Construction method of carbon emission prediction model based on artificial intelligence

Abstract

The invention discloses a carbon emission prediction model construction method based on artificial intelligence, and the method comprises the steps: collecting enterprise production data, carrying out the preprocessing of the data, obtaining time series data, calculating derivative features based on the time series data, obtaining a feature data set, training an LSTM-attention model, and generating a time series prediction model. And extracting distribution parameters from the causal forest through a variational inference algorithm to obtain a cause-fruit tree model, and constructing a carbon emission prediction model through weighted fusion based on a double-layer optimization fusion strategy according to a carbon emission trend prediction result output by the time sequence prediction model and a marginal effect coefficient output by the cause-fruit tree model. According to the method, the time sequence prediction model and the cause fruit tree model are constructed, and a double-layer optimization fusion strategy is combined, so that the carbon emission prediction precision is improved, and the process parameter influence stability is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of greenhouse gas prediction, and particularly to a construction method of a carbon emission prediction model based on artificial intelligence. Background Art

[0002] In recent years, artificial intelligence and big data technologies have deeply empowered the field of carbon emission prediction and management, forming a multi-level technical system. In terms of prediction models, for example, the grey prediction method realizes the qualitative analysis of carbon emission trends through cumulative generation and grey equation construction. With the increasing complexity of data, machine learning and deep learning models have gradually become the mainstream:

[0003] However, despite significant technological progress, current carbon emission prediction and management still face multiple bottlenecks. The heterogeneity of multi-source data leads to difficulties in fusion. The lack of unified standard systems for carbon footprint certification and green power traceability exacerbates the difficulty of enterprises' collaborative carbon reduction. Pure data-driven black-box models are difficult to attribute the impact of extreme climate events, and there are still error accumulations when dealing with long time series. The above problems need to be solved urgently. Summary of the Invention

[0004] The purpose of the present invention is to provide a construction method of a carbon emission prediction model based on artificial intelligence.

[0005] To achieve the above purpose, the present invention is implemented according to the following technical solutions:

[0006] The first aspect of the present invention provides a construction method of a carbon emission prediction model based on artificial intelligence, including:

[0007] A. Collect enterprise production data and preprocess it to obtain time series data, where the enterprise production data includes carbon emission data and production data;

[0008] B. Use a sliding window to segment the time series data to obtain sample data, calculate the derivative features of the sample data, align the time series data and the derivative features according to the window to obtain a feature dataset, and divide the feature dataset into a training set, a validation set, and a test set;

[0009] C. Input the training set into the LSTM-attention model for training to generate a time series prediction model. Input the validation set into the time series prediction model, optimize the weight parameters of the attention layer based on the prediction error. At the same time, extract process parameters and carbon emission data from the training set and input them into the causal tree model, extract distribution parameters through the variational inference algorithm, and obtain the marginal effect coefficient of the process parameters based on the distribution parameters;

[0010] D Based on the carbon emission trend prediction results output by the time series prediction model and the process parameter marginal effect coefficients output by the causal tree model, a carbon emission prediction model is constructed through weighted fusion based on a two-layer optimization fusion strategy, and a visual prediction chart is output. The carbon emission prediction model is optimized using the test set through the root mean square error.

[0011] As a further method, the method for collecting enterprise production data and preprocessing it to obtain time series data includes: The enterprise production data includes carbon emission data and production data. The production data includes using industrial Internet of Things sensors to collect in real time the equipment model, operating status, and process parameters of the enterprise production equipment, and associating product batches with production volumes, collecting equipment energy consumption data, carbon emission data, and carbon emission standards in the production process, obtaining equipment carbon emission factors according to the equipment type and carbon emission standards, forming an original data set with time stamps, cleaning the original data set, merging duplicate records based on batch numbers and time stamps, generating a time axis at a fixed interval of 15 minutes, aggregating and sampling the high-frequency data in the original data set according to the time axis, and associating the low-frequency data to the nearest time point on the time axis; if there is no corresponding low-frequency data at a certain time point, it is filled with adjacent values to obtain time series data.

[0012] Furthermore, the method for aligning the time series data and the derived features by window to obtain a feature data set includes:

[0013] Based on the sampling interval of the time series data, the sliding window size is set to 30 sampling intervals. The time series data is cut into fixed sample lengths by moving the window sequentially based on the sampling interval, and an identifier is generated sequentially for each sample window. For each sample window, derived features are calculated;

[0014] Based on the carbon emission standard and the carbon emission factor, the associated fields directly related to carbon emission prediction in the sample are screened out, the Pearson correlation coefficient between the remaining fields and the carbon emission data is calculated, and the indirectly associated fields with the absolute value of the correlation coefficient greater than 30% are retained. The associated fields and the indirectly associated fields are used as the basic fields of the sample, and the sample data is extracted based on the basic fields to obtain the original time series features;

[0015] The original time series features and the derived features are aligned by the sample window identifier and then merged into the same data table. Each row represents a sample, and each column represents the basic fields of the sample, obtaining a feature data set.

[0016] Furthermore, the method for calculating the derived features for each sample window includes:

[0017] Based on the time series data, the sample data is obtained by using a sliding window to segment it. Based on the operating status of the equipment in the sample data, the actual operating time is obtained, and the planned production time is determined in combination with the product batch. The equipment actual operating time is divided by the planned production time to obtain the time utilization rate;

[0018] Determine the actual production speed based on the equipment operation status, actual operation time, and output. Obtain the theoretical production speed based on the equipment model, and divide the actual production speed of the equipment by the theoretical production speed to obtain the performance operation rate.

[0019] Count the number of good products based on the product batches and divide it by the output to obtain the good product rate. Calculate the product of the good product rate, time operation rate, and performance operation rate to obtain the overall equipment efficiency. Divide the energy consumption data by the output to obtain the energy consumption per unit product, and multiply the energy consumption per unit product by the carbon emission factor to obtain the carbon emission per unit product.

[0020] Further, the method for dividing the feature dataset into a training set, a validation set, and a test set includes:

[0021] Perform noise optimization and hierarchical partitioning on the feature dataset;

[0022] The noise optimization includes calculating the mean and variance of the time series data in the feature dataset. Use the mean value of the time series data as the mean of the Gaussian distribution noise, and select 27% of the variance of the original time series data as the variance of the Gaussian distribution noise. Based on the time series data, superimpose Gaussian distribution noise through the formula X′ = X + ∈·N(0,σ 2 ) where X′ represents the enhanced time series data, X represents the time series data, ∈ represents the intensity coefficient, and N(0,σ 2 ) represents the probability distribution;

[0023] Divide the data into levels according to the production shifts. For each level, generate the training set, validation set, and test set according to the ratio of 7:2:1.

