A cross-regional power spot joint bidding system and method thereof

Abstract

The application discloses a cross-regional power spot joint bidding system and a method thereof. The system comprises: a local client deployed at each participating subject, which is used for privacy preprocessing of original operation data, training of a local price prediction sub-model, and splitting of model gradient parameters into private shares; and a secure multi-party computation (MPC) computing node cluster, which is used for receiving and aggregating each private share, restoring global gradient parameters, and iteratively updating until the model converges. The system also calculates a joint bidding curve through an MPC protocol. The application combines federated learning and secure multi-party computation, realizes cross-regional multi-subject collaborative modeling and bidding decision-making under the premise of protecting data privacy, significantly improves the price prediction accuracy and alliance revenue, can resist gradient leakage and collusion attacks, and meets the privacy protection and efficient transaction needs of the national unified power market.

Description

Technical Field

[0001] This invention relates to the fields of electricity market trading technology, federated learning technology, and secure multi-party computation technology, specifically to a cross-regional electricity spot joint bidding system and method applicable to a nationally unified electricity market system. Background Technology

[0002] With the full establishment of a unified national electricity market system, cross-provincial and cross-regional electricity spot trading has become a core channel for optimizing the allocation of electricity resources. To improve bidding revenue, market participants such as virtual power plants, independent power generators, and load aggregators urgently need to train high-precision cross-regional electricity price prediction models based on multi-dimensional data from the entire market. However, electricity market data is highly commercially sensitive: the cost functions of each participant, the upper limit of unit output, the status of energy storage inventory, and historical bidding strategies are all core commercial secrets. Direct sharing can lead to serious losses and market manipulation risks, forming the "data silos" and "trust barriers" that are prevalent in cross-regional transactions.

[0003] There are three similar solutions in the existing technology: (1) A cross-regional electricity price prediction solution based on single federated learning (Chinese patent application CN120744559A), which only uses federated learning, does not encrypt the model gradient, cannot resist model reverse attack and gradient leakage attack, and stops at the electricity price prediction stage; (2) A transaction settlement solution based on blockchain and MPC (Chinese patent application CN117391754A), which focuses on transaction settlement privacy protection, does not involve electricity price prediction and bidding decision, and is limited to a single region within the province; (3) A single-region bidding solution based on mixed integer programming, which cannot use cross-regional data to improve prediction accuracy, and does not consider the revenue distribution and privacy protection of multi-entity collaborative bidding.

[0004] In summary, existing technologies suffer from insufficient privacy protection capabilities, limited scenario coverage, lack of functional closed loops, and weak collaborative capabilities, and cannot meet all the needs of cross-regional joint bidding under a unified national electricity market. Summary of the Invention

[0005] This invention aims to address three core issues: first, resolving the conflict between data silos and trade secret protection in cross-regional power trading, achieving "data immobilization, knowledge sharing"; second, addressing the insufficient accuracy of electricity price forecasts in single regions, aggregating the value of data from the entire market to improve forecast accuracy; and third, resolving the disconnect between forecasting and bidding decision-making, providing an integrated end-to-end solution from modeling to decision-making. Ultimately, it provides a cross-regional power spot market joint bidding engine with high security, accurate forecasting, strong collaborative capabilities, and high feasibility.

[0006] This invention provides a cross-regional electricity spot market joint bidding system, the specific technical solution of which is as follows:

[0007] A cross-regional electricity spot market joint bidding system includes:

[0008] Local clients deployed across multiple participating entities, each local client containing:

[0009] The data preprocessing module is used to preprocess the local raw operational data and add adaptive Gaussian noise to generate a local training feature set.

[0010] The local model training module is used to train a local price prediction sub-model based on the local training feature set and to calculate the model gradient parameters for the current round.

