Enterprise carbon quota gap prediction and rolling optimization method

Abstract

The invention provides an enterprise carbon quota gap prediction and rolling optimization method, which comprises the following steps: collecting enterprise historical data from an enterprise internal management system, obtaining public data from an external data source, and carrying out standardization processing on the collected data; based on the standardized data, initial time series data are constructed, static features and interactive features are extracted and generated, and time series data are obtained through integration; processing the time sequence data by using the embedded codes and the position codes; using a dynamic time fusion Transform model to predict the carbon emission and the carbon quota distribution amount of the enterprise in a plurality of time periods in the future based on the time sequence data after the embedded codes and the position codes; and based on the prediction results of the carbon emission and the carbon quota distribution amount, carbon quota gap rolling prediction is carried out by using a reinforcement learning algorithm, and a carbon emission management optimization scheme is formed. The method can achieve the dynamic adjustment and prediction of the future carbon quota gap demand of the enterprise, and provides an intelligent decision basis for the carbon emission management of the enterprise.

Description

Technical Field

[0001] The present invention relates to the field of carbon emission management, and particularly to a method for predicting and rolling optimization of enterprise carbon quota gaps based on dynamic TFT and reinforcement learning, which is used for the dynamic management and optimization of enterprise carbon emissions and carbon quotas. Background Art

[0002] As an important means to achieve carbon emission reduction targets, the carbon quota trading mechanism has been widely applied globally. This mechanism assigns a certain carbon emission limit to enterprises, enabling them to both reduce carbon emissions by improving energy efficiency during production and operation and obtain additional benefits by trading unused quotas. However, enterprises face multiple challenges in effectively managing carbon quotas and accurately predicting future carbon emissions in practice.

[0003] Existing carbon quota management methods usually rely on static prediction models. These models mostly conduct carbon emission predictions based on the historical carbon emission data and carbon quota allocation data of enterprises, combined with simple linear regression or empirical rules. However, with the changes in carbon market policies and the complexity of enterprise production environments, static models relying solely on historical data are difficult to cope with the uncertainty problems in practical applications. In addition, existing models are difficult to fully consider the long-term impacts of multi-dimensional factors such as enterprise internal production plans, technology R & D investments, and external policies and market changes on carbon emissions, resulting in obvious limitations in the management and optimization of carbon quotas.

[0004] Specifically, when formulating a carbon emission management plan, enterprises not only need to consider the changing trends of historical carbon emission data but also need to combine external factors such as carbon market fluctuations and policy adjustments. However, due to the dynamic nature of enterprise carbon emissions and the complexity of time series, existing static models have obvious deficiencies in handling long-term predictions and dynamic optimizations, resulting in enterprises facing the risks of expanding carbon quota gaps or exceeding carbon emissions in actual operations. In addition, existing models lack an effective prediction and rolling adjustment mechanism for carbon quota gaps, making it difficult for enterprises to timely adjust production plans and carbon trading strategies and unable to achieve long-term optimized management of carbon emissions.

[0005] With the rapid development of data analysis and artificial intelligence technologies, dynamic prediction and optimization methods have gradually become important means to solve carbon emission management problems. Through deep learning models, the time-dependent features in historical carbon emission data can be fully mined, and combined with external data, dynamic predictions of future carbon emissions and carbon quota gaps can be made. However, when current deep learning models are applied to carbon emission predictions, there are still some technical challenges, such as the integration and processing of multi-dimensional data, the extraction and fusion of time series features, and the adaptive adjustment of models. In particular, the research on rolling optimization of carbon quota gaps within multiple time periods is still immature and difficult to meet the actual needs of enterprises in complex carbon emission management. Summary of the Invention

[0006] Based on the above background, the purpose of the present invention is to provide an enterprise carbon quota gap prediction and rolling optimization method based on dynamic TFT and reinforcement learning. By using the enterprise's historical carbon emission data and carbon quota allocation data, and combining with external macroeconomic and policy data, a dynamic carbon emission prediction model is established. Through the combination of time series analysis, embedding coding and multi-head attention mechanism, the model can accurately predict the enterprise's carbon emissions and carbon quota allocations in multiple future time periods, and conduct rolling prediction of the carbon quota gap through reinforcement learning, optimize the enterprise's production and R & D strategies, and achieve continuous improvement and optimization of carbon emission management.

