A carbon emission management method and system based on federated learning

Abstract

This application proposes a carbon emission management method and system based on federated learning, belonging to the field of carbon emission management. The method includes: training several first carbon emission prediction models on a preset model architecture downloaded from collaborators at several park nodes; updating the first carbon emission prediction models to obtain several second carbon emission prediction models; processing the input parameters of the target time zones of several parks to obtain several first carbon emission prediction values; obtaining several carbon emission statistical values, summing them with the corresponding first carbon emission prediction values ​​to obtain several second carbon emission prediction values, and performing carbon emission management. This application solves the technical problem in existing technologies where carbon emission management ignores indirect carbon emissions generated during the federated learning process, leading to poor carbon emission management accuracy. It achieves the technical effect of incorporating indirect carbon emissions generated by computing and communication power consumption during the federated learning process into management statistics, thereby improving the accuracy of carbon emission management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of carbon emission management, and in particular to a carbon emission management method and system based on federated learning. BACKGROUND

[0002] Traditional carbon emission management mainly relies on statistical reports and empirical formulas, but in the face of complex scenarios of multiple parks and multiple enterprises, there are often problems such as data islands and low prediction accuracy. At present, the development of federated learning technology provides a new idea for solving multi-party collaborative carbon emission management. Federated learning allows multiple participants to collaboratively train machine learning models without sharing raw data, protecting the data privacy of each party and improving model performance using distributed data. Therefore, the carbon emission management method based on federated learning has gradually attracted the attention of researchers and practitioners.

[0003] However, the federated learning process itself consumes a large amount of computing resources and network communication resources, and these computing resource consumptions will produce additional indirect carbon emissions. Specifically, each participating node needs a large amount of CPU or GPU computing when training the model, the encrypted transmission of model parameters requires network devices to run continuously, and the parameter fusion by the collaboration party also requires the server to work for a long time. These processes will produce power consumption and in turn produce indirect carbon emissions. The current carbon emission management only considers the direct carbon emissions at the business level when predicting and calculating carbon emissions, and ignores the indirect carbon emissions produced in the federated learning process. This ignorance leads to poor accuracy of carbon emission management, which cannot comprehensively and accurately reflect the real carbon emission situation, affecting the scientificity and effectiveness of carbon emission management decisions. SUMMARY

[0004] The present application provides a carbon emission management method and system based on federated learning to solve the technical problem of poor accuracy of carbon emission management caused by ignoring indirect carbon emissions produced in the federated learning process in the prior art.

[0005] The technical solution of the present application to solve the above technical problems is as follows:

[0006] In a first aspect, the present application provides a carbon emission management method based on federated learning, comprising: taking a carbon emission source quantity list, a carbon emission source start-up duration list, a unit production carbon emission quantity list, and a unit production duration list as inputs, taking a preset time zone carbon emission quantity monitoring value as a true value, training a plurality of first carbon emission prediction models on a plurality of park nodes based on a preset model architecture downloaded from a collaboration party; uploading a plurality of first model parameters of the plurality of first carbon emission prediction models to the collaboration party for optimization fusion based on a first encryption key downloaded from the collaboration party through the plurality of park nodes, obtaining fusion model parameter encryption, and issuing the fusion model parameter encryption to the plurality of park nodes to update the plurality of first carbon emission prediction models, thereby obtaining a plurality of second carbon emission prediction models; processing a plurality of park target time zone input parameters based on the plurality of second carbon emission prediction models, respectively, to obtain a plurality of first carbon emission quantity prediction values; performing carbon emission sample statistical analysis based on a plurality of list element quantities of the plurality of park nodes, respectively, to obtain a plurality of carbon emission quantity statistical values, and summing up the plurality of first carbon emission quantity prediction values corresponding one by one to obtain a plurality of second carbon emission quantity prediction values; and performing carbon emission management based on the plurality of second carbon emission quantity prediction values.

[0007] In a second aspect, the present application provides a carbon emission management system based on federated learning, comprising: a federated training module configured to take a carbon emission source quantity list, a carbon emission source start-up duration list, a unit production carbon emission quantity list, and a unit production duration list as inputs, take a preset time zone carbon emission quantity monitoring value as a true value, and train a plurality of first carbon emission prediction models on a plurality of park nodes based on a preset model architecture downloaded from a collaboration party; a parameter fusion module configured to upload a plurality of first model parameters of the plurality of first carbon emission prediction models to the collaboration party for optimization fusion based on a first encryption key downloaded from the collaboration party through the plurality of park nodes, obtain fusion model parameter encryption, and issue the fusion model parameter encryption to the plurality of park nodes to update the plurality of first carbon emission prediction models, thereby obtaining a plurality of second carbon emission prediction models; a prediction calculation module configured to process a plurality of park target time zone input parameters based on the plurality of second carbon emission prediction models, respectively, to obtain a plurality of first carbon emission quantity prediction values; a statistical analysis module configured to perform carbon emission sample statistical analysis based on a plurality of list element quantities of the plurality of park nodes, respectively, to obtain a plurality of carbon emission quantity statistical values, and sum up the plurality of first carbon emission quantity prediction values corresponding one by one to obtain a plurality of second carbon emission quantity prediction values; and a management execution module configured to perform carbon emission management based on the plurality of second carbon emission quantity prediction values.

[0008] The present application has the following beneficial effects:

[0009] Using a list of carbon emission sources, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output as input, and a preset time zone carbon emission monitoring value as the true value, several first carbon emission prediction models are trained on a preset model architecture downloaded from collaborators at several park nodes, laying the foundation for subsequent collaborative optimization. Through several park nodes, based on the first encryption key downloaded from collaborators, several first model parameters of several first carbon emission prediction models are encrypted and uploaded to collaborators for optimization and fusion. The resulting fused model parameters are encrypted and distributed to several park nodes to update several first carbon emission prediction models, resulting in several second carbon emission prediction models. This realizes the collaborative mechanism of federated learning, protects the privacy of model parameters in each park through encryption, and improves the overall model performance through optimization and fusion by collaborators, forming an optimized prediction model. Based on several secondary carbon emission prediction models, the input parameters of several target time zones in several parks are processed to obtain several primary carbon emission prediction values. The models trained by federated learning are used to predict carbon emissions in the target time zones, obtaining business-level carbon emission prediction results based on machine learning. According to the number of list elements of several park nodes, carbon emission sample statistical analysis is performed to obtain several carbon emission statistical values. These are summed with the corresponding primary carbon emission prediction values ​​to obtain several secondary carbon emission prediction values. The computational power consumption in the federated learning process is quantified by statistically analyzing the number of list elements in each park, thereby calculating the indirect carbon emissions generated by federated learning. These are then merged with the business prediction values ​​to achieve comprehensive carbon emission statistics. Carbon emission management is performed based on these secondary carbon emission prediction values ​​to ensure management accuracy and comprehensiveness.

[0010] The above technical solution enables the indirect carbon emissions generated during federated learning to be incorporated into the carbon emission management system, solving the problem that existing technologies ignore carbon emissions from computing power consumption and improving the accuracy and completeness of carbon emission management. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a carbon emission management method based on federated learning provided by this invention;

[0012] Figure 2 This is a schematic diagram of a carbon emission management system based on federated learning, provided for the present invention.

[0013] In the attached diagram, the components represented by each number are as follows:

[0014] Federated training module 11, parameter fusion module 12, prediction calculation module 13, statistical analysis module 14, and management execution module 15. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0018] Example 1, as Figure 1 As shown, this embodiment of the invention provides a carbon emission management method based on federated learning, including:

[0019] S1. Using the list of carbon emission source quantity, the list of carbon emission source start-up time, the list of carbon emission per unit output, and the list of production time per unit output as input, and the carbon emission monitoring value of the preset time zone as the true value, train several first carbon emission prediction models on the preset model architecture downloaded from the collaborator at several park nodes.

[0020] Specifically, a distributed architecture design of collaborators and park nodes is adopted. The collaborator acts as the central server for federated learning, responsible for the unified management and coordination of the entire process, including the storage and distribution of pre-defined model architectures. Several park nodes constitute the distributed computing network for federated learning, with each node corresponding to an independent industrial or enterprise park and serving as a participant in the federated learning process. Each park node maintains its own carbon emission data locally, possessing independent data processing and model training capabilities, while also communicating securely with the collaborator through the network. This architecture design ensures the data privacy and security of each park, as data does not need to leave the local environment, meeting the requirements of data protection and privacy computing.

[0021] Before training begins, several park nodes download a unified pre-defined model architecture from collaborators. This pre-defined model architecture is a pre-designed and validated machine learning model structure, specifically employing deep neural networks, random forests, support vector machines, or combinations thereof. This pre-defined model architecture is capable of handling multi-dimensional input parameters and establishing a non-linear mapping relationship between input parameters and carbon emissions. By using a unified pre-defined model architecture, it ensures that the models trained by each park node have the same structural foundation, providing technical support for subsequent model optimization.

