices, used to measure the expected increase in future prices relative to current prices. When the value is positive, it indicates that there is room for price increases, while a negative value indicates that there is a risk of price declines. In the premium attribute section: The green certificate premium rate is expressed as follows: Indicates the price of green certificates; This represents the benchmark electricity price; this indicator is used to characterize the premium of green certificates relative to the electricity price, thereby reflecting their market acceptance and policy-driven added value; for non-green certificate type carbon assets, this item is set to 0 to avoid interference with other assets; In the financial attributes section: It indicates the discount rate for pledged financing and is used to characterize the financing capacity of assets in financial institutions; This indicator is defined as the ratio between the amount of collateralized financing available to an asset and its market valuation. The higher the value, the stronger the asset's financing capacity and financial leverage. This data typically comes from external carbon finance data interfaces and is used to quantify the availability of assets within the financial system.

[0096] The above three parts allow for a comprehensive assessment of the value of carbon assets from different dimensions. The first part reflects market price expectations and is the core source of value potential; the second part reflects the premium advantage of specific asset classes (such as green certificates) under policy and market mechanisms; and the third part introduces financial attributes, making asset valuation not only limited to transaction value but also encompassing its use value in financing scenarios.

[0097] Overall, value potential indicators Integrating price forecasts with asset structure characteristics and financial attributes ensures comparability among different carbon assets under the same evaluation system. A higher value indicates that the asset has significant potential for future appreciation or financial utilization; a lower value indicates that the asset's value growth is limited in the short term or its additional attributes are weak. This indicator can provide an important basis for subsequent asset classification and strategy selection.

[0098] In a dynamic assessment system for carbon assets, liquidity is used to characterize the trading activity and ease of transactions of an asset in the market, and is an important indicator affecting the asset's realizability and trading efficiency. By introducing micro-structural data such as turnover rate, price spread, and market depth, asset liquidity can be quantitatively described from the perspective of trading behavior, thereby providing support for subsequent classification and decision-making.

[0099] For any carbon asset Its liquidity indicator is defined as follows: The meanings of the parameters in the above formula are as follows: Indicates the first A liquidity indicator for carbon assets, used to comprehensively measure the trading activity of the asset in the market; , which is a weighting coefficient used to adjust the contribution ratio of different sub-indicators in liquidity calculation; In each component: The turnover rate is a normalized indicator that reflects the ratio between the trading volume and the holding volume of assets within a certain period of time. The higher the turnover rate, the more frequent the asset trading and the stronger the liquidity. It represents the bid-ask spread, which is the difference between the best selling price and the best buying price in the market. It is used to measure transaction costs and market liquidity. This represents the median price, typically the average of the best buying price and the best selling price, used to normalize the price difference. It represents the relative price difference, used to eliminate differences in price scales and make assets in different price ranges comparable; This is a liquidity inverse indicator. When the price spread is small, this value is large, indicating low transaction costs and good liquidity. The normalized market depth indicator typically reflects the size of buy and sell orders. The greater the depth, the stronger the market's absorption capacity and the higher its liquidity.

[0100] All the above data can be obtained in real time through the carbon exchange API, ensuring the timeliness and accuracy of the data. By normalizing the turnover rate and market depth, the differences in scale and dimensions between different assets can be eliminated, allowing all indicators to be calculated on a unified scale.

[0101] After obtaining liquidity metrics, to achieve a unified comparison among multi-dimensional features, it is necessary to normalize each dimension of the features. Specifically, this includes normalizing the compliance urgency. Value potential and liquidity The original values ​​are Min-Max normalized and mapped to an interval. Thus, the normalized feature vector is obtained, which is expressed as follows: in: Indicates the first The normalized eigenvector of a carbon asset; This indicates the urgency of compliance after normalization; This represents the value potential after normalization. This represents the normalized liquidity.

