A cold chain transportation carbon emission optimization method and system based on dynamic feature adaptive hierarchical modeling

Abstract

The application discloses a cold chain transportation carbon emission optimization method and system based on dynamic feature adaptive hierarchical modeling, relates to the technical field of logistics management, and comprises the following steps: collecting cold chain transportation multi-source operation data, time aligning, abnormity processing and uniformly sampling; screening key features based on historical energy consumption and carbon emission data, and dividing basic parameters and dynamic parameters according to time stability; generating a confidence evaluation, freezing low-confidence features and triggering hierarchical adaptive adjustment; establishing a baseline carbon emission model based on the basic parameters, establishing a disturbance correction model based on the dynamic parameters, and superimposing to obtain real-time carbon emission; inputting a state sequence into a generated cognitive model to complete missing data and predict future carbon emission, and outputting control instructions by optimizing refrigeration parameters and transportation strategies under temperature control constraints. Through the technical scheme of the application, the stability, accuracy and robustness of estimation and prediction are significantly improved, and a low-carbon operation optimization closed loop is realized.

Description

Technical Field

[0001] This invention relates to the field of logistics management technology, and in particular to a method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling, and a system for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling. Background Technology

[0002] To ensure that the temperature inside the cargo compartment remains within the target range, cold chain transportation typically requires refrigeration equipment to continuously adjust its operating parameters based on vehicle conditions and environmental factors throughout the entire transportation process. In actual operation, factors such as vehicle driving status, road conditions, load variations, fluctuations in external temperature and humidity, and door opening during loading and unloading operations all contribute to significant time-varying characteristics in refrigeration load and vehicle energy consumption. This, in turn, leads to substantial fluctuations in carbon emissions across different mission phases, routes, and operating conditions. For cold chain companies, without reliable real-time carbon emission estimates and future trend predictions during mission execution, it will be difficult to implement refined emission reduction control and strategy optimization while meeting temperature control requirements.

[0003] In existing technologies, carbon emission accounting for cold chain transportation often employs fixed emission factor conversion, static parameter models, or empirical regression methods based on a limited number of features. These approaches tend to model influencing factors from different sources in a mixed manner, lacking a hierarchical representation of "inherent relatively stable factors of the task" and "dynamic disturbance factors during operation." This leads to the accumulation of biases in the model output when operating conditions change rapidly or parameters drift, resulting in insufficient stability in real-time carbon emission estimation.

[0004] At the same time, cold chain transportation multi-source operation data generally suffers from problems such as inconsistent sampling frequency, outliers, missing data and noise. Some solutions lack credibility assessment and degradation mechanisms for feature data quality. Low-credibility features are still used for model updates or decision calculations, which can easily amplify estimation errors and reduce the reliability of control commands.

[0005] Furthermore, applications aimed at optimizing carbon emissions control not only require estimating current carbon emissions but also predicting future carbon emissions changes over a period of time in order to adjust refrigeration equipment operating parameters and transportation operation strategies in advance. However, existing solutions often struggle to simultaneously achieve integrated processing of missing data completion and future carbon emissions prediction due to input deficiencies, feature distribution drift, and the effects of multivariate coupling. This results in insufficient usability of the prediction results, making it difficult to support an executable optimization control closed loop. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method and system for optimizing carbon emissions in cold chain transportation based on adaptive hierarchical modeling using dynamic features. The method preprocesses multi-source operational data from cold chain transportation to form standardized feature sequences suitable for modeling. Sensitivity / importance analysis is then performed on historical energy consumption and carbon emission data to screen key features. Furthermore, these key features are adaptively divided into a set of basic parameters and a set of dynamic parameters based on time stability, achieving a hierarchical expression of stable factors and dynamic disturbances. Low-confidence features are frozen based on confidence evaluation. Simultaneously, drift detection and type diagnosis mechanisms identify the drift type of feature variables and trigger differentiated response strategies to reduce the interference of abnormal or missing inputs on model updates and estimation results. Based on this, a baseline carbon emission estimation model and a perturbation carbon emission correction model are established and superimposed to output real-time carbon emission estimates. A generative cognitive model incorporating physical information constraints is then used to complete missing data and predict future carbon emissions. Finally, a hierarchical optimization framework outputs carbon emission optimization control commands for refrigeration equipment parameters and transportation strategies under temperature control constraints, thereby improving the stability of real-time estimation, the usability of predictions, and the executability of decisions.

[0007] To achieve the above objectives, this invention provides a method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling, comprising:

[0008] Collect multi-source operational data during the cold chain transportation process, and perform time alignment, outlier processing, and unified sampling frequency processing on the multi-source operational data to obtain a standardized feature sequence. The multi-source operational data includes at least energy consumption-related data and carbon emission-related data. The carbon emission-related data includes carbon emission measurement data and / or carbon emission conversion parameters.

[0009] Based on historical energy consumption data and historical carbon emission data, sensitivity analysis or feature importance calculation is performed on the feature variables in the standardized feature sequence to screen out a set of key features.

[0010] A time stability analysis is performed on the feature variables in the key feature set, and the key feature set is divided into a basic parameter set and a dynamic parameter set based on the stability analysis results.

[0011] Confidence evaluation results are generated for the feature variables in the key feature set. When the confidence evaluation result of a feature variable is lower than a preset threshold, the feature variable is frozen. Drift detection and drift type diagnosis are performed on the feature variables. The drift types include gradual drift, abrupt drift, periodic drift, and conceptual drift. The corresponding parameter set update strategy is triggered according to the drift type diagnosed, and the basic parameter set and the dynamic parameter set are adaptively adjusted.

[0012] A baseline carbon emission estimation model is established based on the set of basic parameters and the baseline carbon emission is output. A perturbation carbon emission correction model is established based on the set of dynamic parameters and the perturbation carbon emission is output. The baseline carbon emission and the perturbation carbon emission are superimposed to obtain the real-time carbon emission estimate.

[0013] The cognitive model is generated by inputting a state sequence containing the basic parameter values ​​corresponding to the basic parameter set, the dynamic parameter values ​​corresponding to the dynamic parameter set, real-time energy consumption data, and the real-time carbon emission estimate, and outputting a future carbon emission prediction value. The model also completes the state sequence when the input data is missing. The cognitive model introduces physical constraints based on the thermal balance relationship of the cargo compartment and includes an energy conservation consistency correction layer for performing consistency correction on the prediction output.

[0014] Based on the predicted future carbon emissions, the operating parameters of the refrigeration equipment and the transportation operation strategy are optimized and decisions are made, and carbon emission optimization control commands are output. The optimization decision adopts a hierarchical optimization framework and includes rolling time-domain solution based on model predictive control, and an iterative search is performed by constructing a Gaussian process surrogate model in the potential state vector space.

[0015] In the above technical solution, preferably, the multi-source operating data includes vehicle operating parameters, environmental parameters, refrigeration equipment status parameters, and operational behavior parameters;

[0016] The vehicle operating parameters include vehicle speed, acceleration, gradient, road type, and load factor; the environmental parameters include ambient temperature, humidity, solar radiation intensity, air pressure, and wind speed; the refrigeration equipment status parameters include compressor power, operating frequency, current, voltage, and number of start-stop cycles; and the operational behavior parameters include the number of door openings, door opening duration, idling time, and cold start frequency.

[0017] In the above technical solution, preferably, the sensitivity analysis includes: calculating a sensitivity index Si for each characteristic variable, wherein the sensitivity index Si satisfies ,in, The gradient between the feature variable and carbon emissions. The fluctuation range of the characteristic variable, This represents the fluctuation range of carbon emissions.

[0018] In the above technical solution, preferably, the time stability analysis includes calculating the standard deviation and frequency of change of the feature variables based on a sliding window;

[0019] When the standard deviation of the characteristic variable within the sliding window is less than the stability threshold σ b or the frequency of change of the characteristic variable f i Less than the low frequency threshold f bIf the characteristic variable is a parameter that can be configured or measured before the start of the transportation task, then the characteristic variable shall be included in the set of basic parameters.

[0020] Feature variables that do not meet the aforementioned conditions or rely on real-time sensing acquisition are included in the dynamic parameter set.

[0021] In the above technical solution, preferably, the process of generating the confidence evaluation result includes:

[0022] For each feature variable in the set of key features, obtain the corresponding signal quality index, missing ratio index, and noise level index.

[0023] The signal quality index, the missing ratio index, and the noise level index are fused and calculated according to the preset fusion rules to obtain the confidence evaluation result of the feature variables;

[0024] When the confidence evaluation result is lower than a preset threshold, the feature variable is frozen. The freezing process includes stopping the use of the feature variable for model updates or using the frozen value in the calculation.

[0025] In the above technical solution, preferably, the specific process of adaptively adjusting the basic parameter set and the dynamic parameter set includes:

[0026] Drift detection and drift type diagnosis are performed on the feature variables in the key feature set. The drift types include gradual drift, abrupt drift, periodic drift, and conceptual drift. Gradual drift is detected by trend test of sliding window mean, abrupt drift is detected by abrupt change in cumulative sum statistic, periodic drift is detected by frequency domain analysis, and conceptual drift is detected by the change in the correlation between feature variables and carbon emission measurement data.