[0022] First, a carbon emission source characteristic description system is established as input parameters, including a list of carbon emission source quantity, a list of carbon emission source start-up time, a list of carbon emissions per unit output, and a list of production time per unit output. The list of carbon emission source quantity records the specific quantity of various emission sources involved in the carbon emission calculation, such as the number of production equipment, process units, and auxiliary facilities; this list reflects the scale characteristics of emission sources within the park. The list of carbon emission source start-up time records the actual operating time of various carbon emission sources within a preset time zone, affecting the calculation of total carbon emissions; the longer the operating time, the greater the corresponding carbon emissions. The list of carbon emissions per unit output records the carbon emissions generated by various carbon emission sources when producing one unit of product, reflecting the emission intensity characteristics of different emission sources. The list of production time per unit output records the time required for various carbon emission sources to complete one unit of output, which is closely related to production efficiency. The elements in the above four lists are strictly organized according to a one-to-one correspondence; that is, the quantity, start-up time, carbon emissions per unit output, and production time per unit output of the i-th carbon emission source correspond to the i-th element of each of the four lists, ensuring data consistency and completeness.

[0023] Subsequently, a distributed training model using federated learning is employed, executing model training tasks in parallel across several park nodes. Each park node represents an independent data source and computing unit, which can be an industrial park, a corporate group, or a regional carbon emission management center. Each park node communicates with collaborating server via network connection, but the raw data of each node remains within its local environment, ensuring data privacy and security. Specifically, each park node uses a list of carbon emission source quantities, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output as input, and a preset time zone carbon emission monitoring value as the true value, to train a downloaded preset model architecture. The preset time zone can be set according to actual needs, such as hourly, daily, weekly, or monthly time dimensions. Carbon emission monitoring values ​​are obtained in real time through carbon emission monitoring equipment deployed in each park, including but not limited to specialized equipment such as infrared gas analyzers, non-dispersive infrared absorption spectrometers, and laser absorption spectrometers. These monitoring values, after data preprocessing and quality control, serve as the standard answers for model training, ensuring the authenticity and authority of the training data. During training, parameters are continuously adjusted using the backpropagation algorithm to minimize the loss function between the predicted and true values. Each node in the park trains a first carbon emission prediction model suitable for its local carbon emission characteristics based on the characteristics and scale of its local data, thus obtaining several first carbon emission prediction models.

[0024] The distributed training method described above makes full use of the local data resources of each park, protects sensitive data from being leaked, and provides a foundation for subsequent model updates.

[0025] S2. Through the aforementioned park nodes, based on the first encryption key downloaded from the collaborator, the parameters of the aforementioned first carbon emission prediction models are encrypted and uploaded to the collaborator for optimization and fusion. The resulting fused model parameters are encrypted and sent to the aforementioned park nodes to update the aforementioned first carbon emission prediction models, thereby obtaining several second carbon emission prediction models.

[0026] Specifically, after completing local model training, each park node uploads the first model parameters corresponding to the trained first carbon emission prediction model to the collaborator for optimization and fusion. To ensure the security of parameter transmission, each park node first downloads the first encryption key from the collaborator. The first encryption key is generated using symmetric or asymmetric encryption algorithms, possessing high-strength encryption protection capabilities, and can effectively prevent parameters from being maliciously intercepted or tampered with during network transmission.

[0027] Each node in the industrial park extracts the first model parameters of its local primary carbon emission prediction model, including but not limited to key model parameters such as weight matrix, bias vector, and activation function parameters. Subsequently, the first model parameters are encrypted using the downloaded first encryption key to generate an encrypted parameter package. The encrypted parameter package is then uploaded to the collaborating party via a secure communication protocol, ensuring the confidentiality and integrity of the parameter information during transmission.

[0028] After receiving encrypted parameter packets from several park nodes, the collaborating party first decrypts them to recover the initial model parameters of each park node. Then, it performs optimization and fusion, using the prediction accuracy of the initial model parameters of each park node as the fitness evaluation index, and employs a swarm intelligence optimization method to globally optimize each initial model parameter. Specifically, genetic algorithms, particle swarm optimization algorithms, simulated annealing algorithms, or their improved forms can be used as the core of the optimization process, searching for the optimal parameter combination through iterative optimization. During the optimization process, the performance of several initial model parameters of each park node is comprehensively considered, and through operations such as crossover, mutation, and selection, the system gradually converges to the global optimum, ultimately obtaining the fused model parameters. These fused model parameters integrate the advantageous features of each park node, possessing stronger generalization ability and prediction accuracy.

[0029] After optimization and fusion, the collaborating party encrypts the fusion model parameters using the corresponding first encryption key, generating encrypted fusion model parameters. These encrypted parameters are then distributed to each park node. Upon receiving the encrypted fusion model parameters, each park node decrypts them using the corresponding first decryption key to obtain the fusion model parameters. Then, each park node uses the received fusion model parameters to update the parameters of its local first carbon emission prediction model. During the update process, the original parameters of the local first carbon emission prediction model are replaced by the fusion model parameters, thereby achieving model optimization and upgrade. The fusion model parameters, optimized by the collaborating party, integrate the data features and model training experience of multiple park nodes, thus exhibiting superior performance and stronger generalization ability compared to parameters trained independently by a single park node. Through this parameter update process, the first carbon emission prediction models of each park node are optimized and improved, forming a second carbon emission prediction model with superior performance, resulting in several second carbon emission prediction models. These second carbon emission prediction models show significant improvements in prediction accuracy, stability, and applicability compared to the original several first carbon emission prediction models.

[0030] Through the above-mentioned process of aggregating and optimizing encrypted parameters, collaborative optimization of the distributed model was achieved while ensuring data security, laying a solid foundation for improving the accuracy of carbon emission prediction.

[0031] S3. Based on the aforementioned second carbon emission prediction models, process the input parameters of several target time zones in the parks respectively to obtain several first carbon emission prediction values.

[0032] Specifically, each park node sets a target time zone as the prediction time range based on actual carbon emission management needs. The target time zone can be flexibly configured according to management strategies and monitoring requirements, such as a time span of one hour, one day, one week, or one month. For the set target time zone, each park node collects and organizes corresponding input parameters to form several park target time zone input parameters.

[0033] The input parameters for the target time zones of several industrial parks adopt the same data structure as in step S1, including a list of the number of carbon emission sources in the target time zone, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output. The organization of the above input parameters is consistent with that in the training phase to ensure that the model can correctly process and identify input features.

[0034] Subsequently, each park node uses the obtained second carbon emission prediction model to process the input parameters for its local target time zone. Because the second carbon emission prediction model undergoes global optimization through federated learning, it possesses stronger predictive power and generalization performance, accurately identifying the complex mapping relationship between input parameters and carbon emissions. During the prediction process, each park node inputs the target time zone input parameters into its local second carbon emission prediction model. Based on its internal parameter configuration and learned feature representations, the model calculates and outputs the corresponding carbon emission prediction result, i.e., the first carbon emission prediction value. The prediction calculation process involves forward propagation, feature extraction, and nonlinear transformation operations within a multi-layer neural network, ultimately generating the first carbon emission prediction value for the target time zone. After the above prediction processing, each park node obtains its local first carbon emission prediction value, resulting in several first carbon emission prediction values. These several first carbon emission prediction values ​​correspond to the carbon emission prediction results of each park node within the target time zone, reflecting the carbon emission level of each park based on current operating parameters and process configurations.

[0035] It should be noted that the initial carbon emission forecast obtained at this stage only considers the carbon emissions generated by the park's production activities themselves, and does not yet include the additional carbon emissions generated during the federated learning calculation process.

[0036] Through the above prediction process, distributed carbon emission prediction based on the optimization model was achieved, providing accurate data support for carbon emission management decisions in various industrial parks.

[0037] S4. Based on the number of list elements of the aforementioned park nodes, perform carbon emission sample statistical analysis to obtain several carbon emission statistical values, and sum them with the corresponding first carbon emission prediction values ​​to obtain several second carbon emission prediction values.

[0038] Specifically, each park node first extracts the number of list elements from the input parameters of its local target time zone. The number of list elements refers to the number of elements in any one of the following lists: the list of carbon emission source numbers, the list of carbon emission source start-up times, the list of carbon emissions per unit of output, and the list of production time per unit of output. Since these four lists are organized in a one-to-one correspondence, each list has the same number of elements, reflecting the total scale of emission sources participating in the carbon emission calculation.

[0039] The number of list elements directly relates to the computational power consumption in the federated learning process. A larger list indicates a larger data volume to be processed, increasing the computational complexity of model training and prediction, thus leading to higher computational power requirements and corresponding carbon emissions. For example, when the carbon emission source start-up duration list contains more start-up duration data, the model needs to process more feature dimensions, significantly increasing the computational load, and consequently raising energy consumption and carbon emissions.