[0102] By mapping features across all dimensions to the same interval, we can avoid the unbalanced impact of different units or numerical ranges on subsequent calculations, making each dimension comparable in cluster analysis. At the same time, normalization also improves the numerical stability of the algorithm, preventing any single dimension from dominating the overall calculation results due to excessively large values.

[0103] Based on this, subsequent weighted distance calculations and dynamic threshold quantile analyses are both based on the normalized feature vectors. This process ensures that different assets are compared and classified within a unified evaluation space. In this way, a detailed characterization of carbon asset status can be achieved under a multi-dimensional feature system, providing reliable input for dynamic adaptive clustering.

[0104] In the dynamic classification of carbon assets, to achieve similarity measurement in a multi-dimensional feature space, weighted distance is introduced as the core metric to measure the degree of difference between different assets in three dimensions: compliance urgency, value potential, and liquidity. For any two carbon assets... and Its weighted distance is defined as follows: The meanings of each parameter are as follows: Assets With assets Weighted distance in the normalized feature space; Assets The normalized eigenvectors; Assets The normalized eigenvectors; Representing assets With assets The urgency of normalization compliance; These represent the normalized value potential; These represent normalized liquidity; These represent the weight coefficients for the three dimensions, used to adjust the importance of each dimension in distance calculation.

[0105] The weight parameters satisfy the following constraints: This constraint ensures that the sum of the weights of each dimension is 1, keeping the distance calculation results stable and comparable. Through weighting, the influence of different dimensions on the clustering results can be adjusted according to actual needs.

[0106] To determine the weights, an adaptive update mechanism based on historical data is introduced. By collecting data on the execution of the system's recommended strategies over a past period, including records of strategy adoption and rejection and corresponding actual returns, asset samples where the strategy was adopted and yielded positive returns are defined as positive samples, while the remaining samples are defined as negative samples. Based on this sample division, an intra-class scatter matrix is ​​constructed. and inter-class scatter matrix It is used to characterize the distribution characteristics between samples of different categories.

[0107] Based on this, we establish the generalized eigenvalue problem, which is expressed as follows: in: This represents the inter-class scatter matrix, used to measure the degree of dispersion between the centers of different classes; This represents the intra-class scatter matrix, used to measure the degree of dispersion of samples within the same class; This represents the eigenvector, used to indicate the optimal projection direction; Eigenvalues ​​are used to characterize the ability to distinguish projection directions.

[0108] By solving this eigenvalue problem, selecting the eigenvector corresponding to the largest eigenvalue, and performing normalization, the optimal weight combination can be obtained. This process is equivalent to linear discriminant analysis, which aims to maximize the distance between different categories while ensuring compactness of samples within each category, thereby improving the discriminative power of clustering.

[0109] After determining the weights, cluster analysis is performed on the normalized feature vectors based on the current weights. The K-means algorithm is used to partition the asset set, where the number of clusters is... The silhouette coefficient is automatically determined, with a value ranging from 2 to 6. During clustering, weighted distance is used as a similarity measure, ensuring that the classification results reflect the differences in importance across different dimensions.

[0110] For each cluster center, its features are represented as follows: in: Indicates the first The urgency of compliance for each cluster center; Indicates potential value; Indicates liquidity.

[0111] After clustering, quantile analysis was performed on each dimension. For the normalized values ​​of all assets in that dimension, the 70th and 30th quantiles were calculated, and expressed as follows: in: The high threshold (70th percentile) indicating the urgency of compliance. This represents the lower threshold (30th percentile).

[0112] Based on the above threshold, semantic labels can be used to divide cluster centers: when At that time, it corresponds to high urgency; when When the urgency level is low, it corresponds to low urgency; otherwise, it corresponds to medium urgency. Value potential and liquidity dimensions are categorized using the same labeling method, thus forming a multi-dimensional semantic description.

