Federal learning-based cross-regional carbon market risk early warning and price prediction method

Abstract

The application discloses a cross-regional carbon market risk early warning and price prediction method based on federal learning, wherein the method comprises the following steps: a central server initializes a global carbon market prediction and early warning model and sends the model to a plurality of participating nodes; each participating node trains the model by using local carbon market related data to obtain a local model parameter update amount; the parameter update amount is encrypted and uploaded to the central server; the central server performs aggregation operation on the encrypted parameters uploaded by the plurality of participating nodes to update the global model; the training and aggregation process is repeated until the model meets a preset convergence condition; market characteristic data of a to-be-predicted period is input into the converged global model to obtain a carbon price prediction result, a risk index is calculated according to the prediction result, and risk early warning information of a corresponding grade is generated. The application can realize accurate prediction of carbon market price fluctuation and early warning of systematic risk under the premise that the original data of each participant is not disclosed.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a method for risk early warning and price prediction in cross-regional carbon markets based on federated learning. Background Technology

[0002] With the advancement of global carbon neutrality goals, carbon trading markets have become an important economic tool for regulating carbon emissions. However, current carbon market price forecasting and risk warning face two core challenges:

[0003] Data silos and privacy barriers: Carbon market data involves energy structures, core industrial production indicators, and corporate emissions data across provinces and cities, making it highly sensitive. Due to data security and trade secret protection, regional regulatory agencies or trading platforms find it difficult to directly share raw data. This prevents traditional centralized forecasting models from obtaining complete cross-regional macroeconomic data, limiting forecast accuracy.

[0004] Limitations of single-region models: Existing carbon price prediction models are typically trained on historical data from only a single region, neglecting economic linkages, energy complementarity, and policy transmission effects between regions. For example, adjustments to energy policies in region A may affect carbon allowance demand in region B through the supply chain. Single models struggle to capture such complex cross-regional dynamics, leading to delayed or ineffective risk warnings. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] The main objective of this invention is to provide a method for risk early warning and price prediction in cross-regional carbon markets based on federated learning.

[0007] The second objective of this invention is to provide an electronic device.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a method for risk early warning and price prediction in cross-regional carbon markets based on federated learning, comprising: S1. The central server initializes the global carbon market forecasting and early warning model and sends the global carbon market forecasting and early warning model to multiple participating nodes; S2. Each participating node uses locally stored carbon market-related data to train the global carbon market prediction and early warning model to obtain local model parameter updates, wherein the carbon market-related data includes at least historical carbon trading data. S3. Each participating node encrypts the local model parameter update and then uploads it to the central server. S4. The central server performs aggregation calculations on the encrypted parameters uploaded by multiple participating nodes to obtain updated global model parameters, and updates the global carbon market prediction and early warning model accordingly. S5. Repeat steps S2 to S4 until the global carbon market prediction and early warning model meets the preset convergence condition. S6. Input the market characteristic data of the period to be predicted into the converged global carbon market prediction and early warning model to obtain the carbon price prediction result; calculate the risk index based on the carbon price prediction result, and generate the corresponding level of risk early warning information based on the risk index.

[0009] In one embodiment of the present invention, in step S2, the carbon market-related data further includes one or more of electricity load data, macroeconomic indicator data, and meteorological data.

[0010] In one embodiment of the present invention, in step S2, each participating node performs standardization processing on local data before model training and extracts feature variables for model training.

[0011] In one embodiment of the present invention, in step S3, the encryption process employs a homomorphic encryption algorithm or a differential privacy mechanism.

[0012] In one embodiment of the present invention, the homomorphic encryption algorithm is the Paillier homomorphic encryption algorithm.

[0013] In one embodiment of the present invention, in step S4, the central server aggregates the parameters uploaded by each participating node using a weighted aggregation method, wherein the aggregation weight is determined based on the amount of data corresponding to the participating node.

[0014] In one embodiment of the present invention, the global carbon market prediction and early warning model in step S1 is a hybrid neural network model combining a long short-term memory network (LSTM) and a Transformer.