[0007] The present invention realizes the above technical purpose based on the following technical solutions:

[0008] An enterprise carbon quota gap prediction and rolling optimization method includes the following steps:

[0009] S1. Collect the enterprise's historical data from the enterprise internal management system, and at the same time obtain public data from external data sources, and standardize the collected data;

[0010] S2. Based on the standardized data, construct initial time series data, extract and generate static features and interaction features, and integrate them to obtain time series data; process the time series data using embedding coding and positional coding;

[0011] S3. Use the dynamic time fusion Transformer model to predict the enterprise's carbon emissions and carbon quota allocations in several future time periods based on the time series data after embedding coding and positional coding;

[0012] S4. Based on the prediction results of carbon emissions and carbon quota allocations, use the reinforcement learning algorithm to conduct rolling prediction of the carbon quota gap, and form an optimization plan for carbon emission management.

[0013] Further, step S1 specifically includes the following steps:

[0014] Export the enterprise's carbon emission data, production and operation data, and technology R & D investment data from the enterprise internal management system, and store these data in a multi-dimensional matrix structure; through unified data format processing, align each dimension of data on the same time axis;

[0015] Obtain external macroeconomic data and carbon trading market data from the public database through the API interface, map and merge the external data with the enterprise internal data;

[0016] For continuous data, linear interpolation is used to fill in missing values, and for discrete data, zero filling is used; unreasonable data points are identified and removed through anomaly detection algorithms;

[0017] All data is normalized, converting data in different dimensions to the same numerical range and converting it to a time series format to obtain a standardized data set.

[0018] Further, step S2 specifically includes the following steps:

[0019] From the standardized data set, data with time series attributes is extracted, and each feature data is aligned along the unified time axis to obtain the initial time series data;

[0020] Static features in the standardized data are extracted and combined with the initial time series data to form combined data;

[0021] The combined data of multiple data dimensions is processed interactively to generate interaction features and integrated into the initial time series data to obtain time series data;

[0022] The time series data is embedded encoded to reduce the dimensionality of high-dimensional features;

[0023] Position encoding is added to the time series data based on the time step.

[0024] Further, the interactive processing of the combined data of multiple data dimensions to generate interaction features includes the following operating steps:

[0025] Apply the product interaction function to the combined data, and its formula is:

[0026] z ij =x i ·x j

[0027] where z ij is the generated interaction feature, x i and x j are the two combined data to be interacted;

[0028] The generated interaction features are screened, and L1 regularization regression is used as the feature selection method, and its regularization formula is:

[0029]

[0030] where y i is the target output, x ij is the feature, β j is the coefficient to be optimized, and λ is the regularization parameter;

[0031] Integrate the filtered interaction features with the initial time series data into an extended input matrix, and ensure that all data is aligned along a unified time axis to form time series data suitable for the dynamic time fusion Transformer model.

[0032] Further, step S3 specifically includes the following operations:

[0033] Input the time series data after embedding encoding and positional encoding into the input layer of the dynamic time fusion Transformer model;

[0034] Split the input data according to a set time period to generate multiple time series subsets with a fixed time span;

[0035] Use the long short-term memory network layer in the dynamic time fusion Transformer model to extract time-dependent features and identify the historical dependencies of each time period;

[0036] Perform weighted processing on the data at different time steps through the multi-head attention mechanism of the dynamic time fusion Transformer model to generate a weighted feature vector for each time step;

[0037] Input the merged complete weighted feature vector into the fully connected layer to predict the enterprise carbon emissions and carbon quota allocations in multiple future time periods.

