A system and method for predicting and evaluating carbon emissions across the entire aviation fuel production process

Abstract

The present application relates to the field of carbon emission prediction, in particular to a kind of preparation aviation fuel whole process carbon emission prediction evaluation system and method;The method comprises, establish SAF carbon emission database, and the section data acquisition of carbon emission involved in the whole process of SAF produced by HEFA process;The collected section data are processed and constructed;Using the section data of real-time acquisition, construct training dynamic perception-mechanism constraint collaborative network model;Real-time acquisition current new section data, through dynamic perception-mechanism constraint collaborative network model, obtain the carbon emission evaluation value of current time HEFA process and the prediction sequence of future multiple time steps;The total amount of carbon emission prediction value is calculated;The prediction sequence of future multiple time steps and the total amount of carbon emission prediction value are compared with the preset carbon emission target threshold value, when exceed standard, then the section parameter of HEFA process is optimized;The present application realizes the technical effect of providing timely and reliable data support for production process.

Description

Technical Field

[0001] This invention relates to the field of carbon emission prediction, and specifically to a predictive assessment system and method for carbon emissions from the entire process of aviation fuel preparation. Background Technology

[0002] Sustainable aviation fuel (SAF) is a key pathway for the aviation industry to achieve carbon neutrality. Among these technologies, HEFA (hydrogenated ester and fatty acid) production is currently the most mature and widely used SAF production technology. However, accurately assessing its carbon emissions throughout its entire lifecycle, especially achieving real-time, dynamic assessment and prediction, remains a core challenge for the industry.

[0003] Current traditional carbon emission assessment methods rely on static life cycle carbon emission assessments (LCA) based on annual or quarterly average data. These methods lack the means to monitor the dynamic changes in carbon emissions during actual production processes, and therefore cannot provide timely and reliable data support for production processes. Furthermore, existing methods are also inadequate for predicting future carbon emissions during production processes.

[0004] In summary, traditional technologies can no longer meet the relevant requirements for carbon emissions throughout the entire life cycle. Therefore, developing a way to provide timely and reliable data support for the SAF production process and to conduct real-time assessment and prediction of carbon emissions in the future production process has become an urgent technical challenge to be solved in this field.

[0005] For example, a Chinese patent, publication number CN121581881A, publication date February 27, 2026, entitled "A Method for Full Life Cycle Carbon Emission Assessment Based on Substation Prediction Model," describes a method for assessing full life cycle carbon emissions based on a substation prediction model. This method includes: constructing an operational structure tree based on the distribution characteristics and functional relationships of physical entities in a substation, using carbon emission entities at each stage of the life cycle as tree nodes; mapping the physical attributes and dynamic data of each carbon emission entity to a digital twin platform based on the node hierarchy in the operational structure tree to obtain virtual nodes; associating the dynamic data of the virtual nodes with the static data at each stage of the life cycle based on the node identifiers in the operational structure tree to obtain a full life cycle data chain; calculating the carbon emission results at each stage of the life cycle based on the associated data of each tree node in the full life cycle data chain; and aggregating the carbon emission results of the lower-level nodes of each tree node upwards according to the parent-child node relationship in the operational structure tree to obtain the total carbon emission assessment.

[0006] While the aforementioned patents improve the accuracy of carbon emission assessment results, they are still based on static, full life-cycle assessments, which cannot provide timely and reliable data support for the production process, and make it difficult to conduct real-time assessments and predictions of future carbon emissions from the production process. Summary of the Invention

[0007] To address the problems existing in the prior art, this application provides a predictive assessment system and method for carbon emissions throughout the entire process of aviation fuel preparation, which can accurately assess carbon emissions throughout its entire life cycle and make real-time predictions of carbon emissions during future production processes.

[0008] To achieve the above-mentioned technical effects, the specific technical solution of this application is as follows: In a first aspect, a predictive assessment system for carbon emissions throughout the entire process of aviation fuel production includes: The HEFA process data acquisition module is used to establish a SAF carbon emission database and to collect real-time data on carbon emission processes and indicators throughout the entire HEFA process for producing SAF. The HEFA process data processing module is used to perform time alignment, missing value handling, outlier handling, standardization, and dataset construction on real-time acquired process data. The network model construction module is used to construct a dynamic sensing-mechanism constraint collaborative network model; the dynamic sensing-mechanism constraint collaborative network model is used to obtain the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps; The carbon emission prediction module calculates the total predicted carbon emissions based on the current carbon emission assessment value of the HEFA process and the prediction sequence of multiple future time steps. The decision-making and optimization module compares the predicted values ​​from the carbon emission prediction module with the preset carbon emission target thresholds. If the values ​​exceed the standard, the module optimizes the process parameters of the HEFA process.