[0011] The share generation module is used to split the model gradient parameters into N private shares using an additive private sharing protocol, where N is an integer greater than or equal to 2;

[0012] A secure multi-party computation (MPC) computing node cluster, comprising the aforementioned N MPC computing nodes, each MPC computing node being used for:

[0013] Receive the private share corresponding to this node from each local client;

[0014] Add all received private shares to obtain aggregate gradient shares;

[0015] Wherein, after at least T MPC computing nodes have completed their calculations, the global aggregate gradient parameters are jointly recovered by the at least T MPC computing nodes, where T is a preset secret sharing threshold and 2≤T≤N; the global aggregate gradient parameters are distributed to each local client to update the local price prediction sub-model; the above process is repeated until the preset convergence condition is met; the MPC computing node cluster is also used to: securely calculate the joint price curve through the MPC protocol based on the converged global price prediction model and the local constraints of each participating entity.

[0016] Preferably, the local model training module adopts a horizontal federated learning framework and introduces a dynamic weight adjustment mechanism based on transaction size and historical prediction accuracy; wherein the first Aggregate weight of each participating entity The calculation formula is:

[0017] ;

[0018] In the formula, Let be the aggregation weight of the i-th participating entity; Let be the average prediction accuracy of the i-th subject over the last 10 rounds; denoted as , where is the average transaction volume over the past 30 days in the region where the i-th entity is located; N represents the total number of participating entities. This dynamic weighting mechanism effectively improves the accuracy of cross-regional electricity price forecasting and avoids the negative impact of low-quality data on the global model.

[0019] Preferably, the system further includes an abnormal node detection module, which is used to calculate the degree of deviation between each local gradient and the global gradient based on gradient cosine similarity during the model iteration and update process, and automatically identify and remove malicious nodes with gradient abnormalities, thereby ensuring the robustness of model training.

[0020] This invention also provides a method for cross-regional electricity spot joint bidding based on the aforementioned cross-regional electricity spot joint bidding system, comprising the following steps:

[0021] Step S1: Each participating entity preprocesses the original operational data locally and adds adaptive Gaussian noise to generate a local training feature set; this step avoids leakage of the original data from the source through local differential privacy protection.

[0022] Step S2: Each participating entity independently trains a local price prediction sub-model based on the local training feature set and calculates the model gradient parameters for the current round; through horizontal federated learning, "the data remains still while the model moves", protecting local data privacy.

[0023] Step S3: Use the additive private sharing protocol to split the model gradient parameters into N private shares and send them to N MPC computing nodes respectively, where N is an integer greater than or equal to 2; this splitting operation ensures that no single node can restore the complete gradient information.

[0024] Step S4: Each MPC computing node performs an addition operation on all received private shares to obtain the aggregated gradient share; the additive homomorphic property allows the aggregation process to be completed without decryption.

[0025] Step S5: After at least T MPC computing nodes have completed the calculation, the global aggregate gradient parameters are recovered by the at least T MPC computing nodes, where T is a preset secret sharing threshold and 2≤T≤N; this threshold mechanism can resist collusion attacks by no more than T-1 nodes.

[0026] Step S6: Distribute the global aggregated gradient parameters to each participating entity, update the local price prediction sub-model, and repeat steps S2 to S5 until the model converges; iterative convergence ensures the high accuracy of the final prediction model.

[0027] Step S7: Based on the converged global price prediction model, securely calculate the joint bid curve using the MPC protocol. This step achieves end-to-end automation from prediction to bidding, while protecting the cost and constraint information of each entity.

[0028] The method may further include: adopting a dynamic weight adjustment mechanism in step S2 (the formula is the same as above), introducing abnormal node detection in step S6, and selecting a non-cooperative game mode or a cooperative game mode in step S7, and distributing fair benefits based on the Shapley value in the cooperative mode.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] (1) Higher level of privacy and security: Federated learning achieves "data not moving", MPC protects the parameter exchange process, can resist model reverse attack and collusion attack, and the security level meets the national level three security requirements.

[0031] (2) Significantly improved prediction accuracy: Through cross-regional multi-subject collaborative modeling, the accuracy of electricity price prediction is 15%-25% higher than that of single-regional models and 10%-15% higher than that of single federated learning models.

[0032] (3) Full-chain closed loop can be implemented: The entire process from data preprocessing, model training, price prediction to bidding strategy generation and revenue distribution is automated and can be directly connected to the power trading system.

[0033] (4) Strong collaborative capability: It supports both cooperative and non-cooperative modes, achieves fair distribution of benefits through the Sharpe ratio, and effectively incentivizes market participants to participate.

