Method, system, equipment and medium for sharing parameters of nuclear power station water loss accident prediction model

Abstract

The invention provides a nuclear power station water loss accident prediction model parameter sharing method, system, device and medium, a model parameter set of a to-be-stored LOCA prediction model trained by each intelligent prediction node and training data are obtained, the training data comprise an LOCA identifier and LOCA operation data, and the LOCA operation data comprise LOCA core data; performing Hash operation on the model parameter set to obtain a to-be-stored Hash value; and encrypting and storing the model parameter set in a local storage node, and uploading the hash value, the LOCA identifier and the LOCA core data to an alliance chain, so that a secure sharing system of the LOCA prediction model parameters stored on the alliance chain and under the alliance chain can be established among the nuclear power stations, the credibility of the prediction model is verifiable, and the reliability of the prediction model is improved. And the security of model parameters and the resource utilization rate are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nuclear power, and in particular to a method, system, device and medium for sharing parameters of a nuclear power plant loss of coolant accident (LOCA) prediction model. BACKGROUND

[0002] With the development and application of nuclear power technology, how to ensure the safe operation of nuclear power plants has become one of the core focuses of research. The loss of coolant accident (LOCA) of a nuclear power plant refers to the rapid and large loss of coolant caused by the interruption of the primary heat transport (PHT) system, which leads to overheating of the reactor core of the nuclear power plant and causes serious safety hazards. Therefore, it is of great significance to systematically analyze and predict LOCA and provide effective decision support for emergency response to accidents.

[0003] Due to differences in safety strategies, data standards and management permissions between nuclear power plants, each nuclear power plant has long formed a data island, and there is a lack of a unified sharing system for accident prediction weight parameters. The training process of the deep learning model used for accident prediction is closed and lacks a transparent mechanism, the prediction logic and source are not easy to audit, which makes it difficult to establish trust between organizations and the credibility of the model is difficult to verify. The centralized management mode is difficult to ensure the tamper-proofing and privacy protection of the model parameters. Putting all weight parameters on the chain may also bring resource consumption and leakage risks. SUMMARY

[0004] The present application provides a method, system, device and medium for sharing parameters of a nuclear power plant loss of coolant accident (LOCA) prediction model, which can solve one of the problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] In a first aspect, a method for sharing parameters of a nuclear power plant loss of coolant accident (LOCA) prediction model is provided, comprising:

[0007] obtaining a set of model parameters of a LOCA prediction model trained by each intelligent prediction node to be stored, and training data, the training data comprising: a LOCA identifier and LOCA operation data, the LOCA operation data comprising a LOCA core data;

[0008] performing a hash operation on the set of model parameters to obtain a hash value to be stored; and,

[0009] storing the set of model parameters in a local storage node, and storing the hash value, the LOCA identifier and the LOCA core data on a consortium chain.

[0010] Based on the above technical solution, the model parameter set of the LOCA prediction model trained by each intelligent prediction node is obtained, and training data is obtained, the training data including: LOCA identification and LOCA operation data, the LOCA operation data including a LOCA core data; a hash operation is performed on the model parameter set to obtain a hash value to be stored; and the model parameter set is stored in the local storage node in an encrypted manner, and the hash value, the LOCA identification and the LOCA core data are stored on the alliance chain, so that a secure sharing system of the LOCA prediction model parameter stored on the chain and off the chain is established among nuclear power plants, the credibility of the prediction model can be verified, and the security of the model parameter and the resource utilization rate are ensured.

[0011] In a possible design of the first aspect, the sharing method further includes:

[0012] The intelligent prediction node queries the latest version number, the latest hash value and the latest designated storage node of the model parameter set of the LOCA prediction model through the alliance chain interface;

[0013] The intelligent prediction node obtains the latest model parameter set from the latest designated storage node based on the secure link; and

[0014] The intelligent prediction node verifies based on the latest hash value and the latest model parameter set, and when the verification is passed, the intelligent prediction node updates the LOCA prediction model by using the latest model parameter set.

[0015] In a possible design of the first aspect, the sharing method further includes:

[0016] The intelligent prediction node trains the LOCA prediction model by using new LOCA operation data, the new LOCA operation data corresponding to a new LOCA identification, and the new LOCA identification corresponding to a new hash value in a one-to-one manner.