[0113] By using the aforementioned weighted distance definition, adaptive weight update, and clustering and label generation mechanism, dynamic classification of carbon assets can be achieved in a multi-dimensional feature space. This allows the classification results to be adjusted based on historical return feedback and market changes, thereby improving the accuracy and adaptability of the classification.

[0114] In the dynamic management of carbon assets, to achieve an effective mapping from forecast results to specific disposal actions, a strategy matching and execution mechanism is introduced. This mechanism uses a predefined set of strategies and decision rules to respond to different asset states. First, the strategy set is constructed, and its expression is as follows: The meanings of each element are as follows: This represents the set of strategies used to define all possible asset disposal methods; This indicates an immediate sell strategy, suitable for assets with high liquidity and low potential value. This indicates a wait-and-see strategy, suitable for situations where the value trend is unclear or highly volatile. This refers to a pledged financing strategy, which uses carbon assets to obtain working capital. This refers to a replacement optimization strategy, which involves adjusting the asset structure to improve overall profitability or compliance efficiency. This indicates a hedging strategy used to mitigate the risk of price fluctuations.

[0115] In the strategy matching process, prediction confidence is introduced as a key control variable, and its expression is as follows: , representing the prediction confidence level, is used to characterize the reliability of the prediction result. This value is usually calculated from the width of the price prediction interval or the distribution characteristics. When When the forecast is highly uncertain, a manual review mechanism is triggered to avoid potential risks associated with automated decision-making; when... In this case, strategy matching is performed based on the category label to which the asset belongs.

[0116] During strategy selection, different strategies are matched according to preset priorities, which can be set based on historical performance or business rules. Classification tags are derived from clustering results in a multi-dimensional feature space, encompassing information such as compliance urgency, value potential, and liquidity. By mapping classification tags to strategy sets, automatic transformation from asset status to disposal behavior can be achieved.

[0117] After the strategy is determined, the strategy parameters are further calculated, with key parameters (such as the staking ratio) dynamically adjusted based on the prediction confidence level. When the prediction confidence level is high, the risk-bearing capacity can be appropriately increased; when the prediction confidence level is low, the relevant parameters are discounted to control potential risks. This process achieves flexible adjustment of strategy execution, enabling the system to adapt under different uncertainty conditions.

[0118] After completing strategy selection and parameter calculation, a corresponding disposal result is generated, which is presented as a structured information set, typically including asset identification, selected strategy, execution time window, expected return, and risk warnings. This result can serve as a direct basis for subsequent execution or manual decision-making.

[0119] In terms of operational mechanism, the entire process is executed cyclically according to a fixed period. During the data acquisition phase, only data that has stabilized at the end of the day is used to ensure the reliability of the input information; subsequently, multi-source data fusion prediction, dynamic adaptive clustering, and strategy matching are performed sequentially, and the final processing results are output and pushed out. This process forms a closed-loop update mechanism in the time dimension, enabling the model to continuously respond to market changes.

[0120] Furthermore, in the periodic update mechanism, the clustering weights are adaptively updated every certain period of time (e.g., 30 days) to reflect the latest market environment and changes in strategy performance. By continuously incorporating historical feedback information, the weight allocation can be gradually optimized, making the classification and strategy matching more closely aligned with actual return performance.

[0121] Overall, through a set of strategies With prediction confidence The introduction of this technology enables a complete closed loop from forecasting results to actual operation, giving the carbon asset management process automated execution and risk control capabilities, thereby improving decision-making efficiency and robustness in a complex and ever-changing market environment.

[0122] In the construction of carbon asset value prediction models, the model training and parameter determination methods directly affect the accuracy and stability of the prediction results. Firstly, in the model training phase, a CNN-LSTM structure is used to model multi-source time series data. The training data comes from carbon price data from a carbon exchange from 2019 to 2026, along with corresponding market microstructure data, public opinion indicators, search index change rates, and simulated enterprise energy consumption data, thus forming a multi-dimensional input feature set. This dataset covers different market stages, which helps the model learn the mapping relationship between price fluctuations and external factors.

