A regional carbon emission prediction method and device based on multi-source data privacy protection

Abstract

This invention discloses a method and apparatus for regional carbon emission prediction based on multi-source data privacy protection, comprising: a remote sensing party and an energy party generating hash label sets for their respective samples according to the same rules and exchanging them to find their intersection; the remote sensing party extracting a feature submatrix based on the intersection, and the energy party extracting a subset of real labels based on the intersection; the remote sensing party generating candidate split features and thresholds based on the feature submatrix and sending them to the energy party; the energy party calculating the gain based on the subset of real labels and returning it; the remote sensing party selecting the optimal split based on the gain to recursively construct a decision tree; iteratively constructing multiple decision trees to form a vertical federated random forest; the remote sensing party using out-of-bag samples to predict and sending associated hash labels to the energy party; optimizing the model based on the error information returned by the energy party; and the remote sensing party extracting features from new samples and inputting them into the optimized model to obtain carbon emission prediction values. This invention can achieve fine-scale carbon emission prediction and is suitable for privacy-sensitive scenarios such as cross-domain carbon accounting and low-carbon spatial planning.

Description

Technical Field

[0001] This invention belongs to the field of carbon emission technology, specifically relating to a method and apparatus for predicting regional carbon emissions based on multi-source data privacy protection. Background Technology

[0002] Currently, carbon emission accounting has evolved from traditional macro-statistics to intelligent methods that emphasize fine scale, privacy protection, and multi-source fusion. The core technological paths are mainly divided into two categories: one is centralized machine learning modeling based on remote sensing data, which uses publicly available spatial data to depict the spatial distribution of carbon emissions, but it cannot meet the needs of energy privacy data fusion; the other is distributed modeling based on federated learning and differential privacy, which enables collaborative training under privacy protection, but it still has shortcomings in spatiotemporal alignment of multi-source data and fine scale adaptability.

[0003] For example, patent document CN120450123A provides a method and system for intelligent carbon emission prediction based on big data. It generates a carbon emission feature tensor through quantum spatiotemporal coding and spatiotemporal network alignment, constructs a dynamic causal graph network using causal entropy, and generates a causal weight matrix by combining counterfactual intervention and Bayesian falsification of causal relationships. Based on this matrix, a federated learning framework is constructed, aggregating encrypted gradients from edge nodes in the cloud and embedding a causal regularization term in the loss function to achieve cross-regional privacy-preserving collaborative training and dynamic causal intervention.

[0004] Patent document CN118627692A discloses a method and system for heterogeneous data fusion access based on a carbon intelligent management platform. Edge computing devices are deployed on the power supply, load, and energy storage sides to collect and upload multi-source heterogeneous data in real time. A random forest-based fusion algorithm is used in the cloud, combined with offline semi-supervised models and incremental semi-supervised models, to process the received multi-source heterogeneous data, accurately extracting key features affecting carbon emissions and reducing the risk of data leakage.

[0005] The patent document with publication number CN118536662A provides a personalized carbon emission prediction method and device based on federated learning. It uses a federated learning framework for collaborative modeling, protects the sensitive data of each participant from being leaked, ensures the privacy and security of each participant, and predicts carbon emissions based on the individual circumstances of each participant.

[0006] The aforementioned patent literature, starting from the scenario of multi-source data fusion and privacy protection in carbon emission prediction, has constructed methods based on federated learning, random forest, or spatiotemporal coding, which to some extent solve the problems of data isolation, multi-source data integration, or personalized prediction. However, the following shortcomings still exist: (1) Most existing federated learning schemes are horizontal federation or rely on centralized aggregation, lacking design for the scenario of vertical segmentation of features and labels, making it difficult to meet the privacy protection needs of feature data and privacy label data belonging to different holders; (2) The multi-source data fusion and alignment mechanism is insufficient. Remote sensing feature data and energy data are scattered among different entities, with heterogeneous data types and complex nonlinear mapping relationships. The data alignment process relies on strong assumptions or trusted third parties, lacking lightweight, decentralized alignment protocols that can strictly protect the original spatiotemporal privacy, resulting in low adaptation efficiency and the risk of data leakage; (3) Most methods focus on macro-regional prediction or single-point data quality improvement, with coarse accounting scales, making it difficult to characterize the spatial heterogeneity of carbon emissions at fine scales such as 1km grids or urban functional areas, and also unable to effectively handle variable coupling and cross-spatial unit interaction effects, resulting in insufficient model robustness and generalization ability.