[0034] (5) Good robustness and scalability: The distributed architecture has no single point of failure, supports dynamic node addition and removal, and has abnormal node detection and fault tolerance mechanisms. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structure of a cross-regional electricity spot joint bidding system according to an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail below.

[0037] like Figure 1 As shown, the present invention provides a cross-regional electricity spot market joint bidding system, comprising:

[0038] Local clients deployed across multiple participating entities, each local client containing:

[0039] The data preprocessing module is used to preprocess the local raw operational data and add adaptive Gaussian noise to generate a local training feature set.

[0040] The local model training module is used to train a local price prediction sub-model based on the local training feature set and to calculate the model gradient parameters for the current round.

[0041] The share generation module is used to split the model gradient parameters into N private shares using an additive private sharing protocol, where N is an integer greater than or equal to 2;

[0042] A secure multi-party computation (MPC) computing node cluster, comprising the aforementioned N MPC computing nodes, each MPC computing node being used for:

[0043] Receive the private share corresponding to this node from each local client;

[0044] Add all received private shares to obtain aggregate gradient shares;

[0045] Wherein, after at least T MPC computing nodes have completed the calculation, the global aggregate gradient parameters are jointly recovered by the at least T MPC computing nodes, where T is a preset secret sharing threshold, and 2≤T≤N; the global aggregate gradient parameters are sent to each local client to update the local price prediction sub-model;

[0046] Repeat the above process until the preset convergence condition is met;

[0047] The MPC computing node cluster is also used to: securely calculate the joint price curve through the MPC protocol based on the converged global price prediction model and the local constraints of each participating entity.

[0048] The local model training module employs a horizontal federated learning framework and introduces a dynamic weight adjustment mechanism based on transaction size and historical prediction accuracy; wherein, the first Aggregate weight of each participating entity The calculation formula is:

[0049] ;

[0050] In the formula, Let be the aggregation weight of the i-th participating entity; Let be the average prediction accuracy of the i-th subject over the last 10 rounds; is the average transaction size of the region where the i-th entity is located over the past 30 days; N is the total number of participating entities.

[0051] The system also includes an abnormal node detection module, which is used to calculate the degree of deviation between each local gradient and the global gradient based on gradient cosine similarity during the model iteration and update process, and automatically identify and remove malicious nodes with gradient anomalies.

[0052] This invention also provides a method for cross-regional electricity spot joint bidding based on the aforementioned cross-regional electricity spot joint bidding system, comprising the following steps:

[0053] Step S1: Each participating entity preprocesses the original operational data locally and adds adaptive Gaussian noise to generate a local training feature set;

[0054] Step S2: Each participating entity independently trains a local price prediction sub-model based on the local training feature set and calculates the model gradient parameters for the current round;

[0055] Step S3: Use the addition private sharing protocol to split the model gradient parameters into N private shares and send them to N MPC computing nodes respectively, where N is an integer greater than or equal to 2;

[0056] Step S4: Each MPC computing node performs an addition operation on all received private shares to obtain the aggregate gradient share;

[0057] Step S5: After at least T MPC computing nodes have completed the calculation, the global aggregate gradient parameters are recovered by the at least T MPC computing nodes, where T is a preset secret sharing threshold and 2≤T≤N;

[0058] Step S6: Distribute the global aggregated gradient parameters to each participating entity, update the local price prediction sub-model, and repeat steps S2 to S5 until the model converges;

[0059] Step S7: Based on the converged global price prediction model, securely calculate the joint price curve using the MPC protocol.

[0060] In step S2, a horizontal federated learning framework is adopted, and a dynamic weight adjustment mechanism based on transaction size and historical prediction accuracy is introduced; among which, the first Aggregate weight of each participating entity The calculation formula is:

[0061] ;

[0062] In the formula, Let be the aggregation weight of the i-th participating entity; Let be the average prediction accuracy of the i-th subject over the last 10 rounds; is the average transaction size of the region where the i-th entity is located over the past 30 days; N is the total number of participating entities.

[0063] In step S6, during the model iteration and update process, there is also an abnormal node detection step: the deviation between each local gradient and the global gradient is calculated based on the gradient cosine similarity, and malicious nodes with abnormal gradients are automatically identified and removed.