[0015] In one embodiment of the present invention, in step S6, the risk indicators include at least the predicted price volatility and the tail risk indicator, and the risk warning information includes warning information for different risk levels.

[0016] In one embodiment of the present invention, a green warning, a yellow warning, or a red warning is generated based on the comparison result between the risk indicator and a preset risk threshold.

[0017] To achieve the above objectives, a second aspect of this application provides an electronic device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing the method described in the first aspect embodiment.

[0018] The embodiments of the present invention have the following beneficial effects: (1) Achieve cross-regional collaborative utilization of data. Without sharing the original business data, the market rules and regional correlation characteristics contained in the carbon market data of multiple regions are integrated through the federated learning mechanism, which breaks through the information limitations of traditional single-region modeling and improves the model's ability to represent complex market changes.

[0019] (2) Improve the accuracy of carbon price forecasting. By aggregating historical carbon trading data and related influencing factors from different regions, the model can learn cross-regional price transmission relationships and market linkage patterns, thereby improving the accuracy and stability of carbon price forecasting results.

[0020] (3) Enhance risk warning capabilities. By using the trained global model to predict future carbon price trends and combining it with price volatility and tail risk indicators for risk assessment, potential abnormal market fluctuation risks can be identified in advance, improving the timeliness and effectiveness of risk warnings.

[0021] (4) Ensure data security and privacy. Each participating node only uploads the updated model parameters after encryption or privacy protection processing. The original data is always kept locally, which effectively reduces the risk of data leakage and improves the security and feasibility of the system.

[0022] (5) Adapting to non-independent and identically distributed data scenarios. In view of the differences in the distribution of carbon market data in different regions, the model achieves collaborative optimization through federated training and global parameter aggregation mechanism, thereby improving the generalization ability and robustness of the model in complex cross-regional scenarios. Attached Figure Description

[0023] The above-described and additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart illustrating a cross-regional carbon market risk warning and price prediction method based on federated learning, provided as an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0026] This embodiment provides a cross-regional carbon market risk early warning and price prediction method based on federated learning. It can construct a high-performance global prediction model by combining the characteristics of multiple data sources without disclosing the original data of each participant, thereby achieving accurate prediction of carbon market price fluctuations and early warning of systemic risks. Figure 1 As shown, the method includes the following steps: S1. The central server initializes the global carbon market forecasting and early warning model and sends the global carbon market forecasting and early warning model to multiple participating nodes.

[0027] In this embodiment, the central server first constructs a global carbon market forecasting and early warning model for carbon market price prediction and risk warning. The global carbon market forecasting and early warning model adopts a hybrid neural network structure combining Long Short-Term Memory (LSTM) and Transformer. LSTM is used to extract long-term temporal dependencies in carbon market time series data, while Transformer is used to mine global correlation features between multi-source heterogeneous features to improve the model's ability to represent complex market fluctuation patterns.

[0028] Specifically, the central server initializes the model according to pre-defined network structure parameters, including the number of model layers, hidden layer dimensions, number of attention heads, learning rate, loss function, and training epochs, and randomly generates the initial model weight parameters W0. Subsequently, the central server establishes a federated learning communication network and performs identity authentication and node registration for multiple participating nodes in the federated training to form a cross-regional collaborative modeling network.

[0029] Participating nodes can be carbon trading centers, carbon market regulatory agencies, or data management nodes with independent data storage capabilities in different regions. Each participating node stores carbon market-related data within its corresponding region and has local model training capabilities. Due to differences in industrial structure, energy consumption structure, and market activity levels across different regions, the data held by each participating node typically exhibits non-independent and identically distributed (Non-IID) characteristics. Therefore, a federated learning mechanism is needed to achieve cross-regional collaborative modeling.

[0030] After node registration is complete, the central server sends the initialized global carbon market prediction and early warning model and its corresponding training configuration parameters to each participating node. These training configuration parameters include the number of local training rounds, batch size, optimizer parameters, and model update cycle. Upon receiving the global model, participating nodes deploy the model in their local environment and prepare for subsequent local training.

