Electricity-carbon-green certificate multi-mode dynamic coupling pricing and cross-chain clearing method based on federated learning-block chain fusion architecture

Abstract

The invention provides an electricity-carbon-green certificate multi-mode dynamic coupling pricing and cross-chain clearing method based on a federated learning-block chain fusion framework. Data islands are broken through differential privacy horizontal federation aggregation, an LSTM-GRU hybrid model is trained, and global price prediction with the error smaller than 5% is achieved; a TD3 reinforcement learning agent is introduced, a nonlinear linkage coefficient is generated in real time, the policy mutation response speed is increased by 50%, and the arbitrage space is compressed by 40%; a private chain and alliance chain double-chain heterogeneous architecture is constructed, an HTLC is used for binding electricity purchasing, carbon selling and green certificate purchasing as an inseparable transaction, the clearing delay is smaller than or equal to 800 ms, and the failure rate is smaller than 0.01%; zK-SNARK zero-knowledge proof is adopted to ensure that transactions can be audited and privacy is not leaked. The provincial 100-node cluster actual measurement shows that the prediction error is reduced by 75%, the policy delay is shortened by 51%, the combined transaction failure rate is reduced by 99.9%, and safe, efficient and compliant technical support is provided for multi-market collaboration in a high-fluctuation environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of energy financial technology and distributed computing, and specifically relates to a power-carbon-green certificate multi-modal dynamic coupling pricing and cross-chain clearing method based on a federated learning-blockchain fusion architecture. BACKGROUND

[0002] In the critical stage of the digital transformation of the energy market, the limitations of the existing technical system are increasingly prominent, seriously restricting cross-field collaborative development and market efficiency improvement.

[0003] The current technical architecture has three core problems that need to be solved. First, the data barriers formed by the scattered regulatory subjects of the electricity, carbon, and green certificate markets, the traditional centralized model is difficult to obtain global data due to compliance restrictions, and the data island effect is significant, hindering cross-market deep analysis and collaborative decision-making. Second, the existing linkage model relies on a linear parameter adjustment mechanism, which cannot effectively respond to the chaotic effects caused by nonlinear changes in the market when faced with policy mutations such as carbon quota tightening, making it difficult to maintain system dynamic balance and stable operation. Third, the traditional mode of manual reconciliation in cross-market transaction processes leads to a long settlement cycle, not only reducing the efficiency of capital circulation, but also causing the loss of price arbitrage opportunities due to time differences, damaging the enthusiasm of market participants and overall vitality.

[0004] At the same time, there are two key bottlenecks in the technical implementation. On the one hand, a single blockchain architecture cannot reconcile the contradiction between high-frequency trading (second-level response requirements) in the electricity market and strong regulatory requirements in the carbon market, compromising performance and compliance; on the other hand, the LSTM model in processing cross-market heterogeneous data lacks a feature alignment mechanism, resulting in a feature alignment error of more than 15% in the prediction process, severely weakening the reliability and decision-making reference value of the prediction results. These technical defects interweave and form a complex technical obstacle, posing a substantial challenge to the efficient operation and innovative development of the energy market. SUMMARY

[0005] This invention constructs a multi-market collaborative trading architecture for electricity, carbon, and green certificates in a novel power system. It addresses the collaborative optimization issues of atomicity, policy sensitivity, and clearing efficiency in cross-market transactions through a three-layer coupling mechanism. First, differential privacy encryption and horizontal aggregation technology are employed in the federated learning layer to train a global prediction model by integrating dynamic features from multiple market sources. Second, a policy shock response mechanism is constructed through a multimodal reinforcement learning agent in the dynamic coupling layer to accurately quantify market linkages. Finally, a dual-chain heterogeneous architecture and atomic swap protocol are designed in the cross-chain clearing layer to achieve the indivisibility of combined transactions and real-time clearing. This architecture innovatively introduces a meta-learning-based risk hedging engine to support the stable operation of multi-energy markets in highly volatile environments. The specific technical solution is: a multimodal dynamic coupling pricing and cross-chain clearing method for electricity, carbon, and green certificates based on a federated learning-blockchain fusion architecture, comprising the following steps:

[0006] Step 1, Multi-source data federation processing and privacy protection: Each participant adds differential privacy Gaussian noise, AES-256 encryption, and Paillier homomorphic encryption to the electricity load curve, carbon quota auction volume, and green certificate price series locally, and then uploads them to the federation server through a secure channel. The server dynamically weights the data according to the data volume to complete the horizontal federation aggregation and generate a time-aligned multimodal dataset, which serves as the input to the global price prediction model.

