Construction method of CO2 emission prediction model for iron and steel industry

Abstract

The invention discloses a construction method for a CO2 emission prediction model in the iron and steel industry. The construction method comprises the steps of data collection, cleaning and normalization processing; normalized CO2 emission data is input into an empirical mode decomposition (EMD) module, the non-stationarity of the data is reduced, and kernel principal component analysis (KPCA) processing is carried out on each IMF component to extract nonlinear features of the IMF component and reduce the dimensionality of the IMF component. And taking the IMF component after dimension reduction and the residual term as input, constructing a long short-term memory (LSTM) network for training and prediction, modeling and predicting the CO2 emission data after dimension reduction through an LSTM model obtained through training, and finally obtaining a prediction result of the CO2 emission. According to the method, the EMD method, the KPCA method and the LSTM method are combined, and the problems that according to an existing prediction method, data non-stationarity and non-linear feature mining is not deep, the feature extraction efficiency is low, and the long sequence data processing capacity is limited are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of industrial CO2 emission prediction technology, specifically a method for constructing a CO2 emission prediction model for the steel industry. Background Technology

[0002] As a typical example of high energy consumption and high emissions, the steel industry faces enormous pressure to reduce emissions. The steel industry contributes approximately 7% of global CO2 emissions, and this proportion is even higher than 15% in China. CO2 emissions from steel production mainly originate from blast furnace ironmaking, coking, and sintering. With the deepening of global climate governance, countries have introduced strict carbon emission policies, and steel companies urgently need to achieve green and low-carbon transformation through technological innovation and management optimization. Against this backdrop, accurately predicting the CO2 emissions of the steel industry has become a crucial step in formulating emission reduction strategies, optimizing production processes, and evaluating the effectiveness of policies.

[0003] Currently, CO2 emission prediction methods in the steel industry are mainly divided into two categories: mechanistic model-based and data-driven. Mechanism-based methods, such as material flow analysis and energy balance models, accurately calculate CO2 emissions by constructing the material and energy flow relationships in the steel production process. However, these methods require high data quality and specialized knowledge, and are difficult to adapt to complex and ever-changing actual production environments, limiting their widespread application. Data-driven methods, such as regression analysis, support vector machines (SVM), and artificial neural networks (ANN), establish a mapping relationship between input and output by learning from historical data, thereby achieving CO2 emission prediction. Among them, artificial neural networks have been widely used in CO2 emission prediction due to their powerful nonlinear fitting capabilities. However, traditional neural network models (such as BP neural networks) suffer from problems such as being prone to getting trapped in local optima and having insufficient ability to process long-sequence data, making it difficult to fully capture the complex characteristics of CO2 emission data. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a CO2 emission prediction model for the steel industry. By combining EMD, KPCA and LSTM, it effectively solves the problems of existing prediction methods, such as insufficient mining of non-stationary and nonlinear features, low feature extraction efficiency and limited ability to process long-sequence data.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for constructing a CO2 emission prediction model for the steel industry includes the following steps:

[0007] S1: Data collection and preprocessing, including the data source, data type, data time and data frequency, and then cleaning and normalizing the data;

[0008] S2: Input the normalized CO2 emission data into the Empirical Mode Decomposition (EMD) module to decompose the complex non-stationary signal into multiple Intrinsic Mode Function (IMF) components and a residual term, thereby reducing the non-stationarity of the data and providing better input for subsequent feature extraction and modeling.

[0009] S3: Perform kernel principal component analysis (KPCA) on each IMF component to extract its nonlinear features and reduce dimensionality;

[0010] S4: Using the dimensionality-reduced IMF components and residual terms as input, a Long Short-Term Memory (LSTM) network is constructed for training and prediction. The trained LSTM model is then used to model and predict the dimensionality-reduced CO2 emission data, ultimately yielding the predicted CO2 emissions.

[0011] Furthermore, the data sources in S1 include the company's internal production management system, energy management system, and environmental monitoring system; the data types include:

[0012] Production data: steel output, output of each production process, and product types;

[0013] Energy consumption data: Consumption of coal, coke, electricity, and natural gas;

[0014] CO2 emission monitoring data: The CO2 emissions from each emission source can be calculated by monitoring the CO2 concentration and emission flow rate in the emitted gas;

[0015] The collected data covers 3-5 years of historical data, spanning different production conditions and seasonal variations.