[0031] Furthermore, to ensure parameter compatibility during subsequent model aggregation, the central server uniformly distributes model structure definition files to all participating nodes, enabling each participating node to train using the same network structure and parameter dimensions, thereby ensuring that the updated model parameters uploaded subsequently can be uniformly aggregated and calculated.

[0032] As a preferred implementation, the central server also establishes a global verification mechanism, which presets model convergence evaluation indicators during the model initialization stage. The evaluation indicators include one or more of the following: prediction error, loss function value, or model accuracy indicator, and are used to determine whether the global model has reached the preset convergence condition during subsequent federated training.

[0033] Through the above methods, the central server completes the initialization of the global carbon market prediction and early warning model and the construction of the collaborative training environment for participating nodes, providing a foundation for each participating node to carry out local model training and collaborative parameter updates in the future.

[0034] S2. Each participating node uses locally stored carbon market-related data to train the global carbon market prediction and early warning model to obtain local model parameter updates, wherein the carbon market-related data includes at least historical carbon trading data.

[0035] After receiving the global carbon market forecasting and early warning model from the central server, each participating node trains the model locally using locally stored carbon market-related data. This carbon market-related data includes at least historical carbon trading data, which may include historical carbon prices, trading volume, trading value, allowance supply, allowance trading frequency, and other data reflecting the operational status of the regional carbon market.

[0036] Furthermore, the carbon market-related data may also include one or more of the following: electricity load data, macroeconomic indicator data, and meteorological data. Electricity load data reflects regional energy consumption intensity, macroeconomic indicator data reflects regional economic performance, and meteorological data reflects the impact of external environmental factors such as temperature, precipitation, and wind speed on energy demand and carbon market price fluctuations. By introducing these multi-source data, the local training process can simultaneously consider the combined impact of market trading factors, energy demand factors, economic factors, and environmental factors on carbon price changes.

[0037] Before local training, each participating node preprocesses the carbon market-related data stored locally. This preprocessing includes missing value handling, outlier removal, time alignment, data standardization, and feature extraction. Specifically, each participating node can perform time alignment on data from different sources at a daily, weekly, or monthly granularity, and normalize or standardize numerical data to reduce the impact of different data units on the model training results.

[0038] After data preprocessing, each participating node constructs training samples based on the preprocessed local data. Each training sample includes an input feature sequence and a corresponding target output value. The input feature sequence includes carbon market-related features within a preset time window prior to the current moment, and the target output value is the carbon price or carbon price trend within a preset future time period. Thus, the model can learn the mapping relationship between historical market conditions and future price changes.

[0039] During local training, each participating node does not transmit raw carbon market-related data to the central server or other participating nodes. Instead, it independently completes model forward propagation, loss calculation, and backpropagation within its local computing environment. Specifically, participating nodes input their local training samples into the received global carbon market prediction and early warning model to obtain local prediction results. They then calculate a loss function value based on the difference between the local prediction results and the actual target output value. Subsequently, participating nodes update the model parameters based on the loss function value, obtaining the locally trained model parameters or the parameter update amount relative to the global model parameters.

[0040] In some implementations, the local model parameter update can be the gradient update or the difference between the locally trained weight parameters and the initial global model weight parameters. To facilitate subsequent aggregation, each participating node generates its local model parameter update according to the model structure and parameter dimensions uniformly distributed by the central server, enabling the parameter update amounts uploaded by different participating nodes to be aggregated on the central server side.

[0041] Through the aforementioned local training process, each participating node can transform the operational patterns of the local carbon market into model parameter update information without leaking the original data, thus providing a foundation for subsequent cross-regional encrypted parameter aggregation.

[0042] S3. Each participating node encrypts the local model parameter update and then uploads it to the central server.

[0043] After obtaining the local model parameter update, each participating node encrypts the local model parameter update and uploads the encrypted parameter update to the central server. This encryption prevents the central server or other participating nodes from deducing the participating node's original local data from the uploaded content, thereby ensuring the security of local data during the federated training process.

