A machine learning-based carbon footprint accounting method and system

Abstract

The application discloses a carbon footprint accounting method and system based on machine learning, and relates to the technical field of carbon footprint management.The method comprises the following steps: acquiring activity quantity sequences, factor sequences and characteristic sequences of all process parameters of a plurality of historical life cycles of the same product; calculating the isolated significance degree, the interaction significance degree and the characteristic significance degree of all process parameters with respect to each emission factor based on the factor sequences and the characteristic sequences; dividing the characteristic sequences of the process parameters into a plurality of samples and calculating the relative importance degree of each sample; constructing a loss function based on the relative importance degree of the samples, constructing an emission factor prediction model and training the emission factor prediction model based on the loss function; acquiring data of a current life cycle of the product, obtaining a predicted emission factor based on the trained emission factor prediction model, and calculating the total carbon footprint of the product.The application can improve the prediction accuracy of the emission factor and effectively improve the accuracy of the carbon footprint accounting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of carbon footprint management, and particularly relates to a carbon footprint accounting method and system based on machine learning. BACKGROUND

[0002] Carbon footprint accounting refers to the process of quantifying and evaluating the greenhouse gas emissions generated by a product, organization or individual during its life cycle. The main accounting processes include "cradle to gate" and "cradle to grave", the former calculates the emissions from production to factory, and the latter includes the emissions in the use and disposal stages. Currently, carbon footprint can be divided into different levels such as country, city, organization, enterprise, family, product and individual. In order to promote carbon footprint management, relevant departments are establishing a unified carbon footprint accounting standard system to promote green and low-carbon transformation.

[0003] Based on the "cradle to gate" accounting process, carbon footprint accounting for products is a key means for enterprises to respond to regulations, meet supply chain requirements and improve competitiveness. The current mainstream carbon footprint accounting method is based on life cycle assessment. The process is to collect energy and material data at each link from raw materials, production to factory, multiply the corresponding emission factors, and finally quantify the greenhouse gas emissions throughout the product process. The existing carbon footprint accounting method based on life cycle assessment relies on static emission factor databases such as IPCC and Ecoinvent. These static emission factor data are calculated by industry average or fixed model, which cannot dynamically reflect the process fluctuations and interaction effects in actual production.

[0004] In actual production, fixed emission factors cannot accurately adapt to the entire production cycle. The existing emission factor prediction model ignores the interaction effects between process parameters and the importance of samples, which can easily lead to large prediction errors of emission factors and inaccurate carbon footprint accounting. Specifically, the dynamic changes of process parameters (such as furnace temperature and raw material ratio) will cause nonlinear fluctuations of emission factors. Fixed factors cannot capture such dynamic characteristics, resulting in great limitations of static factors in the process of carbon footprint accounting. At the same time, the emission factor prediction model (such as linear regression model and simple MLP model) used in the existing carbon footprint accounting process regards process parameters as independent variables, without considering the synergistic or offsetting effects between parameters. For example, the proportion of coal coke and scrap steel jointly affects the combustion efficiency, so the interaction effect is ignored. Moreover, the importance of samples selected by the existing emission factor prediction model is not balanced. The proportion of stable working condition samples in the training data is high, but the proportion of high fluctuation samples containing key change patterns, such as equipment start-stop and raw material switching, is small, which can easily lead to large prediction deviation of the existing model trained by equal weight. SUMMARY

[0005] In order to improve the accuracy of carbon footprint accounting, solve the defects of large prediction emission factor deviation and inaccurate carbon footprint accounting of the existing emission factor prediction model, realize the accurate prediction of carbon footprint in the production process, so as to timely formulate emission reduction strategies and optimize subsequent production process, the application provides a kind of carbon footprint accounting method and system based on machine learning, and its technical scheme is as follows:

[0006] In the first aspect, the application provides a kind of carbon footprint accounting method based on machine learning, and the steps include: obtaining the activity sequence and factor sequence corresponding to all emission factors of each life stage in the multiple historical life cycles of the same product, and the characteristic sequence of all process parameters of each life stage;Based on the factor sequence and the characteristic sequence, the isolated significance, the interaction significance and the characteristic significance of all process parameters relative to each emission factor are calculated;The characteristic sequence of all process parameters of each life stage is divided into several samples with equal length, and the relative importance of each sample is calculated based on the characteristic significance of the process parameters;Based on the relative importance of the sample, a loss function is constructed, an emission factor prediction model is constructed, and the model is trained based on the sample and the loss function;Obtain the data of the current life cycle of the product, obtain the predicted emission factor based on the trained emission factor prediction model, and calculate the total carbon footprint of the product;

[0007] Wherein, the loss function is a self-defined sample weighted loss function, based on the emission factor prediction model, the predicted value of the next period emission factor of the period where the input sample data is located is obtained, the element value corresponding to the next period in the factor sequence of the corresponding emission factor is taken as the actual measurement value, the relative importance of the sample is normalized to obtain the corresponding normalized relative importance, the product of the absolute value of the difference between the predicted value and the actual measurement value and the corresponding normalized relative importance is taken as the loss value of the sample, based on the sample quantity corresponding to the corresponding emission factor, the cumulative result of the sample loss value is taken as the loss function.