[0040] Subsequently, based on the number of extracted list elements, each park node performed a carbon emission sample statistical analysis. This analysis aims to quantify the carbon emissions generated by federated learning computing power consumption, addressing the lack of carbon emission statistics for the computation process in traditional carbon emission management. During the statistical analysis, each park node established a quantitative relationship model between computing power consumption and carbon emissions based on the local number of list elements, combined with historical computing power consumption data and corresponding carbon emission monitoring records. Through sample statistical methods, such as mean calculation, regression analysis, or machine learning prediction, the computing power consumption carbon emissions corresponding to the current number of list elements were estimated. After statistical analysis, each park node obtained several carbon emission statistics, which accurately reflect the additional carbon emissions generated during federated learning computation.

[0041] Subsequently, each park node sums its obtained carbon emission statistics with the first carbon emission prediction value obtained in step S3. This summation process uses a one-to-one correspondence, meaning the carbon emission statistics of the i-th park node are added to the corresponding i-th first carbon emission prediction value, ensuring the accuracy and completeness of the statistics. Through this summation, each park node obtains several second carbon emission prediction values. These second carbon emission prediction values ​​comprehensively consider carbon emissions generated by park production activities and carbon emissions generated by federated learning computing power consumption, offering greater comprehensiveness and accuracy compared to the first carbon emission prediction value that only considers production activities. This comprehensive statistical method effectively solves the problem of neglecting carbon emissions during the calculation process in traditional carbon emission management, providing a reliable data foundation for developing more precise carbon emission management strategies.

[0042] Through the above statistical analysis and comprehensive prediction calculation of carbon emission samples, a comprehensive quantitative prediction of carbon emissions has been achieved, improving the accuracy of carbon emission management.

[0043] S5. Perform carbon emission management based on the aforementioned second carbon emission prediction values.

[0044] Specifically, each node in the park formulates corresponding carbon emission management strategies based on the obtained second carbon emission forecast. The second carbon emission forecast comprehensively considers carbon emissions from production activities and carbon emissions from federated learning computing power consumption, and can fully reflect the park's actual carbon emission level in the target time zone, providing accurate data support for management decisions.

[0045] Based on the magnitude and trend of the second carbon emission forecast, each node in the park can assess the gap between the current carbon emission status and the preset target, identify carbon emission risk points and optimization space. When the second carbon emission forecast exceeds the preset target, the park needs to take corresponding emission reduction measures; when the second carbon emission forecast is within the forecast range, the current operating mode can be maintained or further optimized and improved.

[0046] Based on the carbon emission management strategy, each node in the industrial park implements specific carbon emission management measures. These measures include, but are not limited to, the following aspects: In terms of production scheduling optimization, production plans and equipment operation strategies are adjusted according to the second carbon emission forecast, optimizing the start-up time and operating intensity of carbon emission sources to minimize carbon emissions while meeting production needs. In terms of equipment operation and maintenance management, high-carbon emission equipment and processes are identified based on the forecast results, and targeted equipment maintenance and technical upgrade plans are developed to improve energy efficiency. In terms of energy structure adjustment, the proportion of clean energy and traditional energy use is allocated according to carbon emission forecasts, gradually reducing reliance on high-carbon emission energy sources.

[0047] During the implementation of management measures, each park node continuously monitors actual carbon emissions, compares the monitoring results with the second carbon emission prediction value, and evaluates the effectiveness of the management measures and the accuracy of the prediction model. By establishing a closed-loop management mechanism of prediction-implementation-monitoring-feedback, the carbon emission management strategy and prediction model performance are continuously optimized. When there is a significant deviation between actual carbon emissions and the second carbon emission prediction value, the park nodes promptly adjust the management strategy or update the model parameters to ensure the dynamic adaptability and continuous improvement of carbon emission management. At the same time, the data and experience accumulated during the management process provide new samples for subsequent federated learning training, forming a cycle of continuous optimization.

[0048] The aforementioned carbon emission management implementation based on predicted values ​​has achieved a complete closed loop from data prediction to management practice, providing effective support for the control and continuous reduction of carbon emissions in the park.

[0049] Furthermore, using a list of carbon emission source quantities, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output as input, and using carbon emission monitoring values ​​for a preset time zone as true values, several first carbon emission prediction models are trained at several park nodes on a preset model architecture downloaded from collaborators, including:

[0050] S11. At the first park node of the plurality of park nodes, the first park node model construction data is collected with the carbon emission source quantity list, carbon emission source start-up time list, unit output carbon emission list, and unit output production time list as inputs and the carbon emission monitoring value of the preset time zone as the true value.

[0051] S12. Obtain the preset model architecture downloaded from the collaborator;

[0052] S13. Using the data constructed by the first park node model, train the preset model architecture to obtain the carbon emission prediction model of the first park node.

[0053] Among them, the carbon emission prediction model of the first park node belongs to the plurality of first carbon emission prediction models.

[0054] Specifically, this paper selects any one of several park nodes as the first park node and details the process of building the first carbon emission prediction model for a single park node. First, the first park node collects and organizes local training data according to a unified data organization standard, using a list of carbon emission source quantities, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output as inputs, and a preset time zone carbon emission monitoring value as the ground truth. Specifically, the first park node extracts relevant data from its local carbon emission management database, including the lists of carbon emission source quantities, carbon emission source start-up times, carbon emissions per unit output, production times per unit output, and preset time zone carbon emission monitoring values, as the model construction data for the first park node. The list of carbon emission source quantities records the actual quantity of various emission sources within the first park; the list of carbon emission source start-up times records the historical operating time of each emission source within the preset time zone; the lists of carbon emissions per unit output and production times per unit output are obtained based on historical production data statistics; and the carbon emission monitoring values ​​are collected by carbon emission monitoring equipment deployed on the first park node and, after data preprocessing, are used as the ground truth for model training. After the above data collection process, the first park node obtains the first park node model construction data, which includes complete input-output sample pairs for training the first carbon emission prediction model.

[0055] Subsequently, the first park node downloads the pre-defined model architecture from the collaborator. The pre-defined model architecture can employ deep neural networks, gradient boosting trees, support vector regression, or combinations thereof; the specific architecture selection is determined based on the characteristics and accuracy requirements of the carbon emission prediction task. This pre-defined model architecture includes complete network structure definitions for the input layer, hidden layers, and output layer, as well as configurations for key components such as activation functions and loss functions. By using a unified pre-defined model architecture, it ensures that the models trained by each park node have the same structural foundation, providing support for subsequent parameter optimization and fusion.

[0056] Then, the first park node uses the collected first park node model construction data to train the downloaded preset model architecture. The training process employs supervised learning. First, the first park node model construction data is divided into training and validation sets according to a preset ratio, typically 70%-80% for the training set and 20%-30% for the validation set, ensuring sufficient training and effective evaluation. During the training phase, the first park node uses the training set to learn model parameters. Specifically, the list of carbon emission source quantities, the list of carbon emission source start-up times, the list of carbon emissions per unit output, and the list of production time per unit output are input into the preset model architecture. The model, through a forward propagation process, undergoes linear transformations and nonlinear activation functions at each network layer, extracting and combining feature information layer by layer, ultimately generating a predicted carbon emission value for the preset time zone at the output layer. Subsequently, a loss function is calculated between the predicted carbon emission value for the preset time zone and the corresponding true value (the monitored carbon emission value for the preset time zone). The loss function can use mean squared error (MSE), mean absolute error (MAE), or Huber loss, among others. Based on the calculated loss function value, the gradient information of the parameters of each layer is calculated using the backpropagation algorithm, and the model parameters are updated using optimization algorithms (such as Adam, SGD, or RMSprop). During the validation phase, the generalization performance of the model is evaluated using a validation set, overfitting during training is monitored, and strategies such as early stopping or learning rate adjustment are implemented based on the validation loss. After multiple rounds of iterative training, the loss function value of the model parameters on the training set gradually decreases, and the prediction accuracy continuously improves. When the loss function converges to a stable state or reaches the preset number of training rounds, the training process ends, and the carbon emission prediction model for the first park node is obtained, serving as the first carbon emission prediction model for the first park node. For example, the preset model architecture adopts a deep feedforward neural network structure, which includes an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer corresponds to the dimension of the input features, i.e., the total number of elements in the four feature lists. The first hidden layer contains 128 neurons and uses the ReLU activation function, which effectively handles the nonlinear combination of input features. The second hidden layer contains 64 neurons, also using the ReLU activation function, to further extract high-level features. The third hidden layer contains 32 neurons and continues to use the ReLU activation function to achieve deep feature abstraction. The output layer contains one neuron and uses a linear activation function to directly output the predicted carbon emission value for the preset time zone. To prevent overfitting, a Dropout layer is added between the first and second hidden layers, with a dropout rate set to 0.2. All layers of the network are fully connected, and the weight parameters are initialized using the Xavier initialization method. This network architecture has good feature learning ability and generalization performance, and can effectively establish a complex nonlinear mapping relationship between input features and carbon emissions.