[0016] The frequency of data collection should be determined based on the actual situation, with monthly or quarterly data being preferable.

[0017] Furthermore, data cleaning in S1 includes:

[0018] Handling missing values: For missing data, interpolation can be used to fill in the missing values, or data from adjacent time periods can be used as substitutes.

[0019] Outlier removal: Outliers are identified by drawing data distribution charts and box plots. Outliers that exceed the reasonable range are removed or corrected.

[0020] Remove duplicate values: Check if there are duplicate records in the data. If so, delete the duplicate parts and keep the unique data.

[0021] Furthermore, the data normalization method in S1 is as follows: The range normalization method is used to map the data to the [0,1] interval, as shown in the following formula:

[0022]

[0023] Where x is the original data, x min and x max These are the minimum and maximum values ​​of the original data, x. norm It is the normalized value.

[0024] Furthermore, the specific steps of EMD decomposition in S2 are as follows:

[0025] S201: Initialization: Using the original signal x(t) as the initial signal, calculate its local maxima and local minima;

[0026] S202: Extracting IMF components: Using cubic spline interpolation, connect the local maxima and local minima respectively to form the upper envelope e. up (t) and lower envelope e down (t); Calculate the mean m1(t) of the upper and lower envelopes, and then subtract the mean from the original signal to obtain h1(t) = x(t) - m1(t); Check whether h1(t) satisfies the conditions of the IMF, that is, the number of local extrema and zero crossings is equal or differs by at most one in the entire data range, and the mean of the local maximum envelope and the local minimum envelope is zero at any time point; if satisfied, then h1(t) is the first IMF component c1(t); otherwise, take h1(t) as a new signal and repeat the above steps until an IMF component that satisfies the conditions is obtained;

[0027] S203: Update residuals: Subtract the extracted IMF components from the original signal to obtain the residual r1(t) = x(t) - c1(t);

[0028] S204: Repeated decomposition: Using the residual r1(t) as the new signal, repeat S201-S203 until the residual r n Until (t) becomes a monotonic function or no new IMF components can be extracted, the original signal is finally represented as Where c i (t) represents the i-th IMF component. r n(t) is the residual term.

[0029] Furthermore, the specific steps in S3 are as follows:

[0030] S301: Data Standardization: Standardize the data for each IMF component to have a mean of 0 and a variance of 1. The standardization formula is:

[0031]

[0032] Where, x ij It is the j-th feature value of the i-th sample, μ i and σ i These are the mean and standard deviation of the j-th feature, respectively. ij It is the standardized value;

[0033] S302: Selecting the kernel function and calculating the kernel matrix: The Gaussian kernel function is selected as the kernel function for KPCA, and its expression is:

[0034]

[0035] Where, x i and x j There are two samples. σ is the bandwidth parameter of the kernel function, used to control the width of the Gaussian kernel. Calculate the kernel matrix K, whose elements K... ij =K(x) i x j );

[0036] S303: Centralized Kernel Matrix: The kernel matrix is ​​centralized, resulting in a centralized kernel matrix K. c Calculated using the following formula:

[0037] K c =K-1 N K-K1 N +1 N K1 N

[0038] Among them, 1 N It is an N×N matrix, where each element is 0. N is the number of samples;

[0039] S304: Solving for eigenvalues ​​and eigenvectors: Solving for the centered kernel matrix K c eigenvalues ​​λ i and the corresponding eigenvector α i Satisfying K c α i =λ i α i ;

[0040] S305: Principal Component Selection: Select the top d largest eigenvalues ​​and their corresponding eigenvectors based on their magnitudes, where d is the dimension after dimensionality reduction. The value of d is typically determined by the contribution rate of the eigenvalues, calculated using the following formula:

[0041]

[0042] S306: Calculate the projection coefficients: For a new data point x, its projection coefficient z on the i-th principal component. i Calculated using the following formula:

[0043]

[0044] Where, α ji It is the j-th element of the i-th feature vector corresponding to the j-th sample;

[0045] S307: Obtain the dimensionality-reduced data: Combine the projection coefficients of all data points on the selected d principal components to obtain the dimensionality-reduced data matrix z, where each row represents the representation of a sample in the dimensionality-reduced feature space.

[0046] Furthermore, S4 uses the Keras deep learning framework to build an LSTM model, the structure of which is as follows:

[0047] Input layer: Used to input the dimensionality-reduced IMF components and residual terms. The input dimension is determined based on the number of features after KPCA dimensionality reduction.