[0027] Based on the drift type diagnosed, a corresponding parameter set update strategy is triggered: when a gradual drift is detected, the weight coefficients of the baseline carbon emission estimation model are updated incrementally online; when a sudden drift is detected, the drifting feature variables are transferred between the basic parameter set and the dynamic parameter set, and the coefficients of the corresponding model are re-estimated; when a periodic drift is detected, time periodic encoding is introduced for the drifting feature variables; when a conceptual drift is detected, the key feature set is re-screened and the basic parameter set and dynamic parameter set are re-partitioned.

[0028] After the parameter set update strategy is executed, the change of carbon emission estimation residual is monitored. When the carbon emission estimation residual does not converge to the preset residual threshold, the response strategy is switched to a higher level according to the preset strategy upgrade rules.

[0029] In the above technical solution, preferably, the baseline carbon emission estimation model includes a steady-state energy consumption model and an energy consumption-carbon emission conversion, wherein the steady-state energy consumption model is used to obtain the baseline energy consumption estimate based on the set of basic parameters. E base ,in ,in For the first in the set of basic parameters i One basic parameter, β i For the corresponding weighting coefficients, β 0 represents the baseline constant term;

[0030] The baseline energy consumption estimate E base The baseline carbon emission is calculated based on the aforementioned carbon emission conversion parameters;

[0031] The perturbation carbon emission correction model is used to obtain the perturbation energy consumption estimate based on the dynamic parameter set. E yn ( t ),in , where γ k ( t ) represents the dynamic sensitivity coefficient, Δ x k ( t ) represents the deviation of the dynamic parameters in the set of dynamic parameters from the reference.

[0032] The estimated disturbance energy consumption E yn ( t The perturbation carbon emissions are obtained based on the aforementioned carbon emission conversion parameters;

[0033] The baseline carbon emissions are superimposed with the perturbed carbon emissions to obtain the real-time carbon emissions estimate.

[0034] In the above technical solution, preferably, the generative cognitive model includes an encoder, a state transition generator, and a decoder; the encoder includes a one-dimensional convolutional layer and a Transformer layer, used to encode the state sequence into a latent state vector; the state transition generator includes a fully connected structure and a self-attention structure, used to generate the latent state at the next time step; the decoder is used to decode and output the predicted values ​​of future carbon emissions, future energy consumption, and future temperature.

[0035] The generative cognitive model employs a training mechanism of pre-training and online incremental updates, and fills in the state sequence based on pseudo-samples when input data is missing.

[0036] The optimization decision includes determining carbon emission optimization control instructions for adjusting the operating parameters of the refrigeration equipment and the transportation operation strategy, provided that the predicted future temperature value meets the preset temperature control requirements.

[0037] The Transformer layer employs a physics-guided attention mechanism, using the state variables corresponding to the cargo compartment's thermal balance relationship as prior constraints for attention weights. The training objective function of the generative cognitive model includes data-driven loss, latent space regularization loss, and physics-constraint loss constructed based on the cargo compartment's thermal balance equation. The decoder's output is equipped with an energy conservation correction layer, used to perform energy consumption-carbon emission consistency correction and temperature-energy consumption consistency correction on future carbon emission predictions, future energy consumption predictions, and future temperature predictions.

[0038] In the above technical solution, preferably, the generative cognitive model inputs the state sequence of the most recent m time steps in real-time operation and recursively outputs the carbon emission prediction value of the next T time steps, where m and T are preset positive integers; the carbon emission prediction value is compared with the corresponding carbon emission measurement data or the carbon emission estimate value obtained by the carbon emission conversion parameter to obtain residual information, and online incremental update is performed based on the residual information;

[0039] The optimization decision-making adopts a hierarchical optimization framework: the high-level strategy optimization determines the route selection, operation mode, and stop plan at the start of the task; the low-level model predictive control, based on the short-term prediction of the generative cognitive model during transportation, aims to minimize carbon emissions in the prediction time domain and uses the temperature control safety range as a hard constraint to solve the rolling time domain optimization problem and output control commands for compressor power and vehicle speed; the low-level model predictive control constructs a Gaussian process surrogate model in the potential state vector space and performs iterative search based on the acquisition function to determine the carbon emission optimization control command under the condition of satisfying the temperature control safety range.

[0040] This invention also proposes a cold chain transportation carbon emission optimization system based on dynamic feature adaptive hierarchical modeling, which applies the cold chain transportation carbon emission optimization method based on dynamic feature adaptive hierarchical modeling disclosed in any of the above technical solutions, including a data acquisition and processing unit, a key feature screening unit, a set analysis and partitioning unit, a feature freezing and adjustment unit, a carbon emission hierarchical modeling unit, a generation and cognitive prediction unit, and an optimization operation decision-making unit, with each unit electrically connected through a communication interface.

[0041] The data acquisition and processing unit is used to collect multi-source operation data during the cold chain transportation process, and to perform time alignment, abnormal data processing and unified sampling frequency processing on the multi-source operation data to generate a standardized feature sequence. The multi-source operation data includes at least energy consumption-related data and carbon emission-related data. The carbon emission-related data includes at least one of carbon emission measurement data and carbon emission conversion parameters.

[0042] The key feature screening unit is used to perform sensitivity analysis or feature importance calculation on the feature variables in the standardized feature sequence based on historical energy consumption data and historical carbon emission data, so as to screen out a set of key features.

[0043] The set analysis partitioning unit is used to perform time stability analysis on the feature variables in the key feature set, so as to divide the key feature set into a basic parameter set and a dynamic parameter set.

[0044] The feature freezing and adjustment unit is used to generate confidence evaluation results for the feature variables in the key feature set; the feature freezing and adjustment unit is also used to perform drift detection and drift type diagnosis on the feature variables, and trigger the corresponding parameter set update strategy according to the drift type obtained by diagnosis, and adaptively adjust the basic parameter set and the dynamic parameter set. The drift type includes gradual drift, abrupt drift, periodic drift and concept drift.

[0045] The carbon emission hierarchical modeling unit is used to establish a baseline carbon emission estimation model based on the basic parameter set and output the baseline carbon emission, establish a perturbation carbon emission correction model based on the dynamic parameter set and output the perturbation carbon emission, and superimpose the baseline carbon emission and the perturbation carbon emission to obtain the real-time carbon emission estimate.

[0046] The cognitive prediction unit is used to input a state sequence containing the basic parameter values ​​corresponding to the basic parameter set, the dynamic parameter values ​​corresponding to the dynamic parameter set, real-time energy consumption related data, and the real-time carbon emission estimate into the cognitive model, output the future carbon emission prediction value, and complete the state sequence when the input data is missing.

[0047] The optimized operation decision unit is used to make optimized decisions on the operating parameters of the refrigeration equipment and the transportation operation strategy based on the predicted future carbon emissions, and output carbon emission optimization control commands.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] (1) By performing time alignment, abnormal data processing and unified sampling frequency processing on multi-source operation data of cold chain transportation, a standardized feature sequence is formed, which enables multi-source data from vehicle operating conditions, environment and refrigeration equipment to be compared and fused under the same time reference, reducing the interference of sampling inconsistencies and outliers on subsequent carbon emission estimation, and improving the input quality and modeling stability of the data side.

[0050] (2) By performing sensitivity analysis or feature importance calculation based on historical energy consumption data and historical carbon emission data, key features are screened and then divided into basic parameter set and dynamic parameter set after time stability analysis of key features. This achieves hierarchical expression of "inherent stable factors of task" and "dynamic disturbance factors of operation", avoiding the accumulation of bias caused by mixing stable factors and disturbance factors in modeling, and improving the adaptability of real-time carbon emission estimation to changes in working conditions and differences in tasks.

[0051] (3) By generating confidence evaluation results for key features, freezing processing is performed on the corresponding features when the confidence is lower than the threshold. The drift detection and type diagnosis mechanism identifies progressive drift, abrupt drift, periodic drift and concept drift. The corresponding parameter set update strategy is triggered according to the drift type, so that the adaptive adjustment is upgraded from passive threshold rule judgment to active intelligent diagnosis and accurate response. This enables the system to have the ability to degrade and correct in scenarios with missing data, noise or signal quality degradation, and improves the robustness and adaptability of the estimation results.

[0052] (4) By establishing a baseline carbon emission estimation model based on the set of basic parameters and establishing a perturbation carbon emission correction model based on the set of dynamic parameters, the two are superimposed to obtain the real-time carbon emission estimate, thereby realizing the interpretable decomposition of carbon emission composition, making the source of carbon emission change traceable, and maintaining the continuity and stability of real-time carbon emission output under rapid operating condition fluctuation scenarios.

[0053] (5) By generating a cognitive model from the input state sequence, the model outputs a predicted value of future carbon emissions. When the input data is missing, the state sequence is completed, thus realizing the integrated processing of "missing data completion + future prediction". This provides a forward-looking evaluation basis for the operating parameters of refrigeration equipment and transportation operation strategies, thereby supporting the formation of executable carbon emission optimization control instructions under the premise of meeting temperature control requirements, and improving the timeliness and effectiveness of emission reduction decisions in cold chain transportation.

[0054] (6) By integrating physical information constraints into the generative cognitive model, including a physical guidance attention mechanism, a physical constraint loss function that integrates the thermal balance equation, and an energy conservation correction layer, the model not only learns data patterns but also follows the thermodynamic physical laws of the cold chain system. It can make predictions that conform to physical intuition in the case of sparse data, distribution extrapolation, or extreme working conditions, which significantly improves the robustness and credibility of future carbon emission predictions.