[0009] Furthermore, the dynamic perception-mechanism constraint collaborative network model includes a dynamic perception module, a physical mechanism constraint module, a dynamic-mechanism interaction attention module, a multi-task output module, and a model accuracy optimization module; the dynamic perception module and the physical mechanism constraint module transmit data to the dynamic-mechanism interaction attention module; the dynamic-mechanism interaction attention module transmits data to the multi-task output module; and the multi-task output module transmits data to the model accuracy optimization module.

[0010] Furthermore, the dynamic perception module uses a temporal convolutional network layer to capture local and long-term dependencies in the HEFA process time-series data. Based on the output of the temporal convolutional network layer, a gated recurrent unit layer is used to learn the dynamic evolution law of the sequence in the time dimension to obtain dynamic feature encoding. The physical mechanism constraint module calculates physical guidance encoding based on the parameters of the HEFA full-process production data. The dynamic-mechanism interaction attention module, based on the dynamic feature encoding and physical guidance encoding, obtains fused feature output through attention weighting and feature fusion. The multi-task output module maps the fused features to the final prediction target, outputting the carbon emission assessment value of the HEFA process at the current moment and the predicted sequence for multiple future time steps. The model accuracy optimization module is used to train the dynamic perception-mechanism constraint collaborative network model end-to-end by minimizing the composite loss function, define the loss function, and determine whether the loss function value reaches the set standard threshold range.

[0011] Secondly, a method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation, based on the predictive and assessment system for carbon emissions throughout the entire process of aviation fuel preparation described in the first aspect, specifically includes the following method steps: Step S1: Establish a SAF carbon emission database and collect real-time data on the carbon emission links and indicators of the entire HEFA process for producing SAF. Step S2: Perform time alignment, missing value handling, outlier handling, standardization, and dataset construction on the real-time collected work segment data; Step S3: Using real-time collected work section data, construct and train a dynamic perception-mechanism constraint collaborative network model; Step S4: Collect the current new process segment data in real time, and obtain the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps through the dynamic perception-mechanism constraint collaborative network model; calculate the total carbon emission prediction value based on the current carbon emission assessment value of the HEFA process and the prediction sequence for multiple future time steps. Step S5: Compare the predicted sequence of future time steps and the predicted total carbon emissions with the preset carbon emission target threshold. If the threshold is exceeded, optimize the process parameters of the HEFA process.

[0012] Furthermore, the process data includes the feedstock section, pretreatment section, hydrotreating section, and isomerization and fractionation section; the feedstock section includes feedstock oil type, feedstock weight, acid value, iodine value, and fatty acid composition; the pretreatment section includes ester by-product quantity, pretreatment temperature, pretreatment pressure, and steam consumption; the hydrotreating section includes hydrotreating reactor temperature, hydrotreating reactor pressure, hydrogen-to-oil ratio, volumetric hourly space velocity, instantaneous fresh hydrogen flow rate, circulating hydrogen compressor power, and reaction furnace fuel consumption; the isomerization and fractionation section includes isomerization and fractionation reactor temperature, isomerization and fractionation reactor pressure, isomerization hydrogen-to-oil ratio, top / bottom temperature of each column, top / bottom pressure of each column, reflux ratio, and SAF production flow rate.

[0013] Furthermore, in step S3, the dynamic perception module of the dynamic perception-mechanism constraint collaborative network model is first used to capture the local and long-term dependencies in the HEFA process time series data through temporal convolutional network layers. Specifically, causal dilated convolution is used to ensure that the output at time step t depends only on the inputs at and before time step t, thus capturing the local and long-term dependencies in the HEFA process time series data. After residual block processing through multiple layers of temporal convolutional networks, the high-level temporal feature tensor H is obtained. tcn .

[0014] Furthermore, the high-level temporal feature tensor H tcn The calculation formula is as follows: ; in, is the output of a single convolutional kernel at time step t; K is the kernel size; d is the dilation factor, which increases exponentially with the number of network layers; For the input time series dataset at time steps The value; is the weight parameter of the convolution kernel at the k-th position; b is the bias term; Output data based on a single convolution kernel The temporal feature matrix M of a single sample is calculated, and the temporal feature tensor H is finally obtained by combining multiple samples. tcn The calculation formula is as follows: ; ; Temporal feature tensor H based on the output of a temporal convolutional network tcn The dynamic evolution of the sequence in the time dimension is learned by using a gated recurrent unit layer: the calculation of the gated recurrent unit at each time step t involves updating the gate. Reset door And candidate states; the calculation formula for the gated recurrent unit layer is as follows: ; ; ; ; in, It is the input of the current time step, i.e., H tcn A slice at time step t; It is the hidden state at time t-1; It is the Sigmoid activation function; W1, W2, and W3 are trainable parameter matrices, and B1, B2, and B3 are trainable vectors; It is a candidate hidden state; It is the final hidden state at the current time step; The hidden state at the last time step is taken as the final output of this module, which is the dynamic feature encoding H. dynamic =h T .