[0064] In step S7, the step of securely calculating the joint quotation curve specifically includes:

[0065] In a non-cooperative game theory model, each participant independently generates a bid based on global price predictions, and the prediction results are securely distributed via the MPC protocol; or

[0066] By choosing a cooperative game theory model, the optimal allocation of resources and pricing strategies among the alliance members are calculated securely using the MPC protocol, thereby maximizing the overall benefits of the alliance.

[0067] The cooperative game model also includes a step of calculating the profit distribution ratio of each entity based on the Shapley value to ensure a fair distribution of profits.

[0068] The following detailed embodiments are based on the framework of the claims and the beneficial effects of each technical feature are explained.

[0069] Example 1: System Architecture and Initialization

[0070] This embodiment provides a cross-regional electricity spot market joint bidding system, which is applied to the Beijing-Tianjin-Hebei cross-regional electricity spot market.

[0071] System Composition: Includes local clients for 10 participating entities (3 virtual power plants in Beijing, 2 thermal power plants in Tianjin, and 4 new energy power plants in Hebei), and 3 neutral MPC computing nodes (jointly deployed by the power grid company, power trading center, and a trusted third-party institution). System Parameters: N=3, T=2, i.e., secret sharing threshold (2,3); the preset convergence condition for federated learning iteration is that the change in loss value is less than 1e-5 for 3 consecutive rounds; the global model adopts an LSTM+Attention architecture.

[0072] In this way, the legitimacy of the participating entities and the security of communication are ensured through digital certificate identity authentication and key distribution; the threshold setting allows any two or more nodes to recover the secret, while a single node cannot restore the original gradient information of any participating entity, which can resist the attack of a single node or a conspiracy attack by at most one node.

[0073] Example 2: Local Data Preprocessing (Step S1)

[0074] Each participating entity extracts historical operating data from the past year, including cost function parameters, unit output limits, energy storage charging and discharging status and inventory, historical pricing strategies, regional load data, and meteorological data. The data preprocessing module performs missing value imputation (using linear interpolation), outlier removal (3σ principle), standardization (Z-score normalization), and adds adaptive Gaussian noise. The noise intensity is dynamically adjusted according to data sensitivity (the higher the sensitivity, the greater the noise variance), ultimately generating a local training feature set.

[0075] In this way, adaptive Gaussian noise achieves local differential privacy protection, making it impossible to accurately reverse engineer the original operational data even if the local dataset is stolen, thus protecting trade secrets from the source.

[0076] Example 3: Local Sub-model Training and Dynamic Weights

[0077] Each entity independently trains an LSTM+Attention time-series prediction model based on its local feature set. The local model training module adopts a horizontal federated learning framework and introduces a dynamic weight adjustment mechanism based on transaction size and historical prediction accuracy. The aggregate weight of the i-th participating entity... The calculation formula is as follows:

[0078] ;

[0079] in, =10 represents the total number of participating entities. The inverse of the average absolute error percentage (after normalization) of the model predictions for the i-th subject over the last 10 rounds. This represents the average transaction volume (in MWh) of the region where the i-th entity is located over the past 30 days. For example, a virtual power plant has an average accuracy rate of 92% over the past 10 rounds and an average daily transaction volume of 500 MWh over the past 30 days. Its weight is significantly higher than that of entities with low accuracy and small transaction volume.

[0080] This dynamic weighting mechanism allows high-precision, large-scale subjects to make a greater contribution to the global model aggregation, effectively suppressing the negative impact of low-quality data and improving the global model prediction accuracy by about 12.7% compared to equal-weighted federated learning.

[0081] Example 4: Gradient Share Splitting and Transmission

[0082] The share generation module uses an additive private sharing protocol: it splits the model gradient parameter G of the current round into 3 random shares. , , Satisfying G The data is sent to three MPC compute nodes (each node receives a different share). This process takes approximately 0.05 seconds in this embodiment.

[0083] After splitting, each node only has a partial share and cannot restore the original gradient, thus resisting model reverse attacks; at the same time, due to the homomorphism of splitting addition, subsequent aggregation does not require decryption, ensuring computational efficiency.