[0007] Therefore, there is an urgent need for a carbon emission prediction method that can achieve fine-scale accounting, meet privacy protection requirements, integrate multi-source data, and adapt to complex mapping relationships. Summary of the Invention

[0008] In view of the above, the purpose of this invention is to provide a regional carbon emission prediction method and apparatus based on multi-source data privacy protection. It realizes the construction of a vertical federated random forest based on privacy-preserving alignment of hash tags, ensuring that the remote sensing party and the energy party are trained collaboratively and that the original data of both parties do not leave the domain. The prediction stage is completed by only the remote sensing party, which can realize fine-scale carbon emission prediction. It is suitable for application scenarios such as precise low-carbon spatial planning, cross-domain carbon accounting and carbon emission supervision under privacy protection.

[0009] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a regional carbon emission prediction method based on multi-source data privacy protection, comprising the following steps: The remote sensing party and the energy party generate their respective hash label sets for the samples according to the same rules. After exchanging the hash label sets, they calculate the intersection. The remote sensing party extracts the feature sub-matrix of the matching samples based on the intersection, and the energy party extracts the true label subset of the matching samples based on the intersection. The remote sensing party generates candidate splitting features and thresholds based on the feature submatrix and sends them to the energy party. The energy party calculates the gain corresponding to each candidate splitting feature and threshold based on the real label subset and returns it. The remote sensing party selects the optimal split based on the gain to recursively construct a decision tree. Multiple decision trees are iteratively constructed to form a vertical federated random forest as a carbon emission prediction model. The remote sensing party makes model predictions based on out-of-bag samples and sends the hash labels associated with the predicted values ​​to the energy party. The model is optimized based on the error information returned by the energy party. For new samples, remote sensing extracts the features of the new samples and inputs them into the optimized carbon emission prediction model to obtain carbon emission prediction values.

[0010] Preferably, the remote sensing party and the energy party generate their respective sample hash label sets according to the same rule, including: The remote sensing party and the energy party pre-agree on the same spatiotemporal grid encoding rules and the same encrypted hash function. After each party encodes the spatiotemporal coordinates of each sample locally according to the spatiotemporal grid encoding rules, they input the encrypted hash function to generate irreversible hash labels, forming a hash label set. The two parties exchange their respective hash label sets through a secure communication protocol.

[0011] Preferably, the remote sensing party generates candidate splitting features and thresholds based on the feature submatrix and sends them to the energy party, including: The remote sensing party uses Bootstrap sampling to select a preset number of candidate splitting features from all features of the feature submatrix, generates several splitting thresholds for each candidate splitting feature according to the value distribution, forms a candidate splitting set containing candidate splitting features and thresholds, and sends it to the energy party.

[0012] Preferably, the energy source calculates and returns the gain corresponding to each candidate splitting feature and threshold based on a subset of real labels, including: The energy side calculates the gain for each candidate splitting feature and threshold using the mean squared error reduction. : , in, Indicates the first One candidate splitting feature, Representation of features The A splitting threshold, This represents the mean square error function. Indicates the first The subset of true labels corresponding to the current node in the decision tree. and These represent the label subsets corresponding to the left and right child nodes after being divided according to the current candidate splitting features and thresholds. This indicates the number of samples in the set; the energy side only returns the gain value result and does not expose the true label of any individual sample.