[0015] Furthermore, in step S3, the physical mechanism constraint module of the dynamic perception-mechanism constraint collaborative network model calculates the physical guidance code based on the parameters of the HEFA full-process production data. The specific calculation formula is as follows:

[0016] in, Theoretical physical baseline values ​​are artificially set for the entire process data of SAF production using the HEFA process. and These are all weights of the model. and Both are model biases, and ReLU is the activation function; The physical boot code for the output; In step S3, the dynamic-mechanism interaction attention module of the dynamic perception-mechanism constraint collaborative network model is used. Based on dynamic feature encoding and physical guided encoding, the fused feature output is obtained through attention weighting and feature fusion. The specific calculation method is as follows: First, perform a linear projection transformation on the dynamic feature encoding and the physical guidance encoding: ; ; ; in, For query vector, For key vectors, It is a value vector; , and All are trainable projection matrices; , and All of these are corresponding bias terms; Next, attention weights are calculated, and the weight distribution is calculated using a scaled dot product attention mechanism;

[0017] Where A is the attention weight matrix; This is a normalization function that makes the sum of the weights of each row equal to 1. This is the scaling factor; For transpose; Finally, feature fusion calculation is performed, where physical features are weighted and fused based on attention weights to obtain the fused features. : .

[0018] Furthermore, in step S3, a multi-task output module of a dynamic perception-mechanism constraint collaborative network model is used to map the fused features to the final prediction target, outputting the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps. The specific calculation formula is as follows: ; ; in, It is an assessment of the carbon emission intensity at the current moment; It is a carbon emission prediction sequence for the next t time steps; These are all trainable parameters of the output layer; In step S3, the model accuracy optimization module of the dynamic perception-mechanism constraint collaborative network model is used to perform end-to-end training by minimizing the composite loss function. The loss function is defined, and it is determined whether the loss function value reaches the set standard threshold range. The loss function definition formula is as follows: ; ; in This is the total loss function; is the mean squared error loss function; N is the total number of samples, i.e., the number of data points involved in the calculation; i is the sample index, indicating that the current sample is the i-th sample; The current value represents the true label for the assessment at the current moment; t represents the time step. This represents the predicted carbon emissions at the i-th future time point. This represents the actual carbon emissions at the i-th future time point; , , These are all hyperparameters used to balance various losses; When the loss function value reaches a manually set standard threshold range, the model is considered to have met the preset conditions. ; in, The accuracy of current carbon emission assessments is set artificially. To improve the accuracy of future carbon emission predictions The threshold for compliance; This represents the total number of future time steps when making multi-step predictions.

[0019] Furthermore, in step S4, the predicted total carbon emissions are calculated using the predicted carbon emission prediction sequence. Based on the corresponding predicted product output sequence, the predicted total carbon emissions for a specified future time period are calculated. The specific calculation formula is as follows: ; In the formula, The meaning is the predicted value of total carbon emissions. The meaning is the predicted carbon emissions at time step t.

[0020] Based on the above technical solution, the beneficial effects of this application are as follows: 1. This invention constructs a dynamic perception-mechanism constraint collaborative network model by employing a network model construction module. Using this model, carbon emissions throughout the entire life cycle are accurately assessed in real time through real-time data collection, and carbon emissions in future production processes are predicted in real time, thereby achieving the technical effect of providing timely and reliable data support for the production process.

[0021] 2. The dynamic sensing module used in this invention combines a temporal convolutional network and a gated recurrent unit to capture the local mutation characteristics and long-term evolution patterns of HEFA process time series data in real time. Furthermore, the physical mechanism constraint module uses physical guided encoding to correct the deviation of pure data prediction.

[0022] 3. This invention establishes a SAF carbon emission database through the HEFA process data acquisition module, and collects real-time data on the carbon emission links and indicators of the entire HEFA process for producing SAF, thereby assisting subsequent steps in solving the technical problem of strong lag caused by the use of static data assessment in the prior art.