[0008] Preferably, based on multiple same products and the same data acquisition frequency, the activity data sequence and the factor data sequence corresponding to all emission factors of each life stage in the multiple historical life cycles of the product are obtained through the CEMS system, and the characteristic data sequence of all process parameters of each life stage in the multiple historical life cycles of the product is obtained through the SCADA system;All elements in all sequences are standardized to eliminate dimension, and the corresponding activity sequence, factor sequence and characteristic sequence of process parameters are obtained.

[0009] Preferably, for the entire life cycle of the product, a discharge factor is selected as a target discharge factor, and any one process parameter in the life stage where the discharge factor is located is selected as a target parameter; the first-order difference processing is performed on the characteristic sequence of all process parameters in the life stage where the target discharge factor is located to obtain the differential characteristic sequence of each process parameter; the first-order difference processing is performed on the factor sequence corresponding to the target discharge factor to obtain the differential factor sequence of the target discharge factor; the sliding window processing is performed on the differential characteristic sequence and the differential factor sequence respectively, the window length and the sliding step are set, and the number of windows divided by sliding is counted; the absolute value difference between the corresponding elements of the differential characteristic sequence and the differential factor sequence in the same window is calculated, the absolute values of all the differences in the same window are taken after being accumulated based on the window length, and the similar significant characteristics of the window are obtained, and then the similar significant characteristics of all the windows are sequentially accumulated to obtain the overall similar significant characteristics, and the natural exponential function is used to perform reverse mapping on the overall similar significant characteristics, and the value of the reverse mapping is taken as the isolated significant degree of the target parameter with respect to the target discharge factor; similarly, the isolated significant degree of each process parameter in each life stage of each life cycle with respect to each discharge factor is sequentially calculated.

[0010] Preferably, the combination of the target parameter and any other process parameter in the same life stage is referred to as a target parameter pair, and a plurality of target parameter pairs are obtained; any target parameter pair is selected, the absolute value of the product between each element value in the differential characteristic sequence of the target parameter and the isolated significant degree of the target parameter is calculated, the absolute value of the product between each element value in the differential characteristic sequence of the non-target parameter and the isolated significant degree of the non-target parameter is calculated, and the sum of the one-to-one corresponding two is taken as each element value of the target parameter pair; the sum of the isolated significant degree of the target parameter and the isolated significant degree of the non-target parameter in the target parameter pair is taken as the significant characteristic value of the target parameter pair, the ratio between each element value of the target parameter pair and the significant characteristic value of the target parameter pair is calculated, the difference between the corresponding element value in the differential factor sequence of the target discharge factor and the absolute value of the corresponding element value is calculated, and the absolute value of the difference is taken; based on the length of the differential characteristic sequence or the differential factor sequence, the absolute values of the differences are accumulated as the interaction characteristics of the corresponding target parameter pair, and based on the number of target parameter pairs, the interaction characteristics of all target parameter pairs are accumulated as the overall interaction characteristics of the target parameter with respect to the target discharge factor; the natural exponential function is used to perform reverse mapping on the overall interaction characteristics, and the value of the reverse mapping is taken as the interaction significant degree of the target parameter with respect to the target discharge factor; similarly, the interaction significant degree of each process parameter in each life stage of each life cycle with respect to each discharge factor is sequentially calculated.

[0011] Preferably, the isolated significance of the target parameter with respect to the target emission factor in each historical life cycle and the interaction significance of the target parameter with respect to the target emission factor in each historical life cycle are obtained, the product between the isolated significance and the interaction significance of the target parameter with respect to the target emission factor in each historical life cycle is sequentially calculated, and the accumulated value of the product is obtained based on the number of the historical life cycles, and the ratio between the accumulated value and the number of the historical life cycles is taken as the characteristic significance of the target parameter with respect to the target emission factor; similarly, the characteristic significance of each process parameter with respect to each emission factor is sequentially calculated.

[0012] Preferably, for each life stage of each emission factor, the characteristic sequences of all process parameters in the corresponding life stage of each historical life cycle are segmented to obtain a plurality of samples, each of which is a data matrix with equal length, the number of rows of each sample is the number of process parameters in the life stage, and the number of columns of each sample is the length of the segmented characteristic sequence segment; the characteristic significance of all process parameters with respect to each emission factor is normalized to obtain the normalized characteristic significance corresponding to each process parameter; for all samples in the sample set of any emission factor, each sample is multiplied by the normalized characteristic significance corresponding thereto to obtain the characteristic matrix of each sample.

[0013] Preferably, the standard deviation of the characteristic sequence segment of each process parameter in each sample is calculated, and the average value of the standard deviation of the characteristic sequence segment of the same process parameter in all samples in the sample set is obtained; the ratio between the standard deviation of each process parameter characteristic sequence segment in any sample and the average value of the standard deviation corresponding to the corresponding process parameter is taken as the relative volatility of the corresponding process parameter in the sample; the product between the relative volatility of each process parameter in any sample, the normalized characteristic significance, and the standard deviation of the characteristic sequence segment is taken as the relative importance of each process parameter in the sample, and the relative importance of all process parameters in the sample is accumulated based on the number of process parameters contained in the sample, and the accumulated value is taken as the relative importance of the sample; similarly, the relative importance of all samples in the sample set is calculated.