[0055] (7) By adopting a hierarchical optimization framework, the high-level strategy optimization determines the task-level transportation scheme, and the low-level model predictive control achieves minute-level fine adjustment. Combined with the efficient search method of potential space, the control command for carbon emission is found quickly under the premise of ensuring the hard constraint of temperature control, so that carbon emission optimization has the characteristics of globality, real-time and safety. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling, as disclosed in one embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram illustrating the implementation process of a cold chain transportation carbon emission optimization method based on dynamic feature adaptive hierarchical modeling, as disclosed in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The present invention will now be described in further detail with reference to the accompanying drawings:

[0060] like Figure 1 and Figure 2 As shown, this invention provides a method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling, which is applicable to the entire process operation data of cold chain transportation tasks.

[0061] First, multi-source operational data of the transportation process is collected. This multi-source operational data includes at least energy consumption-related data and carbon emission-related data, with the carbon emission-related data including carbon emission measurement data and / or carbon emission conversion parameters. Time alignment is performed on the collected data to ensure a consistent time-series correspondence between data from different sources under the same time benchmark. Abnormal data is processed to remove or correct data points that significantly deviate from normal operating conditions. Finally, a unified sampling frequency is applied to all data sources to obtain standardized feature sequences, providing a stable and comparable input foundation for subsequent modeling.

[0062] After completing the construction of the standardized feature sequence, based on historical energy consumption data and historical carbon emission data, sensitivity analysis or feature importance calculation is performed on the feature variables in the standardized feature sequence to screen out the key feature set that contributes more to carbon emission changes, so as to reduce the impact of redundant input on model complexity and real-time computing cost.

[0063] Subsequently, time stability analysis was performed on the feature variables in the key feature set, and based on the stability analysis results, the key feature set was divided into a basic parameter set and a dynamic parameter set, so that the long-term relatively stable factors and the short-term disturbance factors in the operation process were separated in the modeling structure.

[0064] Furthermore, confidence evaluation results are generated for the feature variables in the key feature set. When the confidence evaluation result is lower than the preset threshold, the corresponding feature variable is frozen. Drift detection and drift type diagnosis are performed on the feature variable. Drift types include gradual drift, abrupt drift, periodic drift, and conceptual drift. Based on the drift type diagnosed, the corresponding parameter set update strategy is triggered to adaptively adjust the basic parameter set and dynamic parameter set to suppress the disturbance of estimation and decision caused by unreliable input due to missing data, increased noise, or sensor anomalies.

[0065] In one implementation, a baseline carbon emission estimation model is established based on a set of basic parameters and the baseline carbon emission is output. A perturbation carbon emission correction model is established based on a set of dynamic parameters and the perturbation carbon emission is output. The baseline carbon emission and the perturbation carbon emission are superimposed to obtain a real-time carbon emission estimate, thereby correcting short-term fluctuations while maintaining long-term baseline stability.

[0066] Subsequently, a state sequence containing basic parameter values ​​corresponding to the basic parameter set, dynamic parameter values ​​corresponding to the dynamic parameter set, real-time energy consumption data, and real-time carbon emission estimates is input to generate a cognitive model, which outputs future carbon emission predictions and completes the state sequence when input data is missing to ensure the continuity of the prediction chain.

[0067] Finally, based on future carbon emission forecasts, the operating parameters of the refrigeration equipment and the transportation operation strategy are optimized, and carbon emission optimization control commands are output.

[0068] In this implementation, a closed-loop link is established through key feature screening, stability stratification, confidence freezing and adaptive adjustment, baseline / perturbation stratification estimation and future trend prediction, which realizes stable output of real-time carbon emission estimation and forward-looking support for future trends, thereby improving the accuracy of carbon emission estimation and the feasibility of operational decisions under complex operating conditions.

[0069] In the above embodiments, preferably, the multi-source operational data includes vehicle operating parameters, environmental parameters, refrigeration equipment status parameters, and operational behavior parameters. These four types of parameters are initially cleaned and time-aligned to remove outliers and unify the sampling frequency, forming a standardized input feature matrix. X (t), along with energy consumption-related data and carbon emission-related data, are incorporated into the standardized feature sequence construction process.

[0070] Among them, vehicle operating parameters include vehicle speed, acceleration, gradient, road type, and load factor, which are used to characterize the changes in traction power consumption of the vehicle under different road conditions and load conditions; environmental parameters include ambient temperature, humidity, solar radiation intensity, air pressure, and wind speed, which are used to characterize the impact of the external environment on the cooling load and heat transfer conditions; refrigeration equipment status parameters include compressor power, operating frequency, current, voltage, and number of start-stop cycles, which are used to reflect the operating intensity and energy consumption output of the refrigeration equipment; and operational behavior parameters include the number of door openings, door opening duration, idling time, and cold start frequency, which are used to reflect the contribution of loading and unloading operations and driving behavior to temperature control load and energy consumption disturbances.

[0071] In this implementation, by structurally collecting and uniformly organizing the main driving factors of cold chain transportation, the model's comprehensive response capability to vehicles, environment, equipment and operational behavior is improved, the estimation bias caused by insufficient information in a single dimension is reduced, and more complete data support is provided for subsequent key feature screening and hierarchical modeling.

[0072] In the above embodiments, preferably, the sensitivity analysis calculates a sensitivity index for each feature variable in the standardized feature sequence. S i The initial input consists of N multi-source feature variables from a standardized feature sequence. .

[0073] Therefore, a carbon emission response relationship was constructed based on historical energy consumption data and historical carbon emission data, and the carbon emission amount was obtained under this relationship. C t For each input feature variable x i gradient This is used to characterize the instantaneous impact of changes in characteristic variables on changes in carbon emissions; simultaneously, the characteristic variables are calculated separately. x i standard deviation std( x i ) and carbon emissions C t standard deviation std( C t (), used to reflect the scale of characteristic fluctuations and the scale of carbon emission fluctuations.

[0074] Sensitivity index S i satisfy , representing the relative proportion of carbon emission response when the characteristic changes by one standard deviation, provides a comparable sensitivity measure for different characteristic variables, and is used to rank and filter the characteristic variables accordingly. The gradient between the characteristic variable and carbon emissions can be approximated by linear regression, the SHAP value, or mutual information. The fluctuation range of the characteristic variable. This represents the fluctuation range of carbon emissions.

[0075] In one implementation, high-sensitivity features can be selected based on a threshold δ to form a key feature set. .

[0076] In this implementation, by jointly measuring the "influence intensity (gradient)" and the "fluctuation scale (standard deviation)," the contribution of multi-source features to carbon emissions can be quantitatively identified, reducing the computational burden caused by the introduction of redundant features and improving the interpretability and stability of the key feature set screening.

[0077] In the above embodiments, preferably, the time stability analysis uses a sliding window method to statistically analyze the time series data of each feature variable, and calculates the standard deviation and frequency of change within the sliding window to quantify the fluctuation intensity and rate of change of the feature during task operation.

[0078] When the standard deviation of the feature variable within the sliding window is less than the stability threshold σ b This indicates that the feature fluctuates relatively little within the window scale; when the frequency of change of the feature variable... f i Less than the low frequency threshold f b This indicates that the feature changes slowly over time; when the feature variable is a parameter that can be configured or measured before the start of the transportation task (e.g., vehicle profile, average load factor), it indicates that it has natural stability during task execution.

[0079] When any of the above conditions are met, the characteristic variable is included in the set of basic parameters to represent long-term steady-state characteristics, which can be expressed mathematically as follows: Feature variables that do not meet the aforementioned conditions or rely on real-time sensing data acquisition are included in the dynamic parameter set to accommodate short-term disturbances and real-time changes. Mathematically, this can be expressed as... .

[0080] In this implementation, by using statistical stability and frequency of change as boundaries to achieve basic / dynamic stratification, parameter coupling caused by mixing steady-state factors and disturbance factors in the modeling is avoided, thereby improving the stability of baseline estimation and the responsiveness of disturbance correction.

[0081] In the above embodiments, preferably, the process of generating the confidence evaluation result includes:

[0082] For each feature variable in the key feature set, the corresponding signal quality index, missing proportion index, and noise level index are obtained. The signal quality index reflects the effectiveness and reliability of the sensor signal acquisition; the missing proportion index reflects the impact of sampling interruptions, upload delays, or data loss on temporal continuity; and the noise level index reflects signal instability caused by measurement jitter, environmental interference, or sensor drift.

[0083] According to preset fusion rules, three types of indicators—signal quality, missing ratio, and noise level—are fused and calculated to obtain the confidence evaluation results of the feature variables, ensuring that each feature variable has a definite degree of credibility at the current moment. Mathematically, this can be expressed as: , This indicates the confidence weight of the above three types of indicators.

[0084] When the confidence evaluation result is lower than the preset threshold, the feature variable is frozen. The freezing process includes stopping the use of the feature variable in model updates or using the frozen value in the calculation, thereby avoiding the amplification of errors caused by low-confidence inputs in the model update chain.

[0085] In this implementation, a confidence evaluation and freezing strategy that integrates signal quality, missing data, and noise is used to suppress the disturbance of unreliable features to estimation and prediction in scenarios with sensor anomalies, missing data, or increased noise, thereby improving the robustness and continuity of the overall carbon emission estimation chain.

[0086] In the above embodiments, preferably, the adaptive adjustment of the basic parameter set and the dynamic parameter set includes three stages: drift detection, type diagnosis, and differentiated response.