[0023] 4. This invention solves the technical problems of poor timeliness and strong lag in the prior art by using a multi-task output module to output the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence of multiple future time steps, and obtains the total carbon emission prediction value for a specified time period. Attached Figure Description

[0024] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0026] Example 1 like Figure 1 As shown, a predictive assessment system for carbon emissions throughout the entire process of aviation fuel production includes: The HEFA process data acquisition module is used to establish a SAF carbon emission database and to collect real-time data on carbon emission processes and indicators throughout the entire HEFA process for producing SAF. The HEFA process data processing module is used to perform time alignment, missing value handling, outlier handling, standardization, and dataset construction on real-time acquired process data. The network model building module is used to construct a dynamic sensing-mechanism constraint collaborative network model; the dynamic sensing-mechanism constraint collaborative network model is used to obtain the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps; The carbon emission prediction module calculates the total predicted carbon emissions based on the current carbon emission assessment value of the HEFA process and the prediction sequence of multiple future time steps. The decision-making and optimization module compares the predicted values ​​from the carbon emission prediction module with the preset carbon emission target thresholds. If the values ​​exceed the standard, the module optimizes the process parameters of the HEFA process.

[0027] The dynamic perception-mechanism constraint collaborative network model includes a dynamic perception module, a physical mechanism constraint module, a dynamic-mechanism interaction attention module, a multi-task output module, and a model accuracy optimization module. The dynamic perception module and the physical mechanism constraint module transmit data to the dynamic-mechanism interaction attention module; the dynamic-mechanism interaction attention module transmits data to the multi-task output module; and the multi-task output module transmits data to the model accuracy optimization module.

[0028] The dynamic perception module uses a temporal convolutional network layer to capture local and long-term dependencies in the HEFA process time-series data. Based on the output of the temporal convolutional network layer, a gated recurrent unit layer is used to learn the dynamic evolution of the sequence in the time dimension, resulting in dynamic feature encoding. The physical mechanism constraint module calculates physical guidance encoding based on the parameters of the entire HEFA production data. The dynamic-mechanism interaction attention module, based on the dynamic feature encoding and physical guidance encoding, fuses features through attention weighting to obtain fused feature output. The multi-task output module maps the fused features to the final prediction target, outputting the carbon emission assessment value of the HEFA process at the current moment and the predicted sequence for multiple future time steps. The model accuracy optimization module is used to train the dynamic perception-mechanism constraint collaborative network model end-to-end by minimizing the composite loss function, defining the loss function, and determining whether the loss function value reaches the set standard threshold range.

[0029] Example 2 like Figure 2 As shown, a method for predicting and assessing carbon emissions from the entire process of aviation fuel preparation is implemented based on a predictive and assessment system for carbon emissions from the entire process of aviation fuel preparation as described in Example 1. The method specifically includes the following steps: Step S1: Establish a SAF carbon emission database and collect real-time data on carbon emission processes and indicators throughout the entire HEFA process for SAF production. The establishment of the SAF carbon emission database is based on existing SAF life-cycle carbon emission databases (such as China Carbon Accounting Database CEADS, Swiss Ecoinvent, German Gabi, EU ELCD, etc.), combined with the factory's existing HEFA process SAF production database. Step S2: Perform time alignment, missing value handling, outlier handling, standardization, and dataset construction on the real-time collected work segment data; Step S3: Using real-time collected work section data, construct and train a dynamic perception-mechanism constraint collaborative network model; Step S4: Collect the current new process segment data in real time, and obtain the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps through the dynamic perception-mechanism constraint collaborative network model; calculate the total carbon emission prediction value based on the current carbon emission assessment value of the HEFA process and the prediction sequence for multiple future time steps. Step S5: Compare the predicted sequence of future time steps and the predicted total carbon emissions with the preset carbon emission target threshold. If the threshold is exceeded, optimize the process parameters of the HEFA process.

[0030] Example 3 Based on Example 2, the process data includes the feedstock section, pretreatment section, hydrotreating section, and isomerization and fractionation section. The feedstock section includes feedstock oil type, feedstock weight, acid value, iodine value, and fatty acid composition. The pretreatment section includes ester by-product quantity, pretreatment temperature, pretreatment pressure, and steam consumption. The hydrotreating section includes hydrotreating reactor temperature, hydrotreating reactor pressure, hydrogen-to-oil ratio, volumetric hourly space velocity, instantaneous fresh hydrogen flow rate, circulating hydrogen compressor power, and reaction furnace fuel consumption. The isomerization and fractionation section includes isomerization and fractionation reactor temperature, isomerization and fractionation reactor pressure, isomerization hydrogen-to-oil ratio, top / bottom temperature of each column, top / bottom pressure of each column, reflux ratio, and SAF product flow rate. All index names and data formats are shown in Table 1 below.