[0014] Preferably, the feature matrix of a certain period sample is input into the emission factor prediction model to obtain the predicted value of the emission factor of the next period of the period where the sample is located, and based on the predicted value corresponding to the sample, the actual measured value and the normalized relative importance, the expression of the loss function is constructed; the structure of the emission factor prediction model is constructed based on the MLP model, the feature matrix corresponding to all samples in the sample set of the emission factor is input into the emission factor prediction model in turn, and iterative training is performed based on the loss function until the prediction iteration is completed or the loss function does not decrease significantly, and the repeated training operation is performed on each emission factor respectively, and the model parameters corresponding to each emission factor are obtained after the training is completed.

[0015] Preferably, based on the same data acquisition frequency, the activity data sequence and the factor data sequence corresponding to all emission factors of the same product from the initial period to the current period in the current life cycle are collected in real time through the CEMS system, the feature data sequence of all process parameters in the life stage that has been carried out in the current life cycle of the same product is collected in real time through the SCADA system, all elements in all sequences are standardized to eliminate the dimension, and the current activity sequence, the current factor sequence and the feature sequence of the current process parameter are obtained; based on the current factor sequence and the feature sequence of the current process parameter, the current feature significance of all process parameters relative to each emission factor is calculated and normalized in the same way as the historical life cycle data is processed, the feature sequence of all current process parameters is segmented into a plurality of current samples in the same length, and the current sample is multiplied by the normalized value of the current feature significance corresponding thereto to obtain the current feature matrix; the current feature matrix is input into the trained emission factor prediction model to obtain the predicted emission factor corresponding to the next period; based on the calculation formula of the product carbon footprint, the predicted total carbon footprint of the product produced to the next period is calculated according to the predicted emission factor and the activity data of the next period preset in the production plan, and the actual total carbon footprint of the product produced to the next period is calculated according to the predicted emission factor and the real-time activity data collected in the next period; when the actual total carbon footprint exceeds the predicted total carbon footprint and the deviation is greater than the threshold, the emission reduction strategy is formulated in time and the subsequent production process is optimized.

[0016] In the second aspect, the present application provides a carbon footprint accounting system based on machine learning, which is used to realize the above-mentioned carbon footprint accounting method based on machine learning, and comprises a processor, a memory, a communication interface and an information collection system, the processor stores computer program instructions for realizing the above-mentioned carbon footprint accounting method based on machine learning, and the communication interface is in communication connection with the information collection system.

[0017] Compared with the prior art, the present application has the following beneficial effects:

[0018] The present application obtains the significance of different process parameters on the predicted emission factor by analyzing the interaction effect between different process parameters, identifies the nonlinear interaction relationship between process parameters, prompts the emission factor prediction model to learn the complex nonlinear relationship, and improves the prediction accuracy of the emission factor; then, by analyzing the significance of the process parameters in the sample and the difference between the process parameters of the samples, the relative importance of each sample is obtained, and a loss function is constructed to train the prediction model of the optimization target emission factor, thereby effectively improving the prediction accuracy of the emission factor;

[0019] At the same time, based on the accurately predicted emission factor, the accuracy of carbon footprint accounting can be effectively improved, and the carbon footprint of the next period can be accurately predicted in the production process, and then according to the deviation between the predicted carbon footprint and the actual carbon footprint, corresponding emission reduction measures or adjustment of process parameters can be taken, so as to timely formulate emission reduction strategies and optimize subsequent production processes. Therefore, compared with the prior art, the present application has higher prediction accuracy of the emission factor, can make the accuracy of carbon footprint accounting higher, solves the defects of large deviation of predicted emission factor and inaccurate carbon footprint accounting in the prior art, and plays an important role in promoting the green and low-carbon transformation of enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 An implementation flowchart of a carbon footprint accounting method based on machine learning according to an embodiment of the present application;

[0021] Figure 2 A structural block diagram of a carbon footprint accounting system based on machine learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical features of the present application will be further described in detail below in combination with the drawings so that those skilled in the relevant art can understand.

[0023] A carbon footprint accounting method based on machine learning, the implementation process is as shown in Figure 1 The specific implementation steps are as follows:

[0024] Step S1: Obtain the activity sequence and factor sequence corresponding to all emission factors in each life stage of the same product in multiple historical life cycles, and the characteristic sequence of all process parameters in each life stage;

[0025] Specifically, based on multiple identical products and the same data collection frequency, an exemplary collection frequency is 0.5h, the CEMS system (Continuous-Emission-Monitoring-System, flue gas automatic monitoring system) is used to obtain the activity data sequence and the factor data sequence corresponding to all emission factors in each life stage of the product in multiple historical life cycles, and the SCADA system (Supervisory-Control-And-Data-Acquisition, data acquisition and monitoring control system) is used to obtain the characteristic data sequence of all process parameters in each life stage of the product in multiple historical life cycles; all elements in all sequences are standardized to eliminate the dimension, and the corresponding activity sequence, factor sequence and characteristic sequence of the process parameters are obtained;

