Enterprise carbon emission prediction and decision support method and device fusing lmdi decomposition and lightweight transformer prediction

Abstract

The application provides an enterprise carbon emission prediction and decision support method and device fusing LMDI decomposition and lightweight Transformer prediction, and belongs to the field of carbon emission management; solves the problems of low prediction accuracy, lack of interpretability and difficulty in depicting the collaborative relationship between emission sources of existing enterprise carbon emission prediction; the method comprises the following steps: obtaining data of each emission source of an enterprise, preprocessing the data, inputting the data into an LMDI model for carbon emission decomposition, and adaptively adjusting different enterprise working conditions through a dynamic time window and a decay weight mechanism, and introducing an emission source interaction analysis mechanism to depict the synergistic or offsetting effect between different emission sources; a lightweight Transformer prediction model is used to predict the future carbon emission trend; a multi-dimensional decision support mechanism is used to generate decision suggestion support; the application can realize carbon emission decomposition of multiple emission sources, dynamic trend prediction and driving factor diagnosis, and support enterprises to carry out fine carbon management and decision making.

Description

Technical Field

[0001] This application relates to the cross-application of carbon emission management and artificial intelligence technologies, and in particular to a method and apparatus for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction. Background Technology

[0002] As the core entities in carbon emission management, enterprises have increasingly complex emission characteristics and significant dynamic fluctuations, posing multiple challenges to traditional carbon accounting and forecasting methods in practical applications.

[0003] Currently, the main problems in corporate carbon emission analysis and forecasting are as follows:

[0004] 1. Traditional LMDI methods have poor applicability.

[0005] The Log-Mean Dijkstra Index (LMDI), as a mainstream tool for carbon emission decomposition, typically relies on macroeconomic statistical indicators (such as energy structure, industrial structure, and energy intensity) to construct intermediate effect variables, demonstrating good adaptability at the national or industry level. However, in enterprise scenarios, due to the diversity of emission sources, frequent data updates, and finer granularity, traditional LMDI models are difficult to adapt and are highly dependent on complete economic statistical data.

[0006] 2. The prediction and deconstruction are disconnected, lacking explanatory power.

[0007] Existing forecasting methods mainly employ time series models or black-box models based on neural networks such as LSTM and GRU. While these methods possess a certain level of predictive accuracy, they struggle to identify the driving factors behind carbon emissions and lack interpretability. Although traditional LMDI methods can decompose and attribute causes, they cannot provide predictions of future trends, leading to a separation between prediction and explanation. This makes it difficult for companies to make effective control or optimization decisions based on model results.

[0008] 3. Lack of ability to model inter-source interactions and dynamic effects

[0009] Corporate carbon emissions exhibit significant inter-source synergy characteristics. Multiple emission sources may be coupled, amplified, or offset each other. However, the traditional LMDI method is a static summation, which cannot reflect this dynamic correlation and has not established a mathematical description of the interaction between emission sources.

[0010] Therefore, there is an urgent need for a method that can address enterprise-level multi-source emission structures, balance decomposition and interpretability with trend prediction accuracy, and possess inter-source interaction modeling capabilities, in order to support carbon emission management and emission reduction decisions in high-frequency dynamic scenarios. Summary of the Invention

[0011] To address the problems of low accuracy, lack of interpretability, and difficulty in characterizing the synergistic relationships between emission sources in existing technologies for enterprise carbon emission prediction, this application proposes a method and apparatus for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction.

[0012] The technical solution adopted in this application is: a method for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction, including the following steps:

[0013] Step 1: Obtain historical data and process parameter configurations for each emission source of the enterprise, and perform data preprocessing. Input the preprocessed data into the LMDI model for carbon emission decomposition, and adaptively adjust the different enterprise operating conditions through dynamic time windows and attenuation weight mechanism. At the same time, introduce an emission source interaction analysis mechanism to characterize the synergistic or offsetting effects between different emission sources.

[0014] Step 2: Construct a lightweight Transformer prediction model based on LMDI features, and use the LMDI decomposition results as structured feature vectors to input into the Transformer architecture for future carbon emission trend prediction.

[0015] Step 3: Based on the prediction results, a multi-dimensional decision support mechanism is adopted to provide multi-scale analysis of emission trends, realize dynamic identification of driving factors, quantify the interaction effects between sources, and generate decision support suggestions.

[0016] Furthermore, the historical data of each emission source of the enterprise includes the time-series carbon emissions of each emission source, and the process parameter configuration includes time window parameters and industry-specific parameters. The time window parameters include the process change cycle, minimum sample size requirement and data characteristic cycle, and the industry-specific parameters include the time decay coefficient and the default parameter values ​​for different industries.