[0038] Further, the extraction of time-dependent features using the long short-term memory network layer in the dynamic time fusion Transformer model includes the following operations:

[0039] Input the split time series subsets into the LSTM layer in the dynamic time fusion Transformer model, and extract time-dependent features through the following formula:

[0040] h t = LSTM(x t , h t-1 )

[0041] where x t is the input at the current time step, h t-1 is the time-dependent feature extracted in the previous step, and h t is the current time-dependent feature, which is used to be passed to the subsequent layers of the dynamic time fusion Transformer model.

[0042] Further, the weighted processing of the data at different time steps through the multi-head attention mechanism of the dynamic time fusion Transformer model to generate a weighted feature vector for each time step includes the following operations:

[0043] Input the data of different time steps into the multi - head attention mechanism to calculate the attention weights between each time step and other time steps. The specific calculation formula is as follows:

[0044]

[0045] where \(Q\) is the query vector, \(K\) is the key vector, and \(V\) is the value vector. \(\sqrt{d_k}\) is the scaling factor, and the weighted feature vector of each time step is obtained through calculation;

[0046] Inputting the complete merged weighted feature vector into the fully - connected layer to predict the enterprise carbon emissions and carbon quota allocations in multiple future time periods includes the following operations:

[0047] For the data of each time step, use the multi - head attention mechanism for parallel processing. Each attention head independently calculates a set of attention weights, and the outputs of multiple attention heads are concatenated together through a linear transformation to generate a complete weighted feature vector;

[0048] Input the complete weighted feature vector into the fully - connected layer. After linear transformation and activation function processing, output the carbon emissions and carbon quota allocations in multiple future time periods:

[0049]

[0050] where \(W\) is the weight matrix of the fully - connected layer, \(z\) is the weighted feature vector, and \(b\) is the bias term. \(\hat{y}\) is the finally predicted carbon emissions and carbon quota allocations.

[0051] Furthermore, step S4 specifically includes the following operations:

[0052] Take the predicted value of the enterprise carbon emissions and the predicted value of the carbon quota allocation as the initial input of the reinforcement learning algorithm, and set the carbon quota gap of each time period as the state variable in the reinforcement learning environment;

[0053] Define the action space of the reinforcement learning algorithm, including adjusting the enterprise's production and operation strategies, R & D investment, and carbon trading operations. Each action corresponds to different carbon emission management adjustment measures;

[0054] Set the reward function according to the change trend of the carbon quota gap.

[0055] Adopt the Q - learning or policy gradient method in reinforcement learning. Based on the carbon quota gap and the corresponding reward value in each time period, continuously update the enterprise production strategy and R & D investment decision;

[0056] At the end of each time period, based on the current carbon quota gap, re-forecast the carbon emissions and quotas for the next time period, and combine the rolling forecast results to dynamically adjust the enterprise's production and operation plan and carbon emission management strategy.

[0057] Furthermore, according to the change trend of the carbon quota gap, set a reward function, and use the Q-learning or policy gradient method in reinforcement learning. Based on the carbon quota gap and the corresponding reward value within each time period, continuously update the enterprise's production strategy and R & D investment decision, including the following operations:

[0058] Set a reward function based on the difference between carbon emissions and carbon quotas:

[0059] R t =-|E t -Q t |

[0060] where R t is the reward value at time step t, E t is the predicted carbon emissions, and Q t is the carbon quota allocation;

[0061] At each time step, according to the current state S t and the action A t taken, update the Q value:

[0062]

[0063] where α is the learning rate, γ is the discount factor, and R t is the reward value; update the value of each state-action pair through Q-learning to ensure that the carbon quota gap gradually decreases;

[0064] In the policy gradient method, based on the policy π θ (a|s), update the parameter θ:

[0065]

[0066] Maximize the cumulative reward within each time period by optimizing the gradient of the policy, and continuously update the enterprise's production strategy and R & D investment decision;

[0067] After the end of each time period, according to the carbon quota gap and the corresponding reward feedback, adjust the enterprise's production plan and technology R & D investment, and apply the updated strategy to the next time period for rolling optimization.