[0026] In the carbon footprint accounting, the product life cycle covered by “from cradle to gate” includes three stages of raw material acquisition and processing, production process and internal transportation of the factory; the activity refers to the measurement of the activity that can produce carbon emissions within the collection interval, the emission factor refers to the activity that can produce carbon emissions, the value of the emission factor represents the amount of greenhouse gas emissions per unit activity, and the emission factor includes but is not limited to iron ore emission factor, coke emission factor, power emission factor, blast furnace emission factor, internal combustion engine forklift emission factor, etc.; the process parameters include but are not limited to coal coke ratio, scrap steel ratio, furnace temperature, production capacity, oxygen injection amount and other parameters that affect carbon emissions;

[0027] For example, the value of the internal combustion engine forklift emission factor is 0.05 kgCO2e / ton×km, when the activity of the internal combustion engine forklift is to carry 1.5 tons of iron ore to move 1 km, the carbon footprint generated is 0.075 kgCO2e;

[0028] The activity data sequence contains elements representing the value of the activity in different collection intervals for a certain emission factor; the factor data sequence contains elements representing the value of the emission factor at different times; the characteristic data sequence of the process parameters contains elements representing the value of the process parameters at different times.

[0029] Step S2: Based on the factor sequence and the characteristic sequence, the isolated significance of all process parameters with respect to each emission factor is calculated.

[0030] Since the process parameters in the product life cycle are dynamically changing, the emission factor changes due to the influence of the process parameters, therefore, it is very important to analyze whether the change of the process parameters and the change of the emission factor are similar and significant for predicting the emission factor;

[0031] Specifically, for the entire historical lifecycle of a product, an emission factor is selected as the target emission factor, and any process parameter within the lifecycle stage of the emission factor is selected as the target parameter. First-order differencing is performed on the feature sequences of all process parameters within the lifecycle stage of the target emission factor to obtain the differencing feature sequence for each process parameter. First-order differencing is performed on the factor sequence corresponding to the target emission factor to obtain the differencing factor sequence of the target emission factor. Sliding window processing is then applied to the differencing feature sequence and the differencing factor sequence, respectively. The window length and sliding step size are set, and the number of windows divided by the sliding window is counted. Generally, the window length is set to between one-fifth and one-third of the sequence length, and the sliding step size is set to 1.

[0032] The absolute values ​​of the differences between corresponding elements of the differential feature sequence and the differential factor sequence within the same window are calculated. Based on the window length, the absolute values ​​of all differences within the same window are summed to obtain the similarity saliency features of that window. The similarity saliency features of all windows are then summed sequentially to obtain the overall similarity saliency features. The natural exponential function is used to perform a reverse mapping on the overall similarity saliency features, and the value of the reverse mapping is used as the isolation saliency of the target parameter relative to the target emission factor. Similarly, the isolation saliency of all process parameters relative to each emission factor within each stage of each historical life cycle is calculated sequentially.

[0033] The first-order difference value represents the amount of change. The more similar the magnitude of the change in the target parameter is to the magnitude of the change in the target emission factor, the greater the influence of this process parameter on the emission factor. Absolute values ​​are used to simultaneously consider both positive and negative correlations. The formula for calculating isolated significance is as follows:

[0034]

[0035] In the formula, This indicates the degree of isolated significance of the target parameter relative to the target emission factor. This represents the natural exponential function. Indicates the number of windows in the sliding partition. Indicates the length of the window. Indicates the first In the differential feature sequence within the nth window, the th One element, Indicates the first In the difference factor sequence within the nth window, the th Each element.

[0036] Step S3: Based on the isolated significance of process parameters, calculate the interaction significance of all process parameters with respect to each emission factor;

[0037] Due to the complex physical and chemical processes involved in the product life cycle, there is a nonlinear interaction between the characteristics, for example: in the carbon footprint accounting of blast furnace steelmaking, the interaction between the coal coke ratio and the furnace temperature will amplify the emission factor because the coke burns more fully at high temperatures, and the interaction between the scrap steel ratio and the coal coke ratio will reduce the emission factor, so analyzing the influence of the interaction between process parameters on the emission factor is beneficial to the comprehensive evaluation of the importance of process parameters;

[0038] Specifically, the combination of the target parameter and any other process parameter in the same life stage is referred to as a target parameter pair, and a plurality of target parameter pairs are obtained; for any target parameter pair, the absolute value of the product between each element value in the difference characteristic sequence of the target parameter and the isolated significance of the target parameter is calculated in turn, and then the absolute value of the product between each element value in the difference characteristic sequence of the non-target parameter and the isolated significance of the non-target parameter is calculated in turn, and the sum of the one-to-one corresponding two is taken as each element value of the target parameter pair; the sum of the isolated significance of the target parameter and the isolated significance of the non-target parameter in the target parameter pair is taken as the significant characteristic value of the target parameter pair, and the ratio between each element value of the target parameter pair and the significant characteristic value of the target parameter pair is calculated in turn, and then the difference between the absolute value of the corresponding element value in the difference factor sequence of the target emission factor and each ratio is calculated, and the absolute value of the difference is taken; based on the length of the difference characteristic sequence or the difference factor sequence, the absolute value of each difference value is accumulated as the interaction characteristic of the corresponding target parameter pair, and based on the number of target parameter pairs, the interaction characteristics of all target parameter pairs are accumulated as the overall interaction characteristic of the target parameter with respect to the target emission factor; the overall interaction characteristic is inversely mapped using a natural exponential function, and the value of the inverse mapping is taken as the interaction significance of the target parameter with respect to the target emission factor; similarly, the interaction significance of all process parameters in each life stage of each historical life cycle with respect to each emission factor is calculated in turn;