[0057] All other park nodes employ the exact same training strategy and implementation process as the first park node, ensuring good consistency and comparability of the models trained by each park node. Through this unified training process, each park node obtains its own first carbon emission prediction model, which together constitute several first carbon emission prediction models in the federated learning. Each first carbon emission prediction model fully learns the local data characteristics of its corresponding park while maintaining a unified model structure, providing a foundation for subsequent parameter fusion and global optimization in federated learning.

[0058] Furthermore, by constructing data using the first park node model, the preset model architecture is trained to obtain the first park node carbon emission prediction model, including:

[0059] S131. Data is constructed using the first park node model to train the preset model architecture and obtain an initial carbon emission prediction model.

[0060] S132. Extract the set of output deviation vectors from the initial carbon emission prediction model whose output loss is greater than or equal to the loss threshold.

[0061] S133. Perform LOF outlier minimum value sorting on the output deviation vector set to obtain the first output deviation calibration vector;

[0062] S134. When the magnitude of the first output deviation calibration vector is greater than or equal to the deviation magnitude threshold, the first output deviation calibration vector is used to replace the true value of the first park node model construction data to obtain the first park node model update construction data, and the preset model architecture is trained to obtain the residual compensator.

[0063] S135. Sum the outputs of the residual compensator and the initial carbon emission prediction model to obtain the first compensated carbon emission prediction model architecture.

[0064] S136. Data is constructed using the first park node model to train the first compensation carbon emission prediction model architecture and obtain the first compensation carbon emission prediction model.

[0065] S137. When the magnitude of the second output deviation calibration quantity of the first compensated carbon emission prediction model is less than the deviation magnitude threshold, the first compensated carbon emission prediction model is set as the first park node carbon emission prediction model; otherwise, the residual compensator continues to execute the loop.

[0066] In a preferred embodiment, an iterative optimization mechanism based on residual compensation is used to refine the training process of the carbon emission prediction model for the first park node. By gradually identifying, analyzing and compensating for prediction errors, the prediction accuracy and generalization ability of the carbon emission prediction model for the first park node are improved.

[0067] First, the first park node uses the model construction data of the first park node to perform initial training on the preset model architecture. The training process adopts a supervised learning framework, dividing the model construction data of the first park node into training and validation sets, and minimizing the loss function between the predicted value and the true value through optimization algorithms such as gradient descent. After a preset number of iterative training rounds or after reaching the convergence condition, the initial carbon emission prediction model is obtained. This initial carbon emission prediction model has the basic ability to process inputs such as a list of carbon emission source quantities, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output, and output predicted carbon emission values. However, due to factors such as data noise, model complexity limitations, or uneven distribution of training samples, the initial carbon emission prediction model may have a large prediction bias on some samples.

[0068] Subsequently, error analysis was performed on the initial carbon emission prediction model. Specifically, all samples from the data constructed using the first park node model were used to infer the initial carbon emission prediction model, obtaining the predicted value for each sample. The output loss between the predicted value and the corresponding true value (i.e., the carbon emission monitoring value for the preset time zone) was calculated. The calculation method for this output loss included absolute error, squared error, or relative error. Simultaneously, a loss threshold was set as a screening criterion for high-error samples. This loss threshold could be adjusted according to model performance requirements and data characteristics, typically set to the 75th or 90th quantile of the training sample loss distribution. Then, samples in the initial carbon emission prediction model whose output loss was greater than or equal to the loss threshold were extracted, and the output bias vectors corresponding to these high-error samples were calculated. The output bias vector is the difference between the predicted value and the true value, i.e., output bias vector = predicted value - true value. The direction and magnitude of this output bias vector reflect the prediction bias characteristics of the initial carbon emission prediction model. The output bias vectors of all high-error samples were collected to form an output bias vector set. The output bias vector set contains the main error information of the initial carbon emission prediction model in the current training state, providing a data foundation for subsequent error compensation.

[0069] Subsequently, LOF outlier analysis is performed on the output deviation vector set to identify the most representative error patterns. LOF outlier analysis is a density-based outlier detection method that quantifies the outlier degree of each data point by calculating the ratio of its local density to its k nearest neighbors. In specific implementation, for each output deviation vector in the output deviation vector set, k nearest neighbors are determined (usually k=5 or k=10), and the local reachability density of the output deviation vector is calculated. Then, the Local Outlier Factor (LOF) value of the vector relative to its k nearest neighbors is calculated. The formula for calculating the LOF value is: LOF = Σ(LRD_neighbor) / (k×LRD_point), where LRD_neighbor is the local reachability density of the nearest neighbors, and LRD_point is the local reachability density of the current point. The closer the LOF value is to 1, the more similar the density of the point is to its neighborhood, and the more likely it is a normal point; the LOF value is greater than 1, and the more likely it is an outlier point. In this embodiment, the output deviation vector with the smallest LOF value is selected as the first output deviation calibration vector. The first output deviation calibration vector represents the vector in the output deviation vector set that is closest to the normal error distribution pattern. It can accurately reflect the systematic deviation characteristics of the initial carbon emission prediction model and avoid interference from noise or abnormal samples.

[0070] Then, the magnitude of the first output deviation calibration vector (i.e., the Euclidean norm of the vector) is calculated and compared with a preset deviation magnitude threshold. This deviation magnitude threshold is set according to the accuracy requirements of carbon emission prediction, and is usually 1-2 times the expected prediction error. When the magnitude of the first output deviation calibration vector is greater than or equal to the deviation magnitude threshold, it indicates that the initial carbon emission prediction model has a significant systematic bias, and a residual compensator needs to be constructed for error compensation. At this time, the first output deviation calibration vector is used to replace the true values ​​of the corresponding high-error samples in the first park node model construction data to generate the first park node model update construction data. Specifically, the first park node model update construction data keeps the original input features unchanged (i.e., the list of carbon emission source quantities, the list of carbon emission source start-up times, the list of carbon emissions per unit output, and the list of production time per unit output), but updates the labels of the high-error samples from the original true values ​​to the values ​​of the first output deviation calibration vector. This label replacement strategy enables the subsequently trained residual compensator to specifically learn the error patterns of the initial carbon emission prediction model. Subsequently, the first park node model update construction data is used to train the preset model architecture to obtain the residual compensator. The residual compensator uses the same network architecture as the initial carbon emission prediction model, but its training objective is to predict the output error of the initial model rather than directly predicting carbon emissions.

[0071] Next, the residual compensator is integrated with the initial carbon emission prediction model at the architecture level. Specifically, a first compensated carbon emission prediction model architecture is constructed, which includes two parallel prediction branches: the main prediction branch (initial carbon emission prediction model) and the error compensation branch (residual compensator). The forward inference process of the first compensated carbon emission prediction model architecture is as follows: For the input feature vector (a list of carbon emission source quantities, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output), the basic prediction value is first obtained through the initial carbon emission prediction model, and the error compensation value is obtained through the residual compensator. Finally, the two are summed to obtain the final compensated prediction value, that is: final prediction value = initial carbon emission prediction model output + residual compensator output.

[0072] To illustrate with a concrete example: Suppose the original data of a high-error sample has the following input features: [Number of devices: 5, Operating time: 8 hours, Carbon emissions per unit output: 2 tons per piece, Operating time per unit output: 1 hour per piece], and the corresponding ground truth label is 100 tons of carbon emissions. The initial carbon emission prediction model predicts 80 tons for this sample, resulting in a prediction error of 20 tons. After analysis using the LOF algorithm, the first output bias calibration vector is determined to be 18 tons. When generating the first park node model to update the construction data, the input features of this sample remain unchanged, still [5 devices, 8 hours, 2 tons per piece, 1 hour per piece], but its label is updated from 100 tons to 18 tons. This label replacement strategy allows the subsequently trained residual compensator to specifically learn the error pattern of the initial carbon emission prediction model. When the residual compensator receives the same input features [5 devices, 8 hours, 2 tons per piece, 1 hour per piece], its training objective is to output the error compensation value of 18 tons. Subsequently, the preset model architecture was trained using the updated construction data from the first park node model to obtain the residual compensator. The residual compensator adopts the same network architecture as the initial carbon emission prediction model, but its training objective is to predict the output error of the initial carbon emission prediction model rather than directly predicting the carbon emissions. In practical applications, when the same input features are input, the initial carbon emission prediction model outputs a basic prediction value of 80 tons, and the residual compensator outputs an error compensation value of 18 tons. The sum of the two yields a final prediction result of 98 tons, which significantly improves accuracy compared to the original prediction value of 80 tons and is closer to the actual carbon emissions of 100 tons.

[0073] Then, the first compensated carbon emission prediction model architecture is jointly trained end-to-end using the data constructed from the original first park node model. This training process aims to optimize the collaborative effect between the initial carbon emission prediction model and the residual compensator, ensuring that the two components can work effectively together to improve the overall prediction accuracy. During training, gradient information is backpropagated to the parameters of both the initial carbon emission prediction model and the residual compensator, enabling the two components to optimize collaboratively. After sufficient training, the first compensated carbon emission prediction model is obtained, which has a stronger error correction capability compared to the initial carbon emission prediction model.