[0123] In terms of data partitioning, the samples were divided into training, validation, and test sets according to time sequence, with a ratio of 60% / 20% / 20%. This partitioning method ensures that the causal relationship of the time series is not disrupted, while avoiding future information leakage. During model training, the Adam optimizer was used for parameter updates, with a learning rate set to 0.001 and a batch size of 32. An early stopping strategy was implemented, terminating training when the validation set performance did not show significant improvement over 20 consecutive iterations to prevent overfitting.

[0124] In terms of loss function design, a three-quantile equal-weighted quantile loss function is adopted. By simultaneously optimizing the prediction results of multiple quantiles, the model can learn complete price distribution information, rather than just a single predicted value. This method can effectively characterize prediction uncertainty and provide a foundation for subsequent risk control.

[0125] In terms of model structure, the convolutional neural network is configured as a two-layer structure with kernel sizes of 3 and 5, respectively, to extract local temporal features at different scales. The long short-term memory network is also configured as a two-layer structure with 128 hidden units, used to capture long-term dependencies in the sequence. A dropout parameter of 0.3 is introduced to enhance the model's generalization ability. Furthermore, the market expectation factor encoding network adopts a fully connected structure, mapping from 64 dimensions to 16 dimensions, to compress and represent high-dimensional market features.

[0126] To ensure the reliability of probabilistic prediction results, a probabilistic calibration mechanism is introduced. This involves performing a ProbabilityIntegralTransform test on the predicted distribution on the validation set to evaluate the actual coverage of each quantile prediction. When a significant deviation between the predicted distribution and the ideal uniform distribution is detected (e.g., KL divergence greater than 0.05), an automatic calibration process is triggered. The calibration method employs isotonic regression to post-process the predicted quantiles, ensuring that the actual coverage of the predicted interval remains consistent with the nominal quantiles. The calibrated model parameters are saved and used in series during the inference phase, thereby guaranteeing the consistency and reliability of the prediction results across different stages.

[0127] Regarding the determination of weighting parameters, the weighting parameters are determined for the three dimensions of compliance urgency, value potential, and liquidity. , as well as Without historical data support, default initial values ​​can be used and manually adjusted through interactive methods; when historical data is available, different parameter combinations are evaluated through grid search and backtesting methods, and the parameter combination that can make the simulated total benefit reach the optimal value is selected, thereby achieving adaptive optimization of the weights.

[0128] In the dynamic clustering process, weight parameters The update employs linear discriminant analysis, which solves the generalized eigenvalue problem to achieve the optimization objectives of minimizing intra-class distances and maximizing inter-class distances. This process has low computational complexity, making it suitable for use in periodic update scenarios. Updates can be triggered periodically based on the latest data, ensuring that the clustering results continuously reflect market changes.

[0129] Overall, by combining standardized data partitioning strategies, reasonable model structure design, the introduction of quantile loss functions, and probabilistic calibration with adaptive parameter update mechanisms, the stability and interpretability of prediction results can be improved while ensuring model accuracy, thereby providing reliable support for carbon asset value prediction and subsequent decision-making.

[0130] When deploying a carbon asset value prediction and dynamic classification management process in a private cloud environment of an enterprise in a certain industrial park, it can be automated according to a fixed cycle to achieve closed-loop processing from data collection, predictive analysis to decision output. Preferably, 2:00 AM every day is set as the unified trigger time node to ensure that the previous trading day's and operational data have all stabilized, thereby avoiding data fluctuations from interfering with subsequent analysis.