[0039] wherein the difference values of the two process parameters are fused through the isolated significance to construct the interaction difference value, and the smaller the difference between the interaction difference value and the difference value of the difference factor sequence, the greater the interaction significance; the calculation formula of the interaction significance is as follows:

[0040]

[0041] In the formula, represents the interaction significance of the target parameter with respect to the target emission factor, represents a natural exponential function, represents the number of target parameter pairs, represents the length of the difference characteristic sequence or the difference factor sequence, represents the isolated significance of the target parameter, represents the isolated significance of the non-target parameter, The degree of isolated significance of each objective parameter against non-objective parameters. The first element in the difference feature sequence representing the target parameter One element, Indicates the first Within the difference feature sequence of the target parameter pair and the non-target parameter, the i-th... One element, The first differential factor in the sequence representing the target emission factor Each element.

[0042] Furthermore, compared with existing technologies that only analyze the impact of a single process parameter on emission factors, interaction effects can cause the impact of changes in a single process parameter to be offset. For example, when the proportion of coal and coke increases, the proportion of scrap steel also increases, which can lead to a lower actual change in emission factors. In step S3 of this invention, parameter pairs are formed by the target process parameter and the other process parameters. Interaction features are constructed through parameter pairs, and the differences between changes in interaction features and changes in the target emission factors are analyzed. This can identify the nonlinear interaction relationship between process parameters, comprehensively analyze the synergistic effect between various process parameters, and effectively improve the prediction accuracy of subsequent models.

[0043] Step S4: Based on the isolated significance and interactive significance of the process parameters, calculate the characteristic significance of all process parameters relative to each emission factor;

[0044] Even if a target parameter has a low isolated significance, it may interact with other process parameters, resulting in a large interaction significance. Therefore, it is necessary to comprehensively evaluate the characteristic significance of the target parameter by considering both isolated significance and interaction significance.

[0045] Specifically, the isolated significance of the target parameter relative to the target emission factor in each historical life cycle, as well as the interactive significance of the target parameter relative to the target emission factor in each historical life cycle, are obtained. The product of the isolated significance and interactive significance of the target parameter relative to the target emission factor in each historical life cycle is calculated sequentially, and the products are accumulated based on the number of historical life cycles. The ratio between the accumulated value and the number of historical life cycles is used as the characteristic significance of the target parameter relative to the target emission factor. Similarly, the characteristic significance of all process parameters relative to each emission factor is calculated sequentially.

[0046] For a given target parameter, its importance should be comprehensively assessed by considering both the impact of a single parameter and the interaction between parameters. Furthermore, due to differences in the working environment and the numerical values ​​of specific parameters across different historical lifecycles, a comprehensive assessment across multiple historical lifecycles is necessary. The formula for calculating the significance of a feature is as follows:

[0047]

[0048] In the formula, This indicates the significance of the target parameter relative to the target emission factor. This indicates the number of historical lifecycles collected. Indicates the first The degree of isolated significance of the target parameter relative to the target emission factor in each historical life cycle. Indicates the first The significance of the interaction between the target parameter and the target emission factor throughout a historical life cycle.

[0049] Step S5: Divide the feature sequences of all process parameters in each life stage into several samples of equal length, and calculate the relative importance of each sample based on the feature significance of the process parameters;

[0050] Because the operating conditions are different at different stages of the product's life cycle, and the stability of different samples is different, the interaction relationships and change trends implied by stable operating conditions are relatively simple and contribute little to the model training and learning. They cannot provide information on high emissions or frequently changing conditions, which leads to the dominant model learning and prediction samples ignoring the operating condition trends implied by other samples, which easily results in a large prediction deviation. Moreover, the model is prone to getting stuck in local optimal predictions of stable operating conditions.

[0051] Specifically, for each emission factor's life cycle stage, the feature sequences of all process parameters in the corresponding life cycle stage are segmented to obtain multiple samples. Generally, the segment length ranges from 10 to 20. Each sample is a data matrix of equal length, with the number of rows representing the number of process parameters in that life cycle stage and the number of columns representing the length of the segmented feature sequence. All samples in that life cycle stage within each historical life cycle are called the sample set of the corresponding emission factor. The feature significance of all process parameters relative to each emission factor is normalized to obtain the normalized feature significance of each process parameter. For all samples in the sample set of any emission factor, each sample is multiplied by its corresponding normalized feature significance to obtain the feature matrix of each sample.