[0007] Step 2, Training the global price prediction model: The federated server inputs the multimodal dataset into the LSTM-GRU hybrid model. The LSTM layer captures the long-term policy trend of electricity prices, the GRU layer learns the intraday fluctuations of green certificates, and the price confidence interval is output after the feature fusion of the two branches.

[0008] Step 3, Multimodal reinforcement learning strategy generation: The state space is based on the normalized change in electricity price-carbon price and the volatility of green certificates. The price confidence interval provides the volatility boundary for the state space. The Actor network generates a nonlinear linkage coefficient matrix in real time, and the Critic network uses the negative arbitrage space as the reward function for closed-loop evaluation.

[0009] Step 4, Cross-chain atomic settlement protocol execution: Using non-linear linkage coefficients to guide transaction matching and quota calculation, a heterogeneous dual-chain architecture of private chain and consortium chain is constructed. The private chain carries millisecond-level electricity transactions, and the consortium chain completes carbon / green certificate compliance audits with verifiable functions. The "electricity purchase - carbon sale - green certificate purchase" is bound into the same atomic transaction through hash time lock.

[0010] Step 5, Zero-knowledge proof audit and system verification: Using transaction data, generate audit proofs for three types of constraints: electricity volume ≤ load limit, carbon sales volume ≤ holding quota, and green certificates ∈ renewable energy list, and publicly verify them at regulatory nodes.

[0011] The beneficial effects of this invention are:

[0012] (1) The federal learning layer designed by the application uses differential privacy encryption horizontal aggregation technology (based on Gaussian noise mechanism) to fuse multi-dimensional sensitive data such as power load fluctuation, carbon quota supply-demand gap, green certificate circulation, and train a global LSTM-GRU hybrid prediction model. Compared with traditional centralized modeling, this technology reduces the risk of data privacy leakage by more than 90%, improves model training efficiency by 40%, effectively protects data privacy and improves model training speed.

[0013] (2) The federal learning layer of the application relies on the above-mentioned differential privacy encryption horizontal aggregation technology architecture to realize encrypted collaborative calculation of cross-regional market characteristics, and supports the green certificate demand prediction error of the provincial power grid under typhoon weather to be controlled within 5%. Compared with the traditional modeling method, the prediction accuracy of this architecture under complex weather conditions is significantly superior, and the cross-regional data collaborative calculation capability and special scene prediction reliability are effectively improved.

[0014] (3) The application introduces a multi-modal reinforcement learning (MMRL) agent in the dynamic coupling layer, taking the historical volatility rate of electricity-carbon-green certificate price as the state space, and generating a nonlinear linkage coefficient matrix through a double-delay deep deterministic policy gradient algorithm to accurately depict the energy-environment policy transmission path. Compared with the traditional linear regression model, this mechanism breaks through the limitations of linear relationship modeling and exhibits significant technical advantages in capturing the nonlinear characteristics of policy transmission paths, providing a more accurate quantitative tool for policy impact analysis.

[0015] (4) When the policy impact simulator injects extreme events such as carbon tax step adjustment and green certificate quota mutation, the multi-modal reinforcement learning agent mechanism of the dynamic coupling layer of the application makes the system policy response speed increase by 50% and the cross-market arbitrage space compress by 40%. Compared with the response ability of the traditional model under extreme policy impact, this mechanism significantly enhances the anti-policy disturbance ability of market operation, effectively maintaining market transaction order and stability.