[0024] Further, the method for inputting the training set into the LSTM-attention model for training to generate a time series prediction model includes:

[0025] Convert the training set into a three-dimensional tensor, which includes the number of samples, time steps, and the number of features. Set the carbon emission target value based on the carbon emission data of the training set as the output data. Perform min-max normalization on the three-dimensional tensor and the carbon emission target value, and use the normalized three-dimensional tensor as the input data of the LSTM-attention model;

[0026] Stack two layers of bidirectional LSTM units to extract the basic data at each time step in the input data. Generate position encoding for each time step in the input data through the sine function, add the position encoding to the basic data element by element, and input the time step sequence fused with the position encoding into the bidirectional LSTM for processing to obtain a hidden state sequence containing the feature information of each time step. Calculate the attention weight matrix of the hidden state sequence. The formula for the attention matrix is:

[0027]

[0028] where β n,τ,τ′ is the attention weight of time step τ to τ′ in the nth sample, Query n,τ is the query vector of the nth sample at time step τ, Key n,τ is the key vector of the nth sample at time step τ, PosBias τ,τ′ is the position bias of time step τ to τ′, d k is the dimension corresponding to the scaling factor, k is the abbreviation of Key, T is the total number of time steps, τ and τ′ are time step indices, and τ″ is the time step index in the summation;

[0029] The key time steps of the hidden state sequence are weighted based on the attention weight matrix to obtain the attention output, and the attention output is mapped and transformed through a linear activation function to obtain a 1D carbon emission prediction value, and a time series prediction model is constructed;

[0030] The validation set is converted into a three-dimensional tensor and normalized, and then input into the time series prediction model to obtain the mean absolute error between the validation set prediction result and the carbon emission target value. Based on the error, the backpropagation algorithm is executed to derive the gradients of the parameters in the attention weight matrix, and the time step weight parameters are adjusted according to the gradient direction and magnitude.

[0031] Furthermore, the method for extracting process parameters and carbon emission data from the training set and inputting them into the causal tree model, and extracting distribution parameters through the variational inference algorithm and obtaining the marginal effect coefficient of process parameters based on the distribution parameters includes:

[0032] Based on the training set data table, the process parameters of each column and the carbon emissions per unit product are extracted as sample features, and each row of data is extracted as a sample to construct a decision tree. The decision tree is iterated, and each time the decision tree is iterated, 67% of the total number of samples is randomly selected as the current sample, and the square root of the total number of sample features in the current iteration is obtained as the number of features;

[0033] The process parameters are traversed through the splitting algorithm for their value ranges, and the actual value range and the intermediate values of adjacent values are included in the candidate threshold list. The current sample is divided into a treatment group and a control group according to the candidate threshold, and the distribution parameters and regression coefficients of the causal effect strength are obtained based on the treatment group and the control group through the variational inference algorithm. The variational inference formula is:

[0034]

[0035] where measures the approximation degree of the variational distribution to the true posterior and the data fitting degree, m is the sample index, x m is the process parameter data of the mth sample, t m is the treatment variable of the mth sample, y mis the carbon emission result of the m-th sample, k m is the latent variable of the m-th sample, is the mathematical expectation of the variational distribution, f(x m , t m |k m ) represents the joint distribution of process parameters and treatment variables, f(y m |t m , k m ) represents the conditional distribution of carbon emission results, r(k m |x m , t m , y m ) represents the variational distribution;

[0036] According to the mean difference of the two groups of carbon emissions, calibrate the mean difference using the regression coefficient, obtain the causal effect strength of the candidate threshold using the product of the regression coefficient and the mean difference, compare the causal effect strengths of the candidate thresholds, select the threshold that maximizes the difference in carbon emissions between the treatment group and the control group for node splitting. After training a single decision tree is completed, perform an arithmetic average on the predicted values of all decision trees to obtain the causal effect estimate value;

[0037] Apply perturbations to the process parameters based on their existing values, predict the carbon emissions per unit product before and after the perturbations through the causal forest model to obtain the difference before and after. Repeat the perturbations until the standard deviation of the average change rate of the difference is less than one percent to obtain the marginal effect coefficient. Use the Shapley value decomposition method to traverse all combinations of process parameters, calculate the marginal contribution of each process parameter when added to the parameter combination to the carbon emission prediction value, obtain the Shapley value based on the average of the marginal contributions of the process parameters, screen out the process nodes as the root nodes according to the size of the Shapley value, recursively divide the samples so that the mean carbon emissions per unit product under the process parameter combination corresponds to each leaf node, and concatenate the splitting conditions from the root node to each leaf node into a decision path to construct a causal tree model.

[0038] Furthermore, the method for constructing a carbon emission prediction model through weighted fusion based on the double-layer optimization fusion strategy includes:

[0039] Based on the time series prediction model, output the carbon emission trend prediction result, use the causal tree model to output the process parameter marginal effect coefficient. Design a multi-objective optimization function through the double-layer optimization fusion strategy, use the sequential least squares method to solve the multi-objective optimization function to obtain the weights of the time series prediction result and the weights of the causal tree model result, add the time series prediction result and the influence of the process parameters of the causal tree model according to the weights to obtain the carbon emission prediction value, and construct a carbon emission prediction model. The formula of the multi-objective optimization function is:

[0040]

[0041] where n is the total number of samples, is the predicted value of the i-th sample, the true value of the i-th sample, m is the number of process parameters, β j is the marginal effect coefficient of the j-th process parameter, α is the weight for balancing prediction accuracy and causal stability, ω1, ω2 are the weights of the time series prediction result and the causal tree model result;

[0042] Use the carbon emission prediction model to output a visual prediction chart, and optimize the carbon emission prediction model based on the root mean square error through the test set.