[0084] Example 5: MPC Secure Aggregation and Gradient Recovery

[0085] Each MPC compute node receives a private share corresponding to that node from all 10 subjects (e.g., node 1 receives a private share from all subjects). Each node adds its received share to obtain an aggregated gradient share (k=1,2,3). Once at least T=2 computation nodes have completed their calculations, they exchange their aggregated shares. Recover the global aggregate gradient parameters .

[0086] The additive aggregation process is completed in a dense state, where no single node can obtain the gradient information of any entity; the threshold mechanism can resist collusion attacks by no more than one node. In this embodiment, a single round of aggregation takes 0.32 seconds, meeting the 15-minute reporting time requirement of the current market.

[0087] Example 6: Iterative Update and Convergence Determination

[0088] Recovered global aggregated gradient parameters The global gradient parameter is distributed to all participating entities, and each entity updates its local sub-model parameters using gradient descent based on this global gradient parameter. Steps S2 to S5 are repeated, and the global loss function value is calculated after each round. The model is considered converged when the change in loss value is less than 1e-5 for three consecutive rounds. In this embodiment, convergence occurred after 42 iterations.

[0089] Thus, the converged model has high prediction accuracy, with a global electricity price prediction mean absolute error (MAE) of 8.2 yuan / MWh, which is 21.3% lower than the single-region model and 12.7% lower than the single federated learning model.

[0090] Example 7: Abnormal Node Detection

[0091] During the iteration process, each MPC computing node calculates the cosine similarity between the gradient uploaded by each participating subject and the current global gradient. If the gradient cosine similarity of a subject is lower than a preset threshold (e.g., 0.6), it is marked as an abnormal node, and its gradient share is removed in subsequent aggregations. This embodiment assumes the existence of one malicious node, which is detected and removed, reducing the model convergence rounds from 52 to 42.

[0092] This detection effectively defends against poisoning attacks and abnormal data injection, ensuring the stability and robustness of large-scale distributed systems.

[0093] Example 8: Joint Bidding Decision

[0094] Based on the converged global price prediction model, the system supports two bidding modes:

[0095] Non-cooperative game theory mode: Each participant independently generates its own price curve based on global price forecasts. The MPC protocol is only used to securely distribute the forecast results to each participant (without disclosing information about other participants). This mode is simple, computationally fast, and suitable for scenarios with high information sensitivity.

[0096] Cooperative Game Theory (adopted in this embodiment): Each entity inputs its local constraints (maximum unit output, energy storage charging and discharging power limits, ramp rate, etc.) and cost information via the MPC protocol in the form of private shares. The MPC computing node cluster securely calculates the optimal output allocation and bidding combination for each entity within the alliance, maximizing the overall alliance revenue. Specifically, a combination of linear programming and MPC is used to solve the problem in a dense state. After the bidding is completed, the revenue distribution ratio of each entity is calculated based on the Shapley value: the Shapley value is equal to the average marginal contribution of that entity across all possible alliance sequences.

[0097] As can be seen, the overall revenue of the alliance under the cooperative model increased by 18.6% compared to the non-cooperative model, with each entity achieving varying degrees of revenue growth ranging from 5% to 25%. The Shapley value ensured fairness in distribution and incentivized participation in the cooperative process. The entire bidding decision-making process took 0.15 seconds, meeting the day-ahead market submission timeframe.

[0098] Example 9: Profit Distribution and Incremental Update

[0099] After the transaction is completed, revenue settlement will be completed according to the actual clearing price and the actual power generation / consumption of each entity, based on the Shapeli value ratio. The actual electricity price data and clearing results of this transaction will be added to the local dataset to trigger the next round of incremental model training (federated learning incremental rounds) to continuously optimize prediction accuracy.

[0100] It is evident that the closed-loop update mechanism enables the model to adapt to market changes and maintain high prediction accuracy over the long term.

[0101] The experimental data are summarized in Table 1.