[0087] During the drift detection phase, the system performs multi-dimensional drift detection on each feature variable in the key feature set to identify whether the feature variable deviates from its historical distribution characteristics. Drift detection employs a sliding window mechanism, with the detection window length set to [value missing]. Each sampling period is used to calculate the statistical properties of the feature variables and compare them with historical benchmarks within each detection period.

[0088] During the type diagnosis phase, the system further diagnoses the drift type of the detected drift feature variables. Drift types include gradual drift, abrupt drift, periodic drift, and conceptual drift. The specific diagnostic methods are as follows:

[0089] (1) Progressive drift detection

[0090] Progressive drift refers to a slow, continuous, unidirectional trend in the mean or variance of a characteristic variable. Typical causes include vehicle aging, equipment wear and tear, and reduced cooling efficiency.

[0091] Asymptotic drift is detected by performing a linear trend test on the sliding window mean sequence. Let the characteristic variable be... In continuous The mean sequence within each sliding window is Linear regression was performed on the sequence to obtain the trend slope. The progressive drift index is defined as follows:

[0092] ;

[0093] when When determining the characteristic variables A gradual drift occurs, in which This is the progressive drift threshold.

[0094] (2) Detection of abrupt shifts

[0095] Abrupt drift refers to a significant step change in a characteristic variable within a short period of time. Typical causes include equipment replacement, sensor failure repair, and sudden adjustment of operational strategies.

[0096] Abrupt drift is detected using the Cumulative Sum (CUSUM) algorithm. Let the feature variables be... The historical average benchmark is The allowable offset is Then the positive cumulative sum With negative cumulative sum The calculation formula is:

[0097] ;

[0098] The abrupt shift index is defined as:

[0099] ;

[0100] when When determining the characteristic variables A sudden shift occurred, in which This is the mutation drift threshold. After a mutation is detected, the accumulated values ​​will be... and Reset to zero and update the historical mean benchmark. .

[0101] (3) Periodic drift detection

[0102] Periodic drift refers to the regular periodic fluctuations of characteristic variables. Typical causes include seasonal changes, diurnal temperature variations, and operational differences between weekdays and non-weekdays.

[0103] Periodic drift is detected by performing a Fast Fourier Transform (FFT) on the time series of the feature variables. Let the feature variables be... In the time window The timing within is Perform an FFT transform on it to obtain the spectrum. The periodic drift index is then defined as:

[0104] ;

[0105] in Let be the set of periodic frequencies to be detected, corresponding to daily, weekly, and seasonal cycles, respectively. When When determining the characteristic variables There is periodic drift, in which As a periodic drift threshold, significant periods are recorded simultaneously. ,in .

[0106] (4) Concept drift detection

[0107] Concept drift refers to a change in the relationship between a characteristic variable and carbon emissions, even though the distribution characteristics of the characteristic variable itself may not have changed significantly. Typical reasons include changes in operating models, adjustments to cargo types, and the implementation of new energy policies.

[0108] Concept drift is detected by monitoring changes in the correlation coefficient between characteristic variables and carbon emissions. This is done within a time window. Internally computed characteristic variables With carbon emissions The Pearson correlation coefficient between them is The concept drift index is then defined as:

[0109] ;

[0110] in This represents the time interval for comparing correlation coefficients. When determining the characteristic variables Conceptual drift occurs, in which This is the concept drift threshold.

[0111] In one implementation, the four types of drift detection can be executed in parallel. The same feature variable may trigger multiple drift types at the same time. The system determines the dominant drift type and triggers the corresponding response strategy according to the priority order of "concept drift > abrupt drift > periodic drift > gradual drift".

[0112] During the differentiated response phase, the system triggers the corresponding response strategy based on the drift type obtained from the diagnosis, as follows:

[0113] (1) Gradual drift response strategy: incremental weight update

[0114] To address gradual drift, the system employs an incremental weight update strategy to avoid triggering unnecessary parameter re-division due to slight drift.

[0115] Specifically, for the feature variables included in the basic parameter set When asymptotic drift occurs, the system uses an online gradient descent method to adjust the weighting coefficients of the baseline carbon emission estimation model. Make minor adjustments and update the formula as follows:

[0116] ;

[0117] in For learning rate, The residual loss for carbon emissions is estimated and defined as:

[0118] ;

[0119] in This is a real-time carbon emission estimate. This refers to carbon emission measurement data or carbon emission estimates obtained from carbon emission conversion parameters. To update the number of samples in the window.

[0120] This strategy maintains the overall stability of the model structure while enabling smooth adaptation to gradual changes.

[0121] (2) Abrupt drift response strategy: parameter set repartition

[0122] In response to abrupt drift, the system implements a parameter set repartitioning strategy to respond promptly to structural changes in the system.

[0123] Specifically, when the feature variables in the set of basic parameters When a sudden drift occurs, perform the following actions:

[0124] Step 1: Remove this feature variable from the set of basic parameters. Transfer to dynamic parameter set :

[0125] ;

[0126] Step 2: Re-estimate the weighting coefficients of the baseline carbon emission estimation model and the dynamic sensitivity coefficient of the perturbation carbon emission correction model ;

[0127] Step 3: Update the historical mean benchmark after the mutation This is used for subsequent drift detection.

[0128] Conversely, when the characteristic variables in the dynamic parameter set... Abrupt drift occurs and then tends to stabilize (continuous) Within each testing cycle and When this happens, the feature variable can be transferred to the set of basic parameters.

[0129] (3) Periodic drift response strategy: time period encoding

[0130] To address periodic drift, the system employs a time-period encoding strategy, enabling the model to explicitly perceive and utilize periodic change patterns.

[0131] Specifically, for the characteristic variables that undergo periodic drift. Introducing time period encoding vectors Defined as:

[0132] ;

[0133] in These are the detected significant periods. If multiple significant periods are detected simultaneously... The time-period encoding vector is then expanded as follows:

[0134] ;

[0135] Time period encoding vector Compared with the original feature variables The concatenated data is used as input to the model.

[0136] ;

[0137] In baseline carbon emission estimation models or perturbation carbon emission correction models, independent weight coefficients are assigned to the time period encoding components, enabling the model to learn the impact of periodic changes on carbon emissions.

[0138] (4) Concept drift response strategy: model reconstruction

[0139] To address concept drift, the system implements a model reconstruction strategy to re-establish the mapping relationship between features and carbon emissions.

[0140] Specifically, when concept drift is detected, the system triggers the following refactoring process:

[0141] Step 1: Based on the latest historical energy consumption data and historical carbon emission data (take the most recent one) (Data from each sampling period), re-perform sensitivity analysis or feature importance calculation:

[0142] ;

[0143] Step Two: Based on the new sensitivity index Re-filter the set of key features ;

[0144] Step 3: Perform time stability analysis on the new set of key features and repartition the set of basic parameters. With dynamic parameter set ;

[0145] Step 4: Based on the re-partitioned parameter set, retrain all parameters of the baseline carbon emission estimation model and the perturbation carbon emission correction model.

[0146] The concept drift response strategy has the highest computational cost, so it is only triggered when concept drift is confirmed.

[0147] [Closed-loop feedback and strategy upgrade]

[0148] After the response strategy is executed, the system continuously monitors the changes in the carbon emission estimation residuals to verify the effectiveness of the response strategy.

[0149] Let the root mean square of the carbon emission estimation residuals before the execution of the response strategy be... Post-execution observation period The root mean square of the carbon emission estimation residuals is The residual convergence criterion is:

[0150] ;

[0151] in The convergence threshold is typically set to 0.8, which means that a decrease in residuals of more than 20% is considered convergence.

[0152] When the carbon emission estimation residuals do not converge (i.e. When this occurs, the system will switch to a higher-level response strategy according to the following policy escalation rules:

[0153]

[0154] Through a closed-loop mechanism of detection → diagnosis → response → verification → upgrade, we ensure that adaptive adjustment can effectively cope with various drift scenarios.

[0155] In this implementation, a drift detection and type diagnosis mechanism is used to automatically identify progressive drift, abrupt drift, periodic drift, and conceptual drift. Based on the drift type, four differentiated response strategies are triggered: incremental weight update, parameter set repartition, time periodic encoding, and model reconstruction. The effectiveness of the response is ensured through a closed-loop feedback and strategy upgrade mechanism.

[0156] The energy consumption and carbon emissions of cold chain transport vehicles exhibit a significant dual characteristic across different times, locations, and environments: one type of factor remains stable over the long term, determining the system's "structural energy consumption level"; the other type of factor fluctuates rapidly with weather, traffic, and operating conditions, determining the "short-term energy consumption shift." Traditional models fit both types of factors uniformly, resulting in mixed parameters, difficulty in interpretation, and poor model generalization.

[0157] Therefore, the core improvement proposed in this invention lies in further dividing the "key parameter set" output by the parameter selection module into basic parameters and dynamic parameters, which are used to construct the "steady-state energy consumption model" and the "dynamic disturbance model," respectively. Through this hierarchical modeling, the system can respond to external fluctuations in real time under the premise of long-term structural stability, achieving high accuracy and robustness in carbon emission estimation.

[0158] Specifically, fundamental parameters refer to characteristics that remain stable over a relatively long period of time and can be configured or measured before transportation. They primarily determine the long-term energy consumption level of the system. Typical timescales for fundamental parameters range from several days to several months, and they usually remain unchanged during a cold chain operation. Examples of fundamental parameters are shown in Table 1 below.