[0044] Specifically, each participating node can use a homomorphic encryption algorithm to encrypt its local model parameter updates, enabling the central server to perform aggregation calculations on the encrypted parameters uploaded by multiple participating nodes without decrypting the parameter updates of individual participating nodes. In one implementation, the homomorphic encryption algorithm is the Paillier homomorphic encryption algorithm. Each participating node uses a pre-configured public key to encrypt each parameter component in its local model parameter update, obtaining an encrypted parameter update, and then sends the encrypted parameter update to the central server.

[0045] In another implementation, participating nodes can also employ a differential privacy mechanism to protect the privacy of their local model parameter updates. Specifically, before uploading, participating nodes prune their local model parameter updates and add random noise that meets a preset privacy budget to the pruned parameter updates to reduce the identifiability of the parameter update results from a single training sample or locally sensitive data.

[0046] Furthermore, when uploading the encrypted parameter update, each participating node can also upload auxiliary information for aggregation calculation. This auxiliary information includes one or more of the following: the data volume of the participating node, the number of local training rounds, or the parameter version identifier. This auxiliary information does not include the original local business data and is used by the central server to determine the aggregation weight of the corresponding participating node and to verify whether the parameter update matches the current global model version during subsequent aggregation processes.

[0047] To ensure the reliability of the upload process, each participating node can append its identity identifier, timestamp, and integrity verification information to the encrypted parameter update. Upon receiving the encrypted parameter update, the central server confirms the upload source based on the node identity identifier, determines whether the update corresponds to the current training round based on the parameter version identifier, and checks for missing or tampered data based on the integrity verification information.

[0048] Through the encrypted upload process described above, each participating node only transmits the updated model parameters to the central server after encryption or privacy protection, without uploading local raw data such as historical carbon trading data, electricity load data, macroeconomic indicator data, and meteorological data. This reduces the risk of data leakage while achieving cross-regional collaborative modeling.

[0049] S4. The central server performs aggregation calculations on the encrypted parameters uploaded by multiple participating nodes to obtain updated global model parameters, and updates the global carbon market prediction and early warning model accordingly.

[0050] After receiving the encrypted parameter updates uploaded by multiple participating nodes, the central server performs an aggregation operation on these updates to obtain the global parameter updates used to update the global carbon market prediction and early warning model. During the aggregation process, the central server does not obtain the original carbon market-related data from each participating node, nor does it need to access the local training samples of each participating node.

[0051] Specifically, the central server first verifies the validity of the encrypted parameter updates uploaded by each participating node, determining whether they belong to the current training round, whether they are consistent with the current global model parameter version, and whether the parameter dimensions meet the requirements of a unified model structure. For parameter updates that fail the verification, the central server does not participate in the current round of aggregation to avoid abnormal parameters affecting the global model update results.

[0052] In one implementation, the central server uses a weighted aggregation method to aggregate the parameter updates uploaded by multiple participating nodes. The weights of the weighted aggregation are determined based on the amount of data corresponding to each participating node; that is, participating nodes with larger amounts of local training data receive higher weights during the aggregation process, while participating nodes with smaller amounts of local training data receive lower weights. This method can reduce the impact of differences in data scale among different participating nodes on the stability of model training.

[0053] Specifically, let ΔWi be the local model parameter update amount uploaded by the i-th participating node in the t-th training round, and let αi be the corresponding aggregation weight. Then, the central server calculates the global parameter update amount ΔW based on the encrypted parameter update amounts of each participating node. After aggregation, the central server obtains the updated global model parameters Wt+1 based on the current global model parameters Wt and the global parameter update amount ΔW, and updates the global carbon market prediction and early warning model accordingly.

[0054] When participating nodes upload parameter updates using homomorphic encryption, the central server can perform additive or weighted aggregation operations on the encrypted parameter updates in encrypted form to obtain the global parameter update result. Subsequently, a secure decryption module with decryption privileges can decrypt the aggregated result to obtain the global parameter update. Because the decryption object is the aggregated overall parameter result, rather than the parameter update of a single participating node, the risk of recovering the training information of a single participating node is reduced.

[0055] When participating nodes upload parameter updates using a differential privacy mechanism, the central server directly aggregates the privacy-protected parameter updates and updates the global model parameters based on the aggregation result. Because participating nodes have already pruned and perturbed the parameter updates before uploading, the parameter updates obtained by the central server are difficult to correspond to specific training samples or sensitive local data.