[0016] (5) The cross-chain clearing layer constructed by the application adopts a double-chain heterogeneous architecture: private chain handles millisecond-level power high-frequency transactions, and alliance chain performs carbon / green certificate compliance audit. Through the cross-chain atomic exchange protocol (based on hash time lock HTLC), the "buy electricity-sell carbon-buy green certificate" combined transaction is bound as an indivisible transaction. The test shows that this mechanism reduces the combined transaction failure rate from 12% of the traditional serial processing to <0.01%, and the clearing delay is compressed to within 800ms, and the audit data is verified through zero-knowledge proof. Compared with the traditional serial clearing mode, this architecture significantly improves the reliability and processing efficiency of multi-market collaborative transactions, fully meeting the strict technical requirements of power grids for high-concurrency transaction scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A method flowchart of the present application. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application 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 only used to explain the present application and do not limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other. In order to achieve the above-mentioned purpose, the technical solutions adopted by the present application are as follows.

[0019] The present application provides a method for multi-modal dynamic coupling pricing and cross-chain clearing of electricity-carbon-green certificate based on a federated learning-blockchain fusion architecture, which realizes the collaborative pricing and atomic clearing of the electricity-carbon-green certificate market through a hierarchical architecture. The specific implementation case is as follows:

[0020] Step 1: Multi-source data federalization processing and privacy protection

[0021] In view of the heterogeneous data island problem of the electricity, carbon and green certificate markets, a distributed collaboration network is constructed using differential privacy encryption technology. After each participant (provincial power grid, carbon exchange, green certificate platform) completes the desensitization of the original data locally, the encrypted features are uploaded through a secure channel, and the spatio-temporal alignment and fusion of cross-market dynamic features are realized in the federal server, completely avoiding the risk of sensitive data leakage.

[0022] Step 1.1: Local data encryption

[0023] Electricity node: input 24-hour load curve (such as real-time load of a certain provincial power grid), where, is the real-time load at the hour, and a Gaussian noise satisfying differential privacy is added:

[0024] ;

[0025] where, is a Gaussian noise with mean 0 and variance , and the noise standard deviation must be strictly limited to 2% of the load peak value to ensure that the disturbed curve still retains the actual fluctuation pattern (such as the evening peak feature); carbon node input carbon quota auction quantity (for example, 200 tons of listing quantity), as carbon emission data is highly classified, AES-256 encryption is used to transmit the original value directly to avoid exposure in the middle link; green certificate node input green certificate price sequence (for example, 300 green certificates listing price), the price gradient Encryption is used to support the federal server to directly calculate the statistical characteristics in the encrypted state (such as the average daily volatility).

[0026] Step 1.2 Federated Aggregation

[0027] The federal server dynamically allocates weights according to the data volume: power node = 24 x 100 nk= 24 x 100 (24 hours x 100 samples per minute), carbon / green certificate node Take the number of transaction records for the day. The weighted aggregation formula is:

[0028] ;

[0029] Where, is the data volume of node m (such as the power node 24 hours x 100 sampling points), and the weight Guarantee that nodes with large data volume contribute more. Output time-aligned multi-modal data set as the input of the global prediction model. The technical effect can reduce the risk of data privacy leakage by 90%, and support the prediction of green certificate demand under the typhoon weather of cross-provincial power grid (error < 5%).

[0030] Step 2 Global Price Prediction Model Training

[0031] In view of the long-short term coupling characteristics of cross-market price fluctuations, a LSTM-GRU hybrid model is designed as the global price prediction model. The LSTM layer captures the monthly trend of electricity price affected by policy, and the GRU layer learns the intra-day fluctuations caused by sudden changes in green certificate supply and demand. After feature fusion of the two branches, the price confidence interval is output, providing a quantitative boundary for hedging strategies.

[0032] Step 2.1 LSTM-GRU Hybrid Model Construction

[0033] Input layer: receives the multi-modal data set from the federal server, with data dimensions of 24 time steps x 3 features (power, carbon, green certificate market features), forming a time series input matrix.