[0043] The second aspect of the present invention provides a system for a carbon emission prediction model based on artificial intelligence, including:

[0044] A data acquisition model that acquires enterprise production data and preprocesses it to obtain time series data. The enterprise production data includes carbon emission data and production data;

[0045] A data processing module that uses a sliding window to segment the time series data based on the time series data, calculates the derivative features of the sample data, aligns the time series data and the derivative features by window to obtain a feature data set, and divides the feature data set into a training set, a validation set, and a test set;

[0046] A model training module that inputs the training set into the LSTM-attention model for training to generate a time series prediction model, inputs the validation set into the time series prediction model, optimizes the weight parameters of the attention layer based on the prediction error, extracts the process parameters and carbon emission data from the training set and inputs them into the causal tree model, extracts the distribution parameters through the variational inference algorithm, and obtains the marginal effect coefficient of the process parameters based on the distribution parameters;

[0047] A model construction module that constructs a carbon emission prediction model through weighted fusion based on a two-layer optimization fusion strategy according to the carbon emission trend prediction result output by the time series prediction model and the marginal effect coefficient of the process parameters output by the causal tree model, outputs a visual result, and optimizes the carbon emission prediction model based on the root mean square error through the test set.

[0048] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0049] (1) By using a sliding window to cut the time series data and calculate the derivative features, and combining Gaussian noise injection to enhance data diversity, the present invention improves the robustness of the model to noise data. At the same time, by hierarchically dividing the training set, validation set, and test set, the stability and generalization ability of model training are ensured.

[0050] (2) By constructing an LSTM-attention model and a time series prediction model, and combining the marginal effect analysis of the causal tree model, the double-layer optimization fusion strategy realizes the collaborative optimization of the prediction trend and the influence of process parameters, enhancing the adaptability of the model to abnormal scenarios while ensuring the prediction accuracy.

[0051] (3) Through the causal tree model based on variational inference and Shapley value decomposition, the present invention quantifies the marginal effect of process parameters on carbon emissions, reveals the causal relationship between parameter adjustment and carbon emission changes, and provides interpretable decision-making support for enterprises to optimize production processes and formulate emission reduction strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of the steps of the method for constructing an artificial intelligence-based carbon emission prediction model in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] Referring to Figure 1 as shown, the present invention provides a method for constructing an artificial intelligence-based carbon emission prediction model, including:

[0055] A. Collect enterprise production data and preprocess it to obtain time series data, where the enterprise production data includes carbon emission data and production data;

[0056] Specifically, the method for collecting enterprise production data and preprocessing it to obtain time series data includes: The enterprise production data includes carbon emission data and production data. The production data includes using industrial Internet of Things sensors to collect the equipment model, operating status, and process parameters of enterprise production equipment in real time and associating product batches with production volumes, collecting equipment energy consumption data, carbon emission data, and carbon emission standards in the production process, obtaining equipment carbon emission factors according to equipment types and carbon emission standards, forming an original data set with time stamps, cleaning the original data set, merging duplicate records based on batch numbers and time stamps, generating a time axis at a fixed interval of 15 minutes, aggregating and sampling high-frequency data in the original data set according to the time axis, and associating low-frequency data to the nearest time point on the time axis; if there is no corresponding low-frequency data at a certain time point, use the adjacent value to fill it to obtain time series data.

[0057] In a specific embodiment, based on the collection of production data in the hot rolling workshop of a steel enterprise, K-type thermocouple sensors are installed in the middle, upper and lower parts of the heating furnace cavity to collect temperature every 10 seconds to monitor heating uniformity, a magnetoelectric speed sensor is installed on the main shaft of the rolling mill to collect speed every second, an ultrasonic gas flow meter is installed on the natural gas pipeline to record the fuel consumption rate at intervals of one second, and the production line automatically numbers each batch of steel billets and records the production time information, and collects the qualified product rate; an infrared gas analyzer is used to monitor the CO2 concentration of the exhaust gas of the heating furnace and collect emissions in combination with the flow rate, and the energy consumption is converted into carbon emissions according to the "Guidelines for the Calculation and Reporting of Greenhouse Gas Emissions from China's Steel Production" using a carbon emission factor of 0.85 tons of CO2 / MWh to form a time-stamped original data set;

[0058] The original data set was hashed for deduplication, outliers were removed using the 3σ principle, the time axis was reconstructed at 15-minute intervals, temperature, mill speed, and fuel consumption were aggregated at 15-minute intervals, and previous data were used to fill in low-frequency data such as batch information when there was no update, resulting in 20,160 time series data.

[0059] B. Based on the time series data, use sliding window segmentation to obtain sample data, calculate the derived features of the sample data, align the time series data with the derived features by window, obtain the feature data set, and divide the feature data set into a training set, a validation set, and a test set;

[0060] Specifically, the method of aligning the time series data with the derived features by window to obtain the feature data set includes: based on the sampling interval of the time series data, setting the sliding window size to 30 sampling intervals, moving the window in sequence based on the sampling interval to cut the time series data into fixed sample lengths, sequentially generating an identifier for each sample window, and calculating the derived features for each sample window;

[0061] Based on carbon emission standards and carbon emission factors, we screen out the associated fields in the sample that are directly related to carbon emission prediction, calculate the Pearson correlation coefficient between the remaining fields and the carbon emission data, retain the indirect associated fields with an absolute value of the correlation coefficient greater than 30%, take the associated fields and the indirect associated fields as the basic fields of the sample, and extract the sample data based on the basic fields to obtain the original time series features;

[0062] The original time series features and the derived features are aligned according to the sample window identifier and merged into the same data table. Each row represents a sample and each column represents the basic field of the sample to obtain a feature data set.

[0063] In a specific embodiment, based on the 15-minute sampling interval of the time series data, the sliding window size is set to 30 sampling intervals, i.e., 7.5 hours. 7.5 hours can basically cover one production shift, generating 20,131 samples. 26 features are selected by combining carbon emission standards and Pearson correlation coefficients, including 5 directly related features, 2 indirectly related features, and 19 original time series features. After aligning the original time series features and derived features according to the sample window identifier, they are merged into the same data table. Each row represents a sample, and each column represents the basic fields of the sample, obtaining a feature dataset.