[0017] Furthermore, in step 1, the preprocessed data is processed according to the following timing steps:

[0018] Step 1.1: Dynamic window partitioning: Adaptively determine the analysis window length based on industry characteristics, and perform sliding window partitioning on the data;

[0019] Step 1.2: Zero value preprocessing: Detect and process zero values ​​in the data;

[0020] Step 1.3: Basic LMDI decomposition calculation: Calculate the contribution of each emission source;

[0021] Step 1.4: Time decay weighting: Apply time decay weights to obtain the weighted contribution of each emission source;

[0022] Step 1.5: Emission Source Interaction Analysis: Construct an interaction network and analyze the interaction relationships between various emission sources.

[0023] Furthermore, the interaction network in the emission source interaction analysis is as follows:

[0024] ;

[0025] In the formula, V is the set of emission source nodes; E is the set of interaction relationship edges; and W is the interaction coefficient weight matrix.

[0026] The formula for calculating the interaction coefficient is:

[0027] ;

[0028] In the formula, For interactive weighting factors; , These represent the LMDI decomposition contributions from emission sources s1 and s2, respectively.

[0029] Furthermore, zero-value preprocessing includes minute value substitution, logarithmic calculation protection, and limit case handling.

[0030] Furthermore, the data obtained after step 1 includes:

[0031] (1) Contribution of each emission source, including the original LMDI decomposition contribution of each emission source;

[0032] (2) Weighted LMDI contribution;

[0033] (3) Emission source interaction coefficient matrix, including: the interaction coefficient values ​​between each pair of emission sources. This is used to quantify the degree of influence between sources;

[0034] (4) Interactive network diagram.

[0035] Furthermore, the feature fusion layer of the lightweight Transformer prediction model is used to fuse the hidden state output by the Transformer encoder with the emission source contribution feature vector obtained from LMDI decomposition, and its expression is as follows:

[0036] ;

[0037] in: ; The hidden state output by the Transformer encoder; Contribute feature vectors to emission sources obtained from LMDI decomposition; This is the weight matrix of the feature fusion layer; For bias terms; This represents a vector concatenation operation; For adaptive fusion coefficients.

[0038] 8. A method for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction according to claim 7, characterized in that: the output of the lightweight Transformer prediction model includes:

[0039] (1) Emissions forecast results, including:

[0040] Projected total carbon emissions of enterprises and their confidence intervals;

[0041] Individual predicted values ​​and confidence intervals for each emission source;

[0042] Emissions trend curve within the predicted time window;

[0043] (2) Interpretability analysis results, including:

[0044] a) Analysis of emission source contribution, including:

[0045] The contribution weight of each emission source;

[0046] The temporal trend of contribution;

[0047] Key driver identification results;

[0048] b) Interaction effect analysis, including:

[0049] Interaction coefficient matrix among emission sources;

[0050] Results of identifying significant interaction relationships;

[0051] The dynamic evolution trend of interaction effects.

[0052] Furthermore, the multi-dimensional decision support mechanism in step 3 includes:

[0053] (1) Emissions trend diagnosis:

[0054] Based on the output of the prediction model, the future emission change trend of each emission source is diagnosed, as follows:

[0055] Calculate the predicted value Compared with historical benchmarks The relative deviation between them:

[0056] ;

[0057] Set the judgment threshold The trend can be determined based on the magnitude of the relative deviation.

[0058] like and This is determined to be an upward trend;

[0059] like and This is determined to be a downward trend;

[0060] like This is determined to be a stable trend;

[0061] (2) Driving factor analysis:

[0062] Based on the contribution weight of each emission source w i The emission sources are classified into different levels of driving factors:

[0063] When w i When the value is greater than 0.25, it is identified as a "key driving factor";

[0064] When 0.15 <w i When the value is ≤0.25, it is judged as an "important driving factor";

[0065] When w i When the value is ≤0.15, it is determined to be a "general driving factor";

[0066] (3) Evaluation of interaction effects:

[0067] The interaction intensity is classified based on the absolute value of the interaction coefficient |I| to quantify the cooperative or conflicting relationships between different emission sources:

[0068] When |I|>0.7, it is determined to be "strong interaction";

[0069] When 0.3 < |I ≤ 0.7, it is judged as "medium interaction";

[0070] When |I|≤0.3, it is determined to be "weak interaction";

[0071] Positive interaction values ​​indicate a synergistic amplification effect, while negative interaction values ​​indicate a counteracting inhibition effect.

[0072] (4) Suggested generation, including:

[0073] a) Individual emission source optimization recommendations: automatically generated based on emission trends and drive levels;

[0074] b) Optimization suggestions for interaction coupling: Determine the collaborative optimization method based on the interaction strength and sign:

[0075] Strong positive interaction: It is recommended to jointly adjust operating parameters;

[0076] Strong negative interaction: It is recommended to implement staggered operation or decoupling design;

[0077] Medium-level interaction: It is recommended to optimize runtime sequence and reduce conflict windows;

[0078] (5) Suggested priority ranking:

[0079] Priority is determined through a comprehensive scoring method, calculated using the following formula:

[0080] ;

[0081] in: Contribution weight matrix for emission sources; The magnitude of the trend change; Interaction strength; The difficulty level of implementation; , , , These are the weighting coefficients;

[0082] Priority criteria:

[0083] When P > 0.7, it is determined to be "high priority";

[0084] When 0.4≤P≤0.7, it is determined to be "medium priority";

[0085] When P < 0.4, it is judged as "low priority".