[0068] Further, based on the carbon quota gap and corresponding reward feedback, adjust the production plan and R & D investment of the enterprise, and apply the updated strategy to the next time period for rolling optimization, including the following operations:

[0069] Construct a linear programming model based on the enterprise's historical carbon emission data and carbon quota gap, with the goal of minimizing the gap between carbon emissions and carbon quotas. The specific optimization objective function is:

[0070]

[0071] According to the predicted future carbon quotas and combined with the enterprise's historical carbon emission data, use a time series model to predict the carbon emissions in the next time period. The formula is:

[0072] E t+1 = g(E t , Q t+1 )

[0073] where E t+1 is the predicted future carbon emissions, and the function g represents the relationship between historical carbon emissions and future carbon quotas;

[0074] After each time period ends, repeat the prediction of the above carbon quotas and carbon emissions to predict the carbon emissions and carbon quotas in the next one or more time periods until the carbon emissions meet the following termination conditions:

[0075] E t ≤ E target

[0076] where E target is the national expected carbon emission target for the enterprise; when the predicted carbon emissions are lower than this target, terminate the rolling prediction;

[0077] The final output result is the carbon emissions and corresponding predicted carbon quotas for each future year before reaching the national expected carbon emission target for the enterprise.

[0078] The present invention has the following beneficial effects compared with the prior art:

[0079] The enterprise carbon quota gap prediction and rolling optimization method proposed by the present invention can accurately predict the carbon emissions and carbon quota allocation amounts in several future time periods. Through in-depth analysis of the enterprise's historical carbon emission data, carbon quota allocation amounts, and external policy and economic environment data, the method of the present invention uses a combination of time series features, static features, and interaction features, and combines embedding coding and positional coding technologies to achieve precise processing of multi-dimensional data.

[0080] The prediction results of this method can provide strong support for the future operation and carbon emission management of enterprises. Based on the predicted future carbon emissions and the carbon quota gap, enterprises can dynamically adjust their production and operation strategies and R & D investment to optimize carbon emission management. The introduction of the reinforcement learning algorithm ensures that enterprises can flexibly respond to policy and market fluctuations according to the changing trend of the carbon quota gap and conduct rolling optimization.

[0081] Through the present invention, enterprises can effectively predict future carbon emissions, reasonably manage the carbon quota gap, gradually adjust their operation strategies and carbon emission technologies in a complex policy environment, and ultimately better approach the national carbon emission target. Brief Description of the Drawings

[0082] Figure 1 is a schematic flow chart of an embodiment of the enterprise carbon quota gap prediction and rolling optimization method of the present invention. Specific Embodiment

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0084] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs.

[0085] See Figure 1 , the enterprise carbon quota gap prediction and rolling optimization method shown in the embodiments of the present invention includes the following steps:

[0086] S1. Collect the enterprise historical data from the enterprise internal management system, and at the same time obtain the public data from external data sources, and perform standardized processing on the collected data.

[0087] As a preferred embodiment, step S1 specifically includes:

[0088] S11. Export the enterprise carbon emission data, production and operation data, and technology R & D investment data from the enterprise internal management system, and store these data in a multi-dimensional matrix structure; through unified data format processing, ensure that each dimension of data is aligned on the same time axis;

[0089] S12. Obtain the external macroeconomic data and carbon trading market data from the public database through the API interface, map and merge the external data with the enterprise internal data;

[0090] S13. For continuous data, use linear interpolation to complete missing values, and for discrete data, use zero filling; identify and remove unreasonable data points through anomaly detection algorithms to ensure the integrity and rationality of the data set;

[0091] S14. Normalize all data, convert data in different dimensions to the same numerical range, and convert it into a time series format to obtain a standardized data set.

[0092] S2. Based on the standardized data, construct initial time series data, extract and generate static features and interaction features, and integrate them to obtain time series data; use embedding encoding and positional encoding to process the time series data.