[0159] Table 1. Types and meanings of basic parameters

[0160]

[0161] Dynamic parameters refer to characteristics that change rapidly over time and are affected by the environment and operation, used to reflect short-term energy consumption and carbon emission deviations. The time scale of dynamic parameters is from seconds to minutes, and they are updated in real time according to the operating status, reflecting the instantaneous disturbance characteristics of the system. Examples of dynamic parameters are shown in Table 2 below.

[0162] Table 2. Types and meanings of dynamic parameters

[0163]

[0164] Using basic and dynamic parameters as inputs, a two-layer model structure for energy consumption and carbon emissions is constructed:

[0165] ;

[0166] in: For steady-state energy consumption-carbon emission baseline model; This is a modified model based on dynamic perturbations.

[0167] In the above embodiments, preferably, the baseline carbon emission estimation model adopts a two-stage structure of steady-state energy consumption model and energy consumption-carbon emission conversion. The steady-state energy consumption model takes the basic parameter values ​​in the basic parameter set as input to calculate the baseline energy consumption estimate. E base ,in ,in The first in the basic parameter set i One basic parameter, β i For the corresponding weighting coefficients, β 0 represents the baseline constant term. n b The number of basic parameters.

[0168] Based on baseline energy consumption estimates E base and carbon emission conversion parameters Baseline carbon emissions were obtained through conversion. This ensures that the energy consumption estimation results and the carbon emission output are mapped to the same dimensions.

[0169] The perturbation carbon emission correction model uses a set of dynamic parameters as input to construct an estimate of perturbation energy consumption. E yn ( t ),in , where γ k ( t ) represents the dynamic sensitivity coefficient, Δ x k ( t ) represents the deviation of the dynamic parameters in the dynamic parameter set from the baseline. n d This represents the number of dynamic parameters.

[0170] Based on disturbance energy consumption estimates E yn ( t ) and carbon emission conversion parameters The perturbation carbon emissions are calculated. .

[0171] Overlay baseline carbon emissions with perturbed carbon emissions This allows for real-time carbon emission estimates, enabling real-time estimation to possess both baseline stability and perturbation responsiveness.

[0172] In this implementation, by separating the steady-state layer and the perturbation layer for modeling, the parameter coupling caused by the mixed fitting of factors at different time scales is reduced, the interpretability and generalization performance of the model are improved, and the continuity and stability of real-time carbon emission estimation are maintained when the operating conditions change rapidly.

[0173] In the above embodiments, preferably, the cognitive model includes an encoder, a state transition generator, and a decoder, used to output future carbon emission predictions, future energy consumption predictions, and future temperature predictions under conditions of missing inputs, strong perturbations, and uneven sample distribution.

[0174] During implementation, the following characteristic data are extracted and integrated: basic parameters, dynamic parameters, energy consumption and carbon emission data, and error feedback, to form the state input. .

[0175] The encoder comprises one-dimensional convolutional layers and Transformer layers for learning representations of the state sequence. First, local temporal features are extracted using the one-dimensional convolutional layers, and then long-term dependencies are captured using the Transformer layers, encoding the state sequence into a latent state vector. This allows multi-source temporal information from the transportation process to form a unified representation in the latent space. Specifically, the input temporal data... Encoded as a latent state vector It outputs the system's current "potential cognitive state".

[0176] In the above embodiments, preferably, the Transformer layer in the encoder adopts a physically guided attention mechanism, using the thermodynamic correlation of the cold chain system as a prior constraint for attention weights.

[0177] The temperature change of a cold chain transport cargo compartment follows a heat balance equation, and its physical mechanism can be expressed as: the rate of temperature change in the compartment equals the difference between heat input and heat output divided by the compartment's heat capacity. Based on this physical mechanism, there are clear causal relationships between different state variables, such as: external temperature affects heat transfer, opening doors leads to heat intrusion, and compressor power determines the cooling capacity.

[0178] Physics-guided attention mechanisms construct physical correlation matrices. The aforementioned physical relationships are encoded as a prior distribution of attention weights. Physical Relationship Matrix The element is defined as:

[0179]

[0180] in This is the indirect correlation attenuation coefficient, typically with a value of 0.5; For state variable dimensions.

[0181] In cold chain transportation scenarios, the typical structure of the physical correlation matrix includes:

[0182]

[0183] The attention weight calculation for the Transformer layer has been modified as follows:

[0184] ;

[0185] in These are query, key, and value matrices, respectively. The dimension of the key vector. The physical guiding strength coefficient. A small constant to prevent overflow in logarithmic operations.

[0186] Through a physics-guided attention mechanism, the model learns data-driven feature associations while being constrained to follow a physically reasonable variable dependency structure, enabling it to make predictions that conform to physical intuition in scenarios with sparse data or distributed extrapolation.

[0187] The state transition generator includes fully connected and self-attention structures to learn the evolutionary mapping of latent states at neighboring time steps. Generate the potential state for the next time step and establish a transition mapping of the system state from t→t+1.

[0188] The decoder is used to map the latent state back to the observation space. It decodes and outputs future carbon emission predictions, future energy consumption predictions, and future temperature predictions, enabling carbon emission predictions to have a consistent time-series coordination relationship with energy consumption and temperature control status.

[0189] In the above embodiments, preferably, an energy conservation correction layer is provided at the output end of the decoder to perform physical consistency correction on the future carbon emission prediction value, future energy consumption prediction value and future temperature prediction value output by the decoder.

[0190] The energy conservation correction layer is based on the following physical constraints:

[0191] Constraint 1: Energy consumption-carbon emission consistency constraint

[0192] There is a definite conversion relationship between carbon emissions and energy consumption:

[0193] ;

[0194] in The carbon emission conversion factor (kgCO2 / kWh) depends on the energy type (diesel, electricity, etc.) and its carbon emission factor.

[0195] When the decoder outputs the carbon emission prediction value Energy consumption forecast If the above conversion relationship is not met, perform a correction:

[0196] ;

[0197] Constraint 2: Temperature Variation-Energy Consistency Constraint

[0198] There is a physical relationship between changes in cargo compartment temperature and refrigeration energy consumption. When the cargo compartment temperature drops, the refrigeration system will inevitably consume energy; when the cargo compartment temperature rises (above the set temperature), refrigeration energy consumption should increase to suppress the temperature rise.

[0199] Define the temperature-energy consistency index:

[0200] ;

[0201] in For the predicted cooling energy consumption, This represents the historical average of cooling energy consumption.

[0202] when and When the temperature rises and exceeds the set temperature, it indicates that the predicted cooling energy consumption is too low and needs to be corrected upwards.

[0203] ;

[0204] in For correction factors, Set the temperature for the temperature control.

[0205] Constraint 3: Conservation of Cumulative Energy

[0206] During a time window Internally, cumulative energy consumption should be consistent with cumulative carbon emissions:

[0207] ;

[0208] If the conditions are not met, the predicted values ​​within that time window will be scaled proportionally for correction.

[0209] Through the energy conservation correction layer, the multivariate predicted values ​​(carbon emissions, energy consumption, temperature) output by the decoder are forced to satisfy the physical conservation law while satisfying the data-driven fitting, thereby improving the physical reliability of the prediction results under extreme working conditions or data-sparse scenarios.

[0210] In one implementation, the cognitive model employs a training mechanism of pre-training and online incremental updates. During the pre-training phase, historical transportation task data is used to form a basic mapping. In the online phase, new data during operation is used to incrementally update the model parameters to adapt to changes in environment and operating conditions. When input data is missing, pseudo-sample imputation is used to complete the state sequence, ensuring its continuity at key time steps and thus supporting the stability of the predicted sequence output by the decoder.

[0211] Based on the above prediction output, the optimization decision determines the carbon emission optimization control command for adjusting the operating parameters of the refrigeration equipment and the transportation operation strategy, under the condition that the future temperature prediction value meets the preset temperature control requirements, so that the low-carbon goal and temperature control constraint form a consistent decision closed loop.

[0212] Specifically, in the pre-training phase, long-term historical data (across vehicles and seasons) is used for model pre-training:

[0213] Objective: To learn the universal spatiotemporal patterns of cold chain systems (e.g., temperature changes, energy consumption response patterns).

[0214] Advantages: It can be directly transferred to new scenarios with sparse data, without the need for training from scratch;

[0215] In the above embodiments, preferably, the training objective function for generating the cognitive model integrates data-driven loss and physical constraint loss to form a training paradigm for a Physics-Informed Neural Network.

[0216] The training objective function is defined as: ;

[0217] in:

[0218] (1) Data-driven loss

[0219] Consistency between data-driven loss-constrained model predictions and measured values:

[0220] ;

[0221] in These are the carbon emissions, energy consumption, and cargo compartment temperature predicted by the model, respectively. These are the corresponding measured values.

[0222] (2) Latent space regularization loss

[0223] The latent space regularization loss ensures a smooth distribution of the latent state vector:

[0224] ;

[0225] in Let the latent state posterior distribution be the output of the encoder. This is the prior distribution (usually set as the standard normal distribution). The divergence is Kullback-Leibler.

[0226] (3) Physical constraint loss

[0227] The physical constraint loss is constructed based on the thermal balance equation of the cold chain cargo compartment, and the constraint model predicts that it satisfies the law of conservation of energy.

[0228] The physical expression of the heat balance equation for cold chain cargo compartments is:

[0229] ;

[0230] in:

[0231] The weight of the cargo inside the cargo compartment (kg);

[0232] Specific heat capacity of the cargo (J / (kg·K));

[0233] The temperature of the cargo compartment (K);

[0234] The heat input power (W) includes heat transfer and heat intrusion from door opening;

[0235] It is the heat output power (W), mainly for the heat dissipation from the cargo compartment to the outside (when the temperature of the cargo compartment is higher than that of the outside).