[0086] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0087] The advantages of this application over the prior art are as follows:

[0088] 1. High-precision dynamic decomposition capability: It adopts a direct decomposition method of emission sources, which reduces data dependence and supports dynamic emission analysis at the monthly or even weekly level.

[0089] 2. Strong interpretability and predictive ability: Through the source contribution ranking layer, the model can output emission trend driving factors and their weights.

[0090] In summary, this application, through the integration of multi-source decomposition, interactive modeling, and deep learning, not only improves the accuracy and granularity of carbon emission prediction, but also enhances the interpretability of the model and provides strong support for enterprises to achieve goal-oriented emission reduction. Attached Figure Description

[0091] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0092] Figure 1 A flowchart illustrating the method provided in this application embodiment;

[0093] Figure 2This is a device frame diagram provided for an embodiment of this application. Detailed Implementation

[0094] like Figure 1 and 2 As shown, this application provides a method for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction. It combines emission source decomposition, trend prediction, and decision support, and mainly includes the following steps:

[0095] Step 1: Direct Decomposition of Enterprise-Level LMDI Emission Sources: This step proposes directly applying the Log-Mean Dijkstra Index (LMDI) method to carbon emission decomposition at the emission source level, eliminating the need to construct intermediate effect variables and significantly reducing data requirements and computational complexity. Simultaneously, a dynamic time window and decay weight mechanism are designed to enhance adaptability to changes in enterprise operating conditions, and zero-value handling and numerical stability mechanisms are developed to improve reliability under incomplete data conditions. Furthermore, an emission source interaction analysis mechanism is introduced to characterize the synergistic or offsetting effects between different emission sources.

[0096] Step 2: Constructing a lightweight Transformer prediction model based on LMDI features: Innovatively, the LMDI decomposition results are used as structured feature vectors and input into the Transformer architecture for predicting future carbon emission trends. By designing a feature fusion layer and a source contribution ranking mechanism, interpretable predictions of emission change trends are achieved. The lightweight Transformer model used is optimized through parameter compression, quantization, and knowledge distillation.

[0097] Step 3: Based on the prediction results, a multi-dimensional decision support mechanism is adopted to provide multi-scale analysis of emission trends, realize dynamic identification of driving factors, quantify the interaction effects between sources, and generate decision support suggestions.

[0098] The input data in step 1 includes:

[0099] (1) Historical data of each emission source of the enterprise, including the time-series carbon emissions of each emission source. Its data format is time series data, which includes timestamps and corresponding emissions; time granularity: supports daily, weekly or monthly data;

[0100] (2) Time window parameters, including:

[0101] Process change cycle ;

[0102] Minimum sample size requirement ;

[0103] Data feature period ;

[0104] (3) Industry-specific parameters, including:

[0105] Time decay coefficient λ;

[0106] Default parameter values ​​for different industries.

[0107] The data processing in step 1 is performed according to the following sequence of steps:

[0108] Step 1.1: Dynamic window partitioning: Adaptively determine the analysis window length based on industry characteristics, and perform sliding window partitioning on the data;

[0109] Step 1.2: Zero value preprocessing: Detect and process zero values ​​in the data to ensure the stability of numerical calculations;

[0110] Step 1.3: Basic LMDI decomposition calculation: Calculate the contribution of each emission source;

[0111] Step 1.4: Time decay weighting: Apply time decay weights to highlight the importance of recent data;

[0112] Step 1.5: Emission Source Interaction Analysis: Construct and analyze the interaction relationships between emission sources.

[0113] The following results were obtained after data processing in step 1:

[0114] (1) Contribution of each emission source This includes the original LMDI decomposition contribution of each emission source, which contains positive and negative values ​​to indicate increases or decreases in contribution;

[0115] (2) Weighted LMDI contribution This is a weighted contribution that takes into account time decay, and it better reflects recent trends.

[0116] (3) Emission source interaction coefficient matrix, including: the interaction coefficient values ​​between each pair of emission sources. This is used to quantify the degree of influence between sources;

[0117] (4) Interactive network diagram, including:

[0118] Node V: Represents each emission source;

[0119] Edge set E: Represents interaction relationships;

[0120] Weight W: Represents the interaction coefficient.

[0121] The output will serve as input features for a lightweight Transformer prediction model and can also be directly used to support corporate carbon emission management decisions.

[0122] The specific implementation principle of step 1 is as follows:

[0123] Using the actual emission sources of enterprises as the basic unit, the following model is established:

[0124] ;

[0125] in, Let t be the company's total carbon emissions at time t. Let n be the emission amount of the i-th emission source, and n be the number of emission sources.