[0048] LSTM layer: The number of hidden layer units is set to 64, and the activation function is ReLU. The LSTM layer is used to capture long-term dependencies in time series data and selectively remember or forget information.

[0049] Output layer: The output layer has 1 neuron and is used to predict CO2 emissions. The activation function is linear.

[0050] Furthermore, the training process of the LSTM network in S4 is as follows:

[0051] S401: Initialize parameters: Randomly initialize the weight matrix and bias terms in the LSTM network;

[0052] S402: Forward Propagation: Input the training data into the LSTM network sequentially, perform forward propagation calculations according to the LSTM model structure, and obtain the output at each time step;

[0053] S403: Calculate the loss function: Use the mean squared error (MSE) as the loss function to calculate the difference between the predicted and actual values. The formula for the loss function is:

[0054]

[0055] Among them, y i It is the actual value. These are the predicted values, and N is the number of samples.

[0056] S404: Backpropagation: Based on the gradient of the loss function, the weight matrix and bias terms in the LSTM network are updated using the backpropagation algorithm, including parameter updates using stochastic gradient descent (SGD) or Adam optimization algorithms.

[0057] S405: Iterative Training: Repeat the forward propagation, loss function calculation, and backpropagation steps until the preset number of training rounds is reached or the loss function converges.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] 1. The method for constructing a CO2 emission prediction model for the steel industry in this invention combines Empirical Mode Decomposition (EMD), Kernel Principal Component Analysis (KPCA), and Long Short-Term Memory Network (LSTM) to form a novel hybrid prediction model. This combination approach has not been reported in the field of CO2 emission prediction in the steel industry, providing a completely new technical solution for solving the CO2 emission prediction problem in the steel industry.

[0060] 2. The method for constructing a CO2 emission prediction model for the steel industry according to the present invention first uses EMD to decompose CO2 emission data to reduce data nonstationarity, and then uses KPCA to extract nonlinear features and reduce dimensionality of the decomposed IMF components. This deep integration approach can more effectively mine the potential information in the data, remove redundant information and noise, provide higher quality input data for the LSTM model, and improve the model's ability to capture complex features of CO2 emission data.

[0061] 3. The method for constructing a CO2 emission prediction model for the steel industry in this invention fully considers the non-stationarity and nonlinearity of CO2 emission data in the steel industry, as well as the shortcomings of existing prediction methods in processing such data. Through EMD decomposition and KPCA processing, it effectively solves the prediction problem caused by the complex characteristics of CO2 emission data in the steel industry, improves the accuracy and adaptability of the prediction model in the application scenarios of the steel industry, and provides a more reliable basis for steel enterprises' energy conservation and emission reduction decisions. Attached Figure Description

[0062] Figure 1 This is an EMD exploded view of the experimental case of this invention;

[0063] Figure 2 This is a graph showing the proportion of each feature quantity after KPCA decomposition in the experimental case of this invention;

[0064] Figure 3 The image shows the predicted results of the experimental cases of this invention.

[0065] Figure 4 This is a prediction error graph for experimental cases of the present invention;

[0066] Figure 5 This is a comparison chart of RMSE results for experimental cases of this invention;

[0067] Figure 6 This is a comparison chart of MAE values ​​for experimental cases of this invention;

[0068] Figure 7 This is a comparison chart of MAPE results from experimental cases of this invention;

[0069] Figure 8 R is an experimental example of this invention. 2 Comparison chart. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] This invention provides a method for constructing a CO2 emission prediction model for the steel industry, comprising the following steps:

[0072] S1: Data collection and preprocessing, including the data sources, data types, data time, and data frequency, followed by data cleaning and normalization. Specifically, before implementing the EMD-KPCA-LSTM-based CO2 emission prediction model for the steel industry, relevant data from steel companies needs to be collected. Data sources mainly include the company's internal production management system, energy management system, and environmental monitoring system. The types of data collected are as follows:

[0073] Production data: steel output, output of each production process, product types, etc.

[0074] Energy consumption data: Consumption of energy sources such as coal, coke, electricity, and natural gas.

[0075] CO2 emission monitoring data: The CO2 emissions from each emission source (such as blast furnace, converter, sintering machine, etc.) can be calculated by monitoring the CO2 concentration and emission flow rate in the emitted gas.