[0013] Preferably, the remote sensing method selects the optimal split based on gain to recursively construct a decision tree, including: The remote sensing unit selects the optimal splitting scheme based on the gain values ​​returned by the energy unit, and completes the node sample partitioning: , , in, and These represent the optimal splitting feature and threshold, respectively. Indicates the first Candidate splitting features and its first A split threshold Gain, This represents the set of all candidate splitting features and thresholds. and These represent the sample index sets of the left and right child nodes, respectively. Represents the feature submatrix held by the remote sensing party. The Middle Each sample has the optimal splitting feature eigenvalues ​​on, Indicates the first The Bootstrap sample index set corresponding to each decision tree; remote sensing method according to and The decision tree structure is divided into nodes and continues to grow until the termination condition is met. The decision tree structure is stored and maintained only by the remote sensing party, and the energy party does not store any decision tree structure information.

[0014] Preferably, the remote sensing party performs model predictions based on out-of-bag samples and sends the hash tags associated with the predicted values ​​to the energy party, and optimizes the model based on the error information returned by the energy party, including: The remote sensing unit uses out-of-bag samples that are not involved in the training of the current decision tree to make predictions and sends the hash labels associated with the predicted values ​​to the energy unit. The energy unit calculates the prediction error based on the corresponding local out-of-bag true labels and returns it. The remote sensing unit adjusts the model hyperparameters or the ensemble weights of each decision tree based on the returned error information until the model converges.

[0015] Preferably, the new sample prediction stage is performed only by the remote sensing party. The remote sensing party extracts the feature vector of the new sample locally and inputs it into the optimized carbon emission prediction model. The carbon emission prediction value is obtained by combining the outputs of each decision tree. The energy party does not participate in any calculations in the prediction stage.

[0016] Secondly, the present invention provides a regional carbon emission prediction device based on multi-source data privacy protection, which is implemented using the above-mentioned regional carbon emission prediction method based on multi-source data privacy protection, including: a multi-source privacy data alignment module, a longitudinal decision tree construction module, a federated random forest optimization module, and a unilateral carbon emission prediction module. The multi-source privacy data alignment module is deployed so that the remote sensing party and the energy party generate their respective hash label sets according to the same rules. After exchanging hash label sets, the two parties calculate the intersection. The remote sensing party extracts the feature sub-matrix of the matching samples based on the intersection, and the energy party extracts the true label subset of the matching samples based on the intersection. The vertical decision tree construction module is deployed as follows: the remote sensing party generates candidate split features and thresholds based on the feature submatrix and sends them to the energy party; the energy party calculates the gain corresponding to each candidate split feature and threshold based on the real label subset and returns it; the remote sensing party selects the optimal split based on the gain to recursively construct the decision tree. The Federated Random Forest optimization module is deployed to iteratively build multiple decision trees to form a vertical Federated Random Forest as a carbon emission prediction model. The remote sensing party makes model predictions based on out-of-bag samples and sends the hash labels associated with the predicted values ​​to the energy party. The energy party calculates the error and returns it. The remote sensing party optimizes the model based on the error information. The unilateral carbon emission prediction module is deployed such that, for a new sample, the remote sensing side extracts the features of the new sample and inputs them into the optimized carbon emission prediction model to obtain the carbon emission prediction value.

[0017] Thirdly, an electronic device provided by an embodiment of the present invention includes a memory and one or more processors. The memory is used to store a computer program, and the processor is used to implement the above-mentioned regional carbon emission prediction method based on multi-source data privacy protection when executing the computer program.

[0018] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the aforementioned regional carbon emission prediction method based on multi-source data privacy protection.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) Achieve multi-source privacy data alignment without third party intervention: This invention combines spatiotemporal grid coding with cryptographic hash function to generate irreversible hash labels as sample identifiers. The remote sensing party and the energy party only exchange hash label sets for sample matching, thereby achieving accurate alignment of cross-domain samples without the need for trusted third party intervention or exposure of original spatiotemporal coordinates and business data. At the same time, it improves the adaptation efficiency and overcomes the shortcomings of existing solutions, such as complex alignment mechanisms and high privacy risks.