[0064] Specifically, for each sample window, the method for calculating the derived features includes: using a sliding window to segment the time series data to obtain sample data, obtaining the actual running time based on the device running state of the sample data, determining the planned production time in combination with the product batch, and dividing the actual running time of the device by the planned production time to obtain the time utilization rate;

[0065] Determining the actual production speed based on the device running state, actual running time, and output, obtaining the theoretical production speed based on the device model, and dividing the actual production speed of the device by the theoretical production speed to obtain the performance utilization rate;

[0066] Counting the number of good products based on the product batch and dividing by the output to obtain the good product rate, calculating the product of the good product rate, time utilization rate, and performance utilization rate to obtain the overall equipment effectiveness (OEE); dividing the energy consumption data by the output to obtain the energy consumption per unit product, and multiplying the energy consumption per unit product by the carbon emission factor to obtain the carbon emissions per unit product.

[0067] In a specific embodiment, based on the device running state of the sample data, the running time of the heating furnace is obtained as 6.5 hours, the planned production time is determined as 7.5 hours in combination with the product batch, and the actual running time of the device is divided by the planned production time to obtain a time utilization rate of 85%;

[0068] Based on the device running state, actual running time of 6.5 hours, and output of 520 tons, the actual production speed is determined as 80 tons / hour. Based on the device model, the theoretical production speed is obtained as 100 tons / hour, and the actual production speed of the device is divided by the theoretical production speed to obtain a performance utilization rate of 80%;

[0069] Counting the number of good products based on the product batch and dividing by the output of 520 tons to obtain a good product rate of 95%, calculating the product of the good product rate, time utilization rate, and performance utilization rate to obtain an OEE of 64.6%;

[0070] Using the energy consumption data divided by the output of 520 tons to obtain an energy consumption per unit of 6.25 m 3 / ton, multiplying the energy consumption per unit product by the carbon emission factor of 0.85 tons CO2 / MWh to obtain a carbon emissions per unit of 5.3125 tons CO2.

[0071] Specifically, the method for dividing the feature dataset into a training set, a validation set, and a test set includes: performing noise optimization and hierarchical division on the feature dataset; the noise optimization includes calculating the mean and variance of the time series data in the feature dataset, using the mean of the time series data as the mean of the Gaussian distribution noise, selecting 27% of the variance of the original time series data as the variance of the Gaussian distribution noise, and superimposing Gaussian distribution noise on the time series data through the formula X′ = X + ∈·N(0,σ 2 ), where X′ represents the enhanced time series data, X represents the time series data, ∈ represents the intensity coefficient, and N(0,σ 2 ) represents the probability distribution;

[0072] Divide the data hierarchy according to the production shift. For each level, generate the training set, validation set, and test set according to the ratio of 7:2:1.

[0073] In a specific embodiment, noise optimization is performed on the temperature. The Gaussian noise formula X′ = X + ∈·N(0,σ 2 ) is used. The mean is taken as the temperature mean of 1200 °C, and the variance is 27. The noise amplitude is controlled by the intensity coefficient ∈ = 0.1, and the original temperature value of 1210 °C is enhanced to 1208.5 °C to simulate measurement errors. At the same time, hierarchical division is performed based on the three-shift production schedule. For each level, division is performed according to 7:2:1, obtaining 14,092 samples in the training set, 4,026 samples in the validation set, and 2,013 samples in the test set, covering data from different shifts to ensure the generalization of the model.

[0074] C Input the training set into the LSTM-attention model for training to generate a time series prediction model. Use the validation set to input the time series prediction model, optimize the weight parameters of the attention layer based on the prediction error. At the same time, extract process parameters and carbon emission data from the training set and input them into the causal tree model, extract distribution parameters through the variational inference algorithm, and obtain the marginal effect coefficient of the process parameters based on the distribution parameters;

[0075] Specifically, the method for inputting the training set into the LSTM-attention model for training to generate a time series prediction model includes: converting the training set into a three-dimensional tensor, the three-dimensional tensor includes the number of samples, the number of time steps, and the number of features. Set the carbon emission target value as the output data based on the carbon emission data of the training set, perform min-max normalization on the three-dimensional tensor and the carbon emission target value, and use the normalized three-dimensional tensor as the input data of the LSTM-attention model;

[0076] Stack two layers of bidirectional LSTM units to extract the basic data at each time step of the input data. Generate position encoding for each time step of the input data through a sine function, add the position encoding to the basic data element by element, input the time step sequence fused with the position encoding into the bidirectional LSTM for processing, obtain the hidden state sequence containing the feature information of each time step, and calculate the attention weight matrix of the hidden state sequence. The formula for the attention matrix is:

[0077]

[0078] where β n,τ,τ′ is the attention weight of time step τ to τ′ in the nth sample. Query n,τ is the query vector of the nth sample at time step τ. Key n,τ is the key vector of the nth sample at time step τ. PosBias τ,τ′ is the position bias of time step τ to τ′. d k is the dimension corresponding to the scaling factor. k is the abbreviation of Key. T is the total number of time steps. τ and τ′ are time step indices, and τ″ is the time step index in the summation;

[0079] Weight the key time steps of the hidden state sequence based on the attention weight matrix to obtain the attention output, map and transform the attention output through a linear activation function to obtain a 1D carbon emission prediction value, and construct a time series prediction model;

[0080] Convert the validation set into a three-dimensional tensor and normalize it, then input it into the time series prediction model to obtain the mean absolute error between the validation set prediction result and the carbon emission target value. Based on the error, execute the backpropagation algorithm to derive the gradients of the parameters in the attention weight matrix, and adjust the time step weight parameters according to the gradient direction and magnitude.

[0081] In a specific embodiment, convert the training set into a three-dimensional tensor, construct input data with an input dimension of 14092 sample numbers of the training set × 30 time steps × 26 features. Based on the CO2 emissions per unit time in the training set, set the output target value of the corresponding sample to 15 tons of CO2 / hour. After performing min-max normalization on the input data and output data, input them into the LSTM-attention model;

[0082] Set the training parameters of the LSTM-attention model as the learning rate of 0.001, paired with the Adam optimizer, iterate for 100 epochs, and the loss function where N is the number of samples in the validation set, is the model prediction value, y i is the true value;

[0083] The model uses two stacked bidirectional LSTM layers with 128 units to capture temporal dependencies and extract the basic data at each time step of the input data. Position encoding is generated for each time step of the input data through sine and cosine functions, and the function formula is:

[0084]

[0085] where pos is the time step position, i is the feature dimension index, and d model is the total dimension of the model features;