[0056] Furthermore, after completing this round of aggregation and global model update, the central server generates a new global model version identifier and sends the updated global carbon market forecasting and early warning model to each participating node, allowing them to continue training their local models in the next round of training. Thus, the local training results of multiple participating nodes can be continuously integrated into the global model without sharing the original data, enabling the global carbon market forecasting and early warning model to gradually learn the correlation characteristics and price fluctuation patterns between different regional carbon markets.

[0057] S5. Repeat steps S2 to S4 until the global carbon market prediction and early warning model meets the preset convergence condition.

[0058] After the central server completes a global model parameter update, it sends the updated global carbon market prediction and early warning model back to each participating node. Each participating node continues local training based on the updated global model, obtains new local model parameter updates, and uploads these updates to the central server using the aforementioned encryption method. The central server then aggregates the encrypted parameters uploaded by multiple participating nodes and updates the global model parameters based on the aggregation result.

[0059] The above process is executed cyclically according to a preset training round, enabling the global carbon market prediction and early warning model to continuously integrate the local training results of participating nodes in different regions. In each round of federated training, each participating node only uses its own stored carbon market-related data to complete model training locally, and the central server only receives model parameter updates after encryption or privacy protection processing, thus keeping the original data within the domain during iterative training.

[0060] In some implementations, the preset convergence condition may include one or more of the following: the loss function value of the global model on the validation set is less than a preset loss threshold, the decrease in the loss function over several consecutive training rounds is less than a preset change threshold, the prediction error is less than a preset error threshold, or the number of federated training rounds reaches a preset maximum number of rounds.

[0061] Specifically, after each round of aggregation and update, the central server can use preset verification data to verify the updated global carbon market prediction and early warning model, obtaining model prediction error, loss function value, or other evaluation indicators. When the evaluation indicators meet preset convergence conditions, the central server stops federated training and determines the current global carbon market prediction and early warning model as the converged global model.

[0062] As one implementation method, if the change in the global model's loss function value is less than a preset threshold during multiple training rounds, the model training is determined to have stabilized; or, when the training rounds reach the preset maximum number of rounds, even if the loss function still changes slightly, iteration can be stopped to avoid overtraining and wasting communication resources.

[0063] By repeatedly executing local training, encrypted upload, and parameter aggregation processes, the global carbon market prediction and early warning model can learn the common changing patterns and inter-regional transmission characteristics of carbon markets in multiple regions without centrally storing the original data of each region, thereby improving the reliability of subsequent carbon price prediction and risk early warning.

[0064] S6. Input the market characteristic data of the period to be predicted into the converged global carbon market prediction and early warning model to obtain the carbon price prediction result; calculate the risk index based on the carbon price prediction result, and generate the corresponding level of risk early warning information based on the risk index.

[0065] After the global carbon market forecasting and early warning model meets the preset convergence conditions, the central server or designated forecasting nodes will use the converged global carbon market forecasting and early warning model for carbon price forecasting and risk warning. Specifically, market characteristic data corresponding to the period to be predicted is obtained, and the market characteristic data is input into the converged global carbon market forecasting and early warning model to output the carbon price forecast results for the future preset time period.

[0066] The market characteristic data for the period to be predicted may include one or more of the following: historical carbon prices, trading volume, electricity load, macroeconomic indicators, and meteorological data within the time window preceding the prediction date. The preset future time period can be set to daily, weekly, or monthly according to actual application needs. By adopting the same feature structure and data processing methods as the model training phase, a good match between the input data and model parameters can be ensured, thereby obtaining stable carbon price prediction results.

[0067] After obtaining the carbon price forecast results, risk indicators are further calculated based on these forecasts. These risk indicators may include predicted price volatility and tail risk indicators. The predicted price volatility characterizes the magnitude of carbon price changes over a predetermined future time period, while the tail risk indicator characterizes the potential downside or upside risk to carbon prices under extreme volatility conditions.