[0074] Subsequently, the performance of the first compensated carbon emission prediction model is evaluated, and its prediction error is analyzed using the same methods as in steps S132 and S133. Specifically, the loss between the predicted output and the true value of the first compensated carbon emission prediction model on the data constructed by the first park node model is calculated. Samples with losses greater than or equal to the loss threshold are extracted, and the corresponding output deviation vector set is calculated. The second output deviation calibration vector is obtained by sorting through the minimum outlier factor of LOF. The modulus of the second output deviation calibration vector is calculated and compared with the deviation modulus threshold.

[0075] When the magnitude of the second output deviation calibration quantity is less than the deviation magnitude threshold, it indicates that the prediction error of the first compensated carbon emission prediction model has been reduced to an acceptable range and meets the accuracy requirements. At this time, the first compensated carbon emission prediction model is determined as the final carbon emission prediction model for the first park node.

[0076] When the magnitude of the second output deviation calibration vector is still greater than or equal to the deviation magnitude threshold, it indicates that the first compensated carbon emission prediction model still has systematic errors that need further compensation. At this point, the residual compensation process continues, constructing a second residual compensator based on the second output deviation calibration vector. This compensator is then integrated with the existing initial carbon emission prediction model and the first residual compensator to form a composite model architecture containing multiple levels of residual compensators. This iterative cycle continues until the magnitude of the output deviation calibration vector is less than the deviation magnitude threshold or the preset maximum number of compensators is reached, thus obtaining the first park node carbon emission prediction model.

[0077] Through the iterative optimization mechanism based on LOF outlier detection and residual compensation, the carbon emission prediction error can be accurately identified, analyzed and compensated, thereby improving the prediction accuracy, stability and robustness of the carbon emission prediction model for the first park node.

[0078] Furthermore, when the magnitude of the first output deviation calibration quantity is less than the deviation magnitude threshold, the initial carbon emission prediction model is set as the first park node carbon emission prediction model.

[0079] In one feasible implementation, when the magnitude of the first output deviation calibration quantity is less than the deviation magnitude threshold, it indicates that the systematic deviation of the initial carbon emission prediction model is within an acceptable range, and the prediction accuracy has reached the preset performance requirements. In this case, there is no need to construct an additional residual compensator for error compensation, and the initial carbon emission prediction model can be directly determined as the final first park node carbon emission prediction model.

[0080] Specifically, when the magnitude of the first output deviation calibration quantity is less than the deviation magnitude threshold, it indicates that the most representative prediction error selected through LOF outlier analysis is sufficiently small, and the initial carbon emission prediction model has good overall prediction performance on the data constructed from the first park node model, which can meet the accuracy requirements of carbon emission management. At this time, the residual compensation process in steps S134 to S137 is skipped, and the initial carbon emission prediction model is directly set as the first park node carbon emission prediction model.

[0081] By employing a bias modulus threshold-based processing mechanism, unnecessary computational complexity is avoided. When the initial carbon emission prediction model already possesses sufficient accuracy, there is no need to add an additional residual compensator structure, maintaining model simplicity and computational efficiency, improving training efficiency, and reducing unnecessary iterative training time and computational resource consumption. Furthermore, adding a residual compensator when prediction accuracy has already met the target may lead to excessive model complexity, thereby reducing generalization performance. Therefore, when the modulus of the first output bias calibration quantity is less than the bias modulus threshold, setting the initial carbon emission prediction model to the first park node carbon emission prediction model can prevent the risk of overfitting.

[0082] Furthermore, at the first park node of the plurality of park nodes, using the list of carbon emission source quantity, the list of carbon emission source start-up time, the list of carbon emissions per unit output, and the list of production time per unit output as input, and using the carbon emission monitoring value of a preset time zone as the true value, the model construction data of the first park node is collected, including:

[0083] S111. Through the first park node, extract the first carbon emission source type and the first carbon emission source scale of the first park, up to the Nth carbon emission source type and the Nth carbon emission source scale, where N represents the total number of carbon emission sources in the first park.

[0084] S112. Retrieve the first historical carbon emission monitoring data that meets the first carbon emission source type and the first carbon emission source scale, perform a central value assessment, and obtain the carbon emission per unit output of the first carbon emission source type and the production time per unit output of the first carbon emission source type.

[0085] S113. Until the Nth historical carbon emission monitoring data that meets the first carbon emission source type and the first carbon emission source scale is obtained, perform a central value assessment to obtain the carbon emission per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type.

[0086] S114. Based on the carbon emission per unit output of the first carbon emission source type and the production time per unit output of the first carbon emission source type, up to the carbon emission per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type, cluster the first carbon emission source type up to the Nth carbon emission source type to obtain multiple cluster carbon emission source types, wherein any cluster of carbon emission source types has an intra-cluster quantity identifier, an average carbon emission per unit output identifier, and an average production time per unit output identifier.

[0087] S115. Based on the cluster quantity identifier, construct a carbon emission source quantity list; based on the average start-up time of the cluster carbon emission source within a preset time zone, construct a carbon emission source start-up time list; based on the average carbon emission per unit output identifier, construct a carbon emission per unit output list; based on the average production time per unit output identifier, construct the production time per unit output list; and collect the first park node model construction data with the carbon emission monitoring value of the preset time zone as the true value.

[0088] In a preferred embodiment, the first park node first conducts a comprehensive survey and information extraction of all carbon emission sources within its park. Specifically, for the first park node, the type and scale of each carbon emission source within the park are identified and recorded one by one. The carbon emission source type includes, but is not limited to, classifications such as production equipment type, process unit type, and auxiliary facility type, which accurately describe the technical characteristics and functional attributes of the emission source. The carbon emission source scale reflects quantitative indicators such as capacity, power, and processing capacity. Following the order from the first to the Nth carbon emission source, the first carbon emission source type and scale, the second carbon emission source type and scale, and so on, up to the Nth carbon emission source type and scale. Here, N represents the total number of carbon emission sources in the first park, reflecting the overall scale of carbon emission sources in the park. This systematic information extraction lays the data foundation for subsequent historical data retrieval and feature analysis.

[0089] Then, a precise matching search was performed in the historical carbon emission monitoring database of the first industrial park node, targeting the first carbon emission source type and the first carbon emission source scale. This database contains a large amount of historical operational data and carbon emission monitoring records from carbon emission sources of different types and scales, providing reliable data support for the statistical analysis of characteristic parameters. The first historical carbon emission monitoring data that met the criteria of the first carbon emission source type and the first carbon emission source scale was retrieved. This first historical carbon emission monitoring data included information such as historical carbon emissions, production output, and operating time for carbon emission sources of the same type and scale. Subsequently, a central tendency assessment was performed on the first historical carbon emission monitoring data. Statistical methods such as mean calculation, median analysis, or weighted average were used to remove the influence of outliers and noise data, obtaining the carbon emission per unit output of the first carbon emission source type and the production time per unit output of the first carbon emission source type. The carbon emission per unit output of the first carbon emission source type reflects the emission intensity characteristics of that type of carbon emission source, while the production time per unit output of the first carbon emission source type reflects the production efficiency characteristics of that type of carbon emission source.

[0090] Subsequently, using the same method as in step S112, historical data retrieval and central tendency assessment are performed sequentially for the second to Nth carbon emission source types. Specifically, historical carbon emission monitoring data that meets the requirements of the second carbon emission source type and the second carbon emission source scale are retrieved, and central tendency assessment is performed to obtain the carbon emissions per unit output of the second carbon emission source type and the production time per unit output of the second carbon emission source type. This process is repeated until the carbon emissions per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type are obtained. Through systematic data processing, characteristic parameters for all carbon emission source types are obtained.

[0091] Next, based on the obtained characteristic parameters of each carbon emission source type, cluster analysis was performed on the first carbon emission source type up to the Nth carbon emission source type. The cluster analysis used the carbon emissions per unit output and the production time per unit output of the first carbon emission source type, up to the Nth carbon emission source type, as feature vectors. Clustering algorithms such as K-means, hierarchical clustering, or DBSCAN were used to group carbon emission source types with similar characteristics into the same cluster. The clustering process comprehensively considered the similarity of both carbon emissions per unit output and production time per unit output, grouping the N carbon emission source types into several clusters, resulting in multi-cluster carbon emission source types. Any cluster of carbon emission source types in the multi-cluster carbon emission source types has three key identifiers: a cluster quantity identifier, recording the number of carbon emission source types included in the cluster; a mean carbon emissions per unit output identifier, reflecting the average emission intensity of all carbon emission source types within the cluster; and a mean production time per unit output identifier, reflecting the average production efficiency of all carbon emission source types within the cluster.