[0017] In a possible design of the first aspect, the model parameter set is stored in the local storage node in an encrypted manner, and specifically includes:

[0018] generating a key and an initial vector required for encryption; and

[0019] performing AES-256 symmetric encryption processing on the model parameter set by using the key and the initial vector, and saving the encrypted file in a controlled directory of the local storage node.

[0020] In a possible design of the first aspect, the LOCA prediction model is a long short-term memory network (LSTM), and the set of model parameters includes weight matrices and bias terms of input gates, forget gates and output gates in the LSTM.

[0021] In a possible design of the first aspect, the training data is measured data or simulation data.

[0022] In a possible design of the first aspect, the LOCA operating data includes a pressurizer pressure, a coolant flow, a break area and a position number, a coolant temperature, a system pump rotating speed and an environmental parameter, the LOCA core data is the break area, the training label is a measured value of the pressurizer pressure, and the hash operation adopts an SHA-256 hash algorithm.

[0023] In a second aspect, a LOCA prediction model parameter sharing system is provided, including an application layer, an interface layer, a service layer and a storage layer, where:

[0024] The service layer is configured to obtain a set of model parameters of a LOCA prediction model trained by each intelligent prediction node and training data, the training data including a LOCA identifier and LOCA operating data, the LOCA operating data including a LOCA core data, and perform a hash operation on the set of model parameters to obtain a hash value to be stored;

[0025] The storage layer is configured to store the set of model parameters in a local storage node in an encrypted manner, and store the hash value, the LOCA identifier and the LOCA core data on a consortium chain.

[0026] In a third aspect, an electronic device is provided, including a processor and a memory coupled to the processor, the memory being configured to store a computer program, and the processor being configured to execute the computer program stored in the memory, so that the electronic device performs the sharing method in any possible implementation manner of the first aspect.

[0027] In a fourth aspect, a computer readable storage medium is provided, including a computer program or instructions, when the computer program or instructions are executed on a computer, causing the computer to perform the sharing method in any possible implementation manner of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or related technical descriptions will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0029] Figure 1 is a flow chart of a sharing method of a nuclear power plant LOCA prediction model parameter provided by an embodiment of the present application;

[0030] Figure 2 is a whole architecture diagram of a LOCA accident prediction data security sharing system based on a consortium chain provided by an embodiment two of the present application;

[0031] Figure 3 is an accident prediction and parameter trusted sharing flow chart of a LOCA accident prediction data security sharing system based on a consortium chain provided by an embodiment two of the present application;

[0032] Figure 4 is an enlarged view of the prediction diagram in Figure 3

[0033] Figure 5 is a data storage schematic diagram of a LOCA accident prediction data security sharing system based on a consortium chain provided by an embodiment two of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0035] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flow chart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the order in the flow chart. The terms "first", "second", etc. in the specification and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0037] Embodiment one

[0038] As Figure 1 ​As shown, the embodiment provides a sharing method of a nuclear power plant LOCA prediction model parameter, comprising:

[0039] S101, obtaining a model parameter set of a LOCA prediction model trained by each intelligent prediction node to be stored, and training data, the training data comprising: LOCA identification and LOCA operation data, the LOCA operation data comprising a LOCA core data;

[0040] S102, performing a hash operation on the model parameter set to obtain a hash value to be stored; and,

[0041] S103, encrypting and storing the model parameter set in a local storage node, and storing the hash value, the LOCA identification and the LOCA core data on a consortium chain.

[0042] Specifically, the intelligent prediction node can be arranged at the nuclear power plant side for predicting the LOCA core data. The LOCA core data can be the pressure of the stabilizer.

[0043] The LOCA prediction model used for prediction is a deep learning neural network. In the embodiment, in order to realize the prediction of the LOCA core data in the future time sequence, the LOCA prediction model is preferably a long short-term memory network (LSTM).

[0044] LSTM is a time recurrent neural network, which is specially designed to solve the long-term dependence problem of general recurrent neural networks (RNN). LSTM is a special RNN, compared with the general RNN which has only one state in a single loop structure, LSTM has four states in a single loop structure. LSTM keeps a persistent cell state continuously between loop structures, which is used to determine which information needs to be forgotten or continue to pass down. The LSTM unit is composed of an input gate, a forget gate, an output gate and a cell state, wherein the input gate determines how much of the input data at the current time needs to be saved to the cell state, the forget gate determines how much of the cell state at the last time needs to be retained to the current time, and the output gate controls how much of the current cell state needs to be output to the current output value.