[0052] The normalization process employs a summation-based normalization method. For all process parameters corresponding to each emission factor, their characteristic significance is summed, and the summation value is used as a reference standard. Then, the characteristic significance of each process parameter is divided by the summation value, and the ratio is taken as the normalized characteristic significance of each process parameter. The formula for summation-based normalization is: For an emission factor, the sum of the significance of the normalized characteristics of all its corresponding process parameters is 1;

[0053] Furthermore, the standard deviation of all process parameter feature sequence segments in each sample is calculated, and the average standard deviation of the same process parameter feature sequence segment in all samples of the sample set is obtained. The ratio between the standard deviation of each process parameter feature sequence segment in any sample and the average standard deviation of the corresponding process parameter is taken as the relative volatility of the corresponding process parameter in that sample. The product of the relative volatility, normalized feature significance, and standard deviation of the feature sequence segment of each process parameter in any sample is taken as the relative importance of each process parameter in that sample. Based on the number of process parameters included in the sample, the relative importance of all process parameters in the sample is accumulated, and the accumulated value is taken as the relative importance of the sample. Similarly, the relative importance of all samples in the sample set is calculated.

[0054] The larger the standard deviation of the characteristic sequence segment of the process parameter within the sample, the greater the fluctuation of the process parameter, the more complex the operating conditions it implies, and the more attention the model should pay to that sample. Simultaneously, the ratio to the mean standard deviation represents the relative volatility. When the ratio is greater than 1, it indicates that the relative volatility of the sample is large, and the greater its importance; when it is less than or equal to 1, the relative volatility is small, and it should be given less attention. Summing all process parameters is to avoid a single process parameter dominating the sample importance assessment and to avoid ignoring the influence of other process parameters. The formula for calculating the relative importance of a sample is as follows:

[0055]

[0056] In the formula, Indicates the relative importance of the samples. This indicates the number of process parameters included in the sample. Indicates the first The significance of the normalized characteristics of each process parameter relative to the target emission factor when the process parameter is used as the target parameter. Indicates the first in the sample The standard deviation of a segment of characteristic sequence of process parameters Represents the i-th sample in the sample set. The average of the standard deviations of the characteristic sequence segments of each process parameter.

[0057] Step S6: Construct a loss function based on the relative importance of the samples, build an emission factor prediction model, and train the model based on the samples and the loss function;

[0058] Specifically, for any sample set of emission factors, the relative importance of all samples in the sample set is normalized to obtain the normalized relative importance of each sample. Consistent with the normalization method for feature significance, the normalization process for relative importance also uses summation normalization, with the normalized relative importance of all samples in each sample set summed to 1. The feature matrix of a sample in a certain time period is input into the emission factor prediction model to obtain the predicted value of the emission factor for the next time period in which the sample is located. The element value corresponding to the next time period in the factor sequence of the corresponding emission factor is used as the actual measured value. The absolute value of the difference between the predicted value and the actual measured value is calculated. The product of the absolute value of the difference and the normalized relative importance of the sample is used as the loss value of the sample. Based on the number of samples in the sample set corresponding to the emission factor, the loss values ​​corresponding to all samples in the sample set are accumulated, and the accumulated result is used as the loss function.

[0059] The expression for the loss function is as follows:

[0060]

[0061] In the formula, This represents the value of the loss function used to train the model. This represents the number of samples in the sample set representing the target emission factor. In the sample set, the first... The relative importance of normalization for each sample; In the sample set, the first... When the feature matrix of each sample is used as input, the predicted value of the emission factor is... This represents the actual measured value corresponding to the target emission factor in the factor sequence;

[0062] Compared with the loss function in the prior art, the loss function constructed in this step takes into account the contribution of different samples to the model training, and pays more attention to the prediction error caused by samples with relatively greater importance.

[0063] Furthermore, an emission factor prediction model is constructed based on an MLP (Multi-layer Perceptron) model. The feature matrices corresponding to all samples in the emission factor sample set are sequentially input into the emission factor prediction model, and iterative training is performed based on the loss function until the prediction iteration is completed or the loss function no longer decreases significantly. The iteration is stopped. The training operation is repeated for each emission factor, and the model parameters corresponding to each emission factor are obtained after training. The constructed emission factor prediction model includes an input layer, hidden layer 1, hidden layer 2, and output layer. The two hidden layers use the ReLU function as the activation function, and the output layer uses the Sigmoid function as the activation function.

[0064] Step S7: obtaining data of the current life cycle of the product, obtaining the predicted emission factor based on the trained emission factor prediction model, and calculating the total carbon footprint of the product;

[0065] Specifically, based on the same data collection frequency, the operation process of step S1 is repeated, the activity data sequence and the factor data sequence corresponding to all emission factors in the initial period to the current period of the current life cycle of the same product are collected in real time through the CEMS system, and the feature data sequence of all process parameters in the life stage that has been carried out in the current life cycle of the same product is collected in real time through the SCADA system; all elements in all sequences are standardized to eliminate the dimension, and the current activity sequence, the current factor sequence and the feature sequence of the current process parameters are obtained;

[0066] Wherein, the calculation process of steps 2 to 4 is repeated in the same way as the historical life cycle data is processed and calculated, the current feature significance of all process parameters with respect to the current each emission factor is calculated based on the current factor sequence and the feature sequence of the current process parameters, and is normalized, and all feature sequences of the current process parameters are divided into multiple current samples in the same way as the historical life cycle data is processed, and the current samples are multiplied by the normalized values of the corresponding current feature significance, to obtain the current feature matrix;

[0067] In addition, the current feature matrix is input into the trained emission factor prediction model to obtain the predicted emission factor corresponding to the next period; the total carbon footprint is calculated based on the carbon footprint calculation formula, the carbon footprint calculation formula is the prior art, and the core formula is as follows:

[0068]

[0069] According to the predicted emission factor and the activity data in the next period preset in the production plan, the predicted total carbon footprint of the product produced to the next period is calculated, according to the predicted emission factor and the real-time activity data collected in the next period, the actual total carbon footprint of the product produced to the next period is calculated; when the actual total carbon footprint exceeds the predicted total carbon footprint and the deviation is greater than the threshold, the emission reduction strategy is formulated in time and the subsequent production process is optimized.