[0020] (2) Improve the model’s ability to handle complex mappings: The present invention adopts a vertical federated random forest architecture, in which the remote sensing party holds feature data and the energy party holds carbon emission label data. Both parties are trained together and the original data does not leave the domain, thus avoiding the compliance risk of label data leaving the domain from the design. At the same time, by taking advantage of the good adaptability of random forest to nonlinear and variable coupled data, through candidate split generation, gain calculation and tree integration, a high-precision fitting of the complex mapping relationship between multi-source features and the true value of carbon emissions is achieved, which significantly improves the robustness and generalization ability of the model.

[0021] (3) Achieving fine-scale spatial heterogeneity characterization of carbon emissions: This invention uses spatiotemporal grids as the modeling benchmark, and can map samples to fine grids or urban functional area scales through unified spatiotemporal coding rules. It maintains this spatial resolution throughout the entire process of sample alignment, feature extraction and prediction in model training, thereby overcoming the problem that most existing methods focus on macro-prediction and are difficult to support precise low-carbon spatial planning. It provides a reliable data foundation and technical support for precise low-carbon spatial planning, cross-domain carbon accounting and carbon emission supervision. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the regional carbon emission prediction method based on multi-source data privacy protection provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of the regional carbon emission prediction method based on multi-source data privacy protection provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the structure of a regional carbon emission prediction device based on multi-source data privacy protection provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0025] Carbon emission prediction typically relies on two types of key data: one is remote sensing feature data that reflects regional development intensity, spatial structure changes, and human activity levels; the other is energy consumption and the actual carbon emission values ​​calculated from it. Because remote sensing data and energy data are stored separately by different entities, there are significant differences in their business attributes, access control, and data security levels. In particular, energy consumption and carbon emission labels have strict privacy protection requirements, prohibiting cross-domain exchange. Meanwhile, there is often a complex nonlinear mapping relationship between remote sensing features and carbon emission values, and they are coupled across spatial units, making it difficult to obtain reliable results through simple linear modeling. Therefore, a joint modeling technique is needed that can integrate the advantages of both sets of data, meet privacy protection requirements, and handle high-dimensional, nonlinear features.

[0026] In view of this, the inventive concept of the present invention is as follows: Addressing the problem in existing technologies that it is difficult to simultaneously achieve fine-scale carbon emission prediction, multi-source data alignment, and privacy protection, the embodiments of the present invention provide a regional carbon emission prediction method and apparatus based on multi-source data privacy protection. The remote sensing party holds a feature matrix for carbon emission prediction, while the energy party holds the actual values ​​of energy consumption and carbon emissions, and its data must not leave the domain. Both parties generate spatiotemporal hash labels through consistent rules, and use these hash labels as unique identifiers for sample alignment. Under the condition that the data of both parties does not leave the domain and is not mutually disclosed, a longitudinal federated random forest is used to complete joint training, thereby achieving fine-scale carbon emission prediction while protecting privacy.

[0027] like Figure 1 and Figure 2 As shown in the embodiment, a method for predicting regional carbon emissions based on multi-source data privacy protection is provided, including the following steps: S1, the remote sensing party and the energy party generate their respective hash label sets for the samples according to the same rule. After exchanging the hash label sets, they calculate the intersection. The remote sensing party extracts the feature sub-matrix of the matching samples based on the intersection, and the energy party extracts the true label subset of the matching samples based on the intersection.

[0028] In this embodiment, the remote sensing and energy parties pre-agree on the same set of spatiotemporal grid coding rules and the same cryptographic hash function. Each party applies the cryptographic hash function locally to the identifier of each data sample to generate a fixed-length unique hash label. ( This process is irreversible, ensuring the privacy of the original spatiotemporal coordinates. For example, if the spatiotemporal grid encoding rule is set to "longitude accurate to 0.01°, latitude accurate to 0.01°, time accurate to the natural day", and the original spatiotemporal information of a sample is "longitude 120.12°, latitude 30.30°, time 2024-01-01", after encoding according to the agreed rule of "longitude_latitude_natural day", the identifier "120.12_30.30_20240101" is obtained. Then, the hash label "4fdec111b1c31c3ca2c431f7da1d24c9b9a9559146f0570a9989307fe3191672" is calculated by the SHA-256 cryptographic hash function. This label is only used for sample matching and cannot be used to reverse-engineer the original spatiotemporal information. Other encoding rules also need to be agreed upon, but are not limited to this example.