[0236] The refrigeration power (W) of the refrigeration equipment.

[0237] The physical calculation formulas for each heat component are as follows:

[0238] Heat transfer (heat transfer from the cargo compartment walls):

[0239] ;

[0240] in The overall heat transfer coefficient of the cargo compartment (W / (m²)) 2 ·K)), The area of ​​the cargo box wall (m²) 2 ), The ambient temperature (K).

[0241] Heat intrusion when the door is opened:

[0242] ;

[0243] in The air exchange mass flow rate (kg / s) when the door is open. is the specific heat capacity of air (J / (kg·K)). This is a function to indicate the door's open state (1 when the door is open, 0 when the door is closed).

[0244] Cooling power:

[0245] ;

[0246] in The power of the compressor (W). This is the coefficient of performance for cooling.

[0247] Discretizing the heat balance equation yields the physical constraints on temperature changes:

[0248] ;

[0249] Physical constraint loss is defined as the residual between the model-predicted temperature and the temperature calculated by the physical equations:

[0250] ;

[0251] in To be based on the heat balance equation The physical predicted temperature obtained by recursive calculation of the state at each time step:

[0252]

[0253] In the above formula, Physical parameters can be obtained from equipment files or calibrated using historical data.

[0254] (4) Adaptive adjustment of loss weights

[0255] The weighting coefficients of the three losses An adaptive adjustment strategy can be adopted. In the early stages of training, data-driven loss should be the primary approach. A larger weight allows the model to fit the data quickly; as training progresses, the weights of the physical constraint loss are gradually increased. This allows the model to converge toward physical consistency. The adaptive weight adjustment formula is:

[0256] ;

[0257] in The target value for the physical constraint weights, To adjust the time constant, This is the current training round.

[0258] During the online fine-tuning phase, fine-tuning is performed on-site using limited real-time data:

[0259] When the input data is sparse, "pseudo-sample imputation" is used:

[0260] The fine-tuning objective function is as follows:

[0261] Update the parameters after each time step.

[0262] In the above embodiments, preferably, the physical parameters involved in the physical constraint loss (such as the overall heat transfer coefficient) Coefficient of performance (COD) (etc.) can be calibrated online during operation to adapt to parameter drift caused by equipment aging or environmental changes.

[0263] The physical parameters were calibrated online using the least squares estimation method. The comprehensive heat transfer coefficient was used as the basis for this determination. Taking calibration as an example:

[0264] According to the heat transfer equation of the cargo box wall:

[0265] ;

[0266] Under steady-state conditions (refrigeration equipment off, cargo compartment doors closed, temperature change slow), the heat transfer through the cargo compartment walls is approximately equal to the heat transfer corresponding to the temperature change in the cargo compartment.

[0267] ;

[0268] collection Group steady-state measurement data The heat transfer coefficient is estimated using the least squares method:

[0269] ;

[0270] Similarly, the coefficient of performance (COP) for cooling Calibration can be performed during the stable operation phase of the refrigeration equipment by analyzing the relationship between refrigeration power and temperature changes.

[0271] ;

[0272] The online calibration cycle for physical parameters can be set to be performed once after each transportation task is completed, or when physical constraints are lost. Triggered when the preset threshold is exceeded.

[0273] By calibrating physical parameters online, physical constraint loss can continuously reflect the true physical characteristics of the system, avoiding physical constraint failure caused by parameter inaccuracies.

[0274] When data is sparse or sensing is interrupted, the model can generate a sequence of missing states based on the learned latent distribution. This allows us to maintain the continuity of system cognition even when observations are incomplete.

[0275] In this implementation, the structure of latent space representation, state evolution generation, multivariate decoding, missing data filling, pre-training, and online incremental updates enables carbon emission prediction to remain continuous in scenarios with incomplete data, and explicitly incorporates temperature control constraints into the decision boundary, thereby improving the executability and operational safety of control commands.

[0276] In the above implementation, preferably, the cognitive model takes into account the state sequence of the most recent m time steps during real-time operation and recursively outputs the carbon emission prediction values ​​for the next T time steps, forming a trend prediction output with a future-oriented window, thus shifting decision-making from passive response to forward-looking adjustment. Here, m and T are preset positive integers.

[0277] Specifically, the following closed-loop steps are performed:

[0278] 1. Input the most recent m-step historical state;

[0279] 2. Encode to obtain the latent state ht;

[0280] 3. Predict the potential state at the next time step using a generator: ;

[0281] 4. Decoding yields future observation results and carbon emission predictions: ;

[0282] 5. If there are missing inputs, [the system will handle the situation]. Fill;

[0283] 6. Compare the predicted values ​​with the measured values, calculate the residuals, and update the parameters.

[0284] The multi-step recursive prediction formula is as follows:

[0285] Outputting future T-step carbon emission trends .

[0286] To suppress the cumulative error caused by changes in the model as the scenario changes, the carbon emission prediction value is compared with the corresponding carbon emission measurement data or the carbon emission estimate value obtained from the carbon emission conversion parameter to obtain residual information. Based on the residual information, online incremental updates are performed so that the model can adapt to new operating conditions in a residual closed-loop manner, thereby maintaining the consistency between prediction and actual operation.

[0287] In the above implementation, preferably, the optimization decision adopts a hierarchical optimization framework, which includes two levels: high-level strategy optimization and low-level model predictive control.

[0288] (1) High-level strategy optimization

[0289] High-level strategy optimization is performed at mission initiation to determine the overall strategy for the transportation mission.

[0290] Decision variables for high-level strategy optimization include: route selection Operating Mode And docking plans.

[0291] The objective function for high-level strategy optimization is:

[0292] ;

[0293] in:

[0294] Estimated total carbon emissions (kgCO2) for the entire process;

[0295] Estimated arrival time Delivery deadline;

[0296] Accumulated temperature control violations are defined as the integral of the duration during which the cargo compartment temperature exceeds the safe range.

[0297] : Weighting coefficients, which correspond to the priorities of carbon emissions, aging, and temperature control, respectively.

[0298] High-level strategy optimization constructs carbon emission prediction models for each candidate route based on historical transportation data, and selects routes that... The smallest strategy combination.

[0299] (2) Predictive control of underlying model

[0300] The underlying model predictive control is executed in real time during transportation, and the equipment parameters are optimized in a refined manner based on short-term predictions from the generative cognitive model.

[0301] In each control cycle The underlying MPC solves the following optimization problems:

[0302] ;

[0303] Constraints:

[0304] ;

[0305] in:

[0306] The prediction time domain length, i.e. the number of control cycles predicted forward, typically 10~20;

[0307] Control vector, including compressor power setpoint (kW) and vehicle speed setting value (km / h);

[0308] : Control increments, used to penalize frequent adjustments;

[0309] The first prediction made by the generative cognitive model Constant carbon emissions;

[0310] The first prediction made by the generative cognitive model Constant temperature of the cargo compartment;

[0311] Temperature-controlled safe zone, such as for frozen goods. ;

[0312] Weighting coefficients are determined by the high-level operating mode (economic mode). (relatively large)

[0313] MPC employs a rolling time-domain strategy: the optimal control sequence is obtained through solution. After that, only the first control variable is executed. The solution will be recalculated in the next cycle.

[0314] (3) Efficient search of potential space

[0315] The optimization of the underlying MPC is performed in the latent state vector space of the generated cognitive model to improve search efficiency.

[0316] Specifically, control input Mapped to latent vectors by encoder Potential spatial dimensions Much smaller than the original control space dimension (typical value) ).

[0317] Constructing a surrogate model for carbon emission prediction in the potential space Gaussian process regression was used:

[0318] ;

[0319] in It is a mean function. For the covariance kernel function, the RBF kernel is used:

[0320] ;

[0321] in The signal variance reflects the overall fluctuation range of carbon emission prediction. It is a length scale that controls the correlation between adjacent points.

[0322] Gaussian processes provide predicted mean and forecast uncertainty The desired improvement (EI) acquisition function is used to select the next evaluation point:

[0323] ;

[0324] in , This is the lowest carbon emission value found so far. and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0325] To ensure the temperature control constraints are met, a Gaussian process is also constructed for temperature prediction, and the probability of satisfying the temperature control constraints is calculated:

[0326] ;

[0327] The constrained data acquisition function is:

[0328] ;

[0329] The optimization process is as follows: after initial random sampling 30 times, the optimal control command is output.

[0330] By efficiently searching potential space, the near-optimal control command for carbon emissions can be quickly found while ensuring temperature control safety.

[0331] In this implementation, a hierarchical optimization framework decouples task-level strategies from real-time control. The upper layer determines the overall transportation plan, while the lower layer, MPC, implements refined carbon emission optimization based on a generative cognitive model. Efficient search of the potential space ensures the computational feasibility and temperature control safety of real-time optimization, thus achieving a balance between carbon emission minimization and temperature control constraints.

[0332] This invention also proposes a cold chain transportation carbon emission optimization system based on dynamic feature adaptive hierarchical modeling. It applies the cold chain transportation carbon emission optimization method based on dynamic feature adaptive hierarchical modeling disclosed in any of the above embodiments, including a data acquisition and processing unit, a key feature screening unit, a set analysis and partitioning unit, a feature freezing and adjustment unit, a carbon emission hierarchical modeling unit, a generation and cognitive prediction unit, and an optimization operation decision unit. The units are electrically connected through a communication interface.