[0070] The application also discloses a carbon footprint accounting system based on machine learning, which is used to realize the above-mentioned carbon footprint accounting method based on machine learning, and the system structure is as shown in Figure 2 The system structure is as shown in

[0071] The embodiments contained herein are merely preferred embodiments of the present application and are not limited to the precise structure as has been described and illustrated in the accompanying drawings; various modifications and changes can be made without departing from the scope of the present application; various modifications and changes made by the engineering technicians in the field to the technical solutions of the present application without departing from the design idea of the present application shall fall within the protection scope of the present application.

Claims

1. A machine learning based carbon footprint accounting method, characterized in that: The activity sequence and the factor sequence corresponding to each life stage of all emission factors in each life stage of multiple historical life cycles of the same product are obtained, and the characteristic sequence of all process parameters in each life stage is obtained; based on the factor sequence and the characteristic sequence, the isolated significance, the interaction significance, and the characteristic significance of all process parameters with respect to each emission factor are calculated; The characteristic sequence of all process parameters in each life stage is divided into a plurality of samples with the same length, and the relative importance of each sample is calculated based on the characteristic significance of the process parameters; a loss function is constructed based on the relative importance of the samples, an emission factor prediction model is constructed, and the model is trained based on the samples and the loss function; The data of the current life cycle of the product is obtained, the predicted emission factor is obtained based on the trained emission factor prediction model, and the total carbon footprint of the product is calculated; Wherein, for the entire historical life cycle of the product, an emission factor is selected as a target emission factor, and an arbitrary process parameter in the life stage of the emission factor is selected as a target parameter; the characteristic sequence of all process parameters in the life stage of the target emission factor is subjected to first-order difference processing to obtain the differential characteristic sequence of each process parameter; the factor sequence corresponding to the target emission factor is subjected to first-order difference processing to obtain the differential factor sequence of the target emission factor; the differential characteristic sequence and the differential factor sequence are subjected to sliding window processing, the window length and the sliding step are set, and the number of sliding windows is counted; the combination of the target parameter and any other process parameter in the same life stage is referred to as a target parameter pair, and a plurality of target parameter pairs are obtained. The calculation formula of the isolated significance is as follows: In the formula, represents the isolated significance degree of the target parameter relative to the target emission factor, represents a natural exponential function, represents the number of sliding division windows, represents the length of the window, represents the first element in the differential characteristic sequence in the first window, represents the first element in the differential characteristic sequence in the first window, represents the first element in the differential factor sequence in the first window; similarly, the isolated significance degree of each process parameter relative to each emission factor in each life stage of each historical life cycle is calculated in turn. The calculation formula of the interaction significance is as follows: In the formula, This indicates the significance of the interaction between the target parameter and the target emission factor. This represents the natural exponential function. Indicates the number of target parameter pairs. Indicates the length of the difference feature sequence or difference factor sequence. Indicates the degree of isolated significance of the target parameter. Indicates the first The degree of isolated significance of each objective parameter against non-objective parameters. The first element in the difference feature sequence representing the target parameter One element, Indicates the first Within the difference feature sequence of the target parameter pair and the non-target parameter, the i-th... One element, The first differential factor in the sequence representing the target emission factor Each element; similarly, the interaction significance of all process parameters with respect to each emission factor in each life stage of each historical life cycle is calculated sequentially. The calculation formula of the characteristic significance is as follows: In the formula, This indicates the significance of the target parameter relative to the target emission factor. This indicates the number of historical lifecycles collected. Indicates the first The degree of isolated significance of the target parameter relative to the target emission factor in each historical life cycle. Indicates the first The significance of the interaction between the target parameter and the target emission factor in each historical life cycle; similarly, the significance of the characteristics of all process parameters relative to each emission factor is calculated sequentially. The loss function is a self-defined sample weighted loss function. Based on the emission factor prediction model, the predicted value of the emission factor in the next period is obtained from the input sample data in the period, the element value corresponding to the next period in the factor sequence of the corresponding emission factor is taken as the actual measurement value, the relative importance of the sample is normalized to obtain the corresponding normalized relative importance, the product of the absolute value of the difference between the predicted value and the actual measurement value and the corresponding normalized relative importance is taken as the loss value of the sample, and the cumulative result of the sample loss value is taken as the loss function based on the number of samples corresponding to the corresponding emission factor. 2.The method of accounting carbon footprint based on machine learning according to claim 1, wherein, The activity sequence and the factor sequence corresponding to each life stage of all emission factors in each life stage of multiple historical life cycles of the same product are obtained, and the characteristic sequence of all process parameters in each life stage is obtained; based on multiple same products and the same data collection frequency, the activity data sequence and the factor data sequence corresponding to each life stage of all emission factors in multiple historical life cycles of the product are obtained through the CEMS system, and the characteristic data sequence of all process parameters in each life stage of the product is obtained through the SCADA system; all elements in all sequences are subjected to standardization processing to eliminate the dimension, and the corresponding activity sequence, factor sequence, and characteristic sequence of the process parameters are obtained. 3.The method of accounting carbon footprint based on machine learning according to claim 1, wherein, The feature sequence of all process parameters in each life stage is divided into a plurality of samples with equal length, and the relative importance of each sample is calculated based on the feature significance of the process parameters, including: for each life stage in which the emission factor is located, the feature sequence of all process parameters in the corresponding life stage of each historical life cycle is divided to obtain a plurality of samples, each of which is a data matrix with equal length, the number of rows of each sample is the number of process parameters in the life stage, and the number of columns of each sample is the length of the divided feature sequence segment, and all samples of the life stage in each historical life cycle are referred to as a sample set of the corresponding emission factor; the feature significance of all process parameters with respect to each emission factor is normalized to obtain the normalized feature significance corresponding to each process parameter, and for all samples in the sample set of any emission factor, each sample is multiplied by the normalized feature significance corresponding thereto to obtain a feature matrix of each sample. 4.The method of accounting carbon footprint based on machine learning according to claim 3, characterized in that, The relative importance of each sample is calculated based on the feature significance of the process parameters, and the calculation further includes: calculating the standard deviation of all process parameter feature sequence segments in each sample, and obtaining the average value of the standard deviation of the same process parameter feature sequence segment in all samples in the sample set; the ratio between the standard deviation of each process parameter feature sequence segment in any sample and the average value of the standard deviation corresponding to the corresponding process parameter is taken as the relative volatility of the corresponding process parameter in the sample, and the product of the relative volatility, the normalized feature significance, and the standard deviation of the feature sequence segment of each process parameter in any sample is taken as the relative importance of each process parameter in the sample, and the relative importance of all process parameters in the sample is accumulated based on the number of process parameters included in the sample, and the accumulated value is taken as the relative importance of the sample; similarly, the relative importance of all samples in the sample set is calculated. 5.The method for calculating carbon footprint based on machine learning according to claim 4, wherein, The loss function is constructed based on the relative importance of the sample, the emission factor prediction model is constructed, and the model is trained based on the sample and the loss function, including: inputting the feature matrix of a sample in a certain period to the emission factor prediction model to obtain the predicted value of the emission factor in the next period of the period in which the sample is located, and based on the predicted value, the actual measured value, and the normalized relative importance corresponding to the sample, the expression of the loss function is constructed; the structure of the emission factor prediction model is constructed based on the MLP model, the feature matrix corresponding to all samples in the sample set of the emission factor is input to the emission factor prediction model in turn, the loss function is iteratively trained, and the iteration is stopped when the prediction iteration is completed or the loss function does not decrease significantly, and the training operation is repeated for each emission factor, and the model parameters corresponding to each emission factor are obtained after the training is completed.