[0093] As a preferred implementation, step S2 specifically includes:

[0094] S21. Extract data with time series attributes from the standardized data set, align each feature data according to a unified time axis, and obtain initial time series data;

[0095] S22. Extract static features from the standardized data and combine them with the initial time series data to form combined data;

[0096] S23. Perform interaction processing on the combined data of multiple data dimensions, generate interaction features, and integrate them into the initial time series data to obtain time series data, specifically including:

[0097] Apply the product interaction function to the feature combination, and its formula is:

[0098] z ij =x i ·x j

[0099] where z ij is the generated interaction feature, and x i and x j are two combined data to be interacted;

[0100] Screen the generated interaction features, and use L1 regularization regression (Lasso) as the feature selection method. Its regularization formula is:

[0101]

[0102] where y i is the target output, x ij is the feature, β j is the coefficient to be optimized, and λ is the regularization parameter;

[0103] Integrate the filtered interaction features with the initial time series data into an extended input matrix, and ensure that all data is aligned along a unified time axis to form time series data suitable for the dynamic time fusion Transformer model.

[0104] S24. Perform embedding encoding on the time series data to reduce the dimensionality of high-dimensional features, specifically including:

[0105] Input the high-dimensional time series features into the embedding encoding layer and map them through the embedding matrix:

[0106] e i =W embed ·x i

[0107] where e i is the embedded feature vector, W embed is the embedding matrix, and x i is the original high-dimensional time series feature.

[0108] S25. Add positional encoding to the time series data based on time steps, specifically including the following operation steps:

[0109] Add positional encoding to the data at each time step, and use sine and cosine functions to generate positional encoding, specifically:

[0110]

[0111] where pos is the time step, i is the dimension index, and d is the feature dimension;

[0112] Add the embedding encoding and the positional encoding to form the final time series data:

[0113] z i =e i +PE i 。

[0114] S3. Use the dynamic time fusion Transformer model to predict the carbon emissions and carbon quota allocations of an enterprise within a number of future time periods based on the time series data after embedding encoding and positional encoding.

[0115] As a preferred implementation, step S3 specifically includes:

[0116] S31. Input the time series data after embedding encoding and positional encoding into the input layer of the dynamic time fusion Transformer model;

[0117] S32. Split the input data according to a set time period to generate multiple time series subsets with a fixed time span;

[0118] S33. Use the long short - term memory network layer in the Dynamic Time Fusion Transformer model to extract time - dependent features and identify the historical dependencies of each time period, specifically including:

[0119] Input the segmented time - series subset into the LSTM layer in the Dynamic Time Fusion Transformer model, and extract time - dependent features through the following formula:

[0120] h t = LSTM(x t , h t-1 )

[0121] where x t is the input at the current time step, h t-1 is the time - dependent feature extracted in the previous step, and h t is the current time - dependent feature, which is used to be passed to the subsequent layers of the Dynamic Time Fusion Transformer model.

[0122] S34. Use the multi - head attention mechanism of the Dynamic Time Fusion Transformer model to weight - process the data at different time steps and generate the weighted feature vector for each time step; specifically including:

[0123] Input the data at different time steps into the multi - head attention mechanism, calculate the attention weights between each time step and other time steps, and the specific calculation formula is:

[0124]

[0125] where Q is the query vector, V is the key vector, V is the value vector, is the scaling factor, and the weighted feature vector for each time step is obtained through calculation.

[0126] S35. Input the complete merged weighted feature vector into the fully - connected layer to predict the enterprise carbon emissions and carbon quota allocations in multiple future time periods, specifically including:

[0127] For the data at each time step, use the multi - head attention mechanism for parallel processing. Each attention head independently calculates a set of attention weights, and the outputs of multiple attention heads are concatenated together through a linear transformation to generate the complete weighted feature vector;

[0128] Input the complete weighted feature vector into the fully - connected layer. After linear transformation and activation function processing, output the carbon emissions and carbon quota allocations in multiple future time periods:

[0129]

[0130] Among them, W is the weight matrix of the fully connected layer, z is the weighted feature vector, and b is the bias term. They are the finally predicted carbon emission and carbon quota allocation volume.