[0076] The data collected should cover as long a time span as possible to include factors such as different production conditions and seasonal changes. It is generally recommended to collect 3-5 years of historical data.

[0077] The frequency of data collection can be determined based on the actual situation, but monthly or quarterly data is generally preferable.

[0078] In this step, the collected data may contain missing values, outliers, and duplicate values, requiring data cleaning to improve data quality. Specific methods are as follows:

[0079] Handling missing values: For missing data, interpolation methods (such as linear interpolation, polynomial interpolation, etc.) can be used to fill the gaps, or data from adjacent time periods can be used as substitutes. For example, if CO2 emission data for a certain month is missing, the average of the data from the two months before and after can be used to fill the gaps.

[0080] Outlier removal: Outliers are identified by drawing data distribution maps, box plots, etc. Outliers that exceed reasonable ranges can be removed or corrected. For example, if the CO2 emissions on a certain day are significantly higher than surrounding data without a reasonable explanation, it can be considered an outlier and processed.

[0081] Remove duplicate values: Check if there are duplicate records in the data. If so, delete the duplicate parts and keep the unique data.

[0082] Furthermore, in the above steps, due to the significant differences in the dimensions and numerical ranges among different indicators, data normalization is necessary to eliminate the influence of dimensions and improve the training efficiency and prediction accuracy of the model. Specifically, the range normalization method is used to map the data to the [0,1] interval, as shown in the following formula:

[0083]

[0084] Where x is the original data, x min and x max These are the minimum and maximum values ​​of the original data, x. norm This is the normalized value. Normalization eliminates the influence of different dimensions between indicators, making the data comparable, and also helps improve the convergence speed and prediction accuracy of the model.

[0085] S2: Input the normalized CO2 emission data into the Empirical Mode Decomposition (EMD) module. EMD is an adaptive time series decomposition method that can decompose complex non-stationary signals into multiple intrinsic mode function (IMF) components and a residual term. Each IMF component represents the characteristics of the original signal at different time scales and has different frequencies and amplitudes. The specific steps of EMD decomposition are as follows:

[0086] S201: Initialization: Use the normalized CO2 emission data as the original signal x(t), initialize the residual r0(t) = x(t), and set the stopping criteria (such as stopping the decomposition when the residual becomes a monotonic function or when no new IMF components can be extracted);

[0087] S202: Extract IMF components:

[0088] EMD decomposition is performed on the original signal x(t) to extract multiple IMF components:

[0089] Finding extreme points: Find all local maxima and local minima in the signal x(t).

[0090] Fitting the envelope: Using cubic spline interpolation, connect the local maxima and local minima to form the upper envelope e. up (t) and lower envelope e down (t).

[0091] Calculate the envelope mean: Calculate the mean of the upper and lower envelopes.

[0092] Extracting candidate IMF components: Subtract the envelope mean m1(t) from the signal x(t) to obtain the candidate IMF component h1(t) = x(t) - m1(t).

[0093] Determine the IMF condition: Check if h1(t) satisfies the IMF condition, which is that the number of local extrema and zero-crossings is equal or differs by at most one across the entire data range, and the mean of the local maximum envelope and local minimum envelope is zero at any time point. If satisfied, h1(t) is the first IMF component c1(t); otherwise, treat h1(t) as a new signal and repeat the above steps until an IMF component that satisfies the condition is obtained.

[0094] S203: Update residual: Subtract the extracted IMF component c1(t) from the original signal x(t) to obtain the new residual r1(t) = x(t) - c1(t);

[0095] S204: Repeated Decomposition: Using the residual r1(t) as the new signal, repeat S201-S203 to extract the next IMF component c2(t) and update the residual r2(t). Continue this process until the residual r n (t) continues until it becomes a monotonic function or no new IMF components can be extracted. Finally, the original signal can be expressed as:

[0096]

[0097] Among them, c i (t) represents the i-th IMF component, r n (t) represents the residual term;

[0098] This EMD decomposition step decomposes complex CO2 emission sequences into multiple relatively stationary IMF components and a residual term, thereby reducing the nonstationarity of the data and providing better input for subsequent feature extraction and modeling.