[0086] The first, second, etc. time steps are sequentially sampled from the samples. Among the 26-dimensional features, the 0th and 1st dimensions are calculated using sine, and the 2nd and 3rd dimensions are calculated using cosine, alternating. The generated position encoding, such as (30, 26), is added element-wise to the basic data. The time step sequence fused with the position encoding is input into the bidirectional LSTM for processing to obtain a hidden state sequence containing the feature information of each time step. For example, calculate the attention weight of time step 5 to time step 10 in sample 1 to obtain the attention weight β 1,5,10 ≈0.00389. The closer the weight value is to 1, the larger it is. An attention matrix is constructed based on the attention weights, and the key time step "high-speed operation stage of the rolling mill" in the hidden state sequence is weighted based on the attention weight matrix to obtain the attention output [0.1233, 0.0450, -0.1546, 0.1023, 0.0022, 0.0058, 0.0173, -0.0878, 0.1013, -0.0624]. The attention output is mapped and transformed through the ReLU linear activation function, and the output is connected to the fully connected layer. Finally, a 1D carbon emission prediction value of 0.734 tons CO2 / hour is output to construct a time series prediction model;

[0087] The 4026 samples of the normalized validation set are converted into three-dimensional tensors and then input into the trained time series prediction model to obtain the carbon emission prediction results. The mean absolute error MAE is 0.38 tons CO2 / hour through the loss function. Based on the calculated error, the backpropagation algorithm is executed. Starting from the output layer of the model and deriving backward, it is calculated that the temperature feature of the time step in the 15th sample in the attention weight matrix contributes more to the error. The gradient value of its corresponding weight parameter is derived through the chain rule of differentiation to be approximately 0.02, and the weight of this time step is reduced.

[0088] Specifically, a method for extracting process parameters and carbon emission data from a training set and inputting them into a causal tree model, and obtaining the marginal effect coefficient of process parameters based on distribution parameters through a variational inference algorithm includes: extracting the process parameters and carbon emissions per unit product of each column as sample features from the training set data table, extracting each row of data as a sample, constructing a decision tree, iterating the decision tree, randomly selecting 67% of the total number of samples as the current sample each time the decision tree iterates, and taking the square root of the total number of sample features in the current iteration to obtain the number of features;

[0089] Traverse the value range of the process parameters through a splitting algorithm, include the actual value range and the intermediate value of adjacent values in the candidate threshold list, divide the current sample into a treatment group and a control group according to the candidate threshold, and obtain the distribution parameters and regression coefficients of the causal effect intensity based on the treatment group and the control group through a variational inference algorithm. The variational inference formula is:

[0090]

[0091] where measures the approximation degree of the variational distribution to the true posterior and the data fitting degree, m is the sample index, x m is the process parameter data of the m-th sample, t m is the treatment variable of the m-th sample, y m is the carbon emission result of the m-th sample, k m is the latent variable of the m-th sample, is the mathematical expectation of the variational distribution, f(x m ,t m |k m ) represents the joint distribution of the process parameter and the treatment variable, f(y m |t m ,k m ) represents the conditional distribution of the carbon emission result, r(k m |x m ,t m ,y m ) represents the variational distribution;

[0092] According to the mean difference of the carbon emissions of the two groups, calibrate the mean difference using the regression coefficient, obtain the causal effect intensity of the candidate threshold using the product of the regression coefficient and the mean difference, compare the causal effect intensities of the candidate thresholds, select the threshold that maximizes the difference in carbon emissions between the treatment group and the control group for node splitting. After training a single decision tree is completed, take the arithmetic mean of the predicted values of all decision trees to obtain the causal effect estimate value;

[0093] Apply perturbations to the process parameters based on their existing values. Predict the carbon emissions per unit product before and after the perturbations using the causal forest model, and obtain the difference before and after. Repeat the perturbations until the standard deviation of the average change rate of the difference is less than one percent to obtain the marginal effect coefficient. Use the Shapley value decomposition method to traverse all combinations of process parameters, calculate the marginal contribution of each process parameter when added to the parameter combination to the predicted carbon emission value, obtain the Shapley value based on the average of the marginal contributions of the process parameters, screen out the process nodes as the root nodes according to the size of the Shapley value, recursively divide the samples so that each leaf node corresponds to the average value of the carbon emissions per unit product under the combination of process parameters, and concatenate the splitting conditions from the root node to each leaf node into a decision path to construct a causal tree model.

[0094] It should be noted that in the variational inference of the causal tree model, the latent variable k m represents the influence coefficient of a certain process parameter on carbon emissions, that is, the regression coefficient. Through the variational inference algorithm, we approximate the posterior distribution of the latent variable k m . First, assume that k m follows a certain prior distribution, and then use the variational distribution to approximate the true posterior distribution of k m . By optimizing the variational parameters μ and σ and minimizing the difference between the variational distribution and the true posterior, the finally obtained k m is the marginal effect coefficient of the process parameter, which is used to quantify the impact of process parameter changes on carbon emissions.

[0095] In a specific embodiment, extract temperature, rotational speed, fuel consumption, OEE parameters, and carbon emissions per unit product from the training set as sample features, extract each row of data as a sample, construct a decision tree, select 9432 samples for the current iteration, obtain 2 sample features of temperature and fuel consumption for splitting. Traverse the value range of the process parameter such as temperature, 1000 - 1300 °C, generate intermediate values 1050 °C and 1100 °C as candidate thresholds, and divide them into a treatment group with a temperature greater than 1100 °C and a control group with a temperature less than 1100 °C according to the threshold. Obtain the average carbon emission of the treatment group as 12 tons per hour and that of the control group as 10 tons per hour. Based on the treatment group and the control group, use the variational inference algorithm to obtain the mean μ = 1.5 and standard deviation σ = 0.2 of the distribution parameters of the causal effect intensity, and the regression coefficient k m = 1.8. According to the mean difference of 2 tons per hour in carbon emissions between the two groups, use the product of the regression coefficient and the mean difference to obtain the causal effect intensity of the candidate threshold as 3.6 tons per hour. Compare the causal effect intensities of the candidate thresholds, and determine the threshold 1100 °C that maximizes the difference as the splitting node. After completing the training of a single decision tree, take the arithmetic mean of the predicted values of 100 decision trees to obtain the estimated value of the causal effect of the production conditions in the steel hot rolling production scenario as 13.6 tons per hour;