[0131] During the data acquisition phase, different types of data are uniformly acquired and integrated through multiple source interfaces. Specifically, the closing price data of the previous day's carbon allowances, CCERs, and green certificates are obtained through the exchange's API interface, while intraday trading micro-structure data, including bid-ask spreads, order volume, and price volatility, are collected simultaneously. Carbon market-related public opinion heat indicators, sentiment scores, and search index change rates are obtained from a public opinion database to characterize market sentiment and attention changes. The previous day's energy load data is obtained from the company's internal ERP system to reflect the company's production activities and energy consumption. Simultaneously, basic energy prices such as coal, gas, and electricity prices are obtained from open market data sources. All of the above data is then uniformly processed to provide standardized input for subsequent processing.

[0132] In the predictive analysis phase, a pre-trained CNN-LSTM quantile regression model and its corresponding probability calibration module are loaded. Using historical data from the past 90 days as input, the model generates carbon asset price predictions for the next 30 days. The model output includes the median price and its corresponding 80% confidence interval, reflecting price trends and uncertainty. Based on this, emissions are estimated in conjunction with future production plans of enterprises. Monte Carlo sampling is then performed based on the residuals of the emissions prediction model to obtain the distribution of the supply-demand gap, including the median and its range. This process achieves unified modeling for price prediction and supply-demand analysis, providing more comprehensive data support for subsequent decision-making.

[0133] During the dynamic classification phase, the optimal weight parameters for the current period are loaded to construct normalized feature vectors for each carbon asset, and K-means clustering analysis is performed based on weighted distance. During clustering, the number of clusters is automatically determined based on feature distribution, and the performance of each asset type is categorized according to compliance urgency, value potential, and liquidity. Furthermore, a dynamic thresholding method is used to label the clustering results, assigning each asset type a clear semantic label to provide a basis for strategy matching.

[0134] During the strategy execution phase, a predicted confidence level is calculated for each asset, and processing is tiered based on the confidence level threshold. When the confidence level is low, a manual review process is triggered to reduce the risk of automated decision-making. When the confidence level meets the requirements, a predefined strategy is matched according to the asset's tag, and the strategy parameters are adjusted based on the confidence level. Finally, a structured disposal result is generated, including asset identification, recommended strategy, execution time window, expected returns, and risk warnings, and is pushed out through channels such as WeChat for real-time delivery of decision results.

[0135] Regarding the periodic update mechanism, the system's performance is reviewed every 30 days, collecting strategy adoption data and corresponding profit and loss figures. Based on this data, the clustering weight parameters are updated using linear discriminant analysis to gradually align with actual returns. Simultaneously, the probability calibration module of the prediction model is updated regularly to maintain consistency between the predicted distribution and actual results. Through this periodic update mechanism, continuous optimization of the model and strategies can be achieved, enabling the system to maintain stable operation and adaptability in dynamic market environments.

[0136] This invention constructs a multi-dimensional input feature system by introducing multi-source market expectation factors such as high-frequency market microstructure data, public opinion intensity, and search index change rates. This system can indirectly and promptly capture the impact of policy releases and emergencies on carbon asset prices. Compared to methods that directly rely on text semantic modeling, it effectively avoids the problems of semantic parsing lag and sample sparsity. Furthermore, by combining quantile regression output with a probability integral transformation calibration mechanism, the prediction results not only possess point estimation capabilities but also stable and reliable interval prediction capabilities. This makes the confidence interval coverage closer to the theoretical nominal value and reduces prediction errors to a certain extent.

[0137] This invention introduces a weight adaptive update mechanism based on linear discriminant analysis in the asset classification process. This allows the weights of each dimension to be dynamically adjusted based on historical return feedback during clustering, thereby making the classification results more closely aligned with actual return targets. By uniformly normalizing compliance urgency, value potential, and liquidity, and combining this with dynamic quantile thresholds to generate classification labels, the influence of different units of measurement is effectively eliminated. This ensures that feature weights can truly reflect changes in corporate preferences at different stages, thereby improving the clustering purity and identification accuracy of high-value assets in the classification process.