[0333] The data acquisition and processing unit is used to collect multi-source operation data in the cold chain transportation process, and to perform time alignment, abnormal data processing and unified sampling frequency processing on the multi-source operation data to generate standardized feature sequences. The multi-source operation data includes at least energy consumption-related data and carbon emission-related data. The carbon emission-related data includes at least one of carbon emission measurement data and carbon emission conversion parameters.

[0334] The key feature screening unit is used to perform sensitivity analysis or feature importance calculation on feature variables in the standardized feature sequence based on historical energy consumption data and historical carbon emission data, so as to screen out a set of key features;

[0335] The set analysis partitioning unit is used to perform time stability analysis on the feature variables in the key feature set, so as to divide the key feature set into a basic parameter set and a dynamic parameter set;

[0336] The feature freeze adjustment unit is used to generate confidence evaluation results for feature variables in the key feature set; the feature freeze adjustment unit is also used to perform drift detection and drift type diagnosis on feature variables, and trigger the corresponding parameter set update strategy according to the drift type obtained by diagnosis, and adaptively adjust the basic parameter set and dynamic parameter set. The drift types include gradual drift, abrupt drift, periodic drift and concept drift.

[0337] The carbon emission hierarchical modeling unit is used to establish a baseline carbon emission estimation model based on the basic parameter set and output the baseline carbon emission, establish a perturbation carbon emission correction model based on the dynamic parameter set and output the perturbation carbon emission, and superimpose the baseline carbon emission and the perturbation carbon emission to obtain the real-time carbon emission estimate.

[0338] The cognitive prediction unit is used to input a state sequence containing basic parameter values ​​corresponding to the basic parameter set, dynamic parameter values ​​corresponding to the dynamic parameter set, real-time energy consumption data, and real-time carbon emission estimates into the cognitive model, output future carbon emission prediction values, and complete the state sequence when input data is missing.

[0339] The optimized operation decision unit is used to make optimized decisions on the operating parameters of refrigeration equipment and transportation operation strategies based on future carbon emission forecasts, and output carbon emission optimization control commands.

[0340] The cold chain transportation carbon emission optimization system based on dynamic feature adaptive hierarchical modeling disclosed in the above embodiments has units whose functions correspond to the steps of the cold chain transportation carbon emission optimization method based on dynamic feature adaptive hierarchical modeling disclosed in the above embodiments. Through the division of labor and collaboration among units and the closed-loop flow of data, multi-source data processing, key feature screening, hierarchical modeling, future trend prediction, and optimization decision-making are integrated into the same technical link, improving the end-to-end stability and engineering implementation consistency of carbon emission estimation, prediction, and control under complex cold chain conditions. During implementation, the above embodiments are followed, and will not be elaborated further here.

[0341] The cold chain transportation carbon emission optimization method and system based on dynamic feature adaptive hierarchical modeling disclosed in the above embodiments mainly includes a data acquisition layer, a feature analysis layer, a hierarchical modeling layer, a cognitive prediction layer, and an optimization control layer. The functions and implementation processes of each layer are as follows.

[0342] First, the data acquisition layer obtains multi-source data during transportation through vehicle-side sensors and energy consumption monitoring devices, including information such as vehicle speed, load, outside temperature, cooling power, and door opening / closing frequency, and performs standardized processing and time synchronization on the data. All data is transmitted to edge nodes for preliminary filtering and noise correction.

[0343] Secondly, the feature analysis layer calculates the sensitivity indicators and time stability parameters of each input variable based on historical energy consumption and carbon emission data. The system automatically filters key features that significantly affect carbon emissions and divides them into a basic feature set and a dynamic feature set based on their volatility. This process can be periodically updated through a built-in algorithm to adapt to environmental changes.

[0344] Next, a layered modeling layer constructs a steady-state energy consumption model using basic parameters to determine the vehicle's carbon emission baseline under standard conditions. Simultaneously, a perturbation model is established using dynamic parameters to correct for energy consumption deviations caused by environmental changes and operational behaviors. The two models work together in real time to output instantaneous energy consumption and carbon emission estimates.

[0345] Subsequently, the cognitive prediction layer utilizes the temporal characteristics of the generative cognitive network structure learning system to map historical states into a latent variable space, thereby predicting future carbon emission trends. When data is missing or sensor anomalies occur, the model can automatically generate missing inputs based on the latent states, achieving continuous prediction.

[0346] Finally, the optimization control layer provides adjustment suggestions based on the prediction results, such as adjusting the power of refrigeration equipment, optimizing vehicle speed and stop duration, and selecting transportation routes with lower energy consumption, thereby achieving dynamic optimization control of carbon emissions. This layer can also interface with the transportation management system to form a data-driven scheduling assistance function.

[0347] The key technology of this invention lies in establishing a cold chain transportation carbon emission modeling and optimization mechanism that can automatically adjust according to the environment and operating conditions. Its core innovations are mainly reflected in the following aspects:

[0348] First, this invention introduces a dynamic feature recognition and classification mechanism. Through sensitivity analysis and time-varying stability determination, it automatically distinguishes between long-term stable basic features and short-term fluctuating dynamic features, thereby enabling the model to adaptively adjust according to environmental changes.

[0349] Secondly, this invention constructs a hierarchical energy consumption and carbon emission modeling framework, which separates the modeling of steady-state energy consumption and dynamic disturbances, and realizes the simultaneous description of the long-term energy consumption law and short-term fluctuation characteristics of the cold chain transportation process, which significantly improves the accuracy and interpretability of the model.

[0350] Furthermore, this invention proposes a dynamic learning and prediction mechanism. By introducing a generative cognitive network structure, it models the temporal characteristics of carbon emission changes and has the ability to maintain prediction continuity under conditions of incomplete data or sensor anomalies.

[0351] Furthermore, this invention establishes a carbon emission trend prediction and optimization feedback link, realizing a closed-loop process from data acquisition and model calculation to energy-saving control, enabling the system to dynamically adjust transportation strategies and equipment parameters based on prediction results during operation.

[0352] Compared to existing technologies that primarily focus on carbon emission monitoring and data storage, the modeling approach of this invention is more dynamic and adaptive. By introducing feature sensitivity analysis and time-varying stability determination mechanisms, the system can automatically identify the degree of influence of different parameters on carbon emissions and classify them according to their fluctuation characteristics. This allows the model to be updated in real time as the environment, equipment, and operations change, thereby improving modeling accuracy and response speed.

[0353] This invention employs a hierarchical energy consumption modeling framework, modeling long-term stable structural features and short-term changing perturbation features separately. This effectively reduces the computational complexity caused by feature coupling and improves the interpretability and scalability of the model. Furthermore, through confidence weights and a dynamic reclassification mechanism, the system can maintain stable operation even when data collection is incomplete or anomalies exist, enhancing the robustness and reliability of the model.

[0354] Furthermore, this invention establishes a complete logical chain from feature acquisition and model construction to carbon emission trend prediction and strategy optimization, transforming carbon emission analysis from passive monitoring to proactive modeling and decision support. This method not only possesses strong theoretical universality but is also compatible with existing cold chain monitoring systems, scheduling platforms, or energy-saving management modules, providing a new technical approach for improving energy efficiency and achieving low-carbon operation in cold chain transportation.

[0355] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling, characterized in that, include: Collect multi-source operational data during the cold chain transportation process, and perform time alignment, outlier processing, and unified sampling frequency processing on the multi-source operational data to obtain a standardized feature sequence. The multi-source operational data includes at least energy consumption-related data and carbon emission-related data. The carbon emission-related data includes carbon emission measurement data and / or carbon emission conversion parameters. Based on historical energy consumption data and historical carbon emission data, sensitivity analysis or feature importance calculation is performed on the feature variables in the standardized feature sequence to screen out a set of key features. A time stability analysis is performed on the feature variables in the key feature set, and the key feature set is divided into a basic parameter set and a dynamic parameter set based on the stability analysis results. Confidence evaluation results are generated for the feature variables in the key feature set. When the confidence evaluation result of a feature variable is lower than a preset threshold, the feature variable is frozen. Drift detection and drift type diagnosis are performed on the feature variables. The drift types include gradual drift, abrupt drift, periodic drift, and conceptual drift. The corresponding parameter set update strategy is triggered according to the drift type diagnosed, and the basic parameter set and the dynamic parameter set are adaptively adjusted. A baseline carbon emission estimation model is established based on the set of basic parameters and the baseline carbon emission is output. A perturbation carbon emission correction model is established based on the set of dynamic parameters and the perturbation carbon emission is output. The baseline carbon emission and the perturbation carbon emission are superimposed to obtain the real-time carbon emission estimate. The cognitive model is generated by inputting a state sequence containing the basic parameter values ​​corresponding to the basic parameter set, the dynamic parameter values ​​corresponding to the dynamic parameter set, real-time energy consumption related data, and the real-time carbon emission estimate, and outputs the future carbon emission prediction value. The model also completes the state sequence when the input data is missing. Based on the predicted future carbon emissions, the operating parameters of the refrigeration equipment and the transportation operation strategy are optimized and decisions are made, and carbon emission optimization control commands are output.