[0099] S3: Perform kernel principal component analysis (KPCA) on each IMF component to extract its nonlinear features and reduce dimensionality. KPCA is a nonlinear dimensionality reduction technique based on kernel methods. It effectively handles nonlinear data by mapping the original data to a high-dimensional feature space where principal component analysis (PCA) is performed, thus mapping the original data to a high-dimensional feature space. The specific steps are as follows:

[0100] S301: Data Standardization: Standardize the data for each IMF component to have a mean of 0 and a variance of 1. The standardization formula is:

[0101]

[0102] Where, x ij It is the j-th feature value of the i-th sample, μ j and σ i These are the mean and standard deviation of the j-th feature, respectively. ij It is the standardized value;

[0103] S302: Selecting the kernel function and calculating the kernel matrix: The Gaussian kernel function is selected as the kernel function for KPCA, and its expression is:

[0104]

[0105] Where, x i and x i There are two samples. σ is the bandwidth parameter of the kernel function, used to control the width of the Gaussian kernel. Calculate the kernel matrix K, whose elements K... ij =K(x) i ,x j );

[0106] S303: Centralized Kernel Matrix: The kernel matrix is ​​centralized, resulting in a centralized kernel matrix K. c Calculated using the following formula:

[0107] K c =K-1 N K-K1 N +1 N K1 N

[0108] Among them, 1 N It is an N×N matrix, where each element is 0. N is the number of samples;

[0109] S304: Solving for eigenvalues ​​and eigenvectors: Solving for the centered kernel matrix K c eigenvalues ​​λ i and the corresponding eigenvector α i Satisfying Kc α i =λ i α i ;

[0110] S305: Principal Component Selection: Select the top d largest eigenvalues ​​and their corresponding eigenvectors based on their magnitudes, where d is the dimension after dimensionality reduction. The value of d is typically determined by the contribution rate of the eigenvalues, calculated using the following formula:

[0111]

[0112] Principal components with a cumulative contribution rate of 85%-95% are generally selected to ensure that sufficient information is retained.

[0113] S306: Calculate the projection coefficients: For a new data point x, its projection coefficient z on the i-th principal component. i Calculated using the following formula:

[0114]

[0115] Where, α ji It is the j-th element of the i-th feature vector corresponding to the j-th sample;

[0116] S307: Obtain the dimensionality-reduced data: Combine the projection coefficients of all data points on the selected d principal components to obtain the dimensionality-reduced data matrix Z, where each row represents the representation of a sample in the dimensionality-reduced feature space;

[0117] In this step, KPCA processing can effectively extract key features from each IMF component, remove redundant information and noise, reduce data dimensionality, and retain the main information in the data, thus providing a simpler and more effective input for subsequent LSTM modeling.

[0118] S4: Using the dimensionality-reduced IMF components and residuals as input, a Long Short-Term Memory (LSTM) network is constructed for training and prediction. The trained LSTM model is then used to model and predict the dimensionality-reduced CO2 emission data, ultimately yielding the predicted CO2 emissions. LSTM is a special type of Recurrent Neural Network (RNN) that effectively addresses the vanishing and exploding gradient problems inherent in traditional RNNs when processing long-sequence data. By introducing memory units and gating mechanisms, LSTM can selectively remember or forget information, thereby capturing long-term dependencies in time-series data.

[0119] This step uses the Keras deep learning framework to build an LSTM model, the structure of which is as follows:

[0120] Input layer: Used to input the dimensionality-reduced IMF components and residual terms. The input dimension is determined based on the number of features after KPCA dimensionality reduction.

[0121] LSTM layer: The number of hidden layer units is set to 64, and the activation function is ReLU. The LSTM layer is used to capture long-term dependencies in time series data and selectively remember or forget information.

[0122] Output layer: The output layer has 1 neuron and is used to predict CO2 emissions. The activation function is linear.

[0123] For a detailed explanation of its working principle, please refer to: The basic structure of LSTM includes an input gate, a forget gate, an output gate, and a memory unit.

[0124] Forget gate: Determines how much information from the cell's previous state needs to be forgotten. The output of the forget gate is denoted by f. t The calculation formula is as follows:

[0125] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0126] Among them, W f It is the weight matrix, b f It is the bias term, σ is the sigmoid activation function, and h t-1 It is the hidden state from the previous moment, x t It is the input at the current moment.