[0031] Table 1: All indicator names and data formats

[0032] The specific methods for performing time alignment, missing value handling, outlier handling, standardization, and dataset construction on the real-time collected work segment data in step S2 are as follows: For time alignment, linear interpolation is used to unify the raw data of the entire HEFA process into the same equally spaced time series, providing a consistent basis for subsequent time series modeling. The linear interpolation calculation formula is as follows: ; in: It is the target timestamp (the point in time on the baseline). It is in the original data and Adjacent timestamps before and after; yes The original data value corresponding to the time step; It is calculated Interpolation at time points.

[0033] In terms of missing value handling, linear interpolation (the calculation formula is the same as the above formula) is used to fill in missing values ​​in the original data caused by various reasons, so as to ensure the integrity of the data sequence.

[0034] For outlier handling, the 3σ principle is used to calculate the mean and standard deviation of each feature column. Data points that deviate from the mean by more than 3σ are considered outliers. The formula for calculating the 3σ principle is as follows: ; in: The data points to be detected; This is the mean of all data for this feature; The standard deviation of all data for this feature Standardization: Z-score standardization is applied to the original dataset to eliminate differences in units and ranges between different features. The calculation formula is as follows: ; in, These are the original eigenvalues; This is the mean of the feature in the original dataset; The standard deviation of this feature on the original dataset; This is the standardized value.

[0035] Dataset construction: Based on the standardized data, a HEFA process time series dataset and the corresponding full-process carbon emission values ​​under the time series were established.

[0036] The HEFA process time series dataset format is as follows: ={Time t; [Raw Material Section: Raw material type, raw material input weight, fatty acid composition, acid value, iodine value]; [Pretreatment Section: Ester by-product quantity, pretreatment temperature, pressure steam consumption]; [Hydrogenation Section: Reactor temperature and pressure, hydrogen-to-oil ratio...]; [Isomerization and Fractionation Section: Reactor temperature and pressure, isomerization hydrogen-to-oil ratio...]}, where The dataset is for time step t.

[0037] Corresponding HEFA carbon emission values ​​(Y) under the time series t Based on HEFA process time-series data, calculations are performed using the established Life Cycle Assessment (LCA) method. The final HEFA carbon emission value (Y) for the corresponding time series is obtained. t ), which is the carbon equivalent emission value (gCO2e / MJ) of a unit SAF product (one megajoule MJ).

[0038] Example 4 Based on Example 3, in step S3, the dynamic perception module of the dynamic perception-mechanism constraint collaborative network model is first used to capture the local and long-term dependencies in the HEFA process time series data through the temporal convolutional network layer. The specific method is as follows: causal dilated convolution is used to ensure that the output of time step t depends only on the input of time step t and before, thereby capturing the local and long-term dependencies in the HEFA process time series data. After residual block processing by the multi-layer temporal convolutional network, the high-level temporal feature tensor H is obtained. tcn .

[0039] High-level temporal feature tensor H tcn The calculation formula is as follows: ; in, is the output of a single convolutional kernel at time step t; K is the kernel size, which can be set and modified manually; d is the dilation factor, which increases exponentially with the number of network layers; For the input time series dataset at time steps The value; is the weight parameter of the convolution kernel at the k-th position; b is the bias term; Output data based on a single convolution kernel The temporal feature matrix M of a single sample is calculated, and the temporal feature tensor H is finally obtained by combining multiple samples. tcn The calculation formula is as follows: ; .

[0040] Temporal feature tensor H based on the output of a temporal convolutional network tcn The dynamic evolution of the sequence in the time dimension is learned by using a gated recurrent unit layer: the calculation of the gated recurrent unit at each time step t involves updating the gate. Reset door And candidate states; the calculation formula for the gated recurrent unit layer is as follows: ; ; ; ; in, It is the input of the current time step, i.e., H tcn A slice at time step t; It is the hidden state at time t-1; It is the Sigmoid activation function; W1, W2, and W3 are trainable parameter matrices, and B1, B2, and B3 are trainable vectors; It is a candidate hidden state; It is the final hidden state at the current time step; The hidden state at the last time step is taken as the final output of this module, which is the dynamic feature encoding H. dynamic =h T .