[0102] Table 1 Summary of Experimental Data

[0103]

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cross-regional electricity spot market joint bidding system, characterized in that, include: Local clients deployed across multiple participating entities, each local client containing: The data preprocessing module is used to preprocess the local raw operational data and add adaptive Gaussian noise to generate a local training feature set. The local model training module is used to train a local price prediction sub-model based on the local training feature set and to calculate the model gradient parameters for the current round. The share generation module is used to split the model gradient parameters into N private shares using an additive private sharing protocol, where N is an integer greater than or equal to 2; A secure multi-party computation (MPC) computing node cluster, comprising the aforementioned N MPC computing nodes, each MPC computing node being used for: Receive the private share corresponding to this node from each local client; Add all received private shares to obtain aggregate gradient shares; Wherein, after at least T MPC computing nodes have completed the calculation, the global aggregate gradient parameters are jointly recovered by the at least T MPC computing nodes, where T is a preset secret sharing threshold, and 2≤T≤N; the global aggregate gradient parameters are sent to each local client to update the local price prediction sub-model; Repeat the above process until the preset convergence condition is met; The MPC computing node cluster is also used to: securely calculate the joint price curve through the MPC protocol based on the converged global price prediction model and the local constraints of each participating entity.

2. The system according to claim 1, characterized in that, The local model training module employs a horizontal federated learning framework and introduces a dynamic weight adjustment mechanism based on transaction size and historical prediction accuracy; wherein, the first Aggregate weight of each participating entity The calculation formula is: ； In the formula, Let be the aggregation weight of the i-th participating entity; Let be the average prediction accuracy of the i-th subject over the last 10 rounds; is the average transaction size of the region where the i-th entity is located over the past 30 days; N is the total number of participating entities.

3. The system according to claim 1, characterized in that, The system also includes an abnormal node detection module, which is used to calculate the degree of deviation between each local gradient and the global gradient based on gradient cosine similarity during the model iteration and update process, and automatically identify and remove malicious nodes with gradient anomalies.

4. A method for cross-regional electricity spot joint bidding based on the cross-regional electricity spot joint bidding system according to any one of claims 1-3, characterized in that, Includes the following steps: Step S1: Each participating entity preprocesses the original operational data locally and adds adaptive Gaussian noise to generate a local training feature set; Step S2: Each participating entity independently trains a local price prediction sub-model based on the local training feature set and calculates the model gradient parameters for the current round; Step S3: Use the addition private sharing protocol to split the model gradient parameters into N private shares and send them to N MPC computing nodes respectively, where N is an integer greater than or equal to 2; Step S4: Each MPC computing node performs an addition operation on all received private shares to obtain the aggregate gradient share; Step S5: After at least T MPC computing nodes have completed the calculation, the global aggregate gradient parameters are recovered by the at least T MPC computing nodes, where T is a preset secret sharing threshold and 2≤T≤N; Step S6: Distribute the global aggregated gradient parameters to each participating entity, update the local price prediction sub-model, and repeat steps S2 to S5 until the model converges; Step S7: Based on the converged global price prediction model, securely calculate the joint price curve using the MPC protocol.

5. The method according to claim 4, characterized in that, In step S2, a horizontal federated learning framework is adopted, and a dynamic weight adjustment mechanism based on transaction size and historical prediction accuracy is introduced; wherein, the first Aggregate weight of each participating entity The calculation formula is: ； In the formula, Let be the aggregation weight of the i-th participating entity; Let be the average prediction accuracy of the i-th subject over the last 10 rounds; is the average transaction size of the region where the i-th entity is located over the past 30 days; N is the total number of participating entities.

6. The method according to claim 4, characterized in that, In step S6, during the model iteration update process, an abnormal node detection step is also included: calculating the degree of deviation between each local gradient and the global gradient based on gradient cosine similarity, and automatically identifying and removing malicious nodes with gradient abnormalities.

7. The method according to claim 4, characterized in that, In step S7, the step of securely calculating the joint quotation curve specifically includes: In a non-cooperative game theory model, each participant independently generates a bid based on global price predictions, and the prediction results are securely distributed via the MPC protocol; or By choosing a cooperative game theory model, the optimal allocation of resources and pricing strategies among the alliance members are calculated securely using the MPC protocol, thereby maximizing the overall benefits of the alliance.

8. The method according to claim 7, characterized in that, The cooperative game model also includes a step of calculating the profit distribution ratio of each subject based on the Shapley value in order to carry out a fair distribution of profits.

Images