[0131] S4. Based on the prediction results of carbon emissions and carbon quota allocation volume, use the reinforcement learning algorithm to conduct rolling prediction of the carbon quota gap, and form an optimized carbon emission management plan.

[0132] As an optimal implementation plan, step S4 specifically includes:

[0133] S41. Take the predicted carbon emission value and carbon quota allocation value of the enterprise as the initial input of the reinforcement learning algorithm, and set the carbon quota gap of each time period as the state variable in the reinforcement learning environment.

[0134] S42. Define the action space of the reinforcement learning algorithm, including adjusting the production and operation strategies, R & D investment, and carbon trading operations of the enterprise. Each action corresponds to different carbon emission management adjustment measures.

[0135] S43. According to the change trend of the carbon quota gap, set the reward function, and use the Q-learning or policy gradient method in reinforcement learning. Based on the carbon quota gap and the corresponding reward value in each time period, continuously update the enterprise production strategy and R & D investment decision; specifically include:

[0136] S431. Set the reward function based on the difference between carbon emissions and carbon quota:

[0137] R t =-|E t -Q t |

[0138] Among them, R t is the reward value at time step t, E t is the predicted carbon emission, and Q t is the carbon quota allocation volume; this reward function promotes the optimization of the strategy in the direction of reducing the quota gap by minimizing the difference between carbon emissions and carbon quota.

[0139] S432. At each time step, update the Q value according to the current state S t and the action A t taken:

[0140]

[0141] Among them, α is the learning rate, γ is the discount factor, and R t is the reward value; update the value of each state-action pair through Q-learning to ensure that the carbon quota gap gradually decreases.

[0142] S433. In the policy gradient method, based on the policy π θ (a|s) update the parameter θ:

[0143]

[0144] By optimizing the gradient of the policy, maximize the cumulative reward in each time period, and continuously update the enterprise production policy and R & D investment decision;

[0145] S434. After each time period ends, according to the carbon quota gap and the corresponding reward feedback, adjust the enterprise's production plan and technology R & D investment, and apply the updated policy to the next time period for rolling optimization; specifically including:

[0146] Based on the enterprise's historical carbon emission data and carbon quota gap, construct a linear programming model, with the goal of minimizing the gap between carbon emissions and carbon quota. The specific optimization objective function is:

[0147]

[0148] According to the predicted future carbon quota, combined with the enterprise's historical carbon emission data, use a time series model to predict the carbon emissions in the next time period. The formula is:

[0149] E t+1 = g(E t , θ t+1 )

[0150] where E t+1 is the predicted future carbon emissions, and the function g represents the relationship between historical carbon emissions and future carbon quota.

[0151] After each time period ends, repeat the above prediction of carbon quota and carbon emissions, predict the carbon emissions and carbon quota in the next one or more time periods until the carbon emissions meet the following termination condition:

[0152] E t ≤ E target

[0153] where E target is the national expected enterprise carbon emission target; when the predicted carbon emissions are lower than this target, terminate the rolling prediction.

[0154] The final output result is the carbon emissions and the corresponding predicted carbon quota in each future year before reaching the national expected enterprise carbon emission target.

[0155] The description of the above embodiments is only used to help understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting and rolling optimization of enterprise carbon quota gap, characterized in that It includes the following steps: S1. Collect the enterprise historical data from the enterprise internal management system, and at the same time obtain the public data from external data sources, and perform standardized processing on the collected data; S2. Based on the data after standardized processing, construct the initial time series data, extract and generate static features and interaction features, and integrate them to obtain the time series data; process the time series data using embedding encoding and positional encoding; S3. Use the dynamic time fusion Transformer model to predict the carbon emissions and carbon quota allocations of the enterprise in the next several time periods based on the time series data after embedding encoding and positional encoding; S4. Based on the prediction results of carbon emissions and carbon quota allocations, use the reinforcement learning algorithm to perform rolling prediction of the carbon quota gap, and form an optimized carbon emission management plan.