[0138] This invention introduces prediction confidence as a decision control variable to construct a risk-aware strategy matching mechanism. When prediction uncertainty is high, it automatically triggers a manual review process, effectively avoiding misjudgments in high-risk situations in automated decision-making. Simultaneously, it dynamically adjusts strategy parameters when the confidence level meets the criteria, enabling flexible control of the decision-making process based on prediction reliability. This balances automation efficiency with decision security, reducing potential losses due to prediction bias.

[0139] This invention employs a modular design with a high degree of decoupling between functional units, allowing for independent optimization and upgrades of different components. Furthermore, the system relies on stable data sources with relatively low computational complexity, eliminating the need for high-cost natural language processing models and facilitating deployment in enterprise private cloud environments. In addition, an automated closed-loop mechanism, formed through model probability calibration and cluster weight updates, enables continuous optimization of system performance without frequent manual intervention, thereby reducing operational costs and enhancing long-term system stability.

[0140] This invention provides, for example Figure 3 The carbon asset value prediction and dynamic classification management system shown includes a multi-source data fusion prediction module, a dynamic adaptive clustering module, and a risk perception strategy matching module. The multi-source data fusion prediction module acquires historical carbon price time series, carbon trading microstructure indicators, public opinion heat and sentiment scores, search index change rate, enterprise energy load data, and macro energy price data corresponding to carbon assets. It normalizes and aligns various types of data to construct input data sequences. By feature encoding carbon trading microstructure indicators, public opinion heat and sentiment scores, and search index change rate, it obtains market expectation factors. The input data sequences and market expectation factors are input into a hybrid neural network model for calculation, outputting the quantile prediction results and corresponding confidence intervals of carbon asset prices within the future time window. Based on the enterprise emission prediction results and carbon asset holdings, it calculates the supply and demand gap distribution. The dynamic adaptive clustering module constructs a feature vector for each carbon asset based on quantile prediction results, confidence intervals, and supply-demand gap distribution. This vector includes compliance urgency, value potential, and liquidity. After normalizing the feature vector, a weighted distance is used for clustering calculation. The weights are obtained by constructing samples from historical strategy adoption results and return feedback and solving them through linear discriminant analysis. This results in a dynamic classification of carbon assets that changes with the quantile prediction results. The risk perception strategy matching module uses the dynamic classification results and confidence intervals of carbon assets to determine the prediction confidence level. When the prediction confidence level is lower than the preset threshold, it outputs a manual review instruction. When the prediction confidence level is not lower than the preset threshold, it matches the disposal rules according to the dynamic classification results of carbon assets and adjusts the disposal parameters, and outputs the carbon asset disposal results.

[0141] The carbon asset value prediction and dynamic classification management method provided in this embodiment of the invention is implemented through the aforementioned carbon asset value prediction and dynamic classification management system. For details of the specific methods and processes of the carbon asset value prediction and dynamic classification management system, please refer to the embodiments of the aforementioned carbon asset value prediction and dynamic classification management method, which will not be repeated here.

[0142] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for carbon asset value prediction and dynamic classification management, characterized in that, Includes the following steps: This study acquires historical carbon price time series, carbon trading microstructure indicators, public opinion heat and sentiment scores, search index change rate, enterprise energy load data, and macro energy price data corresponding to carbon assets. It normalizes and aligns these data types to construct an input data sequence. Market expectation factors are obtained by feature encoding the carbon trading microstructure indicators, public opinion heat and sentiment scores, and search index change rate. The input data sequence and market expectation factors are then fed into a hybrid neural network model for calculation. The model outputs the quantile prediction results and corresponding confidence intervals for carbon asset prices within future time windows. Finally, the supply-demand gap distribution is calculated based on enterprise emission predictions and carbon asset holdings. Based on quantile prediction results, confidence intervals, and supply-demand gap distribution, a feature vector containing compliance urgency, value potential, and liquidity is constructed for each carbon asset. After normalization of the feature vector, a weighted distance is used for clustering calculation. The weights are obtained by constructing samples from historical strategy adoption results and return feedback and solving them through linear discriminant analysis, thereby obtaining dynamic classification results of carbon assets that change with quantile prediction results. The prediction confidence level is determined by using the dynamic classification results and confidence intervals of carbon assets. When the prediction confidence level is lower than the preset threshold, a manual review instruction is output. When the prediction confidence level is not lower than the preset threshold, the disposal rules are matched according to the dynamic classification results of carbon assets and the disposal parameters are adjusted, and the carbon asset disposal results are output.