[0029] The two parties exchange only their complete sets of hashed labels using a secure communication protocol. Subsequently, each party locally calculates the intersection of the received label set and its own label set, determining this intersection as the set of sample identifiers used for joint training. .

[0030] Remote sensing based on intersection Extracting feature submatrices from local feature matrices Energy source based on intersection From local real tag collection Extract the corresponding real label subset Thus, accurate alignment of cross-domain samples is achieved without requiring the original data to go out of domain.

[0031] S2, the remote sensing party generates candidate splitting features and thresholds based on the feature submatrix and sends them to the energy party. The energy party calculates the gain corresponding to each candidate splitting feature and threshold based on the true label subset and returns it. The remote sensing party selects the optimal split based on the gain to recursively construct a decision tree.

[0032] In this embodiment, during the training of the longitudinal federated random forest, the construction of each tree requires collaboration between the two parties, specifically including the following sub-steps: S2.1 Tree Training Mechanism (1) Bootstrap Sample Extraction (Remote Sensing Side): The intersection of the aligned remote sensing sides Perform Bootstrap sampling to obtain the first Training sample index set of decision trees : , Scale and original intersection The sizes are consistent, and the remote sensing will Corresponding hash tag The data is sent to the energy provider, who then extracts the corresponding subset of real tags. .

[0033] (2) Node splitting candidate generation (remote sensing side): randomly select from the feature subset of the current decision node. For each of the candidate features, a split threshold is generated based on its value distribution: , in, Indicates the first One candidate splitting feature, Representation of features The A splitting threshold, For aligned eigenmatrices The complete set of features, Select feature All values ​​in the current node sample This means randomly selecting from the entire feature set. One candidate feature, Indicates the selected feature The value is determined by Quantile generation A splitting threshold, ultimately forming For a candidate split set containing candidate split features and thresholds, the remote sensing side will... Send to the energy provider.

[0034] (3) Splitting gain calculation (energy side): The splitting gain is measured by the reduction in mean square error. The method is as follows: , in, The mean square error of the label set ( (mean of labels) , According to The label subsets of the left and right child nodes after splitting. This indicates the number of samples in the set. In this process, the energy provider only provides the gain value result, and the remote sensing provider only selects the optimal split based on the gain value. The entire process does not involve any privacy-labeled data.

[0035] (4) Optimal splitting selection and node partitioning (remote sensing side): Based on the gain values ​​returned by the energy side, the optimal splitting scheme is selected, and the node sample partitioning is completed: , , in, and These represent the optimal splitting feature and threshold, respectively. and These represent the sample index sets of the left and right child nodes, respectively. Represents the feature submatrix held by the remote sensing party. The Middle Each sample has the optimal splitting feature Eigenvalues ​​on the surface. Remote sensing method. and Divide the nodes and continue to grow the tree structure.

[0036] (5) Definition of terminal node sample set (remote sensing side): sample In the After multiple rounds of splitting in the tree, the final node is a unique terminal node. The set of training samples contained in this node is determined by the Bootstrap sampling samples and the splitting rules, denoted as... For the target sample, For training samples, then the terminal node Represented as: , in, Indicates the first The number of split levels in the decision tree. and They represent the first Optimal splitting features and thresholds of layers Indicates sample In the The branch interval to which the layer splits (specifically, if) hour Conversely ), Indicates sample In the eigenmatrix of the remote sensing side The Middle The values ​​that can be taken on each feature; correspondingly, Indicates training samples In remote sensing eigenmatrix The Middle Values ​​on each feature; terminal node yes All samples The set of training samples filtered through the same splitting path.