[0092] Then, a standardized data list required for model training is constructed based on the cluster analysis results. Specifically, based on the cluster quantity identifier, the number of emission sources included in each cluster's carbon emission source type is counted, and a carbon emission source quantity list is constructed, reflecting the quantity distribution characteristics of different types of carbon emission sources. Based on the average start-up time of carbon emission sources within a preset time zone, a carbon emission source start-up time list is constructed. The average start-up time is obtained by analyzing the historical operation records of each cluster's carbon emission sources within the preset time zone, reflecting the typical operation mode of different types of carbon emission sources. Based on the average carbon emission per unit output identifier, a carbon emission per unit output list is directly constructed, which includes standardized emission intensity parameters for each cluster's carbon emission source type. Similarly, based on the average production time per unit output identifier, a production time per unit output list is constructed, which includes standardized production efficiency parameters for each cluster's carbon emission source type.

[0093] Subsequently, the data collection process for the first park node model was completed by combining the carbon emission monitoring values ​​of the preset time zone as ground truth labels. The data for the first park node model included structured input features (a list of carbon emission source numbers, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output) and corresponding ground truth labels, providing a high-quality and standardized data foundation for subsequent model training.

[0094] Through the above-mentioned systematic data configuration process, an effective transformation from raw carbon emission source information to standardized training data was achieved. This not only ensured the integrity and accuracy of the data, but also optimized the data structure through cluster analysis, laying a data foundation for building a high-precision carbon emission prediction model.

[0095] Furthermore, through the aforementioned park nodes, based on the first encryption key downloaded from the collaborator, several first model parameters of the aforementioned first carbon emission prediction models are encrypted and uploaded to the collaborator for optimization and fusion. The resulting fused model parameters are then encrypted and distributed to the aforementioned park nodes to update the aforementioned first carbon emission prediction models, thereby obtaining several second carbon emission prediction models, including:

[0096] S21. Verify the several second carbon emission prediction models through the several park nodes to obtain several verification loss values;

[0097] S22. Freeze the parameters of the second carbon emission prediction model for the park node whose verification loss value is less than the verification loss threshold among the plurality of verification loss values, and wait for the operation response.

[0098] S23. The second carbon emission prediction model of the park node with a verification loss value greater than or equal to the verification loss threshold among the plurality of verification loss values ​​is cyclically trained.

[0099] In a preferred embodiment, during the optimization and fusion process, the collaborating parties employ a swarm intelligence optimization algorithm to globally optimize the first model parameters uploaded by each park node. Specifically, the prediction accuracy of the first model parameters of each park node is used as the fitness evaluation index to construct a fitness function. The fitness function is based on the model's prediction accuracy on validation data, loss function value, or other performance indicators; a higher value indicates better model performance. Taking a genetic algorithm as an example, the specific implementation process of optimization and fusion is as follows: First, the first model parameters of each park node are considered as individuals in the genetic algorithm, each individual containing complete neural network weights and bias parameters. Subsequently, a selection operation is performed based on the fitness value of each individual; parameter combinations with higher fitness have a higher probability of being selected. In the crossover operation, new parameter combinations are generated through parameter fusion, such as weighted averaging or linear combination of corresponding parameters from two high-fitness individuals. In the mutation operation, the parameters are subjected to small-amplitude random perturbations to increase the diversity of parameter search. After multiple generations of iterative optimization, the algorithm gradually converges to the globally optimal fused model parameters.

[0100] It is important to note that when each park node adopts a residual compensator structure, the model structure may differ because the models of different park nodes may contain different numbers of residual compensators. In this case, the collaborating parties adopt a hierarchical optimization and fusion strategy to ensure the effective fusion of structured models. Specifically, the hierarchical optimization and fusion proceeds sequentially according to the hierarchical order of model components: First, the initial carbon emission prediction model parameters of all park nodes are optimized and fused. Since all nodes' initial carbon emission prediction models use the same preset model architecture, parameter-level swarm intelligence optimization can be directly performed to obtain the fused initial model parameters. Subsequently, for park nodes with the first residual compensator, the parameters of its first residual compensator are extracted, and a second round of optimization and fusion is performed to obtain the fused first residual compensator parameters. For park nodes with the second residual compensator, a third round of optimization and fusion is performed to obtain the fused second residual compensator parameters. This process is repeated layer by layer until the parameter fusion of all residual compensator levels is completed.

[0101] During the parameter distribution phase, the collaborating parties return the corresponding fused model parameters based on the specific model structure of each park node. For example, for a park node that only contains the initial carbon emission prediction model, only the fused initial model parameters are distributed; for a park node that contains one residual compensator, the fused initial model parameters and the first residual compensator parameters are distributed; for a park node that contains multiple residual compensators, the complete parameter set for the corresponding level is distributed.

[0102] After obtaining the second carbon emission prediction model, each park node performs performance validation and evaluation. Specifically, each park node uses its local validation set to test the second carbon emission prediction model for inference, calculating the loss function between the model's predicted output and the true values ​​on the validation set. The validation set consists of test samples independent of the training data, objectively reflecting the model's generalization performance and prediction accuracy. Through validation and evaluation, each park node obtains the validation loss value corresponding to its local second carbon emission prediction model. The validation loss value represents the degree of prediction error of the model on the validation data; a smaller value indicates better model performance. Several validation loss values ​​reflect the overall performance distribution of several second carbon emission prediction models of each park node after federated learning updates.

[0103] Next, the park nodes that have met the performance verification criteria undergo parameter freezing. Specifically, the collaborating parties set a verification loss threshold as a performance evaluation standard, which is determined based on the accuracy requirements of carbon emission prediction and the application scenario. For park nodes whose verification loss values ​​are less than the verification loss threshold, it indicates that their secondary carbon emission prediction model has reached the expected performance requirements. At this point, the parameters of the secondary carbon emission prediction model for these park nodes are frozen. Parameter freezing means setting the model parameters to a non-updateable state to prevent parameter changes during subsequent training. Park nodes with frozen parameters enter a waiting state for job response and can directly use the current secondary carbon emission prediction model to perform carbon emission prediction and management tasks without participating in subsequent iterative training processes.

[0104] For park nodes whose validation performance fails to meet the standards, iterative training is performed. Specifically, for park nodes whose validation loss value is greater than or equal to the validation loss threshold among several validation loss values, it indicates that their second carbon emission prediction model still needs further optimization. These park nodes will continue to participate in the iterative training process of federated learning, re-executing steps such as local training, parameter uploading, optimization fusion, and parameter distribution, until the model performance meets the requirements or reaches the preset maximum number of training rounds.

[0105] By integrating hierarchical management and hierarchical optimization based on verification performance, adaptive processing of models with different performance levels and effective integration of complex model structures are achieved. This not only ensures the prediction accuracy of the overall system but also optimizes the utilization efficiency of computing resources, providing a robust and reliable technical solution for the practical application of federated learning in carbon emission management.

[0106] Furthermore, based on the number of list elements in the aforementioned park nodes, carbon emission sample statistical analysis is performed to obtain several carbon emission statistical values, including:

[0107] S41. Extract the number of list elements in the first park, perform analysis through the first park computing power predictor, and obtain the predicted value of the computing power loss of the first park carbon emission prediction. The first park computing power predictor is generated by machine learning training through multiple sets of data. Each set of multiple sets of data includes the number of list elements in the input data of the second carbon emission prediction model of the first park and a label that identifies the monitoring value of the computing power loss of the carbon emission prediction.

[0108] S42. Using the carbon emission prediction loss computing power prediction value of the first park as a constraint, retrieve the carbon emission sample set, perform a central value assessment, obtain the carbon emission statistics of the first park, and add it to the several carbon emission statistics.

[0109] In a preferred embodiment, firstly, any one of the several park nodes is selected as the first park node, and the prediction process of computing power consumption carbon emissions is explained in detail. First, the number of elements in the first park list is extracted, that is, the number of elements in the input data of the second carbon emission prediction model for the first park, including the list of carbon emission source numbers, the list of carbon emission source start-up times, the list of carbon emissions per unit output, and the list of production time per unit output. Since these lists are organized in a one-to-one correspondence manner, the number of elements in any one list represents the number of elements in the first park list.

[0110] Subsequently, the number of extracted list elements from the first campus was analyzed using a pre-trained first campus computing power predictor. The first campus computing power predictor is a machine learning model specifically designed to predict computing power consumption during federated learning computations, trained using historical data. This predictor receives the number of list elements as input features and outputs the corresponding predicted computing power loss value for carbon emission prediction in the first campus. Specifically, the first campus computing power predictor is trained based on multiple sets of data, obtained through long-term computing power monitoring and carbon emission statistics accumulation. Each set of data contains the number of list elements in the input data of the second carbon emission prediction model for the first campus and a label identifying the monitored computing power loss value for carbon emission prediction. The number of list elements in the input data of the second carbon emission prediction model for the first campus serves as the input feature, reflecting the data scale of a specific prediction task; the monitored computing power loss value for carbon emission prediction is the supervision label, recording the actual computing power resources consumed when performing a prediction task of this scale. The carbon emission prediction computing power consumption monitoring value is obtained by monitoring indicators such as power consumption, CPU utilization, memory usage, and network transmission volume of computing devices in real time over a historical period. Combined with the energy consumption-carbon emission conversion coefficient, the carbon emissions generated by performing a specific-scale carbon emission prediction task are calculated. Through a large number of such input-output sample pairs, the computing power predictor in the first campus has learned the quantitative relationship between the number of list elements and the carbon emissions from computing power consumption.