[0126] Based on the LMDI-I additive decomposition principle, the contribution of each emission source to emission changes is calculated as follows:

[0127] ;

[0128] In the formula, Let be the carbon emissions from the i-th emission source in the base period (starting time). Let be the carbon emissions of the i-th emission source in the current period (comparison time).

[0129] The logarithmic mean function is defined as follows:

[0130] ;

[0131] To improve sensitivity to recent data, a time decay weighting function is introduced:

[0132] ;

[0133] In the formula, t0 is the current time, which is the reference time corresponding to the calculation of the decay weight for the historical time t;

[0134] Thus, the weighted contribution is obtained:

[0135] .

[0136] Furthermore, the designed sliding window adaptive mechanism is as follows:

[0137] ;

[0138] Where: N is the final determined analysis window length (number of samples). For data period, For the process change cycle, This is the minimum sample size requirement.

[0139] To address potential synergistic or offsetting effects among multiple sources, interaction coefficients are calculated and an interaction network is constructed.

[0140] (1) Calculation of interaction coefficient:

[0141] ;

[0142] In the formula, For interactive weighting factors; , The contribution of LMDI decomposition to emission sources s1 and s2.

[0143] (2) Construction of interactive network:

[0144] ;

[0145] In the formula, V is the set of emission source nodes; E is the set of interaction relationship edges; and W is the interaction coefficient weight matrix.

[0146] To ensure numerical stability under extreme value and missing data conditions, a zero-value handling mechanism is set up:

[0147] (1) Small value substitution strategy:

[0148] ;

[0149] in, It is a configurable, tiny positive number. It is a very small positive number, and when the original emission is detected to be 0 at a certain moment, it is used as a substitute;

[0150] (2) Logarithmic calculation protection:

[0151] ;

[0152] (3) Handling extreme cases:

[0153] when In this case, the L'Hospital rule is applied:

[0154] ,

[0155] in, To judge Whether it is close enough to a custom-defined tiny threshold of 0. If the difference in emissions between two times is lower than this threshold, it is considered that there is a risk of numerical instability and needs to be approximated using L'Hospital's rule.

[0156] The input data for the lightweight Transformer prediction model based on LMDI features in step 2 includes:

[0157] (1) LMDI decomposition results, including:

[0158] Contribution of each emission source ;

[0159] Weighted LMDI contribution ;

[0160] Emission source interaction coefficient matrix ;

[0161] Interactive network diagram ;

[0162] (2) Historical emissions data, including:

[0163] Time series of total corporate carbon emissions;

[0164] Historical emission data for each emission source;

[0165] The time granularity is consistent with LMDI decomposition;

[0166] (3) Model parameter configuration

[0167] Predict the size of the time window;

[0168] Number of attention heads h (default 4);

[0169] Attention dimension dk (default 32);

[0170] Model dimension d_model (default 128).

[0171] The model is executed according to the following steps:

[0172] (1) Feature vector construction, including:

[0173] Integrate LMDI decomposition features;

[0174] Constructing time-series feature sequences;

[0175] Normalization processing;

[0176] (2) Transformer encoding, including:

[0177] Multi-head self-attention computation;

[0178] Location encoding;

[0179] Feedforward neural network processing;

[0180] (3) Feature fusion, including:

[0181] LMDI features are fused with Transformer output;

[0182] Adaptive weight calculation;

[0183] Feature enhancement;

[0184] (4) Source contribution ranking, including:

[0185] Calculate the contribution weight of emission sources;

[0186] Generate interpretable analysis results;

[0187] Sorting layer output;

[0188] (5) Predicted output, including:

[0189] Weighted forecast calculation;

[0190] Multi-objective loss optimization;

[0191] Prediction results are generated.

[0192] The model's output includes:

[0193] (1) Emissions forecast results, including:

[0194] Projected total carbon emissions of enterprises and their confidence intervals;

[0195] Individual predicted values ​​and confidence intervals for each emission source;

[0196] Emissions trend curve within the predicted time window;

[0197] (2) Interpretability analysis results, including:

[0198] a) Analysis of emission source contribution, including:

[0199] The contribution weight of each emission source;

[0200] The temporal trend of contribution;

[0201] Key driver identification results;

[0202] b) Interaction effect analysis, including:

[0203] Interaction coefficient matrix among emission sources;

[0204] Results of identifying significant interaction relationships;

[0205] The dynamic evolution trend of interaction effects.

[0206] Model Architecture: This module is based on the concept of time series modeling and uses historical emission data and LMDI decomposition results to predict future trends.

[0207] (1) Model structure, including

[0208] a) Multi-head self-attention mechanism:

[0209] ;

[0210] in:

[0211] ;

[0212] ;

[0213] Variable description:

[0214] Q, K, and V are the query, key, and value matrices, respectively.

[0215] h represents the number of attention heads (h=4 in this method);

[0216] , , , The weight matrix is ​​a learnable matrix;

[0217] For the attention dimension.