[0045] Training the LSTM with the training data actually completes the iteration of the weight matrix and the bias corresponding to the input gate, the forget gate and the output gate.

[0046] The training data can be measured data or simulation data. When a real accident occurs, the measured data corresponding to the real accident can be used to iteratively train the LSTM, and when the prediction performance of the LSTM decreases, the simulation data can be used to iteratively train the LSTM.

[0047] The prediction performance of the LSTM can be analyzed based on mean relative error (MRE), mean square error (MSE), and mean absolute error (MAE).

[0048] In this embodiment, the LOCA operation data includes: pressurizer pressure, coolant flow, break area and position number, coolant temperature, system pump speed, and environmental parameters, etc., the LOCA core data is the break area, the output value of the LOCA prediction model is the predicted value of the pressurizer pressure, the training label is the measured value of the pressurizer pressure, and the model parameters are updated by measuring the difference between the predicted value and the measured value.

[0049] The hash operation can convert data into a flag that is closely related to each byte of the source data, and the hash operation also has the feature that it is difficult to find the inverse rule, so it is difficult to be cracked inversely.

[0050] In this embodiment, the hash operation adopts the SHA-256 hash algorithm. The SHA-256 hash algorithm is a 256-bit encryption hash algorithm, which mainly represents the data characteristics through a fixed-size hash value, and the principle is to generate a 256-bit binary hash value.

[0051] The local storage node can be the same physical entity as the intelligent prediction node, or a separate physical entity.

[0052] In one possible implementation, the step S103 of storing the model parameter set in the local storage node in an encrypted manner specifically includes:

[0053] generating a key and an initial vector (IV) required for encryption; and,

[0054] performing AES-256 symmetric encryption processing on the model parameter set using the key and the initial vector, and saving the encrypted file in the controlled directory of the local storage node.

[0055] Specifically, the symmetric encryption technology is a cryptography method in which the sender and the receiver use the same key to complete data encryption and decryption operations, and typical algorithms include data encryption standard (DES), advanced encryption standard (AES), etc. AES-256 is the use of a 256-bit key in the implementation of the AES algorithm. In order to improve security, AES is usually combined with IV, and IV is a randomly generated value that should be different each time encryption is performed to ensure that even if the same data is encrypted with the same key, the results of multi-level encryption are different.

[0056] The local storage node can store the model parameters in the JSON file format which has clear structure and strong universality, supports efficient data management and cross-platform sharing, and is convenient for parameter verification and synchronization between multiple nuclear power plant intelligent prediction nodes.

[0057] By storing the encrypted files in a controlled directory on a local storage node and utilizing access control and access log recording mechanisms, the security of the encrypted files can be guaranteed.

[0058] Consortium blockchains are a type of blockchain.

[0059] Blockchain is a distributed system that typically does not distinguish between master and slave nodes. All participants use peer-to-peer communication technology and are assumed to have equal status. Blockchain consensus algorithms encompass two layers of meaning: data consensus and business consensus. Data consensus refers to the consistency of the ledgers among nodes achieved through software and hardware algorithms. Business consensus refers to which participants should endorse and guarantee the business meaning represented by the on-chain data, ensuring the authenticity and trustworthiness of the business data.

[0060] Consortium blockchains are limited to consortium members. Because they only offer all or some of their functionality to members, read / write permissions and ledger rules are customized according to consortium rules. The consensus process on a consortium blockchain is controlled by pre-selected nodes. The consortium blockchain is jointly maintained by participating member institutions and provides a full suite of security management functions, including management, authentication, authorization, monitoring, and auditing, for participating members.

[0061] By storing the hash value, LOCA identifier, and LOCA core data on the consortium blockchain, only the hash value is stored instead of the complete model parameters. This not only ensures data traceability but also significantly reduces the on-chain data load.