[0096] It should be noted that the marginal effect coefficient obtained from the perturbation experiment is used as the input of the causal tree model, which can be used to correct the time series prediction results. The Shapley value originates from cooperative game theory, and its core idea is to fairly allocate the contribution of each participant to the overall result. In the causal tree model, the Shapley value attributes the change in carbon emissions to specific process parameters. For example, in the hot rolling of steel, through calculation, it is found that the contribution of temperature is 0.12 tons / hour and the contribution of rotational speed is 0.08 tons / hour, enabling engineers to intuitively understand how temperature and rotational speed jointly affect carbon emissions and preventing the model from becoming a black box. After quantifying the parameter contributions, high-impact parameters can be preferentially optimized. For instance, if the Shapley value of temperature is significantly higher than that of rotational speed, then focus on the temperature control of the production process to improve the emission reduction efficiency. In the production scenario with multi-parameter interaction, the Shapley value can decompose the joint influence of parameters. Even when temperature and rotational speed change simultaneously, it can clearly show their independent contributions to carbon emissions. Finally, by combining the Shapley value with the decision-making path of the causal tree model, for the process parameters arranged in descending order of the Shapley value, the process parameter with the largest Shapley value is preferentially selected as the root node of the causal tree model to ensure that the splitting condition has the greatest impact on carbon emissions.

[0097] In a specific embodiment, in the initial production scenario, when the temperature is 1200 °C, the carbon emissions per unit product are 10 tons / hour. A perturbation is applied to the temperature. When the temperature rises to 1201 °C, the predicted carbon emissions through the causal forest model are 10.08 tons / hour, and the difference is 0.08 tons / hour. Repeat the perturbation to adjust the temperature to 1202 °C and 1199 °C multiple times, and calculate the average change rate of the difference until its standard deviation is less than 1%, obtaining a marginal effect coefficient of 0.08 tons of CO2 per degree. If only the rotational speed is adjusted, the marginal contribution of the rotational speed is obtained as 0.05 tons of CO2 / hour. Traverse the process parameter combinations to calculate the marginal contributions and obtain the Shapley values. For example, for the combination of temperature and rotational speed, only the temperature contributes 0.4, and the total contribution is 0.6 after adding the rotational speed, with the marginal contribution of the rotational speed being 0.2. Calculate the Shapley values based on the mean of the marginal contributions: temperature 0.45, fuel consumption 0.32, rotational speed 0.23. Using the temperature with the larger Shapley value as the root node, recursively divide the samples. When the temperature is greater than 1100 °C and the fuel consumption is greater than 50 m 3 / h, the corresponding mean is 12 tons / hour. String together the above splitting conditions into a decision-making path to construct a causal tree model.

[0098] D Based on the carbon emission trend prediction results output by the time series prediction model and the process parameter marginal effect coefficients output by the causal tree model, a carbon emission prediction model is constructed through weighted fusion based on a two-layer optimization fusion strategy, and a visual prediction chart is output. The carbon emission prediction model is optimized using the test set through the root mean square error.

[0099] It should be noted that the search range of α can be initially defined as [0.1, 0.3, 0.5, 0.6, 0.8] according to business requirements. For each value of α, the multi-objective optimization function is solved by combining the sequential least squares method, and the root mean square error on the validation set is calculated. The value of α that minimizes the root mean square error is selected. Therefore, α = 0.6 is determined. In subsequent calculations, the search range and search accuracy of α can be gradually improved. The sequential least squares method optimizes through iteration. Each time, other variables of the multi-objective optimization function are fixed, and a single weight is updated to gradually approach the minimum value of the objective function. This iterative strategy can handle the non-convexity and non-linearity of complex functions. Initialize a value for ω1 and ω2, fix other variables, and optimize a single weight in turn. For example, first fix ω2, take the partial derivative of ω1 and update it; then fix ω1 and ω2, and repeat the iteration until the objective function converges. Finally, ω1 and ω2 are obtained, so that the prediction model can simultaneously focus on prediction accuracy and causal stability, and realize the weighted fusion of the time series prediction result and the causal tree model result.

[0100] Specifically, a method for constructing a carbon emission prediction model by weighted fusion based on a double-layer optimization fusion strategy includes: outputting a carbon emission trend prediction result based on a time series prediction model, using a causal tree model to output the marginal effect coefficient of process parameters, designing a multi-objective optimization function through a double-layer optimization fusion strategy, using the sequential least squares method to solve the multi-objective optimization function, obtaining the weight of the time series prediction result and the weight of the causal tree model result, adding the time series prediction result and the influence of the process parameters of the causal tree model according to the weights to obtain a carbon emission prediction value, and constructing a carbon emission prediction model. The formula of the multi-objective optimization function is:

[0101]

[0102] where n is the total number of samples, is the predicted value of the i-th sample, the true value of the i-th sample, m is the number of process parameters, β j is the marginal effect coefficient of the j-th process parameter, α is the weight for weighing prediction accuracy and causal stability, and ω1 and ω2 are the weights of the time series prediction result and the causal tree model result;

[0103] Output a visual prediction chart using the carbon emission prediction model, and optimize the carbon emission prediction model based on the root mean square error on the test set.

[0104] In a specific embodiment, based on the output of the time series prediction model, the carbon emission trend prediction result is 0.734 tons of CO2 per hour. Using the causal tree model, the marginal effect coefficient of the rolling mill roll speed is 0.08 tons of CO2 per degree. Through the double-layer optimization fusion strategy, a multi-objective optimization function is designed. Setting α = 0.6, based on the total number of samples 14092, a multi-objective optimization function is constructed. Using the sequential least squares method to solve the multi-objective optimization function, the weight ω1 of the time series prediction result is obtained as 0.7, and the weight ω2 of the causal tree model result is 0.3. Adding the time series prediction result and the influence of the process parameters of the causal tree model according to the weights, the carbon emission prediction value is obtained as 0.5378 tons of CO2 per hour, and the construction of the carbon emission prediction model is completed;

[0105] Use the carbon emission prediction model to output the visualization result;

[0106] Input the test set into the carbon emission prediction model, calculate the root mean square error as 0.071 tons per hour, which does not reach the expected accuracy of 0.05 tons per hour, and optimize the prediction model by adjusting α, ω1, and ω2.