[0218] b) Feature fusion layer:

[0219] ;

[0220] in:

[0221] ;

[0222] Variable description:

[0223] The hidden state output by the Transformer encoder;

[0224] Contribute feature vectors to emission sources obtained from LMDI decomposition;

[0225] This is the weight matrix of the feature fusion layer;

[0226] For bias terms;

[0227] This represents a vector concatenation operation;

[0228] For adaptive fusion coefficients (value range [0,1]).

[0229] c) Emission source contribution weights and prediction output:

[0230] ;

[0231] ;

[0232] Variable description:

[0233] Let i be the feature vector of the i-th emission source;

[0234] This is the weight matrix for the sorting layer;

[0235] For bias terms;

[0236] The contribution weight of the i-th emission source;

[0237] Let be the predicted value for the i-th emission source;

[0238] n represents the total number of emission sources.

[0239] The lightweight design of the model in this embodiment is reflected in:

[0240] Parameter compression: Number of heads reduced to 4, dimensions reduced to 128;

[0241] Structured pruning: Remove connections with a contribution of <0.01;

[0242] INT8 Quantization and Knowledge Distillation: Improving Inference Efficiency and Adapting to Edge Deployment.

[0243] A multi-objective loss function is introduced during model training to balance prediction accuracy and interpretability:

[0244] ;

[0245] in:

[0246] (a) MSE error:

[0247] ;

[0248] (b) Quantile error:

[0249] ;

[0250] (c) Contribution consistency error:

[0251] ;

[0252] Variable description:

[0253] , , These are the weighting coefficients for the loss function (default values ​​are 1.0, 0.5, and 0.3, respectively).

[0254] The actual value;

[0255] This is a predicted value;

[0256] N is the number of samples;

[0257] Q is the set of quantiles (default is {0.1, 0.5, 0.9});

[0258] The actual LMDI contribution of the i-th emission source;

[0259] The predicted LMDI contribution of the i-th emission source.

[0260] The learning rate scheduling strategy during model training is as follows:

[0261] Learning rate using cosine annealing:

[0262] ;

[0263] Variable description:

[0264] Let be the learning rate at time t;

[0265] The initial learning rate is 0.001.

[0266] Minimum learning rate ( );

[0267] T represents the total number of training rounds (default 500 rounds);

[0268] t is the current training round number;

[0269] Pi is the mathematical constant of a circle.

[0270] Building upon the completion of carbon emission decomposition and prediction, this system further enhances decision-making support for enterprise carbon emission management. Through trend identification, driving factor assessment, and interaction effect analysis, it automatically generates emission optimization suggestions and, combined with a priority scoring mechanism, provides actionable strategy outputs.

[0271] The decision support mechanisms in step 3 include:

[0272] (1) Emissions trend diagnosis:

[0273] Based on the output of the prediction model, the future emission change trend of each emission source is diagnosed, as follows:

[0274] Calculate the relative deviation between the predicted value and the historical benchmark value (average of the previous period):

[0275] ;

[0276] Set the judgment threshold (Default 5%), the trend is determined based on the magnitude of the relative deviation:

[0277] like and This is determined to be an upward trend;

[0278] like and This is determined to be a downward trend;

[0279] like This is determined to be a stable trend;

[0280] (2) Driving factor analysis:

[0281] Based on the contribution weight of each emission source The emission sources are classified into different levels of driving factors:

[0282] when When the value is greater than 0.25, it is identified as a "key driving factor";

[0283] When 0.15 < When the value is ≤0.25, it is judged as an "important driving factor";

[0284] when When the value is ≤0.15, it is determined to be a "general driving factor".

[0285] The identification results are used to screen priority control targets.

[0286] (3) Evaluation of interaction effects:

[0287] The interaction intensity is classified based on the absolute value of the interaction coefficient |I| to quantify the cooperative or conflicting relationships between different emission sources:

[0288] When |I|>0.7, it is determined to be "strong interaction";

[0289] When 0.3 < |I ≤ 0.7, it is judged as "medium interaction";

[0290] When |I|≤0.3, it is determined to be "weak interaction".

[0291] Positive interaction values ​​indicate a synergistic amplification effect, while negative interaction values ​​indicate a counteracting inhibition effect.

[0292] (4) Suggested generation, including:

[0293] a) Optimization recommendations for individual emission sources: Strategies as shown in Table 1 are automatically generated based on emission trends and drive levels.

[0294] Table 1 Optimization Recommendations for Individual Emission Sources

[0295]

[0296] b) Optimization suggestions for interaction coupling: Determine the collaborative optimization method based on the interaction strength and sign:

[0297] Strong positive interaction: It is recommended to jointly adjust operating parameters;

[0298] Strong negative interaction: It is recommended to implement staggered operation or decoupling design;

[0299] Medium-level interaction: It is recommended to optimize runtime sequence and reduce conflict window.