[0037] (6) Calculation of predicted value for a single tree (remote sensing method): Decision trees for samples Predicted value The mean of the true labels of the training samples within the terminal node is used for calculation, as follows: , in, This indicates the training samples held by the energy party. The corresponding actual carbon emission labels are indirectly generated from the average of the terminal node labels, and the energy provider does not expose any individual sample labels. The remote sensing party only obtains the final prediction results and does not access the actual label data.

[0038] (7) Recursive growth tree structure (remote sensing): Growth stops when a node meets the preset termination condition: , in, Indicates the number of nodes. Indicates the depth of the decision tree. This represents the minimum number of samples for a node. Indicates the maximum depth of the decision tree. This represents the maximum gain of all candidate splits for the current node; if the termination condition is not met, repeat the above steps (2) to (6) to recursively process the child nodes. The decision tree structure is stored and maintained independently by the remote sensing party, and the energy party does not participate in node partitioning and tree growth decision.

[0039] S2.2, Forest Integration Repeat the single-tree training process in step S2.1 to construct a shared tree. A vertical federated random forest is formed from independent decision trees as a carbon emission prediction model. The results of each tree are weighted and fused during the ensemble prediction. , in, Indicates the first The original weights of the trees ( This indicates the out-of-bag error of the tree. To avoid the minimum value where the denominator is 0). Indicates the first The normalized weights of the trees, Indicates the first The predicted value of the trees, This represents the ensemble prediction results of the longitudinal federated random forest.

[0040] S3. Iteratively construct multiple decision trees to form a vertical federated random forest as a carbon emission prediction model. The remote sensing party makes model predictions based on out-of-bag samples and sends the hash labels associated with the predicted values ​​to the energy party. The model is optimized based on the error information returned by the energy party.

[0041] In the embodiment, the remote sensing party uses samples outside the bag (not involved in the first step) to... The remote sensing unit (using training samples from 100 trees) performs model predictions and sends the hash labels associated with the predicted values ​​to the energy unit. The energy unit calculates the error based on the local out-of-bag labels and returns the error information. The remote sensing unit then adjusts the model hyperparameters or forest size accordingly, completing iterative optimization until the model converges.

[0042] S4. For new samples, the remote sensing side extracts the features of the new samples and inputs them into the optimized carbon emission prediction model to obtain the carbon emission prediction value.

[0043] In this embodiment, for a newly input prediction sample, the remote sensing data provider extracts a feature vector locally. The data is then fed into a pre-trained federated random forest model, and the predicted carbon emissions are obtained by averaging the outputs of each tree. Energy data providers do not need to participate in the forecasting phase.

[0044] In summary, the regional carbon emission prediction method based on multi-source data privacy protection provided by the embodiments of the present invention, through a vertical federated random forest structure, keeps the feature data of the remote sensing party and the label data of the energy party locally in the feature-label separation mode, performs sample alignment through a unified spatiotemporal label hash value, and completes the construction of the carbon emission prediction model without disclosing the original data of either party, achieving high-precision, scalable and privacy-friendly modeling effect.

[0045] Based on the same inventive concept, such as Figure 3 As shown, this embodiment of the invention also provides a regional carbon emission prediction device 300 based on multi-source data privacy protection, including: a multi-source privacy data alignment module 310, a vertical decision tree construction module 320, a federated random forest optimization module 330, and a unilateral carbon emission prediction module 340.

[0046] The multi-source privacy data alignment module 310 is deployed to generate hash label sets for the remote sensing party and the energy party respectively according to the same rules. After exchanging hash label sets, the two parties calculate the intersection. The remote sensing party extracts the feature sub-matrix of the matching samples based on the intersection, and the energy party extracts the true label subset of the matching samples based on the intersection.

[0047] The vertical decision tree construction module 320 is deployed as a remote sensing party to generate candidate split features and thresholds based on the feature submatrix and send them to the energy party. The energy party calculates the gain corresponding to each candidate split feature and threshold based on the true label subset and returns it. The remote sensing party selects the optimal split based on the gain to recursively construct the decision tree.