2. The enterprise carbon quota gap prediction and rolling optimization method according to claim 1, characterized in that Step S1 specifically includes the following steps: Export the enterprise carbon emission data, production and operation data, and technology R & D investment data from the enterprise internal management system, and store these data in a multi-dimensional matrix structure; through unified data format processing, align each dimension of data on the same time axis; Obtain the external macroeconomic data and carbon trading market data from the public database through the API interface, map and merge the external data with the enterprise internal data; Use the linear interpolation method to fill in the missing values for continuous data, and use the zero filling method to process discrete data; identify and remove unreasonable data points through the anomaly detection algorithm; Perform normalization processing on all data, convert the data of different dimensions into the same numerical range, and convert it into the time series format to obtain a standardized data set.

3. The enterprise carbon quota gap prediction and rolling optimization method according to claim 2, wherein Step S2 specifically includes the following steps: Extract the data with time series attributes from the standardized data set, align each feature data according to the unified time axis, and obtain the initial time series data; Extract the static features from the standardized data, combine them with the initial time series data to form combined data; Perform interaction processing on the combined data of multiple data dimensions, generate interaction features, and integrate them into the initial time series data to obtain the time series data; Perform embedding encoding on the time series data to reduce the dimension of the high-dimensional features; Add positional encoding to the time series data based on the time step.

4. The enterprise carbon quota gap prediction and rolling optimization method according to claim 3, characterized in that The interaction processing of the combined data of multiple data dimensions to generate interaction features includes the following operation steps: Apply the product interaction function to the combined data, and its formula is: z ij = x i · x j Among them, z ij is the generated interaction feature, x i and x j are two combined data to be interacted; Screen the generated interaction features, and use L1 regularization regression as the feature selection method, and its regularization formula is: where y i is the target output, x ij is the feature, β j is the coefficient to be optimized, and λ is the regularization parameter; Integrate the screened interaction features with the initial time series data into an extended input matrix, and ensure that all data is aligned according to the unified time axis to form the time series data applicable to the dynamic time fusion Transformer model.

5. The enterprise carbon quota gap prediction and rolling optimization method according to claim 4, characterized in that Step S3 specifically includes the following operations: Input the time series data after embedding encoding and positional encoding into the input layer of the dynamic time fusion Transformer model; Segment the input data according to the set time period to generate multiple time series subsets with a fixed time span; Extract time-dependent features using the long short-term memory network layer in the Dynamic Time Fusion Transformer model to identify historical dependencies for each time period; Perform weighted processing on data at different time steps through the multi-head attention mechanism of the Dynamic Time Fusion Transformer model to generate weighted feature vectors for each time step; Input the merged complete weighted feature vectors into the fully connected layer to predict the enterprise's carbon emissions and carbon quota allocations over multiple future time periods.

6. The enterprise carbon quota gap prediction and rolling optimization method according to claim 5, characterized in that The extraction of time-dependent features using the long short-term memory network layer in the Dynamic Time Fusion Transformer model includes the following operations: Input the segmented time series subsets into the LSTM layer in the Dynamic Time Fusion Transformer model to extract time-dependent features through the following formula: h t = LSTM(x t , h t-1 ) where x t is the input at the current time step, h t-1 is the time-dependent feature extracted in the previous step, and h t is the current time-dependent feature, which is used to be passed to the subsequent layers of the dynamic time fusion Transformer model.