[0300] (5) Suggested priority ranking:

[0301] Priority is determined through a comprehensive scoring method, calculated using the following formula:

[0302] ;

[0303] in: : Contribution weight matrix of emission sources, with values ​​ranging from [0,1]; The magnitude of the trend change is normalized to the [0,1] interval; Interaction strength, normalized to the [0,1] interval; : Implementation difficulty coefficient, with a value range of [0,1]; , , , These are the weighting coefficients, with default values ​​of 0.4, 0.3, 0.2, and 0.1, respectively.

[0304] Priority criteria:

[0305] When P > 0.7, it is determined to be "high priority";

[0306] When 0.4≤P≤0.7, it is determined to be "medium priority";

[0307] When P < 0.4, it is judged as "low priority".

[0308] (6) Output specifications:

[0309] Each suggestion should include the following elements:

[0310] a) Recommendation type: Distinguish between source optimization type and interaction optimization type;

[0311] b) Target objects: Identify specific emission sources or combinations of emission sources;

[0312] c) Recommendation content: Provide a detailed explanation of specific improvement measures;

[0313] d) Priority: Determined based on the priority calculation results.

[0314] This decision support mechanism transforms multi-dimensional analysis results into actionable recommendations through standardized analysis methods and recommendation generation processes. The mechanism considers both the characteristics of individual emission sources and inter-source interactions, and ensures maximum effectiveness of the recommendations through a scientific prioritization method.

[0315] The implementation process of the method described in this application is explained in detail below based on an example of carbon emission prediction and decision support for a steel enterprise.

[0316] Implementation Background: A steel company has an annual output of approximately 5 million tons, mainly including the following carbon emission sources:

[0317] Sintering process;

[0318] Ironmaking process;

[0319] Steelmaking process;

[0320] Steel rolling mill worker;

[0321] Self-owned power plant.

[0322] Data preparation:

[0323] (1) Emission source data:

[0324] Time span: January 1, 2022 to December 31, 2023;

[0325] Data granularity: daily data;

[0326] Data type: Carbon emissions from each emission source.

[0327] (2) Industry-specific parameter configuration:

[0328] a) Specific parameters:

[0329] Process change cycle ( ): 90 days;

[0330] Time decay coefficient (λ): 0.08;

[0331] Minimum sample size ( ): 60;

[0332] Forecast time window: 30 days.

[0333] b) Basis for parameter selection:

[0334] Process change cycle: Determined based on industry production characteristics and process adjustment frequency;

[0335] Time decay coefficient: determined based on the degree of influence of historical data;

[0336] Minimum sample size: The minimum amount of data required to ensure effective model training;

[0337] Forecast time window: determined based on the company's decision-making cycle.

[0338] (3) Parameter configuration:

[0339] Process change cycle ;

[0340] Minimum sample size requirement ;

[0341] Data feature period .

[0342] LMDI decomposition process:

[0343] (1) Dynamic window division:

[0344] ;

[0345] (2) Example of calculating the contribution of emission sources (taking December 2023 data as an example, as shown in Table 2 below).

[0346] Table 2 Emission Source Contribution

[0347]

[0348] (3) Results of interaction analysis of emission sources:

[0349] Significant positive interactions:

[0350] Ironmaking process and self-owned power plant: +0.85 (strong positive correlation);

[0351] Sintering process and ironmaking process: +0.72 (strong positive correlation).

[0352] Significant negative interactions:

[0353] Steelmaking process and self-owned power plant: -0.45 (moderate negative correlation).

[0354] Transformer prediction process:

[0355] (1) Feature vector construction:

[0356] LMDI characteristic dimensions: 5 (number of emission sources) × 3 (contribution, interaction coefficient, time weight);

[0357] Time series characteristics: 300 dimensions (5 emission sources × 60 days);

[0358] Input the normalized data into the model.

[0359] (2) Model configuration:

[0360] Attention count: 4;

[0361] Attention dimension: 32;

[0362] Model dimensions: 128;

[0363] Number of training rounds: 500;

[0364] Batch size: 64.

[0365] (3) Training process configuration:

[0366] a) Data partitioning:

[0367] Training set: 70% (data from the last 18 months);

[0368] Validation set: 20% (data from the next 5 months);

[0369] Test set: 10% (data from the last two months).

[0370] b) Verification strategy:

[0371] Validation frequency: Validation is performed once every 10 epochs;

[0372] Early stopping condition: The loss on the validation set has not improved after 5 consecutive validation set iterations;

[0373] Validation metrics: MSE, MAE, R 2 coefficient.

[0374] c) Learning rate adjustment:

[0375] Initial learning rate: 0.001;

[0376] Attenuation strategy: cosine annealing;

[0377] Minimum learning rate: 1e-6.

[0378] d) Regularization strategy:

[0379] Dropout rate: 0.1;

[0380] L2 regularization coefficient: 0.0001.

[0381] e) Model quantization:

[0382] Quantization method: Dynamic Range Quantization;

[0383] Bit width: INT8;

[0384] Calibration set size: 1000 samples;

[0385] Precision loss threshold after quantization: 1%.