[0048] The Federated Random Forest Optimization Module 330 is deployed to iteratively build multiple decision trees to form a vertical Federated Random Forest as a carbon emission prediction model. The remote sensing party makes model predictions based on out-of-bag samples and sends the hash labels associated with the predicted values ​​to the energy party. The energy party calculates the error and returns it. The remote sensing party optimizes the model based on the error information.

[0049] The unilateral carbon emission prediction module 340 is deployed so that for new samples, the remote sensing side extracts the features of the new samples and inputs them into the optimized carbon emission prediction model to obtain the carbon emission prediction value.

[0050] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described regional carbon emission prediction method based on multi-source data privacy protection when executing the computer program.

[0051] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the aforementioned regional carbon emission prediction method based on multi-source data privacy protection.

[0052] It should be noted that the regional carbon emission prediction device, electronic device and computer-readable storage medium based on multi-source data privacy protection provided in the above embodiments all belong to the same inventive concept as the regional carbon emission prediction method based on multi-source data privacy protection. For details of their specific implementation process, please refer to the embodiments of the regional carbon emission prediction method based on multi-source data privacy protection, which will not be repeated here.

[0053] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A regional carbon emission prediction method based on multi-source data privacy protection, characterized in that, Includes the following steps: The remote sensing party and the energy party generate their respective hash label sets for the samples according to the same rules. After exchanging the hash label sets, they calculate the intersection. The remote sensing party extracts the feature sub-matrix of the matching samples based on the intersection, and the energy party extracts the true label subset of the matching samples based on the intersection. The remote sensing party generates candidate splitting features and thresholds based on the feature submatrix and sends them to the energy party. The energy party calculates the gain corresponding to each candidate splitting feature and threshold based on the real label subset and returns it. The remote sensing party selects the optimal split based on the gain to recursively construct a decision tree. Multiple decision trees are iteratively constructed to form a vertical federated random forest as a carbon emission prediction model. The remote sensing party makes model predictions based on out-of-bag samples and sends the hash labels associated with the predicted values ​​to the energy party. The model is optimized based on the error information returned by the energy party. For new samples, remote sensing extracts the features of the new samples and inputs them into the optimized carbon emission prediction model to obtain carbon emission prediction values.

2. The regional carbon emission prediction method based on multi-source data privacy protection according to claim 1, characterized in that, The remote sensing party and the energy party generate their respective hash label sets for the samples according to the same rule, including: The remote sensing party and the energy party pre-agree on the same spatiotemporal grid encoding rules and the same encrypted hash function. After each party encodes the spatiotemporal coordinates of each sample locally according to the spatiotemporal grid encoding rules, they input the encrypted hash function to generate irreversible hash labels, forming a hash label set. The two parties exchange their respective hash label sets through a secure communication protocol.

3. The regional carbon emission prediction method based on multi-source data privacy protection according to claim 1, characterized in that, The remote sensing party generates candidate splitting features and thresholds based on the feature submatrix and sends them to the energy party, including: The remote sensing party uses Bootstrap sampling to select a preset number of candidate splitting features from all features of the feature submatrix, generates several splitting thresholds for each candidate splitting feature according to the value distribution, forms a candidate splitting set containing candidate splitting features and thresholds, and sends it to the energy party.

4. The regional carbon emission prediction method based on multi-source data privacy protection according to claim 1, characterized in that, The energy source calculates and returns the gain corresponding to each candidate splitting feature and threshold based on a subset of real labels, including: The energy side calculates the gain for each candidate splitting feature and threshold using the mean squared error reduction. : ， in, Indicates the first One candidate splitting feature, Representation of features The A splitting threshold, This represents the mean square error function. Indicates the first The subset of true labels corresponding to the current node in the decision tree. and These represent the label subsets corresponding to the left and right child nodes after being divided according to the current candidate splitting features and thresholds. This indicates the number of samples in the set; the energy side only returns the gain value result and does not expose the true label of any individual sample.