[0041] In step S3, the physical mechanism constraint module of the dynamic perception-mechanism constraint collaborative network model is used to calculate the physical guidance code based on the parameters of the HEFA full-process production data. The specific calculation formula is as follows: ; in, The theoretical physical baseline values ​​are artificially set for the entire process data of SAF production using the HEFA process (such as setting the physical baseline for hydrotreating temperature to 800℃~1200℃ and the physical baseline for isomerization hydrogen-to-oil ratio to 100:1~300:1). and These are all weights of the model. and Both are model biases, and ReLU is the activation function; This is the physical boot code for the output. In step S3, the dynamic-mechanism interaction attention module of the dynamic perception-mechanism constraint collaborative network model is used. Based on dynamic feature encoding and physical guided encoding, the fused feature output is obtained through attention weighting and feature fusion. The specific calculation method is as follows: First, perform a linear projection transformation on the dynamic feature encoding and the physical guidance encoding: ; ; ; in, For query vector, For key vectors, It is a value vector; , and All are trainable projection matrices; , and All of these are corresponding bias terms; Next, attention weights are calculated, and the weight distribution is calculated through the scaling dot product attention mechanism.

[0042] ; Where A is the attention weight matrix; The normalization function ensures that the sum of the weights for each row is 1; This is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product; For transpose; Finally, feature fusion calculation is performed, where physical features are weighted and fused based on attention weights to obtain the fused features. : .

[0043] In step S3, the multi-task output module of the dynamic perception-mechanism constraint collaborative network model is used to map the fused features to the final prediction target, outputting the carbon emission assessment value of the HEFA process at the current time and the prediction sequence for multiple future time steps. The specific calculation formula is as follows: ; ; in, It is an assessment of the carbon emission intensity at the current moment; It is a carbon emission prediction sequence for the next t time steps; All of these are trainable parameters of the output layer.

[0044] In step S3, the model accuracy optimization module of the dynamic perception-mechanism constraint collaborative network model is used. End-to-end training is performed by minimizing the composite loss function. The loss function is defined, and it is determined whether the loss function value reaches the set standard threshold range. The formula for the loss function is as follows: ; ; in This is the total loss function; is the mean squared error loss function; N is the total number of samples, i.e., the number of data points involved in the calculation; i is the sample index, indicating that the current sample is the i-th sample; The current value represents the true label for the assessment at the current moment; t represents the time step. This represents the predicted carbon emissions at the i-th future time point. This represents the actual carbon emissions at the i-th future time point; , and These are all hyperparameters used to balance various losses.

[0045] When the loss function value reaches a manually set standard threshold range, the model is considered to have met the preset conditions. That is: ; in, The accuracy of current carbon emission assessments is set artificially. To improve the accuracy of future carbon emission predictions To predict the threshold for achieving the target loss in accordance with physical laws, the reference thresholds in this embodiment are 0.5, 1.0, and 0.1; This represents the total number of future time steps when making multi-step predictions.

[0046] In step S4, the total carbon emissions are predicted, and the predicted carbon emissions sequence is used. Based on the corresponding predicted product output sequence, the predicted total carbon emissions for a specified future time period are calculated. The specific calculation formula is as follows: ; In the formula, The meaning is the predicted value of total carbon emissions. The meaning is the predicted carbon emissions at time step t.

[0047] In step S5, the time-series carbon emission value and total carbon emission prediction value obtained based on the aforementioned HEFA production process carbon emission prediction are compared with the preset carbon emission target threshold (such as the international SAF full life cycle carbon emission standard, industry legal and regulatory requirements, and enterprise internal standards).

[0048] If carbon emissions meet the standards, the system's decision-making and optimization module automatically determines that the current carbon emissions comply with the standards, generates and outputs a SAF production carbon emission compliance report, and continuously monitors carbon emissions over time. If carbon emissions do not meet the standards, the system issues an over-standard warning and enters the HEFA process optimization process. Based on the current HEFA process conditions and prediction model, the decision-making and optimization module will sequentially adjust the parameters of the hydrotreating section (reactor temperature, pressure, hydrogen-to-oil ratio, and volumetric hourly space velocity, etc.) and the parameters of the isomerization / fractionation section (reactor temperature, pressure, isomerization hydrogen-to-oil ratio, and reflux ratio, etc.), and import the data of the adjusted parameters into the DyMeC-Net model for carbon emission prediction to assess whether carbon emissions meet the standards. If the target is not met, the above parameters will be optimized again until the target is met; if the target is met, the feasibility and rationality of the process will be assessed manually. If it is not reasonable, the process parameters will be corrected and optimized manually, and the prediction sequence and total carbon emission prediction value for multiple future time steps will be predicted again; if it is reasonable, the process can be carried out and carbon emission time series monitoring will be carried out continuously. This optimization part can be completed by those skilled in the art using existing conventional technical knowledge.

[0049] The above description is a detailed description of the preferred embodiments of this application. However, the embodiments are not intended to limit the scope of this patent application. All equivalent changes or modifications made under the technical spirit of this application should fall within the patent scope covered by this application.