[0034] LSTM layer: responsible for capturing long-term trend features, selectively retaining historical information through the gating mechanism (forget gate, input gate, output gate), especially suitable for capturing monthly trends of electricity price affected by policy adjustments, seasonal changes, etc. The hidden state update formula is:

[0035] ;

[0036] Where, are the outputs of the forget gate, input gate, and output gate, respectively; are the cell state and hidden state; is the Sigmoid activation function, and ⊙ is the element-wise multiplication; is a weight matrix, is a bias term; is a candidate cell state.

[0037] GRU layer: optimized for short-term fluctuations, improved computational efficiency through simplified gating structure (reset gate, update gate), quickly learned the characteristics of intraday price fluctuations triggered by sudden changes in green certificate supply and demand. The output state formula is:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] wherein, controls the proportion of historical information retained, controls the degree of forgetting of historical information, is a candidate hidden state.

[0043] Output layer: fuse the features extracted by LSTM and GRU through a fully connected network to generate price confidence interval , providing market participants with quantitative boundaries containing price fluctuation ranges, facilitating the development of more robust hedging strategies.

[0044] wherein, is the output hidden state of the GRU layer at 24 time steps, is the output hidden state of the LSTM layer at 24 time steps, is the lowest predicted price of the confidence interval of multi-market coordinated pricing, is the highest predicted price of the confidence interval of multi-market coordinated pricing.

[0045] Step 2.2 Differential Privacy Guarantee

[0046] During model training, to prevent gradient information leakage, add Laplace noise to the gradient: , noise variance , ( is the failure probability).

[0047] wherein, the noise variance is determined according to the privacy budget ( =0.8) (satisfies the industrial-level privacy standard) This mechanism ensures the training efficiency of the model is improved by 40% while protecting the privacy of the training data, and the coverage rate of the output price confidence interval reaches 95%, which is significantly higher than the 80% of traditional models, effectively covering more actual price fluctuation scenarios.

[0048] Step 3 Multi-modal reinforcement learning strategy generation

[0049] The nonlinear coupling relationship between electricity price, carbon price, and green certificate price fluctuations is constructed. The TD3 algorithm takes the cross-market price difference as the state, generates a hedging coefficient matrix in real time through the Actor network, and evaluates the inhibitory effect of the strategy on market imbalance through the Critic network, forming a closed-loop optimization.

[0050] Step 3.1 State space construction

[0051] Define the normalized state vector:

[0052] ;

[0053] Where, is the change in electricity price and carbon price; is the 30-day average electricity price, reflecting long-term trends; is the green certificate price volatility, and max(σ) is the historical maximum volatility. Capture the intensity of short-term market fluctuations. After normalization, all state variables have the same input dimension, making it easier for the reinforcement learning model to process.

[0054] Step 3.2 TD3 algorithm generation

[0055] The Actor network outputs the linkage coefficient:

[0056] ;

[0057] Example coefficients: ( = 0.03, = 0.7) (trained from historical data) to describe the price transmission relationship between different markets. For example, α can represent the influence coefficient of carbon price changes on electricity price, and β represents the feedback strength of green certificate price fluctuations on the carbon market.

[0058] The Critic network evaluates the action value:

[0059] ;

[0060] Where, is the conditional expectation operator based on the next market state, is the reward discount factor, is the target Actor network.

[0061] The reward function is designed as (r t = | electricity arbitrage space |), which punishes market imbalance behavior through a negative reward mechanism and guides the strategy to optimize in the direction of reducing arbitrage space and promoting market stability.

[0062] Step 3.3 Extreme event response

[0063] When the policy simulator injects an extreme event such as "carbon tax increase by 20%", the system triggers the following response process:

[0064] 1. Detect carbon price mutation: (Threshold configurable) Real-time detection of market anomalies through preset thresholds triggers emergency response mechanisms.

[0065] 2. Dynamically update Actor network parameters: Relearn market linkage after carbon tax increase to ensure strategy adaptation to new market environment. Where, is the parameter set of the Actor network, is the learning rate of the Actor network, is the gradient of the Actor.