[0107] The second aspect of the present invention also provides a construction method of a carbon emission prediction model based on artificial intelligence, including: a data acquisition model that acquires enterprise production data and preprocesses it to obtain time series data, where the enterprise production data includes carbon emission data and production data;

[0108] A data processing module that uses a sliding window to segment the time series data based on the time series data to calculate the derivative features of the sample data, aligns the time series data and the derivative features according to the window to obtain a feature data set, and divides the feature data set into a training set, a validation set, and a test set;

[0109] A model training module that inputs the training set into the LSTM-attention model for training to generate a time series prediction model, inputs the validation set into the time series prediction model, optimizes the weight parameters of the attention layer based on the prediction error, extracts process parameters and carbon emission data from the training set and inputs them into the causal tree model, extracts distribution parameters through the variational inference algorithm, and obtains the marginal effect coefficient of the process parameters based on the distribution parameters;

[0110] A model construction module that, according to the carbon emission trend prediction result output by the time series prediction model and the marginal effect coefficient of the process parameters output by the causal tree model, constructs a carbon emission prediction model through weighted fusion based on the double-layer optimization fusion strategy, outputs a visualization prediction chart, and optimizes the carbon emission prediction model using the test set through the root mean square error.

[0111] The above content is only an example and explanation of the structure of the present invention. Those skilled in the art of this technology can make various modifications or supplements to the described specific embodiments or use similar methods for substitution. As long as they do not deviate from the structure of the invention or exceed the scope defined by this claims, they shall fall within the protection scope of the present invention.

Claims

1. A method for constructing an artificial intelligence-based carbon emission prediction model, characterized in that, It includes the following steps: A. Collect enterprise production data and perform preprocessing to obtain time-series data. The enterprise production data includes carbon emission data and production data. B. Based on the time-series data, use a sliding window to segment and obtain sample data. Calculate the derivative features of the sample data, align the time-series data with the derivative features according to the window, obtain a feature dataset, and divide the feature dataset into a training set, a validation set, and a test set. C. Input the training set into the LSTM-attention model for training to generate a time-series prediction model. Input the validation set into the time-series prediction model, optimize the weight parameters of the attention layer based on the prediction error. At the same time, extract process parameters and carbon emission data from the training set and input them into the causal tree model. Extract distribution parameters through the variational inference algorithm, and obtain the marginal effect coefficient of the process parameters based on the distribution parameters. D. According to the carbon emission trend prediction result output by the time-series prediction model and the marginal effect coefficient of the process parameters output by the causal tree model, construct a carbon emission prediction model through weighted fusion based on a two-layer optimization fusion strategy, output a visualization result, and optimize the carbon emission prediction model using the test set through the root mean square error.

2. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, wherein The method for collecting enterprise production data and performing preprocessing to obtain time-series data includes: The enterprise production data includes carbon emission data and production data. The production data includes using industrial Internet of Things sensors to collect the equipment model, operating status, and process parameters of enterprise production equipment in real time, and associating product batches and yields. Collect equipment energy consumption data, carbon emission data, and carbon emission standards in the production process, obtain equipment carbon emission factors according to the equipment type and carbon emission standards, form an original dataset with timestamps, clean the original dataset, merge duplicate records based on batch numbers and timestamps, generate a time axis at a fixed interval of 15 minutes, perform aggregated sampling on the high-frequency data in the original dataset according to the time axis, and associate the low-frequency data with the nearest time point on the time axis. If there is no corresponding low-frequency data at a certain time point, use the neighboring value to fill it to obtain time-series data.

3. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, characterized in that The method for aligning the time-series data with the derivative features according to the window to obtain a feature dataset includes: Based on the sampling interval of the time-series data, set the sliding window size to 30 sampling intervals. Move the window sequentially based on the sampling interval to cut the time-series data into a fixed sample length, generate an identifier for each sample window in sequence, and calculate the derivative features for each sample window. Screen out the associated fields in the sample that are directly related to carbon emission prediction based on the carbon emission standard and carbon emission factor. Calculate the Pearson correlation coefficient between the remaining fields and the carbon emission data, and retain the indirectly associated fields with the absolute value of the correlation coefficient greater than 30%. Use the associated fields and indirectly associated fields as the basic fields of the sample, and extract sample data based on the basic fields to obtain the original time-series features. Align the original time-series features and the derivative features according to the sample window identifier and merge them into the same data table. Each row represents a sample, and each column represents the basic fields of the sample to obtain a feature dataset.

4. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, characterized in that The method for calculating the derivative features for each sample window includes: Sample data is obtained by using a sliding window segmentation based on time series data. According to the device running state of the sample data, the actual running time is obtained, and the planned production time is determined in combination with the product batch. The time utilization rate is obtained by dividing the actual running time of the device by the planned production time; The actual production speed is determined according to the device running state, actual running time and output. The theoretical production speed is obtained based on the device model. The performance utilization rate is obtained by dividing the actual production speed of the device by the theoretical production speed; The number of good products is counted based on the product batch and divided by the output to obtain the good product rate. The overall equipment efficiency is obtained by calculating the product of the good product rate, time utilization rate and performance utilization rate; The unit product energy consumption is obtained by dividing the energy consumption data by the output, and the unit product carbon emission is obtained by multiplying the unit product energy consumption by the carbon emission factor.

5. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, wherein, The method for dividing the feature data set into a training set, a validation set, and a test set includes: Performing noise optimization and hierarchical partitioning on the feature data set; The noise optimization includes calculating the mean and variance of the time series data in the feature dataset, taking the mean value of the time series data as the mean of the Gaussian distribution noise, selecting 27% of the variance of the original time series data as the variance of the Gaussian distribution noise, and superimposing the Gaussian distribution noise on the time series data through the formula X′ = X + ∈·N(0,σ 2 ), where X′ represents the enhanced time series data, X represents the time series data, ÷ represents the intensity coefficient, and N(0,σ 2 ) represents the probability distribution; The data levels are divided according to the production shift. For each level, the training set, validation set, and test set are generated according to the ratio of 7:2:

1.

6. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, characterized in that The method for inputting the training set into the LSTM-attention model for training to generate a time series prediction model includes: Converting the training set into a three-dimensional tensor, where the three-dimensional tensor includes the number of samples, the number of time steps, and the number of features. The carbon emission target value is set based on the carbon emission data of the training set as the output data. The three-dimensional tensor and the carbon emission target value are normalized by maximum and minimum normalization, and the normalized three-dimensional tensor is used as the input data of the LSTM-attention model; Stack two layers of bidirectional LSTM units to extract the basic data of each time step in the input data. Generate a position encoding for each time step in the input data through a sine function, add the position encoding and the basic data element by element, and input the time step sequence fused with the position encoding into the bidirectional LSTM for processing to obtain a hidden state sequence containing the feature information of each time step. Calculate the attention weight matrix of the hidden state sequence. The formula for the attention matrix is: where β n,τ,τ′ is the attention weight of time step τ to τ′ in the n-th sample, Query n,τ is the query vector of the n-th sample and time step τ, Key n,τ is the key vector of the n-th sample and time step τ, PosBias τ,τ′ is the position bias of time step τ to τ′, d k is the dimension corresponding to the scaling factor, k is the abbreviation of Key, T is the total number of time steps, τ and τ′ are time step indices, τ ″ is the time step index in the summation; Weight the key time steps of the hidden state sequence based on the attention weight matrix to obtain an attention output. Map and transform the attention output through a linear activation function to obtain a 1D carbon emission prediction value, and construct a time series prediction model; Convert the validation set into a three-dimensional tensor and normalize it, then input it into the time series prediction model to obtain the mean absolute error between the validation set prediction result and the carbon emission target value. Based on the error, execute the backpropagation algorithm to deduce the gradients of the parameters in the attention weight matrix, and adjust the time step weight parameters according to the gradient direction and magnitude.

7. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, characterized in that The method for extracting process parameters and carbon emission data from the training set and inputting them into the causal tree model, and extracting distribution parameters through the variational inference algorithm and obtaining the marginal effect coefficient of the process parameters based on the distribution parameters includes: Extract the process parameters of each column and the unit product carbon emission in the training set data table as sample features, extract each row of data as a sample, construct a decision tree, iterate the decision tree, and randomly select 67% of the total number of samples as the current sample each time the decision tree iterates. Take the square root of the total number of sample features in the current iteration to obtain the number of features; Traverse the value range of process parameters through the splitting algorithm, incorporate the actual value range and the intermediate values of adjacent values into the candidate threshold list, divide the current sample into a treatment group and a control group according to the candidate threshold, and obtain the distribution parameters and regression coefficients of the causal effect intensity through the variational inference algorithm based on the treatment group and the control group. The variational inference formula is as follows: Among them Measure the approximation degree of the variational distribution to the true posterior and the data fitting degree. m is the sample index, and x m is the process parameter data of the m-th sample, t m is the treatment variable of the m-th sample, y m is the carbon emission result of the m-th sample, k m is the latent variable of the m-th sample, is the mathematical expectation of the variational distribution, f(x m ,t m |k m ) represents the joint distribution of process parameters and treatment variables, f(y m |t m ,k m ) represents the conditional distribution of carbon emission results, r(k m |x m ,t m ,y m ) represents the variational distribution; According to the mean difference of carbon emissions between the two groups, calibrate the mean difference using the regression coefficient, obtain the causal effect intensity of the candidate threshold using the product of the regression coefficient and the mean difference, compare the causal effect intensities of the candidate thresholds, and select the threshold that maximizes the difference in carbon emissions between the treatment group and the control group for node splitting. After training a single decision tree, perform an arithmetic average on the predicted values of all decision trees to obtain the causal effect estimate; Apply perturbations to the process parameters based on their existing values, predict the carbon emissions per unit product before and after the perturbations through the causal forest model, obtain the difference before and after, repeat the perturbations until the standard deviation of the average change rate of the difference is less than one percent, obtain the marginal effect coefficient, use the Shapley value decomposition method to traverse all combinations of process parameters, calculate the marginal contribution of each process parameter to the carbon emission prediction value when it is added to the parameter combination, obtain the Shapley value based on the average value of the marginal contributions of the process parameters, screen out the process nodes as the root nodes according to the magnitude of the Shapley values, recursively divide the samples, make the average carbon emissions per unit product under the process parameter combination corresponding to each leaf node, and concatenate the splitting conditions from the root node to each leaf node into a decision path to construct a causal tree model.

8. The method for constructing an artificial intelligence-based carbon emission prediction model according to claim 1, characterized in that The method for constructing a carbon emission prediction model through weighted fusion based on the double-layer optimization fusion strategy includes: Based on the output of the time series prediction model for the carbon emission trend prediction result, use the causal tree model to output the process parameter marginal effect coefficient, design a multi-objective optimization function through the double-layer optimization fusion strategy, use the sequential least squares method to solve the multi-objective optimization function, obtain the weights of the time series prediction result and the result of the causal tree model, add the time series prediction result and the influence of the process parameters of the causal tree model according to the weights to obtain the carbon emission prediction value, and construct a carbon emission prediction model. The multi-objective optimization function formula is as follows: where n is the total number of samples, is the predicted value of the i-th sample, the true value of the i-th sample, m is the number of process parameters, β j is the marginal effect coefficient of the j-th process parameter, α is the weight for balancing prediction accuracy and causal stability, ω1, ω2 are the weights of the time series prediction result and the causal tree model result; Use the carbon emission prediction model to output a visual prediction chart, and optimize the carbon emission prediction model based on the root mean square error of the test set.

9. A system for an artificial intelligence-based carbon emission prediction model, which is used to execute the method for constructing an artificial intelligence-based carbon emission prediction model according to any one of claims 1 to 8, characterized in that, The system includes: A data acquisition model that collects enterprise production data and preprocesses it to obtain time series data. The enterprise production data includes carbon emission data and production data; A data processing module that uses a sliding window to segment the time series data based on the time series data to calculate the derived features of the sample data, align the time series data and the derived features according to the window to obtain a feature data set, and divide the feature data set into a training set, a validation set, and a test set; A model training module that inputs the training set into the LSTM-attention model for training to generate a time series prediction model, inputs the validation set into the time series prediction model, optimizes the weight parameters of the attention layer based on the prediction error, extracts the process parameters and carbon emission data from the training set and inputs them into the causal tree model, extracts the distribution parameters through the variational inference algorithm, and obtains the process parameter marginal effect coefficient based on the distribution parameters; The model construction module constructs a carbon emission prediction model through weighted fusion based on a double-layer optimization fusion strategy according to the carbon emission trend prediction results output by the time series prediction model and the process parameter marginal effect coefficients output by the causal tree model, outputs visual prediction results, and optimizes the carbon emission prediction model using the test set through the root mean square error.

Images