[0386] (4) Key indicators of the training process:

[0387] MSE loss: decreased from 0.185 to 0.042;

[0388] Quantile loss: decreased from 0.156 to 0.035;

[0389] Contribution consistency loss: decreased from 0.225 to 0.058;

[0390] Optimal epoch for validation set: 320;

[0391] The size of the quantized model is 23% of the original size.

[0392] (5) Example of prediction results (next 30 days), as shown in Table 3 below:

[0393] Table 3. Carbon emission prediction results from emission sources

[0394]

[0395] Decision recommendation generation results:

[0396] (1) The results of the driving factor identification are shown in Table 4 below.

[0397] Table 4 Ranking of Emission Sources by Importance

[0398]

[0399] (2) Strategy Example A:

[0400] Recommended for: Sintering process;

[0401] Recommendation type: Runtime recommendation;

[0402] Recommendation: Given that the sintering process is currently the most critical emission driver and its emissions have shown a good downward trend, it is recommended to prioritize maintaining and deepening existing energy-saving control measures to ensure the stability of emission reduction results.

[0403] Priority: High.

[0404] Recommendation based on:

[0405] The dynamic weight is the highest (0.35), making it the key driving factor (according to Table 4).

[0406] The emissions trend is continuously declining, indicating that the current control measures are effective and should be maintained (based on Table 3).

[0407] There is a strong positive interaction with the ironmaking process (+0.72), and its stable operation has a positive impact on the carbon emissions of downstream processes (based on the emission source interaction analysis results in the instruction manual).

[0408] (3) Strategy Example B:

[0409] Recommended for: Ironmaking process - self-owned power plant;

[0410] Recommendation type: Collaborative optimization recommendation;

[0411] Recommendation: It is recommended to implement joint regulation and control, and optimize production planning and energy dispatch to avoid both operating at high loads simultaneously, thereby reducing the synergistic emission increase effect.

[0412] Priority: Medium.

[0413] Recommendation based on:

[0414] The two are strongly positively correlated (+0.85), with a significant synergistic effect (based on the emission source interaction analysis results in the instruction manual).

[0415] Among them, the ironmaking process is the key driving factor, and the self-owned power plant is a general driving factor (according to Table 4).

[0416] Both emissions trends are rising in the same direction. If coordinated management is not implemented, there is a risk that total emissions will be pushed up together in the future (based on Table 3).

[0417] Summary of Implementation Results: This embodiment demonstrates that the method of this application can effectively complete enterprise-level multi-source emission decomposition and trend prediction; it can identify key emission sources and synergistic relationships, and has strong interpretability. The decision recommendations are clear and operational, helping enterprises to carry out refined carbon emission management and emission reduction strategy formulation.

[0418] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the method.

[0419] Figure 2A structural block diagram of a computer device according to an embodiment of this application is shown. As shown in Figure 2, the computer device includes a memory and a processor, wherein the memory stores instructions executable on the processor. When the processor executes the instructions, it implements the methods described in the above embodiments. The number of memories and processors can be one or more. This computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0420] The computer device may also include a communication interface for communicating with external devices and exchanging data. The devices are interconnected using different buses and can be mounted on a common motherboard or otherwise as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). The bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, Figure 2 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0421] Optionally, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0422] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting advanced RISC machines (ARM) architecture.

[0423] This application provides a computer-readable storage medium (such as the memory described above) that stores computer instructions, which, when executed by a processor, implement the method provided in this application.

[0424] Optionally, the memory may include a stored program area and a stored data area, wherein the stored program area may store the operating system and application programs required for at least one function; the stored data area may store data created based on the use of the computer device for mapping. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the computer device for mapping via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0425] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction, characterized in that: Includes the following steps: Step 1: Obtain historical data and process parameter configurations for each emission source of the enterprise, and perform data preprocessing. Input the preprocessed data into the LMDI model for carbon emission decomposition, and adaptively adjust the different enterprise operating conditions through dynamic time windows and attenuation weight mechanism. At the same time, introduce an emission source interaction analysis mechanism to characterize the synergistic or offsetting effects between different emission sources. Step 2: Construct a lightweight Transformer prediction model based on LMDI features, and use the LMDI decomposition results as structured feature vectors to input into the Transformer architecture for future carbon emission trend prediction. Step 3: Based on the prediction results, a multi-dimensional decision support mechanism is adopted to provide multi-scale analysis of emission trends, realize dynamic identification of driving factors, quantify the interaction effects between sources, and generate decision support suggestions.

2. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 1, characterized in that: Historical data for each emission source of the enterprise includes the time-series carbon emissions of each emission source. Process parameter configurations include time window parameters and industry-specific parameters. Time window parameters include process change cycle, minimum sample size requirement and data characteristic cycle. Industry-specific parameters include time decay coefficient and default parameter values ​​for different industries.

3. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 2, characterized in that: In step 1, the preprocessed data is processed according to the following timing steps: Step 1.1: Dynamic window partitioning: Adaptively determine the analysis window length based on industry characteristics, and perform sliding window partitioning on the data; Step 1.2: Zero value preprocessing: Detect and process zero values ​​in the data; Step 1.3: Basic LMDI decomposition calculation: Calculate the contribution of each emission source; Step 1.4: Time decay weighting: Apply time decay weights to obtain the weighted contribution of each emission source; Step 1.5: Emission Source Interaction Analysis: Construct an interaction network and analyze the interaction relationships between various emission sources.

4. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 3, characterized in that: The interaction network in the emission source interaction analysis is as follows: ； In the formula, V is the set of emission source nodes; E is the set of interaction relationship edges; and W is the interaction coefficient weight matrix. The formula for calculating the interaction coefficient is: ； In the formula, For interactive weighting factors; , These represent the LMDI decomposition contributions from emission sources s1 and s2, respectively.

5. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 3, characterized in that: Zero-value preprocessing includes minute value substitution, logarithmic calculation protection, and limit case handling.

6. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 4, characterized in that: The data obtained after step 1 includes: (1) Contribution of each emission source, including the original LMDI decomposition contribution of each emission source; (2) Weighted LMDI contribution; (3) Emission source interaction coefficient matrix, including: the interaction coefficient values ​​between each pair of emission sources. This is used to quantify the degree of influence between sources; (4) Interactive network diagram.

7. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 1, characterized in that: The feature fusion layer of the lightweight Transformer prediction model is used to fuse the hidden state output by the Transformer encoder with the emission source contribution feature vector obtained from LMDI decomposition. Its expression is as follows: ； in: ; The hidden state output by the Transformer encoder; Contribute feature vectors to emission sources obtained from LMDI decomposition; This is the weight matrix of the feature fusion layer; For bias terms; This represents a vector concatenation operation; For adaptive fusion coefficients.

8. The enterprise carbon emission prediction and decision support method integrating LMDI decomposition and lightweight Transformer prediction as described in claim 7, characterized in that: The output of the lightweight Transformer prediction model includes: (1) Emissions forecast results, including: Projected total carbon emissions of enterprises and their confidence intervals; Individual predicted values ​​and confidence intervals for each emission source; Emissions trend curve within the predicted time window; (2) Interpretability analysis results, including: a) Analysis of emission source contribution, including: The contribution weight of each emission source; The temporal trend of contribution; Key driver identification results; b) Interaction effect analysis, including: Interaction coefficient matrix among emission sources; Results of identifying significant interaction relationships; The dynamic evolution trend of interaction effects.

9. A method for enterprise carbon emission prediction and decision support that integrates LMDI decomposition and lightweight Transformer prediction according to claim 8, characterized in that: The multi-dimensional decision support mechanism in step 3 includes: (1) Emissions trend diagnosis: Based on the output of the prediction model, the future emission change trend of each emission source is diagnosed, as follows: Calculate the predicted value Compared with historical benchmarks The relative deviation between them: ； Set the judgment threshold The trend can be determined based on the magnitude of the relative deviation. like and This is determined to be an upward trend; like and This is determined to be a downward trend; like This is determined to be a stable trend; (2) Driving factor analysis: Based on the contribution weight of each emission source w i The emission sources are classified into different levels of driving factors: When w i When the value is greater than 0.25, it is identified as a "key driving factor"; When 0.15 <w i When the value is ≤0.25, it is judged as an "important driving factor"; When w i When the value is ≤0.15, it is determined to be a "general driving factor"; (3) Evaluation of interaction effects: The interaction intensity is classified based on the absolute value of the interaction coefficient |I| to quantify the cooperative or conflicting relationships between different emission sources: When |I|>0.7, it is determined to be "strong interaction"; When 0.3 < |I ≤ 0.7, it is judged as "medium interaction"; When |I|≤0.3, it is determined to be "weak interaction"; Positive interaction values ​​indicate a synergistic amplification effect, while negative interaction values ​​indicate a counteracting inhibition effect. (4) Suggested generation, including: a) Individual emission source optimization recommendations: automatically generated based on emission trends and drive levels; b) Optimization suggestions for interaction coupling: Determine the collaborative optimization method based on the interaction strength and sign: Strong positive interaction: It is recommended to jointly adjust operating parameters; Strong negative interaction: It is recommended to implement staggered operation or decoupling design; Medium-level interaction: It is recommended to optimize runtime sequence and reduce conflict windows; (5) Suggested priority ranking: Priority is determined through a comprehensive scoring method, calculated using the following formula: ； in: Contribution weight matrix for emission sources; The magnitude of the trend change; Interaction strength; The difficulty level of implementation; , , , These are the weighting coefficients; Priority criteria: When P > 0.7, it is determined to be "high priority"; When 0.4≤P≤0.7, it is determined to be "medium priority"; When P < 0.4, it is judged as "low priority".

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images