[0066] 3. Generate a new hedging matrix based on updated parameters Guide market participants to adjust trading strategies to reduce the risk brought by policy mutations. The TD3 algorithm increases the response speed of policy mutations by 50% and compresses arbitrage space by 40%, significantly enhancing the system's ability to adapt to sudden policies.

[0067] Step 4 Cross-chain atomic clearing protocol execution

[0068] Adopt a dual-chain heterogeneous architecture to separate high-frequency trading from compliance auditing, and use Hash Time Lock Contract (HTLC) to ensure the atomicity of portfolio transactions, ensuring the consistency and reliability of electricity-carbon-green certificate transactions.

[0069] Step 4.1 Dual-chain architecture design

[0070] Separate high-frequency trading and compliance auditing functions through a dual-chain heterogeneous architecture to meet the performance and regulatory needs of different markets. Private chains use efficient consensus mechanisms to handle power transactions at the second level, and consortium chains use verifiable functions to perform environmental compliance reviews.

[0071] Private chain (electricity transaction chain): Adopt PBFT (Practical Byzantine Fault Tolerance) consensus mechanism to realize three-phase commit (preparation, preparation, and submission), with transaction delay controlled within 100ms, meeting the low-latency needs of electricity market high-frequency trading. Define electricity transaction transactions as a four-tuple:

[0072] ;

[0073] Where, represents the unique identifier of the transaction, represents the transaction type (fixed as "electricity"), represents the transaction power, ​Represents the transaction price. The consensus mechanism adopts the PBFT three-phase commit protocol, which meets the delay constraint:

[0074] ;

[0075] Consensus chain (carbon / green certificate chain): Define the quota verification function as a Boolean mapping:

[0076] ; The specific implementation is:

[0077] ;

[0078] Where: is the address of the transaction party, represents the carbon transaction volume, is the mapping of the address to the carbon quota balance, is the carbon quota balance mapping function.

[0079] Step 4.2 HTLC atomic exchange

[0080] Transaction binding: Generate an encrypted hash lock:

[0081] ;

[0082] Binding electricity, carbon, and green certificate transactions to achieve cross-chain atomic exchange, the specific steps are as follows:

[0083] 1. Asset locking and hash lock publication: The private chain first locks the funds corresponding to the electricity transaction, and publishes the hash lock generated based on the SHA256 algorithm (commonly encrypted by preimage, electricity transaction, carbon transaction, and green certificate transaction, binding the three types of transactions); At the same time, the consortium chain freezes the carbon quota and green certificate involved in this transaction to prevent assets from being transferred in advance.

[0084] 2. Audit waiting and compliance verification: The private chain enters a waiting state, waiting for the audit results of the consortium chain; The consortium chain then conducts environmental compliance verification (such as verifying whether the transaction party's carbon quota is sufficient and whether the green certificate is from renewable energy), and generates a signed confirmation after verification, and feeds back to the private chain.

[0085] 3. Key verification and asset release: After receiving the audit pass feedback from the consortium chain, the private chain needs to verify the preimage provided by the transaction party (the unique key to unlock the hash lock ), and if the verification is valid, the locked electricity funds will be released; the consortium chain synchronously verifies the same preimage, and if the verification is valid, the frozen carbon quota and green certificate will be released, completing the entire asset delivery. If the preimage is invalid or not provided within the time limit, both chains will unlock the locked funds and assets, respectively, to avoid partial transactions.

[0086] This mechanism ensures that the combination transaction is either completely successful or completely rolled back, avoiding asset imbalance caused by partial execution. The technical effect shows that the clearing delay is ≤800ms, and the combination transaction failure rate is <0.01%, significantly improving the reliability of cross-market transactions.

[0087] Step 5: Zero-knowledge proof audit and system verification

[0088] Using the ZK-SNARK (Zero-Knowledge Simple Non-Interactive Argument of Knowledge) method, the verifiability of the clearing process is realized without revealing sensitive business data, meeting regulatory requirements while protecting enterprise privacy.