[0111] Subsequently, based on the obtained predicted computing power loss value for carbon emission prediction in the first park, a statistical analysis of carbon emission samples under constraints is performed. Specifically, using the predicted computing power loss value for carbon emission prediction in the first park as the retrieval constraint, a matching search is conducted in a pre-constructed carbon emission sample set. This carbon emission sample set contains a large number of carbon emission monitoring records corresponding to different computing power consumption levels. These records originate from historical federated learning computation processes, data center operation monitoring, or carbon emission statistics from similar computational tasks. The retrieval process searches for historical samples that match or are close to the predicted computing power loss value for carbon emission prediction in the first park, forming a subset of candidate carbon emission samples. Then, a central tendency evaluation is performed on the retrieved subset of candidate carbon emission samples. The central tendency evaluation uses statistical methods, such as mean calculation, median analysis, weighted average, or average calculation after removing outliers, to eliminate noise and outlier effects in the sample data and obtain stable and reliable statistical results. Through the central tendency evaluation, a statistical value of carbon emissions in the first park is obtained, which accurately reflects the carbon emissions generated by the first park performing federated learning computation tasks with the current number of list elements. Then, the carbon emission statistics obtained from the first park are added to a set of carbon emission statistics. Using the same method, other park nodes will also calculate their own carbon emission statistics, ultimately forming a complete set of carbon emission statistics that includes the carbon emission statistics generated by all park nodes during the federated learning computation process.

[0112] Through carbon emission statistical analysis based on computing power predictors, the carbon emissions of the federated learning computing process have been accurately quantified, filling the gap in traditional carbon emission management regarding the statistics of carbon emissions during the computing process, and providing support for comprehensive carbon emission statistics and precise management.

[0113] Example 2, as Figure 2 As shown, based on the same inventive concept as the federated learning-based carbon emission management method provided in Embodiment 1, this embodiment of the invention also provides a federated learning-based carbon emission management system, including:

[0114] The federated training module 11 is used to train several first carbon emission prediction models on several park nodes using a list of carbon emission source numbers, a list of carbon emission source start-up time, a list of carbon emissions per unit output, and a list of production time per unit output as inputs, and a preset time zone carbon emission monitoring value as the true value.

[0115] The parameter fusion module 12 is used to encrypt and upload several first model parameters of several first carbon emission prediction models to the collaborators through the several park nodes based on the first encryption key downloaded from the collaborators for optimization and fusion, and then encrypt and send the fused model parameters to the several park nodes to update the several first carbon emission prediction models to obtain several second carbon emission prediction models.

[0116] Prediction and calculation module 13 is used to process several target time zone input parameters of the park based on the several second carbon emission prediction models to obtain several first carbon emission prediction values.

[0117] The statistical analysis module 14 is used to perform carbon emission sample statistical analysis according to the number of list elements of the several park nodes, obtain several carbon emission statistical values, and sum them with the several first carbon emission prediction values ​​that correspond one-to-one to obtain several second carbon emission prediction values.

[0118] The management execution module 15 is used to perform carbon emission management based on the aforementioned second carbon emission prediction values.

[0119] Furthermore, the Federated Training Module 11 includes the following execution steps:

[0120] In the first park node of the plurality of park nodes, the data for model construction of the first park node is collected with the input of the carbon emission source quantity list, the carbon emission source start-up time list, the carbon emission per unit output list, and the carbon emission monitoring value of the preset time zone as the true value.

[0121] Obtain the preset model architecture downloaded from the collaborator;

[0122] By constructing data using the first park node model, the preset model architecture is trained to obtain the carbon emission prediction model for the first park node.

[0123] Among them, the carbon emission prediction model of the first park node belongs to the plurality of first carbon emission prediction models.

[0124] Furthermore, the Federated Training Module 11 also includes the following execution steps:

[0125] Data is constructed using the first park node model, and the preset model architecture is trained to obtain an initial carbon emission prediction model.

[0126] Extract the set of output bias vectors from the initial carbon emission prediction model whose output loss is greater than or equal to the loss threshold;

[0127] The output deviation vector set is sorted by the minimum LOF outlier factor to obtain the first output deviation calibration vector;

[0128] When the magnitude of the first output deviation calibration vector is greater than or equal to the deviation magnitude threshold, the first output deviation calibration vector is used to replace the true value of the first park node model construction data to obtain the first park node model update construction data, and the preset model architecture is trained to obtain the residual compensator.

[0129] The outputs of the residual compensator and the initial carbon emission prediction model are summed to obtain the first compensated carbon emission prediction model architecture.

[0130] Data is constructed using the first park node model, and the first compensation carbon emission prediction model architecture is trained to obtain the first compensation carbon emission prediction model.

[0131] When the magnitude of the second output deviation calibration quantity of the first compensated carbon emission prediction model is less than the deviation magnitude threshold, the first compensated carbon emission prediction model is set as the carbon emission prediction model of the first park node; otherwise, the residual compensator continues to execute the loop.

[0132] Furthermore, the Federated Training Module 11 also includes the following execution steps:

[0133] When the magnitude of the first output deviation calibration quantity is less than the deviation magnitude threshold, the initial carbon emission prediction model is set as the carbon emission prediction model of the first park node.

[0134] Furthermore, the Federated Training Module 11 includes the following execution steps:

[0135] Through the first park node, extract the first carbon emission source type and the first carbon emission source scale of the first park, up to the Nth carbon emission source type and the Nth carbon emission source scale, where N represents the total number of carbon emission sources in the first park;

[0136] Retrieve first historical carbon emission monitoring data that meets the first carbon emission source type and the first carbon emission source scale, perform central tendency assessment, and obtain carbon emission per unit output of the first carbon emission source type and production time per unit output of the first carbon emission source type;

[0137] Until the Nth historical carbon emission monitoring data that meets the first carbon emission source type and the first carbon emission source scale is obtained, a central value assessment is performed to obtain the carbon emission per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type.

[0138] Based on the carbon emissions per unit output of the first carbon emission source type and the production time per unit output of the first carbon emission source type, up to the carbon emissions per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type, the first carbon emission source type up to the Nth carbon emission source type are clustered to obtain multiple cluster carbon emission source types. Among them, any cluster of carbon emission source types has an intra-cluster quantity identifier, an average carbon emission per unit output identifier, and an average production time per unit output identifier.

[0139] Based on the cluster quantity identifier, a list of carbon emission source quantities is constructed. Based on the average start-up time of the cluster carbon emission source within a preset time zone, a list of carbon emission source start-up time is constructed. Based on the average carbon emission per unit output identifier, a list of carbon emission per unit output is constructed. Based on the average production time per unit output identifier, a list of production time per unit output is constructed. Using the carbon emission monitoring value of the preset time zone as the true value, data for constructing the first park node model is collected.

[0140] Furthermore, the parameter fusion module 12 includes the following execution steps:

[0141] The second carbon emission prediction models are validated through the aforementioned park nodes to obtain several validation loss values.

[0142] For the second carbon emission prediction model of the park node whose verification loss value is less than the verification loss threshold among the aforementioned verification loss values, the parameters are frozen and the operation is awaited.

[0143] The second carbon emission prediction model for the park node with a verification loss value greater than or equal to the verification loss threshold is cyclically trained.

[0144] Furthermore, the statistical analysis module 14 includes the following execution steps:

[0145] Extract the number of list elements in the first park, perform analysis through the first park computing power predictor, and obtain the predicted value of the computing power loss for carbon emission prediction in the first park. The first park computing power predictor is generated by machine learning training through multiple sets of data. Each set of data includes the number of list elements in the input data of the second carbon emission prediction model in the first park and a label that identifies the monitoring value of the computing power loss for carbon emission prediction.

[0146] Using the carbon emission prediction loss computing power prediction value of the first park as a constraint, the carbon emission sample set is retrieved, the central value assessment is performed, the carbon emission statistics of the first park are obtained, and added to the several carbon emission statistics.