2. The method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling according to claim 1, characterized in that, The multi-source operational data includes vehicle operating parameters, environmental parameters, refrigeration equipment status parameters, and operational behavior parameters; The vehicle operating parameters include vehicle speed, acceleration, gradient, road type, and load factor; the environmental parameters include ambient temperature, humidity, solar radiation intensity, air pressure, and wind speed; the refrigeration equipment status parameters include compressor power, operating frequency, current, voltage, and number of start-stop cycles; and the operational behavior parameters include the number of door openings, door opening duration, idling time, and cold start frequency.

3. The method for optimizing carbon emissions in cold chain transportation according to claim 1, characterized in that, The sensitivity analysis includes: calculating a sensitivity index Si for each characteristic variable, wherein the sensitivity index Si satisfies... ,in, The gradient between the feature variable and carbon emissions. The fluctuation range of the characteristic variable. This represents the fluctuation range of carbon emissions.

4. The method for optimizing carbon emissions in cold chain transportation according to claim 1, characterized in that, The time stability analysis includes calculating the standard deviation and frequency of change of the characteristic variables based on a sliding window. When the standard deviation of the characteristic variable within the sliding window is less than the stability threshold σ b or the frequency of change of the characteristic variable f i Less than the low frequency threshold f b If the characteristic variable is a parameter that can be configured or measured before the start of the transportation task, then the characteristic variable shall be included in the set of basic parameters. Feature variables that do not meet the aforementioned conditions or rely on real-time sensing acquisition are included in the dynamic parameter set.

5. The method for optimizing carbon emissions in cold chain transportation according to claim 1, characterized in that, The process of generating the confidence evaluation results includes: For each feature variable in the set of key features, obtain the corresponding signal quality index, missing ratio index, and noise level index. The signal quality index, the missing ratio index, and the noise level index are fused and calculated according to the preset fusion rules to obtain the confidence evaluation result of the feature variables; When the confidence evaluation result is lower than a preset threshold, the feature variable is frozen. The freezing process includes stopping the use of the feature variable for model updates or using the frozen value in the calculation.

6. The method for optimizing carbon emissions in cold chain transportation according to claim 5, characterized in that, The specific process of adaptively adjusting the basic parameter set and the dynamic parameter set includes: Drift detection and drift type diagnosis are performed on the feature variables in the key feature set. The drift types include gradual drift, abrupt drift, periodic drift, and conceptual drift. Gradual drift is detected by trend test of sliding window mean, abrupt drift is detected by abrupt change in cumulative sum statistic, periodic drift is detected by frequency domain analysis, and conceptual drift is detected by the change in the correlation between feature variables and carbon emission measurement data. Based on the drift type diagnosed, a corresponding parameter set update strategy is triggered: when a gradual drift is detected, the weight coefficients of the baseline carbon emission estimation model are updated incrementally online; when a sudden drift is detected, the drifting feature variables are transferred between the basic parameter set and the dynamic parameter set, and the coefficients of the corresponding model are re-estimated; when a periodic drift is detected, time periodic encoding is introduced for the drifting feature variables; when a conceptual drift is detected, the key feature set is re-screened and the basic parameter set and dynamic parameter set are re-partitioned. After the parameter set update strategy is executed, the change of carbon emission estimation residual is monitored. When the carbon emission estimation residual does not converge to the preset residual threshold, the response strategy is switched to a higher level according to the preset strategy upgrade rules.

7. The method for optimizing carbon emissions in cold chain transportation according to claim 1, characterized in that, The baseline carbon emission estimation model includes a steady-state energy consumption model and an energy consumption-carbon emission conversion. The steady-state energy consumption model is used to obtain baseline energy consumption estimates based on the set of basic parameters. E base ,in ,in For the first in the set of basic parameters i One basic parameter, β i For the corresponding weighting coefficients, β 0 represents the baseline constant term; The baseline energy consumption estimate E base The baseline carbon emission is calculated based on the aforementioned carbon emission conversion parameters; The perturbation carbon emission correction model is used to obtain the perturbation energy consumption estimate based on the dynamic parameter set. E yn ( t ),in , where γ k ( t ) represents the dynamic sensitivity coefficient, Δ x k ( t ) represents the deviation of the dynamic parameters in the set of dynamic parameters from the reference. The estimated disturbance energy consumption E yn ( t The perturbation carbon emissions are obtained based on the aforementioned carbon emission conversion parameters; The baseline carbon emissions are superimposed with the perturbed carbon emissions to obtain the real-time carbon emissions estimate.

8. The method for optimizing carbon emissions in cold chain transportation according to claim 1, characterized in that, The generative cognitive model includes an encoder, a state transition generator, and a decoder; the encoder includes a one-dimensional convolutional layer and a Transformer layer, used to encode the state sequence into a latent state vector; the state transition generator includes a fully connected structure and a self-attention structure, used to generate the latent state at the next time step. The decoder is used to decode and output future carbon emission predictions, future energy consumption predictions, and future temperature predictions. The generative cognitive model employs a training mechanism of pre-training and online incremental updates, and fills in the state sequence based on pseudo-samples when input data is missing. The optimization decision includes determining carbon emission optimization control instructions for adjusting the operating parameters of the refrigeration equipment and the transportation operation strategy, provided that the predicted future temperature value meets the preset temperature control requirements. The Transformer layer employs a physics-guided attention mechanism, using the state variables corresponding to the cargo compartment's thermal balance relationship as prior constraints for attention weights. The training objective function of the generative cognitive model includes data-driven loss, latent space regularization loss, and physics-constraint loss constructed based on the cargo compartment's thermal balance equation. The decoder's output is equipped with an energy conservation correction layer, used to perform energy consumption-carbon emission consistency correction and temperature-energy consumption consistency correction on future carbon emission predictions, future energy consumption predictions, and future temperature predictions.

9. The method for optimizing carbon emissions in cold chain transportation according to claim 8, characterized in that, The generative cognitive model takes the state sequence of the most recent m time steps as input during real-time operation and recursively outputs the carbon emission prediction value for the next T time steps, where m and T are preset positive integers; the carbon emission prediction value is compared with the corresponding carbon emission measurement data or the carbon emission estimate value obtained from the carbon emission conversion parameter to obtain residual information, and online incremental updates are performed based on the residual information; The optimization decision-making adopts a hierarchical optimization framework: the high-level strategy optimization determines the route selection, operation mode, and stop plan at the start of the task; the low-level model predictive control, based on the short-term prediction of the generative cognitive model during transportation, aims to minimize carbon emissions in the prediction time domain and uses the temperature control safety range as a hard constraint to solve the rolling time domain optimization problem and output control commands for compressor power and vehicle speed; the low-level model predictive control constructs a Gaussian process surrogate model in the potential state vector space and performs iterative search based on the acquisition function to determine the carbon emission optimization control command under the condition of satisfying the temperature control safety range.

10. A cold chain transportation carbon emission optimization system based on dynamic feature adaptive hierarchical modeling, characterized in that, The method for optimizing carbon emissions in cold chain transportation based on dynamic feature adaptive hierarchical modeling as described in any one of claims 1 to 9 includes a data acquisition and processing unit, a key feature screening unit, a set analysis and partitioning unit, a feature freezing and adjustment unit, a carbon emission hierarchical modeling unit, a cognitive prediction generation unit, and an optimized operation decision-making unit, with each unit electrically connected through a communication interface. The data acquisition and processing unit is used to collect multi-source operation data during the cold chain transportation process, and to perform time alignment, abnormal data processing and unified sampling frequency processing on the multi-source operation data to generate a standardized feature sequence. The multi-source operation data includes at least energy consumption-related data and carbon emission-related data. The carbon emission-related data includes at least one of carbon emission measurement data and carbon emission conversion parameters. The key feature screening unit is used to perform sensitivity analysis or feature importance calculation on the feature variables in the standardized feature sequence based on historical energy consumption data and historical carbon emission data, so as to screen out a set of key features. The set analysis partitioning unit is used to perform time stability analysis on the feature variables in the key feature set, so as to divide the key feature set into a basic parameter set and a dynamic parameter set. The feature freezing and adjustment unit is used to generate confidence evaluation results for the feature variables in the key feature set; the feature freezing and adjustment unit is also used to perform drift detection and drift type diagnosis on the feature variables, and trigger the corresponding parameter set update strategy according to the drift type obtained by diagnosis, and adaptively adjust the basic parameter set and the dynamic parameter set. The drift type includes gradual drift, abrupt drift, periodic drift and concept drift. The carbon emission hierarchical modeling unit is used to establish a baseline carbon emission estimation model based on the basic parameter set and output the baseline carbon emission, establish a perturbation carbon emission correction model based on the dynamic parameter set and output the perturbation carbon emission, and superimpose the baseline carbon emission and the perturbation carbon emission to obtain the real-time carbon emission estimate. The generative cognitive prediction unit is used to input a state sequence containing the basic parameter values ​​corresponding to the basic parameter set, the dynamic parameter values ​​corresponding to the dynamic parameter set, real-time energy consumption related data, and the real-time carbon emission estimate into the generative cognitive model, output the future carbon emission prediction value, and complete the state sequence when the input data is missing; wherein, the generative cognitive model introduces physical constraints based on the cargo compartment thermal balance relationship, and includes an energy conservation consistency correction layer for performing consistency correction on the prediction output; The optimized operation decision unit is used to make optimized decisions on the operating parameters of the refrigeration equipment and the transportation operation strategy based on the predicted future carbon emissions, and output carbon emission optimization control commands; wherein, the optimization decision adopts a hierarchical optimization framework and includes rolling time-domain solution based on model predictive control, and constructs a Gaussian process surrogate model in the potential state vector space for iterative search.

Images