[0089] Step 5.1: Audit proof generation

[0090] Constructing circuit constraints ensures transaction compliance, including:

[0091] 1. Power transaction volume ≤ grid load upper limit, avoiding overloading transactions that cause system risks;

[0092] 2. Carbon sale volume ≤ holding quota, preventing short selling;

[0093] 3. Green certificate source ∈ renewable energy list, ensuring green certificate authenticity.

[0094] Generate zero-knowledge proof :

[0095] (phi,text{private input});

[0096] Where, is the core function of generating audit proof in the zero-knowledge simple non-interactive argument of knowledge protocol, and phi,text{private input} is the compliance constraint set of the electricity-carbon-green certificate cross-chain clearing audit.

[0097] Step 5.2: Full-link performance verification

[0098] Deployed to a provincial power trading center test environment (100-node cluster):

[0099]

[0100] Through zero-knowledge verifier public audit:

[0101] ;

[0102] The system realizes 100% transaction auditability, ensuring that each transaction can be verified by regulatory agencies for compliance, while system stability is improved by 95%, providing a solid technical guarantee for the coordinated operation of the electricity-carbon-green certificate market.

[0103] As Figure 1 shown, the present application provides a power-carbon-green certificate multi-modal dynamic coupling pricing and cross-chain clearing system based on a federal learning-blockchain fusion architecture, comprising:

[0104] The federal learning layer is used for differential privacy encryption fusion of multi-source data, training of global LSTM-GRU model, privacy protection, and improvement of prediction accuracy and training efficiency.

[0105] The dynamic coupling layer uses MMRL agent to generate nonlinear linkage coefficients, describes policy transmission path, improves extreme event response speed and compresses arbitrage space.

[0106] The cross-chain clearing layer uses a double-chain architecture + HTLC protocol to realize combined transaction atomic clearing, zero-knowledge proof audit, reduce failure rate and compress delay.

[0107] The federal learning layer includes: local data encryption module, horizontal federal aggregation module, LSTM-GRU model training module, differential privacy protection module.

[0108] The dynamic coupling layer includes: state space construction module, TD3 algorithm execution module (including Actor / Critic network), extreme event response module.

[0109] The cross-chain clearing layer includes: double-chain heterogeneous module (private chain power transaction, alliance chain compliance audit), HTLC atomic exchange module, zero-knowledge proof audit module.

[0110] Implementation summary:

[0111] This scheme solves the three core problems of data silos, policy nonlinear impact, and clearing non-atomicity in the energy financial market through the three-level cooperation of the federal learning layer (steps 1-2), the dynamic coupling layer (step 3), and the cross-chain clearing layer (steps 4-5). As Figure 1 shown, the system successfully realizes the real-time closed loop of "purchasing electricity-selling carbon-purchasing green certificates" combined transaction in the provincial power trading center, providing stable support for the energy market in a high volatility environment.

Claims

1. A multimodal dynamic coupling pricing and cross-chain settlement method for electricity-carbon-green certificates based on a federated learning-blockchain fusion architecture, characterized in that, Includes the following steps: Step 1, Multi-source data federation processing and privacy protection: Each participant adds differential privacy Gaussian noise, AES-256 encryption, and Paillier homomorphic encryption to the electricity load curve, carbon quota auction volume, and green certificate price series locally, and then uploads them to the federation server through a secure channel. The server dynamically weights the data according to the data volume to complete the horizontal federation aggregation and generate a time-aligned multimodal dataset, which serves as the input to the global price prediction model. Step 2, Training the global price prediction model: The federated server inputs the multimodal dataset into the LSTM-GRU hybrid model. The LSTM layer captures the long-term policy trend of electricity prices, the GRU layer learns the intraday fluctuations of green certificates, and the price confidence interval is output after the feature fusion of the two branches. Step 3, Multimodal reinforcement learning strategy generation: The state space is based on the normalized change in electricity price-carbon price and the volatility of green certificates. The price confidence interval provides the volatility boundary for the state space. The Actor network generates a nonlinear linkage coefficient matrix in real time, and the Critic network uses the negative arbitrage space as the reward function for closed-loop evaluation. Step 4, Cross-chain atomic settlement protocol execution: Using non-linear linkage coefficients to guide transaction matching and quota calculation, a heterogeneous dual-chain architecture of private chain and consortium chain is constructed. The private chain carries millisecond-level electricity transactions, and the consortium chain completes carbon / green certificate compliance audits with verifiable functions. The "electricity purchase - carbon sale - green certificate purchase" is bound into the same atomic transaction through hash time lock. Step 5, Zero-knowledge proof audit and system verification: Using transaction data, generate audit proofs for three types of constraints: electricity volume ≤ load limit, carbon sales volume ≤ holding quota, and green certificates ∈ renewable energy list, and publicly verify them at regulatory nodes.