[0147] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0148] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0153] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A carbon emission management method based on federated learning, characterized in that, include: Using the list of carbon emission sources, the list of carbon emission source start-up time, the list of carbon emissions per unit output, and the list of production time per unit output as inputs, and the carbon emission monitoring value of the preset time zone as the true value, several first carbon emission prediction models are trained on the preset model architecture downloaded from the collaborator at several park nodes. Through the aforementioned park nodes, based on the first encryption key downloaded from the collaborator, the parameters of the aforementioned first carbon emission prediction models are encrypted and uploaded to the collaborator for optimization and fusion. The resulting fused model parameters are encrypted and sent to the aforementioned park nodes to update the aforementioned first carbon emission prediction models, thereby obtaining several second carbon emission prediction models. Based on the aforementioned second carbon emission prediction models, the input parameters of several target time zones of several parks are processed respectively to obtain several first carbon emission prediction values. Based on the number of list elements of the aforementioned park nodes, carbon emission sample statistical analysis is performed to obtain several carbon emission statistical values. These values ​​are then summed with the corresponding first carbon emission prediction values ​​to obtain several second carbon emission prediction values. The number of list elements refers to the number of elements in any one of the following lists: the list of carbon emission source quantity, the list of carbon emission source start-up time, the list of carbon emissions per unit output, and the list of production time per unit output. Carbon emission management is carried out based on the aforementioned second carbon emission forecasts; Specifically, based on the number of list elements in the aforementioned park nodes, carbon emission sample statistical analysis is performed to obtain several carbon emission statistical values, including: The first list element count of the plurality of list elements of the plurality of park nodes is extracted, and the analysis is performed by the first park computing power predictor to obtain the predicted value of the computing power loss of the first park carbon emission prediction. The first park computing power predictor is generated by machine learning training through multiple sets of data. Each set of multiple sets of data includes the list element count of the input data of the second carbon emission prediction model of the first park and a label that identifies the monitoring value of the computing power loss of the carbon emission prediction. Using the carbon emission prediction loss computing power prediction value of the first park as a constraint, the carbon emission sample set is retrieved, the central value assessment is performed, the carbon emission statistics of the first park are obtained, and added to the several carbon emission statistics.

2. The method as described in claim 1, characterized in that, Using a list of carbon emission sources, a list of carbon emission source start-up times, a list of carbon emissions per unit of output, and a list of production times per unit of output as input, and using carbon emission monitoring values ​​for a preset time zone as true values, several first carbon emission prediction models are trained at several park nodes on a preset model architecture downloaded from collaborators, including: In the first park node of the plurality of park nodes, the data for model construction of the first park node is collected with the input of the carbon emission source quantity list, the carbon emission source start-up time list, the carbon emission per unit output list, and the carbon emission monitoring value of the preset time zone as the true value. Obtain the preset model architecture downloaded from the collaborator; By constructing data using the first park node model, the preset model architecture is trained to obtain the carbon emission prediction model for the first park node. Among them, the carbon emission prediction model of the first park node belongs to the plurality of first carbon emission prediction models.

3. The method as described in claim 2, characterized in that, By constructing data using the first park node model, and training the preset model architecture, a carbon emission prediction model for the first park node is obtained, including: Data is constructed using the first park node model, and the preset model architecture is trained to obtain an initial carbon emission prediction model. Extract the set of output bias vectors from the initial carbon emission prediction model whose output loss is greater than or equal to the loss threshold; The output deviation vector set is sorted by the minimum LOF outlier factor to obtain the first output deviation calibration vector; When the magnitude of the first output deviation calibration vector is greater than or equal to the deviation magnitude threshold, the first output deviation calibration vector is used to replace the true value of the first park node model construction data to obtain the first park node model update construction data, and the preset model architecture is trained to obtain the residual compensator. The outputs of the residual compensator and the initial carbon emission prediction model are summed to obtain the first compensated carbon emission prediction model architecture. Data is constructed using the first park node model, and the first compensation carbon emission prediction model architecture is trained to obtain the first compensation carbon emission prediction model. When the magnitude of the second output deviation calibration quantity of the first compensated carbon emission prediction model is less than the deviation magnitude threshold, the first compensated carbon emission prediction model is set as the carbon emission prediction model of the first park node; otherwise, the residual compensator continues to execute the loop.

4. The method as described in claim 3, characterized in that, Also includes: When the magnitude of the first output deviation calibration quantity is less than the deviation magnitude threshold, the initial carbon emission prediction model is set as the carbon emission prediction model of the first park node.

5. The method as described in claim 2, characterized in that, At the first park node of the plurality of park nodes, using the list of carbon emission source quantity, the list of carbon emission source start-up time, the list of carbon emissions per unit output, and the list of production time per unit output as input, and using the carbon emission monitoring value of a preset time zone as the true value, data for model construction of the first park node is collected, including: Through the first park node, extract the first carbon emission source type and the first carbon emission source scale of the first park, up to the Nth carbon emission source type and the Nth carbon emission source scale, where N represents the total number of carbon emission sources in the first park; Retrieve first historical carbon emission monitoring data that meets the first carbon emission source type and the first carbon emission source scale, perform central tendency assessment, and obtain carbon emission per unit output of the first carbon emission source type and production time per unit output of the first carbon emission source type; Until the Nth historical carbon emission monitoring data that meets the first carbon emission source type and the first carbon emission source scale is obtained, a central value assessment is performed to obtain the carbon emission per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type. Based on the carbon emissions per unit output of the first carbon emission source type and the production time per unit output of the first carbon emission source type, up to the carbon emissions per unit output of the Nth carbon emission source type and the production time per unit output of the Nth carbon emission source type, the first carbon emission source type up to the Nth carbon emission source type are clustered to obtain multiple cluster carbon emission source types. Among them, any cluster of carbon emission source types has an intra-cluster quantity identifier, an average carbon emission per unit output identifier, and an average production time per unit output identifier. Based on the cluster quantity identifier, a list of carbon emission source quantities is constructed. Based on the average start-up time of the cluster carbon emission source within a preset time zone, a list of carbon emission source start-up time is constructed. Based on the average carbon emission per unit output identifier, a list of carbon emission per unit output is constructed. Based on the average production time per unit output identifier, a list of production time per unit output is constructed. Using the carbon emission monitoring value of the preset time zone as the true value, data for constructing the first park node model is collected.

6. The method as described in claim 1, characterized in that, Through the aforementioned park nodes, based on the first encryption key downloaded from the collaborator, several first model parameters of the aforementioned first carbon emission prediction models are encrypted and uploaded to the collaborator for optimization and fusion. The resulting fused model parameters are then encrypted and distributed to the aforementioned park nodes to update the aforementioned first carbon emission prediction models, resulting in several second carbon emission prediction models, including: After obtaining the second carbon emission prediction model, each node in the park will conduct performance verification and evaluation: The second carbon emission prediction models are validated through the aforementioned park nodes to obtain several validation loss values. For the second carbon emission prediction model of the park node whose verification loss value is less than the verification loss threshold among the aforementioned verification loss values, the parameters are frozen and the operation is awaited. The second carbon emission prediction model for the park node with a verification loss value greater than or equal to the verification loss threshold is cyclically trained.

7. A carbon emission management system based on federated learning, characterized in that, For implementing the method as described in any one of claims 1 to 6, comprising: The federated training module is used to train several first carbon emission prediction models on several park nodes using a list of carbon emission source numbers, a list of carbon emission source start-up times, a list of carbon emissions per unit output, and a list of production times per unit output as inputs, and a preset time zone carbon emission monitoring value as the true value. The parameter fusion module is used to encrypt and upload several first model parameters of several first carbon emission prediction models to the collaborators through the several park nodes based on the first encryption key downloaded from the collaborators for optimization and fusion, and then encrypt and distribute the fused model parameters to the several park nodes to update the several first carbon emission prediction models to obtain several second carbon emission prediction models. The prediction calculation module is used to process the input parameters of several target time zones of the park based on the several second carbon emission prediction models to obtain several first carbon emission prediction values. The statistical analysis module is used to perform carbon emission sample statistical analysis according to the number of list elements of the several park nodes, obtain several carbon emission statistical values, and sum them with the corresponding several first carbon emission prediction values ​​to obtain several second carbon emission prediction values. The number of list elements refers to the number of elements in any one of the lists: the carbon emission source quantity list, the carbon emission source start-up time list, the unit output carbon emission list, and the unit output production time list. A management execution module is used to perform carbon emission management based on the aforementioned second carbon emission prediction values; Specifically, based on the number of list elements in the aforementioned park nodes, carbon emission sample statistical analysis is performed to obtain several carbon emission statistical values, including: The first list element count of the plurality of list elements of the plurality of park nodes is extracted, and the analysis is performed by the first park computing power predictor to obtain the predicted value of the computing power loss of the first park carbon emission prediction. The first park computing power predictor is generated by machine learning training through multiple sets of data. Each set of multiple sets of data includes the list element count of the input data of the second carbon emission prediction model of the first park and a label that identifies the monitoring value of the computing power loss of the carbon emission prediction. Using the carbon emission prediction loss computing power prediction value of the first park as a constraint, the carbon emission sample set is retrieved, the central value assessment is performed, the carbon emission statistics of the first park are obtained, and added to the several carbon emission statistics.

Images