[0062] In one possible implementation, the sharing method further includes:

[0063] The intelligent prediction node queries the latest version number, latest hash value, and latest designated storage node of the LOCA prediction model and model parameter set through the alliance link port;

[0064] The intelligent prediction node obtains the latest model parameter set from the latest designated storage node based on a secure link; and

[0065] The intelligent prediction node performs verification based on the latest hash value and the latest model parameter set. When the verification is successful, the intelligent prediction node updates the LOCA prediction model using the latest model parameter set.

[0066] Specifically, a secure link refers to data encryption and decryption methods based on technologies such as AES-256 symmetric encryption. When a smart prediction node needs to update its local LOCA prediction model, it obtains the latest hash value of the LOCA prediction model from the consortium blockchain, decrypts it from the latest designated storage node to obtain the latest model parameter set, calculates the hash value to be compared for the latest model parameter set, and verifies the trustworthiness of the latest model parameter set by comparing the latest hash value with the hash value to be compared. If the latest model parameter set is trustworthy, it is used to update the local LOCA prediction model.

[0067] In one possible implementation, the sharing method further includes:

[0068] The intelligent prediction node uses the new LOCA running data to train the LOCA prediction model. The new LOCA running data corresponds to a new LOCA identifier, and the new LOCA identifier corresponds one-to-one with a new hash value.

[0069] Specifically, the smart contracts of the consortium blockchain can ensure that the LOCA representation corresponds one-to-one with the new hash value and retain historical data before the iteration update, thereby achieving unique traceability of the entire training and prediction record process.

[0070] This embodiment also provides a system for sharing LOCA prediction model parameters, including: an application layer, an interface layer, a service layer, and a storage layer, wherein:

[0071] The service layer is used to obtain the set of model parameters of the LOCA prediction model trained by each intelligent prediction node, and the training data to be stored. The training data includes: LOCA identifier and LOCA running data, and the LOCA running data includes LOCA core data. The service layer is also used to perform a hash operation on the set of model parameters to obtain the hash value to be stored.

[0072] The storage layer is used to encrypt and store the model parameter set on a local storage node, and to upload the hash value, LOCA identifier, and LOCA core data to the consortium blockchain.

[0073] Specifically, the application layer manages the aforementioned consortium blockchain, event response, and user operations. The service layer provides services for consortium blockchain smart contracts, consensus algorithms, deployment, hash calculations, model training, weight processing, prediction, and invocation services. The storage layer stores LSTM model parameters, smart contract deployment addresses, and on-chain contract data. The interface layer provides multiple functional interfaces for data integration.

[0074] Based on the above technical solution, a set of model parameters for the LOCA prediction model trained by each intelligent prediction node, and training data are obtained to be stored. The training data includes LOCA identifiers and LOCA operation data, which includes LOCA core data. A hash operation is performed on the set of model parameters to obtain a hash value to be stored. The set of model parameters is then encrypted and stored on a local storage node. The hash value, LOCA identifier, and LOCA core data are uploaded to the consortium blockchain. In this way, a secure sharing system for LOCA prediction model parameters, with on-chain notarization and off-chain storage, can be established among nuclear power plants. The reliability of the prediction model is verifiable, and the security of the model parameters and resource utilization are guaranteed.

[0075] Example 2

[0076] like Figure 2 As shown, this embodiment proposes a secure data sharing system and method for LOCA incident prediction based on consortium blockchain, the technical implementation of which includes the following:

[0077] 1. Secure Shared System Architecture. The solution proposed in this embodiment adopts a layered architecture design, constructing a system architecture that includes an application layer, an interface layer, a service layer, and a storage layer, such as... Figure 2 As shown in the diagram. This layered architecture design helps improve the maintainability, scalability, and reusability of the system, providing a solid foundation for subsequent system development and optimization.

[0078] The application layer, serving as the interface between the system and the nuclear power plant, undertakes several important functions. In terms of blockchain management, it provides a convenient management interface and operating tools for the intelligent prediction model nodes of the nuclear power plant. Through this layer, operations such as initial configuration of the blockchain, node management, and permission settings can be performed, thereby gaining comprehensive control over the blockchain's operational status. Regarding interactive operations, the application layer provides an intuitive and easy-to-use command-line interface (CLI), through which the nuclear power plant can input data, initiate requests, and execute tasks.