2. The carbon asset value prediction and dynamic classification management method according to claim 1, characterized in that, By introducing market behavior characteristics to enhance the informational capacity of price forecast inputs and constructing a unified form for representing market expectations, the steps are as follows: Collect micro-structure indicators of carbon trading, public opinion heat, sentiment scores and search index change rates, perform timestamp alignment, outlier removal and missing value imputation on various types of data, and form an original feature set according to a unified time granularity; The data from different sources in the original feature set are standardized, each feature is mapped to a unified numerical range, and vector concatenation is performed according to a preset dimension order to generate the market expectation feature vector. The market expectation feature vector is input into an encoding network containing a multi-layer linear transformation structure, and matrix multiplication, bias superposition and nonlinear activation are performed sequentially to obtain a low-dimensional compressed representation vector. The low-dimensional representation vector is combined with historical price time series data in terms of feature dimensions to construct a unified input tensor, which then participates in the subsequent price quantile prediction calculation process.

3. The carbon asset value prediction and dynamic classification management method according to claim 2, characterized in that, To construct a multi-quantile output structure based on the representation of the future price distribution of carbon assets, and to generate interval-based prediction results, the steps are as follows: Construct a prediction structure containing multiple parallel output nodes, each node corresponding to a preset quantile, and assign an independent set of parameters to each node; A linear transformation is performed on the fused features, and weight matrix multiplication and bias superposition are performed for different quantiles to obtain the corresponding quantile prediction values; The output results of multiple quantiles are organized and arranged in the order of time steps to form a multidimensional prediction sequence covering future time windows; The median forecast is extracted as a price reference sequence, and the upper and lower quantile results are combined to generate a continuous interval representation for subsequent processing.

4. The carbon asset value prediction and dynamic classification management method according to claim 3, characterized in that, The supply and demand relationship is characterized by combining changes in emissions with asset holdings, and then extended to express the uncertainty characteristics in an interval form. The steps are as follows: Obtain emission forecasts for future timeframes and current carbon asset holdings data, and perform unified and time-aligned processing on data from different sources; For each time step, the difference between emissions and holdings is calculated to form a continuous supply and demand gap sequence, which is then numerically standardized. A random perturbation set is constructed for the emission prediction residuals, multiple emission paths are generated through repeated sampling, and the corresponding gap result is calculated for each path. Statistical analysis was performed on all gap samples to extract the median and quantile intervals, forming a distribution representation of the supply and demand gap.

5. The carbon asset value prediction and dynamic classification management method according to claim 4, characterized in that, A carbon asset urgency index is constructed based on compliance pressure and time factors to characterize the role of different assets in the compliance process. The steps are as follows: Read the median of the supply-demand gap, the quantity of individual assets, and the offset ratio parameters to calculate the maximum scale of assets that can be used to fulfill obligations; The calculation results are subjected to non-negativity constraints and proportional normalization to obtain the gap ratio and limit its range. Obtain information on the total duration and remaining time of the contract fulfillment cycle, and perform proportional conversion on the time variable to obtain the time progress indicator; The gap ratio, time schedule, and penalty items are linearly weighted and combined to form a unified numerical expression of compliance urgency.