[0127] Input gate: Determines how much information from the current input needs to be added to the cell state. The output of the input gate is represented by i. t The calculation formula is as follows:

[0128] i t =σ(W i ·[h t-1 x t ]+b i )

[0129] Simultaneously, calculate the candidate cell state. The formula is:

[0130]

[0131] Cell state update: Update cell state C based on the outputs of the forget gate and the input gate. t The formula is:

[0132]

[0133] Output gate: Determines how much information from the current cell state needs to be output to the current hidden state. The output of the output gate is represented by o. t The calculation formula is as follows:

[0134] o t =σ(W o ·[h t-1 x t ]+b o )

[0135] The hidden state h at the current moment t The calculation formula is:

[0136] h t =o t ·tanh(C t ).

[0137] Specifically, the training process of an LSTM network is as follows:

[0138] S401: Initialize parameters: Randomly initialize the weight matrix and bias terms in the LSTM network;

[0139] S402: Forward Propagation: Input the training data into the LSTM network sequentially, perform forward propagation calculations according to the LSTM model structure, and obtain the output at each time step;

[0140] S403: Calculate the loss function: Use the mean squared error (MSE) as the loss function to calculate the difference between the predicted and actual values. The formula for the loss function is:

[0141]

[0142] Among them, y i It is the actual value. These are the predicted values, and N is the number of samples.

[0143] S404: Backpropagation: Based on the gradient of the loss function, the weight matrix and bias terms in the LSTM network are updated using the backpropagation algorithm, including parameter updates using stochastic gradient descent (SGD) or Adam optimization algorithms.

[0144] S405: Iterative Training: Repeat the forward propagation, loss function calculation, and backpropagation steps until the preset number of training rounds is reached or the loss function converges.

[0145] In this step, the model compilation uses mean squared error (MSE) as the loss function, Adam as the optimization algorithm, and a learning rate of 0.001. The loss function and optimization algorithm are specified during model compilation. Model training: Training data is input into the LSTM model for training. The number of training epochs is set to 200, and the batch size to 32. During training, the loss values ​​of the training and validation sets are monitored to prevent overfitting. An early stopping callback function is used to stop training early when the validation set loss value no longer decreases within 10 consecutive epochs, saving training time and avoiding overfitting.

[0146] In the above steps, the LSTM network model constructed in this invention has 64 hidden layer units, a learning rate of 0.001, and 200 training epochs. The trained LSTM model can effectively model and predict the dimensionality-reduced CO2 emission data, ultimately obtaining the predicted CO2 emission amount.

[0147] To further explain the model constructed in the above steps, this embodiment also evaluates and optimizes the model, with the specific steps as follows:

[0148] (I) Model Evaluation:

[0149] Using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and fit coefficient (R²) 2 The trained LSTM model is evaluated using metrics such as [insert metrics here]. The calculation formula is as follows:

[0150]

[0151] Among them, y i It is the actual value. It is a predicted value. It is the average of the true values, and N is the sample size.

[0152] (II) Model Optimization:

[0153] Based on the model evaluation results, the model can be optimized. This can be achieved by adjusting the model structure (e.g., increasing or decreasing the number of units in the LSTM layer, increasing the model depth), optimizing hyperparameters (e.g., adjusting the learning rate, batch size), increasing the amount of training data, or using data augmentation techniques.

[0154] To verify the effectiveness of the model of this invention, the following specific experimental cases are provided for illustration:

[0155] 1. Data Sources and Preprocessing: Common steel production processes involve multiple stages such as iron ore reduction and coal combustion, generating carbon dioxide emissions. Here, we consider four influencing factors: energy consumption (E), production scale (S), technological level (T), and raw material quality (R). The following nonlinear formula is used to calculate carbon dioxide emissions (CO2):

[0156] CO2=(αE+βS 2 )×(1-γT)×(1-δR)+εE×S×T×R

[0157] In the formula: α, β, γ, δ, and ε are all influencing factors, which need to be set according to specific circumstances.

[0158] The experimental data mentioned above comes from the CO2 emission monitoring system of a large steel company, covering a period of 72 months from January 2015 to December 2020. The data includes monthly CO2 emissions, steel production, energy consumption, and other indicators. The raw data was cleaned and normalized to eliminate the influence of dimensions.