Claims

1. A predictive assessment system for carbon emissions throughout the entire process of aviation fuel production, characterized in that, include: The HEFA process data acquisition module is used to establish a SAF carbon emission database and to collect real-time data on carbon emission processes and indicators throughout the entire HEFA process for producing SAF. The HEFA process data processing module is used to perform time alignment, missing value handling, outlier handling, standardization, and dataset construction on real-time acquired process data. The network model construction module is used to construct a dynamic sensing-mechanism constraint collaborative network model; the dynamic sensing-mechanism constraint collaborative network model is used to obtain the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps; The carbon emission prediction module calculates the total predicted carbon emissions based on the current carbon emission assessment value of the HEFA process and the prediction sequence of multiple future time steps. The decision-making and optimization module compares the predicted values ​​from the carbon emission prediction module with the preset carbon emission target thresholds. If the values ​​exceed the standard, the module optimizes the process parameters of the HEFA process.

2. The prediction and assessment system for the entire process of aviation fuel preparation according to claim 1, characterized in that: The dynamic perception-mechanism constraint collaborative network model includes a dynamic perception module, a physical mechanism constraint module, a dynamic-mechanism interaction attention module, a multi-task output module, and a model accuracy optimization module. The dynamic perception module and the physical mechanism constraint module transmit data to the dynamic-mechanism interaction attention module. The dynamic-mechanism interaction attention module transmits data to the multi-task output module. The multi-task output module transmits data to the model accuracy optimization module.

3. The prediction and assessment system for the entire process of aviation fuel preparation according to claim 2, characterized in that: The dynamic perception module uses a temporal convolutional network layer to capture local and long-term dependencies in the HEFA process time series data. Based on the output of the temporal convolutional network layer, a gated recurrent unit layer is used to learn the dynamic evolution law of the sequence in the time dimension to obtain dynamic feature encoding. The physical mechanism constraint module calculates physical guidance encoding based on the parameters of the HEFA full-process production data. The dynamic-mechanism interaction attention module, based on dynamic feature encoding and physical guided encoding, obtains fused feature output through attention weighting and feature fusion; The multi-task output module maps the fused features to the final prediction target, outputting the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps; the model accuracy optimization module is used to train the dynamic perception-mechanism constraint collaborative network model end-to-end by minimizing the composite loss function, define the loss function, and determine whether the loss function value reaches the set standard threshold range.

4. A method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation, implemented using the predictive and assessment system for carbon emissions throughout the entire process of aviation fuel preparation as described in any one of claims 1 to 3, characterized in that, Specifically, the methods and steps are as follows: Step S1: Establish a SAF carbon emission database and collect real-time data on the carbon emission links and indicators of the entire HEFA process for producing SAF. Step S2: Perform time alignment, missing value handling, outlier handling, standardization, and dataset construction on the real-time collected work segment data; Step S3: Using real-time collected work section data, construct and train a dynamic perception-mechanism constraint collaborative network model; Step S4: Collect the current new process segment data in real time, and obtain the carbon emission assessment value of the HEFA process at the current moment and the prediction sequence for multiple future time steps through the dynamic perception-mechanism constraint collaborative network model; calculate the total carbon emission prediction value based on the current carbon emission assessment value of the HEFA process and the prediction sequence for multiple future time steps. Step S5: Compare the predicted sequence of future time steps and the predicted total carbon emissions with the preset carbon emission target threshold. If the threshold is exceeded, optimize the process parameters of the HEFA process.

5. The method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation according to claim 1, characterized in that: The process data includes the feedstock section, pretreatment section, hydrotreating section, and isomerization and fractionation section. The feedstock section includes feedstock oil type, feedstock weight, acid value, iodine value, and fatty acid composition. The pretreatment section includes ester by-product quantity, pretreatment temperature, pretreatment pressure, and steam consumption. The hydrotreating section includes hydrotreating reactor temperature, hydrotreating reactor pressure, hydrogen-to-oil ratio, volumetric hourly space velocity, instantaneous fresh hydrogen flow rate, circulating hydrogen compressor power, and reaction furnace fuel consumption. The isomerization and fractionation section includes isomerization and fractionation reactor temperature, isomerization and fractionation reactor pressure, isomerization hydrogen-to-oil ratio, top / bottom temperatures of each column, top / bottom pressures of each column, reflux ratio, and SAF production flow rate.

6. The method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation according to claim 1, characterized in that: In step S3, the dynamic perception module of the dynamic perception-mechanism constraint collaborative network model is first used to capture the local and long-term dependencies in the HEFA process time series data through temporal convolutional network layers. Specifically, causal dilated convolution is used to ensure that the output at time step t depends only on the inputs at and before time step t, thus capturing the local and long-term dependencies in the HEFA process time series data. After residual block processing through multiple layers of temporal convolutional networks, the high-level temporal feature tensor H is obtained. tcn .