2. The method according to claim 1, characterized in that, In step 1, input the 24-hour load curve. ,in, For the first For hourly real-time loads, add Gaussian noise to satisfy differential privacy: ； in, With a mean of 0 and a variance of 0, Gaussian noise.

3. The method according to claim 1, characterized in that, In step 1, the federated server dynamically allocates weights based on data volume: power nodes The sampling rate is 24 x 100, or 100 samples per minute over 24 hours, for carbon / green certificate nodes. The weighted aggregation formula is as follows: (This is based on the number of transaction records for the day.) ； in The data volume of node m, weight Nodes with larger datasets contribute more, and output time-aligned multimodal datasets. .

4. The method according to claim 1, characterized in that, In step 2, LSTM is used to capture long-term trend features, selectively retaining historical information through a gating mechanism. The hidden state update formula is: ； in, These are the outputs of the forget gate, input gate, and output gate, respectively. The states are cellular and hidden. Here, ⊙ represents the Sigmoid activation function; This is the weight matrix. For bias terms; This represents the candidate cell state.

5. The method according to claim 1, characterized in that, The GRU layer is optimized for short-term fluctuations, and its output state formula is: ； ； ； ； in, Control the proportion of historical information retained. Controlling the degree of forgetting historical information This is the candidate hidden state.

6. The method according to claim 5, characterized in that, The output layer fuses the features extracted by LSTM and GRU through a fully connected network to generate price confidence intervals. ,in, This refers to the hidden state output by the GRU layer at 24 time steps. This refers to the hidden state of the LSTM layer output at 24 time steps. It is the lowest predicted price within the confidence interval of multi-market coordinated pricing. It is the highest predicted price within the confidence interval of multi-market coordinated pricing.

7. The method according to claim 1, characterized in that, In step 3, Define the normalized state vector: ； in, These represent changes in electricity and carbon prices. This is the 30-day average of electricity prices, reflecting long-term trends. Let σ be the price volatility of the green certificate, and max(σ) be the historical maximum volatility.

8. The method according to claim 7, characterized in that, In step 3, Actor network output linkage coefficients: ； This can represent the coefficient of influence of carbon price changes on electricity prices. This indicates the strength of the response of green certificate price fluctuations to the carbon market.

9. The method according to claim 8, characterized in that, In step 3, Critic network evaluates the value of actions: ； in, It is a conditional expectation operator based on the market state at the next moment. It is a reward discount factor. It is a target Actor network.

10. The method according to claim 1, characterized in that, In step 4, Private blockchain, also known as electricity trading blockchain: It adopts a practical Byzantine fault-tolerant consensus mechanism and defines electricity trading transactions as quadruples: ； in, Represents a unique identifier for a transaction. Indicates the transaction type (fixed to "electricity"). Represents the amount of electricity traded. The consensus mechanism for electricity trading adopts the PBFT three-phase commit protocol, which satisfies latency constraints. ； Consortium blockchain: Defines the quota verification function as a boolean mapping: The specific implementation is as follows: ； in: For the address of the transacting party, Indicates carbon trading volume. Mapping from address to carbon allowance balance It is a carbon quota balance mapping function; Transaction binding: Generate cryptographic hash lock: ； By linking electricity, carbon, and green certificate transactions, cross-chain atomic swaps can be achieved. This refers to secret random numbers, electricity trading transactions, carbon trading transactions, and green certificate trading transactions.

Images