[0068] Specifically, a predicted return series can be calculated based on the predicted carbon price values ​​at multiple consecutive prediction points, and the predicted price volatility can be calculated based on the predicted return series. When the predicted price volatility exceeds a preset volatility threshold, it indicates that the future carbon price may experience significant fluctuations. Furthermore, a tail risk indicator can be calculated based on the predicted return distribution to identify the risk of abnormal price fluctuations that may occur in the carbon market under extreme circumstances.

[0069] After calculating the risk indicators, they are compared with preset risk thresholds, and corresponding risk warning information is generated based on the comparison results. As one implementation method, a green warning is generated when the risk indicator is below the first risk threshold, indicating that the carbon market is in a relatively stable state; a yellow warning is generated when the risk indicator reaches the first risk threshold but is below the second risk threshold, indicating that the carbon market faces price volatility risk; and a red warning is generated when the risk indicator reaches or exceeds the second risk threshold, indicating that the carbon market faces high systemic risk and requires close monitoring by relevant participating nodes.

[0070] Furthermore, the central server can send the carbon price forecast results and risk warning information to each participating node, enabling carbon trading centers or regulatory nodes in different regions to obtain cross-regional collaborative forecast results in a timely manner. The risk warning information may include the warning level, triggering risk indicators, predicted price range, warning time, and corresponding risk descriptions, so that participating nodes can take market monitoring, trading alerts, or regulatory intervention measures based on the warning results.

[0071] Through the above methods, the converged global carbon market forecasting and early warning model can not only output carbon price forecasts for a preset time period, but also identify potential abnormal price fluctuation risks based on the forecast results, thus achieving early warning of cross-regional carbon market risks.

[0072] To implement the methods of the above embodiments, the present invention also provides an electronic device, which includes a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the various steps of the methods described above.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0074] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0075] Furthermore, 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 technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for risk early warning and price prediction in cross-regional carbon markets based on federated learning, characterized in that, Includes the following steps: S1. The central server initializes the global carbon market forecasting and early warning model and sends the global carbon market forecasting and early warning model to multiple participating nodes; S2. Each participating node uses locally stored carbon market-related data to train the global carbon market prediction and early warning model to obtain local model parameter updates, wherein the carbon market-related data includes at least historical carbon trading data. S3. Each participating node encrypts the local model parameter update and then uploads it to the central server. S4. The central server performs aggregation calculations on the encrypted parameters uploaded by multiple participating nodes to obtain updated global model parameters, and updates the global carbon market prediction and early warning model accordingly. S5. Repeat steps S2 to S4 until the global carbon market prediction and early warning model meets the preset convergence condition. S6. Input the market characteristic data of the period to be predicted into the converged global carbon market prediction and early warning model to obtain the carbon price prediction result; calculate the risk index based on the carbon price prediction result, and generate the corresponding level of risk early warning information based on the risk index.

2. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 1, characterized in that, In step S2, the carbon market-related data also includes one or more of the following: electricity load data, macroeconomic indicator data, and meteorological data.

3. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 1, characterized in that, In step S2, each participating node standardizes its local data and extracts feature variables for model training before model training.

4. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 1, characterized in that, In step S3, the encryption process employs a homomorphic encryption algorithm or a differential privacy mechanism.

5. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 4, characterized in that, The homomorphic encryption algorithm is the Paillier homomorphic encryption algorithm.

6. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 1, characterized in that, In step S4, the central server uses a weighted aggregation method to aggregate the parameters uploaded by each participating node, where the aggregation weight is determined based on the amount of data corresponding to the participating node.

7. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 1, characterized in that, The global carbon market prediction and early warning model in step S1 is a hybrid neural network model that combines a long short-term memory network (LSTM) with a Transformer.

8. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 1, characterized in that, In step S6, the risk indicators include at least the predicted price volatility and the tail risk indicator, and the risk warning information includes warning information for different risk levels.

9. The method for risk early warning and price prediction of cross-regional carbon markets based on federated learning according to claim 8, characterized in that, A green, yellow, or red alert is generated based on the comparison between the risk indicators and the preset risk thresholds.

10. An electronic device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-9.

Images