[0079] As a crucial link in the system architecture, the interface layer enables efficient collaboration between the application layer and the service layer through standardized functional interfaces. This layer adopts a modular design, including core functional interfaces such as system initialization, event registration (RegisterIncident), and configuration deployment (DeployAddr), and also includes a separate emergency response interface (TriggerEmergency) to handle special scenarios. Through unified interface specifications and a loosely coupled architecture, the system significantly improves functional scalability and maintenance efficiency while ensuring communication reliability.

[0080] The service layer, as the core functional implementation module of the system, deeply integrates blockchain and deep learning technologies to build a closely collaborative service system: In terms of blockchain, the smart contract service deploys automated business logic based on preset rules, achieving efficient and accurate process handling; the consensus algorithm service adopts the PBFT mechanism to ensure data consistency and tamper-proof capabilities among nodes; and the hash calculation service generates unique identifiers for data, ensuring secure data transmission and storage. In terms of deep learning, the LSTM model training service constructs a high-precision prediction model through time-series data feature extraction and parameter optimization. The prediction service then uses this model to achieve real-time inference analysis, forming a closed-loop learning system. The weight processing service manages the LSTM weight parameters throughout their entire lifecycle, supporting dynamic updates and version control to optimize system performance. Furthermore, the deployment service enables flexible deployment of functional modules, and the calling service provides standardized interfaces, facilitating the combination and reuse of system functions. The various service modules intelligently collaborate, collectively endowing the system with autonomous learning and decision-making capabilities.

[0081] As the foundational support layer of the system architecture, the storage layer employs a hybrid storage strategy to achieve persistent data storage and secure management. This layer uses a smart contract storage module to fully record contract code and state data on the blockchain, ensuring data traceability and tamper-proof characteristics. Simultaneously, it utilizes a local weight file storage mechanism to efficiently store LSTM weight parameters in JSON file format, guaranteeing both the real-time performance of model calls and optimizing blockchain resource utilization. Furthermore, an on-chain contract data storage module based on blockchain distributed ledger technology provides a decentralized storage solution for critical data such as transaction records and contract execution results, significantly improving data security and reliability. This layered and categorized storage architecture effectively meets the system's differentiated storage needs for different types of data.

[0082] 2. Blockchain and Predictive Model Integration Mechanism. This embodiment, based on FISCO BCOS consortium blockchain technology, constructs a consortium blockchain network consisting of four nuclear power plants, achieving secure sharing and real-time synchronization of LOCA accident weight parameters. Considering the high sensitivity and security of nuclear power plant operation data, traditional centralized data sharing methods face risks such as data tampering, loss, and leakage. The consortium blockchain structure ensures the secure flow of data across nuclear power plants through mechanisms such as member controllability, data traceability, and fine-grained control of operational permissions.

[0083] The FISCO BCOS platform provides fine-grained access control, supporting precise authorization of user accounts, contract methods, and on-chain data, ensuring that unauthorized nodes cannot access or tamper with sensitive information. Its on-chain data is immutable and auditable, ensuring the authenticity and reliability of timestamps, content hashes, and uploaded identity information for all accident prediction records, thus building a trustworthy prediction data infrastructure in the nuclear power field.

[0084] The system employs the Byzantine Fault Tolerance (PBFT) consensus algorithm, possessing excellent fault tolerance capabilities. Even if some nodes malfunction, it can still ensure that the vast majority of normal nodes achieve a high degree of consistency, guaranteeing the security and real-time synchronization of LOCA weight parameter sharing. The consortium blockchain network generates node keys, certificates, and genesis blocks through automated configuration. Nodes interconnect through preset P2P, RPC, and front-end interface ports, ensuring efficient and stable transaction synchronization and state updates.

[0085] Based on this, this system designs and deploys fully functional smart contracts, using the FISCO BCOS Table contract mechanism to achieve structured storage of incident data. Key fields include the incident unique identifier (incident_id), the break size (break_size_cm2), and the incident weight parameter (weight). The contract defines complete incident management interfaces for adding, querying, updating, deleting, and triggering emergency responses. Through events such as IncidentReported, IncidentUpdated, IncidentDeleted, and EmergencyTriggered, the system ensures on-chain transparency and traceability of the operation process.