5. The regional carbon emission prediction method based on multi-source data privacy protection according to claim 1 or 4, characterized in that, The remote sensing method selects the optimal split based on gain to recursively construct a decision tree, including: The remote sensing unit selects the optimal splitting scheme based on the gain values ​​returned by the energy unit, and completes the node sample partitioning: ， ， in, and These represent the optimal splitting feature and threshold, respectively. Indicates the first Candidate splitting features and its first A split threshold Gain, This represents the set of all candidate splitting features and thresholds. and These represent the sample index sets of the left and right child nodes, respectively. Represents the feature submatrix held by the remote sensing party. The Middle Each sample has the optimal splitting feature eigenvalues ​​on, Indicates the first The Bootstrap sample index set corresponding to each decision tree; remote sensing method according to and The decision tree structure is divided into nodes and continues to grow until the termination condition is met. The decision tree structure is stored and maintained only by the remote sensing party, and the energy party does not store any decision tree structure information.

6. The regional carbon emission prediction method based on multi-source data privacy protection according to claim 1, characterized in that, The remote sensing party performs model predictions based on out-of-bag samples and sends the hash tags associated with the predicted values ​​to the energy party. It then optimizes the model based on the error information returned by the energy party, including: The remote sensing unit uses out-of-bag samples that are not involved in the training of the current decision tree to make predictions and sends the hash labels associated with the predicted values ​​to the energy unit. The energy unit calculates the prediction error based on the corresponding local out-of-bag true labels and returns it. The remote sensing unit adjusts the model hyperparameters or the ensemble weights of each decision tree based on the returned error information until the model converges.

7. The regional carbon emission prediction method based on multi-source data privacy protection according to claim 1, characterized in that, The new sample prediction phase is performed solely by the remote sensing party. The remote sensing party extracts the feature vectors of the new samples locally and inputs them into the optimized carbon emission prediction model. The carbon emission prediction value is obtained by combining the outputs of each decision tree. The energy party does not participate in any calculations during the prediction phase.

8. A regional carbon emission prediction device based on multi-source data privacy protection, implemented using the regional carbon emission prediction method based on multi-source data privacy protection as described in any one of claims 1 to 7, characterized in that, include: Multi-source privacy data alignment module, vertical decision tree construction module, federated random forest optimization module, and unilateral carbon emission prediction module; The multi-source privacy data alignment module is deployed so that the remote sensing party and the energy party generate their respective hash label sets according to the same rules. After exchanging hash label sets, the two parties calculate the intersection. The remote sensing party extracts the feature sub-matrix of the matching samples based on the intersection, and the energy party extracts the true label subset of the matching samples based on the intersection. The vertical decision tree construction module is deployed as follows: the remote sensing party generates candidate split features and thresholds based on the feature submatrix and sends them to the energy party; the energy party calculates the gain corresponding to each candidate split feature and threshold based on the real label subset and returns it; the remote sensing party selects the optimal split based on the gain to recursively construct the decision tree. The Federated Random Forest optimization module is deployed to iteratively build multiple decision trees to form a vertical Federated Random Forest as a carbon emission prediction model. The remote sensing party makes model predictions based on out-of-bag samples and sends the hash labels associated with the predicted values ​​to the energy party. The energy party calculates the error and returns it. The remote sensing party optimizes the model based on the error information. The unilateral carbon emission prediction module is deployed such that, for a new sample, the remote sensing side extracts the features of the new sample and inputs them into the optimized carbon emission prediction model to obtain the carbon emission prediction value.

9. An electronic device comprising a memory and one or more processors, the memory for storing a computer program, characterized in that, The processor is used to implement the regional carbon emission prediction method based on multi-source data privacy protection as described in any one of claims 1 to 7 when executing a computer program.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a computer, it implements the regional carbon emission prediction method based on multi-source data privacy protection as described in any one of claims 1 to 7.

Images