[0159] 2. Experiment setup, including:

[0160] (1) EMD decomposition: The normalized CO2 emission data were decomposed into EMD components, yielding 7 IMF components and 1 residual term, such as Figure 1 As shown;

[0161] (2) KPCA dimensionality reduction: Perform KPCA processing on each IMF component, select the Gaussian kernel function, and reduce the dimension to 1 dimension, such as... Figure 2 As shown;

[0162] (3) LSTM modeling: Construct an LSTM network model, set the number of hidden layer units to 64, the learning rate to 0.001, and the number of training rounds to 200;

[0163] (4) Comparison Models: Single LSTM model, EMD-LSTM model and KPCA-LSTM model were selected as comparison models to verify the superiority of the model in this paper;

[0164] 3. Evaluation metrics: Root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of fit (R²) are used. 2 Evaluate the predictive performance of this paper:

[0165]

[0166] In the formula, n is the number of samples, y i It is the true value of the i-th sample. It is the predicted value of the i-th sample. It is the average of the true values, that is

[0167]

[0168] 4. Results Analysis:

[0169] The RMSE of the model in this invention is reduced by 55.1% and the MAPE is reduced by 59.5% compared with the single LSTM, demonstrating its superiority in predicting non-stationary and nonlinear CO2 emission data.

[0170]

[0171] from Figure 3 , Figure 4 The prediction results and error graphs of each model are shown. It can be seen that the prediction curve of the EMD-KPCA-LSTM model is closest to the actual curve, especially during periods of large fluctuations in CO2 emissions, it can still capture the changing trend well, indicating that the model has strong robustness and generalization ability.

[0172] from Figures 5-8 The results show a comparison of the prediction performance of each model. It can be seen that the EMD-KPCA-LSTM model proposed in this invention performs better in terms of RMSE, MAE, MAPE, and R... 2 It outperforms other comparative models in all three metrics, indicating that it has higher prediction accuracy and better fitting effect.

[0173] In summary, the present invention provides a method for constructing a CO2 emission prediction model for the steel industry, offering a more accurate and efficient solution for CO2 emission prediction in the steel industry. By combining EMD, KPCA, and LSTM, it effectively addresses the shortcomings of existing prediction methods, such as insufficient mining of non-stationary and nonlinear features, low feature extraction efficiency, and limited ability to process long-sequence data.

[0174] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a CO2 emission prediction model for the steel industry, characterized in that, Includes the following steps: S1: Data collection and preprocessing, including the data source, data type, data time and data frequency, and then cleaning and normalizing the data; S2: Input the normalized CO2 emission data into the Empirical Mode Decomposition (EMD) module to decompose the complex non-stationary signal into multiple Intrinsic Mode Function (IMF) components and a residual term, thereby reducing the non-stationarity of the data and providing better input for subsequent feature extraction and modeling. S3: Perform kernel principal component analysis (KPCA) on each IMF component to extract its nonlinear features and reduce dimensionality; S4: Using the dimensionality-reduced IMF components and residual terms as input, a Long Short-Term Memory (LSTM) network is constructed for training and prediction. The trained LSTM model is then used to model and predict the dimensionality-reduced CO2 emission data, ultimately yielding the predicted CO2 emissions.

2. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 1, characterized in that: Data sources in S1 include the company's internal production management system, energy management system, and environmental monitoring system; data types include: Production data: steel output, output of each production process, and product types; Energy consumption data: Consumption of coal, coke, electricity, and natural gas; CO2 emission monitoring data: The CO2 emissions from each emission source can be calculated by monitoring the CO2 concentration and emission flow rate in the emitted gas; The collected data covers 3-5 years of historical data, spanning different production conditions and seasonal variations. The frequency of data collection should be determined based on the actual situation, with monthly or quarterly data being preferable.

3. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 1, characterized in that: Data cleaning in S1 includes: Handling missing values: For missing data, interpolation can be used to fill in the missing values, or data from adjacent time periods can be used as substitutes. Outlier removal: Outliers are identified by drawing data distribution charts and box plots. Outliers that exceed the reasonable range are removed or corrected. Remove duplicate values: Check if there are duplicate records in the data. If so, delete the duplicate parts and keep the unique data.

4. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 1, characterized in that: The data normalization method in S1 is as follows: The range normalization method is used to map the data to the [0,1] interval, as shown in the following formula: Where x is the original data, x min and x max These are the minimum and maximum values ​​of the original data, x. norm It is the normalized value.

5. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 1, characterized in that: The specific steps for EMD decomposition in S2 are as follows: S201: Initialization: Using the original signal x(t) as the initial signal, calculate its local maxima and local minima; S202: Extracting IMF components: Using cubic spline interpolation, connect the local maxima and local minima respectively to form the upper envelope e. up (t) and lower envelope e down (t); Calculate the mean m1(t) of the upper and lower envelopes, and then subtract the mean from the original signal to obtain h1(t) = x(t) - m1(t); Check whether h1(t) satisfies the conditions of the IMF, that is, the number of local extrema and zero crossings is equal or differs by at most one in the entire data range, and the mean of the local maximum envelope and the local minimum envelope is zero at any time point; if satisfied, then h1(t) is the first IMF component c1(t); otherwise, take h1(t) as a new signal and repeat the above steps until an IMF component that satisfies the conditions is obtained; S203: Update residuals: Subtract the extracted IMF components from the original signal to obtain the residual r1(t) = x(t) - c1(t); S204: Repeated decomposition: Using the residual r1(t) as the new signal, repeat S201-S203 until the residual r n Until (t) becomes a monotonic function or no new IMF components can be extracted, the original signal is finally represented as Where c i (t) represents the i-th IMF component, r n (t) represents the residual term.

6. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 1, characterized in that: The specific steps in S3 are as follows: S301: Data Standardization: Standardize the data for each IMF component to have a mean of 0 and a variance of 1. The standardization formula is: Where, x ij It is the j-th feature value of the i-th sample, μ i and σ j These are the mean and standard deviation of the j-th feature, respectively. ij It is the standardized value; S302: Selecting the kernel function and calculating the kernel matrix: The Gaussian kernel function is selected as the kernel function for KPCA, and its expression is: Where, x i and x j There are two samples. σ is the bandwidth parameter of the kernel function, used to control the width of the Gaussian kernel. Calculate the kernel matrix K, whose elements K... ij =K(x) i ,x j ); S303: Centralized Kernel Matrix: The kernel matrix is ​​centralized, resulting in a centralized kernel matrix K. c Calculated using the following formula: K c =K-1 N K-K1 N +1 N K1 N Among them, 1 N It is an N×N matrix, where each element is 0. N is the number of samples; S304: Solving for eigenvalues ​​and eigenvectors: Solving for the centered kernel matrix K c eigenvalues ​​λ i and the corresponding eigenvector α i Satisfying K c α i =λ i α i ; S305: Principal Component Selection: Select the top d largest eigenvalues ​​and their corresponding eigenvectors based on their magnitudes, where d is the dimension after dimensionality reduction. The value of d is typically determined by the contribution rate of the eigenvalues, calculated using the following formula: S306: Calculate the projection coefficients: For a new data point x, its projection coefficient z on the i-th principal component. i Calculated using the following formula: Where, α ji It is the j-th element of the i-th feature vector corresponding to the j-th sample; S307: Obtain the dimensionality-reduced data: Combine the projection coefficients of all data points on the selected d principal components to obtain the dimensionality-reduced data matrix Z, where each row represents the representation of a sample in the dimensionality-reduced feature space.

7. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 1, characterized in that: S4 uses the Keras deep learning framework to build an LSTM model, the structure of which is as follows: Input layer: Used to input the dimensionality-reduced IMF components and residual terms. The input dimension is determined based on the number of features after KPCA dimensionality reduction. LSTM layer: The number of hidden layer units is set to 64, and the activation function is ReLU. The LSTM layer is used to capture long-term dependencies in time series data and selectively remember or forget information. Output layer: The output layer has 1 neuron and is used to predict CO2 emissions. The activation function is linear.

8. The method for constructing a CO2 emission prediction model for the steel industry as described in claim 7, characterized in that: The training process of the LSTM network in S4 is as follows: S401: Initialize parameters: Randomly initialize the weight matrix and bias terms in the LSTM network; S402: Forward Propagation: Input the training data into the LSTM network sequentially, perform forward propagation calculations according to the LSTM model structure, and obtain the output at each time step; S403: Calculate the loss function: Use the mean squared error (MSE) as the loss function to calculate the difference between the predicted and actual values. The formula for the loss function is: Among them, y i It is the actual value. These are the predicted values, and N is the number of samples. S404: Backpropagation: Based on the gradient of the loss function, the weight matrix and bias terms in the LSTM network are updated using the backpropagation algorithm, including parameter updates using stochastic gradient descent (SGD) or Adam optimization algorithms. S405: Iterative Training: Repeat the forward propagation, loss function calculation, and backpropagation steps until the preset number of training rounds is reached or the loss function converges.

Images