7. The method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation according to claim 6, characterized in that: High-level temporal feature tensor H tcn The calculation formula is as follows: ； in, is the output of a single convolution kernel at time step t; K is the kernel size; d is the dilation factor; For the input time series dataset at time steps The value; is the weight parameter of the convolution kernel at the k-th position; b is the bias term; Output data based on a single convolution kernel The temporal feature matrix M of a single sample is calculated, and the temporal feature tensor H is finally obtained by combining multiple samples. tcn The calculation formula is as follows: ； ； Temporal feature tensor H based on the output of a temporal convolutional network tcn The dynamic evolution of the sequence in the time dimension is learned by using a gated recurrent unit layer: the calculation of the gated recurrent unit at each time step t involves updating the gate. Reset door And candidate states; the calculation formula for the gated recurrent unit layer is as follows: ； ； ； ； in, It is the input of the current time step, i.e., H tcn A slice at time step t; It is the hidden state at time t-1; It is the Sigmoid activation function; W1, W2, and W3 are trainable parameter matrices, and B1, B2, and B3 are trainable vectors; It is a candidate hidden state; It is the final hidden state at the current time step; Taking the hidden state at the last time step as the final output, which is the dynamic feature encoding H. dynamic =h T .

8. The method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation according to claim 7, characterized in that: In step S3, the physical mechanism constraint module of the dynamic perception-mechanism constraint collaborative network model is used to calculate the physical guidance code based on the parameters of the HEFA full-process production data. The specific calculation formula is as follows: ； in, Theoretical physical baseline values ​​are artificially set for the entire process data of SAF production using the HEFA process. and These are all weights of the model. and Both are model biases, and ReLU is the activation function; The physical boot code for the output; In step S3, the dynamic-mechanism interaction attention module of the dynamic perception-mechanism constraint collaborative network model is used. Based on dynamic feature encoding and physical guided encoding, the fused feature output is obtained through attention weighting and feature fusion. The specific calculation method is as follows: First, perform a linear projection transformation on the dynamic feature encoding and the physical guidance encoding: ； ； ； in, For query vector, For key vectors, It is a value vector; , and All are trainable projection matrices; , and All of these are corresponding bias terms; Next, attention weights are calculated, and the weight distribution is calculated using a scaled dot product attention mechanism; ； Where A is the attention weight matrix; This is a normalization function that makes the sum of the weights of each row equal to 1. This is the scaling factor; For transpose; Finally, feature fusion calculation is performed, where physical features are weighted and fused based on attention weights to obtain the fused features. : 。 9. The method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation according to claim 8, characterized in that: In step S3, the multi-task output module of the dynamic perception-mechanism constraint collaborative network model is used to map the fused features to the final prediction target, outputting the carbon emission assessment value of the HEFA process at the current time and the prediction sequence for multiple future time steps. The specific calculation formula is as follows: ； ； in, It is an assessment of the carbon emission intensity at the current moment; It is a carbon emission prediction sequence for the next t time steps; These are all trainable parameters of the output layer; In step S3, the model accuracy optimization module of the dynamic perception-mechanism constraint collaborative network model is used to perform end-to-end training by minimizing the composite loss function. The loss function is defined, and it is determined whether the loss function value reaches the set standard threshold range. The loss function definition formula is as follows: ； ； in, This is the total loss function; is the mean squared error loss function; N is the total number of samples, i.e., the number of data points involved in the calculation; i is the sample index, indicating that the current sample is the i-th sample; The current value represents the true label for the assessment at the current moment; t represents the time step. Let be the predicted carbon emissions at the i-th time point in the future; This represents the actual carbon emissions at the i-th future time point; , , These are all hyperparameters used to balance various losses; When the loss function value reaches a manually set standard threshold range, the model is considered to have met the preset conditions. ； in, The accuracy of current carbon emission assessments is set artificially. To improve the accuracy of future carbon emission predictions The threshold for compliance; This represents the total number of future time steps when making multi-step predictions.

10. The method for predicting and assessing carbon emissions throughout the entire process of aviation fuel preparation according to claim 9, characterized in that: In step S4, the predicted total carbon emissions are used to obtain the predicted carbon emissions sequence. Based on the corresponding predicted product output sequence, the predicted total carbon emissions for a specified future time period are calculated. The specific calculation formula is as follows: ； In the formula, The meaning is the predicted value of total carbon emissions. The meaning is the predicted carbon emissions at time step t.

Images