7. The enterprise carbon quota gap prediction and rolling optimization method according to claim 5, characterized in that The weighted processing of data at different time steps through the multi-head attention mechanism of the Dynamic Time Fusion Transformer model to generate weighted feature vectors for each time step includes the following operations: Input the data at different time steps into the multi-head attention mechanism to calculate the attention weights between each time step and other time steps. The specific calculation formula is: where Q is the query vector, V is the key vector, and V is the value vector, is the scaling factor, and the weighted feature vector at each time step is obtained by calculation; The input of the merged complete weighted feature vectors into the fully connected layer to predict the enterprise's carbon emissions and carbon quota allocations over multiple future time periods includes the following operations: Perform parallel processing on the data for each time step using the multi-head attention mechanism. Each attention head independently calculates a set of attention weights, and the outputs of multiple attention heads are concatenated together through a linear transformation to generate a complete weighted feature vector; Input the complete weighted feature vectors into the fully connected layer. After linear transformation and activation function processing, output the carbon emissions and carbon quota allocations over multiple future time periods: where W is the weight matrix of the fully connected layer, z is the weighted feature vector, and b is the bias term, are the finally predicted carbon emissions and carbon quota allocations.

8. The enterprise carbon quota gap prediction and rolling optimization method according to any one of claims 1-7, characterized in that Step S4 specifically includes the following operations: Use the predicted carbon emissions value and carbon quota allocation value of the enterprise as the initial input of the reinforcement learning algorithm, and set the carbon quota gap for each time period as the state variable in the reinforcement learning environment; Define the action space of the reinforcement learning algorithm, including adjusting the enterprise's production and operation strategies, R & D investment, and carbon trading operations. Each action corresponds to different carbon emission management adjustment measures; Set the reward function according to the change trend of the carbon quota gap, Adopt the Q-learning or policy gradient method in reinforcement learning. Based on the carbon quota gap and the corresponding reward value in each time period, continuously update the enterprise's production strategy and R & D investment decision; At the end of each time period, re-predict the carbon emissions and quotas for the next time period according to the current carbon quota gap, and combine the rolling prediction results to dynamically adjust the enterprise's production and operation plan and carbon emission management strategy.

9. The enterprise carbon quota gap prediction and rolling optimization method according to claim 8, wherein The setting of the reward function according to the change trend of the carbon quota gap, the adoption of the Q-learning or policy gradient method in reinforcement learning, and the continuous update of the enterprise's production strategy and R & D investment decision based on the carbon quota gap and the corresponding reward value in each time period include the following operations: Set the reward function based on the difference between carbon emissions and carbon quotas: R t = -|E t -Q t | Among them, R t is the reward value at time step t, E t is the predicted carbon emissions, Q t is the carbon quota allocation; At each time step, based on the current state S t and the action A taken t , update the Q value: where α is the learning rate, γ is the discount factor, and R t is the reward value; update the value of each state-action pair through Q-learning to ensure that the carbon quota gap gradually decreases; In the policy gradient method, based on the policy π θ (a|s) update the parameter θ: Optimize by the gradient of the strategy to maximize the cumulative reward within each time period, and continuously update the enterprise production strategy and R & D investment decision; After the end of each time period, adjust the enterprise's production plan and R & D investment according to the carbon quota gap and the corresponding reward feedback, and apply the updated strategy to the next time period for rolling optimization.

10. The enterprise carbon quota gap prediction and rolling optimization method according to claim 9, characterized in that, The adjustment of the enterprise's production plan and R & D investment according to the carbon quota gap and the corresponding reward feedback, and the application of the updated strategy to the next time period for rolling optimization include the following operations: Construct a linear programming model based on the enterprise's historical carbon emission data and carbon quota gap, with the goal of minimizing the gap between carbon emissions and carbon quota. The specific optimization objective function is: According to the predicted future carbon quota and combined with the enterprise's historical carbon emission data, use a time series model to predict the carbon emissions in the next time period. The formula is: E t+1 = g(E t , Q t+1 ) Among them, E t+1 is the predicted future carbon emissions, and the function g represents the relationship between historical carbon emissions and future carbon quotas; After the end of each time period, repeat the prediction of the above carbon quota and carbon emissions to predict the carbon emissions and carbon quota in the next one or more time periods until the carbon emissions meet the following termination conditions: E t ≤ E target Among them, E target is the carbon emission target expected by the country; when the predicted carbon emissions are lower than this target, the rolling prediction is terminated; The final output result is the carbon emissions and the corresponding predicted carbon quota in each future year before reaching the national expected enterprise carbon emission target.

Images