6. The carbon asset value prediction and dynamic classification management method according to claim 5, characterized in that, By combining price trends and asset-related attributes to construct a method for expressing return potential, and to characterize the future value changes of different carbon assets, the steps are as follows: Obtain the median price at the end of the forecast time window and the current market price data, and perform alignment and consistency checks on the price series; The difference between the predicted price and the current price is calculated and the ratio is normalized to obtain a series of price change rate indicators. Obtain green certificate prices and benchmark electricity price data and calculate the premium rate, while also obtaining information on the discount rate for pledged financing corresponding to the assets; A weighted combination calculation is performed on the price change rate, premium rate, and financing discount rate to form a unified value potential indicator output.

7. The carbon asset value prediction and dynamic classification management method according to claim 6, characterized in that, The feature weights in the classification process are dynamically adjusted based on historical feedback information to reflect the influence of different dimensions on the results. The steps are as follows: Collect historical strategy execution records and corresponding profit results, filter and process the data, and form a labeled sample set; The samples are classified according to the profit results, and positive and negative sample sets are generated and stored in a structured manner. The intra-class scatter matrix and inter-class scatter matrix are calculated based on the sample set, and the distribution of samples of different categories is expressed in a matrix form. The corresponding eigenvectors are obtained by solving the eigenvalue problem, and the updated set of weight parameters is obtained by performing normalization.

8. The carbon asset value prediction and dynamic classification management method according to claim 6, characterized in that, Based on the feature vector results during the weight parameter update process, the weights of each dimension are normalized proportionally, and the weight parameters are assigned to the corresponding dimensions of compliance urgency, value potential, and liquidity according to the preset mapping relationship. At the same time, the updated weight parameters are written into the clustering calculation process to participate in subsequent feature distance calculation and classification processing.

9. The carbon asset value prediction and dynamic classification management method according to claim 6, characterized in that, A disposal process is constructed based on the reliability of the prediction results and asset classification information to complete strategy selection and output processing. The steps are as follows: Build a set of strategies that include multiple handling methods, and configure corresponding parameters and execution rules for each strategy; Calculate the prediction confidence level and compare it with a preset threshold to form the judgment conditions for different decision paths; Under different judgment conditions, perform manual review processes or perform mapping and matching between classification labels and strategies; The matched strategy results are adjusted in terms of parameters and structured processing information is generated for output.

10. A carbon asset value prediction and dynamic classification management system, used to implement the carbon asset value prediction and dynamic classification management method according to any one of claims 1-9, characterized in that, It includes a multi-source data fusion prediction module, a dynamic adaptive clustering module, and a risk perception strategy matching module: The multi-source data fusion prediction module acquires historical carbon price time series, carbon trading microstructure indicators, public opinion heat and sentiment scores, search index change rate, enterprise energy load data, and macro energy price data corresponding to carbon assets. It normalizes and aligns various types of data to construct input data sequences. By feature encoding carbon trading microstructure indicators, public opinion heat and sentiment scores, and search index change rate, it obtains market expectation factors. The input data sequences and market expectation factors are input into a hybrid neural network model for calculation, outputting the quantile prediction results and corresponding confidence intervals of carbon asset prices within the future time window. Based on the enterprise emission prediction results and carbon asset holdings, it calculates the supply and demand gap distribution. The dynamic adaptive clustering module constructs a feature vector for each carbon asset based on quantile prediction results, confidence intervals, and supply-demand gap distribution. This vector includes compliance urgency, value potential, and liquidity. After normalizing the feature vector, a weighted distance is used for clustering calculation. The weights are obtained by constructing samples from historical strategy adoption results and return feedback and solving them through linear discriminant analysis. This results in a dynamic classification of carbon assets that changes with the quantile prediction results. The risk perception strategy matching module uses the dynamic classification results and confidence intervals of carbon assets to determine the prediction confidence level. When the prediction confidence level is lower than the preset threshold, it outputs a manual review instruction. When the prediction confidence level is not lower than the preset threshold, it matches the disposal rules according to the dynamic classification results of carbon assets and adjusts the disposal parameters, and outputs the carbon asset disposal results.

Images