6. The machine learning based carbon footprint accounting method according to any one of claims 1 to 5, characterized in that, The data of the current life cycle of the product is obtained, the predicted emission factor is obtained based on the trained emission factor prediction model, and the total carbon footprint of the product is calculated, including: based on the same data collection frequency, the activity data sequence and the factor data sequence corresponding to all emission factors in the initial period to the current period in the current life cycle of the same product are collected in real time through the CEMS system, the characteristic data sequence of all process parameters in the life stage that has been carried out in the current life cycle of the same product is collected in real time through the SCADA system, all elements in all sequences are standardized to eliminate dimensions, and the current activity sequence, the current factor sequence, and the characteristic sequence of the current process parameters are obtained; based on the current factor sequence and the characteristic sequence of the current process parameters, the current characteristic significance of all process parameters relative to the current each emission factor is calculated and normalized in the same way as the historical life cycle data is processed and calculated, the characteristic sequence of all current process parameters is segmented into multiple current samples in the same length, and the current sample is multiplied by the normalized value of the corresponding current characteristic significance, to obtain the current feature matrix; the current feature matrix is input into the trained emission factor prediction model to obtain the predicted emission factor corresponding to the next period; based on the calculation formula of the product carbon footprint, the predicted emission factor and the activity data in the next period preset in the production plan are used to calculate the predicted total carbon footprint of the product produced to the next period, and the predicted emission factor and the real-time activity data collected in the next period are used to calculate the actual total carbon footprint of the product produced to the next period; when the actual total carbon footprint exceeds the predicted total carbon footprint and the deviation is greater than a threshold value, a reduction strategy is timely formulated and the subsequent production process is optimized. 7.A machine learning based carbon footprint accounting system, characterized in that: The system comprises a processor, a memory, a communication interface and an information acquisition system, the processor stores computer program instructions for implementing the carbon footprint accounting method based on machine learning according to any one of claims 1 to 6, and the communication interface is in communication connection with the information acquisition system.

Images