[0086] The smart contract employs conditional queries, data verification, and anomaly rollback mechanisms to ensure the legality of data operations and system stability. To achieve efficient interaction with the smart contract, a client module based on the FISCO BCOS Java SDK was developed. The client utilizes the Spring framework to initialize on-chain connections and authentication, supporting contract deployment, address management, and full lifecycle operations on accident data. During invocation, it incorporates on-chain event listening to provide real-time feedback on operation results, forming a closed-loop control. The client uses the BigInteger type to handle weight parameters, avoiding floating-point errors, and possesses a robust anomaly handling mechanism to ensure system robustness. Through close collaboration between the smart contract and the client, this system achieves secure sharing, consistency verification, and rapid response of LOCA weight parameters among nuclear power plants. It fully leverages the decentralized, tamper-proof, and traceable characteristics of blockchain, meeting the high security requirements of nuclear emergency decision-making and cross-organizational collaboration, and promoting the integrated application of deep learning prediction models and blockchain technology.

[0087] 3. Trusted Sharing Operation Process and Working Mechanism. This system addresses the safe sharing requirements of predicted parameters for Loss of Coolant Accidents (LOCA) in nuclear power plants. It designs and implements a comprehensive, end-to-end safe sharing system integrating "prediction-proofing-verification-updating," such as... Figure 3 and Figure 4As shown, the system combines a Long Short-Term Memory (LSTM) neural network model with blockchain technology to achieve trusted management of accident prediction results and model parameters, ensuring traceability of data sources and verifiability of parameters, thereby improving the security and transparency of cross-organizational prediction collaboration. Its main process includes the following four stages:

[0088] (1) Accident data collection and model prediction

[0089] The industrial simulation platform models coolant loss accidents under different breach sizes and operating conditions, generating multiple sets of time-series data. The system calls a pre-trained LSTM model to predict key parameters such as regulator pressure, outputting a continuous prediction sequence to support accident evolution trend assessment and response judgment.

[0090] (2) Model parameter extraction and summary calculation

[0091] After model training is complete, the system automatically extracts nine key parameters from the LSTM model that are relevant to this prediction. The LSTM has three gates used to control the information transmission path, which are:

[0092] Forgotten Gate f t It controls how much information is forgotten from the internal state of the previous hidden layer with a certain probability. It looks at the previous output h. t-1 and the current input x t Determine the previous state c t-1 Information that needs to be forgotten.

[0093] Input gate i t : Responsible for controlling how much information from the current input needs to be stored in the candidate states. At this point, the new state c... t It can be obtained through formula (1). (1)

[0094] in: This represents the current cell state vector. This represents the cell state at the previous moment; Candidate cell state; This indicates element-wise multiplication (Hadamard product). , These are the forget gate and the input gate vector, respectively.

[0095] Output gate o t : Responsible for controlling how much information the current hidden layer needs to output to the external state. The current internal state c t Normalization is performed using the tanh function, followed by element-wise multiplication by the output gate o. t The output value h is obtained. tThe calculation process is shown in formula (2). (2)

[0096] in: This is the hidden layer output vector at the current time step; The hyperbolic tangent activation function maps the input values ​​to... Interval.

[0097] The gate values ​​of LSTM are between (0,1), indicating that information can be allowed to pass through at a certain proportion. The calculation formulas for the three gates are as follows. (3) (4) (5)

[0098] in: : is the current input vector; : Output of the hidden layer at the previous time step; : is the Sigmoid activation function, which maps the input to the (0,1) interval; : These are the input weight matrices corresponding to the input gate, forget gate, and output gate, respectively; : These are the hidden layer weight matrices corresponding to the input gate, forget gate, and output gate, respectively; These represent the bias terms for the three gates. From this, we can derive nine important model parameters: the weight matrices and bias terms for the input gate, forget gate, and output gate, i.e., W. i W f W o U i U f U o b i b f b o These parameters are continuously updated during training and are the core elements for the model to learn time series features and make effective predictions. Subsequently, the SHA-256 hash algorithm is used to calculate a digest of this parameter set, generating a globally unique hash value, which serves as a unique fingerprint of the prediction results and model structure.

[0099] (3) Storage strategy

[0100] In this embodiment, as Figure 5As shown, the system comprises four intelligent prediction nodes for nuclear power plants, which need to collaboratively process LOCA accident data and the weight parameters of the LSTM model. Considering the inherent limitations of on-chain storage capacity and transmission bandwidth in blockchain systems, directly uploading the complete model parameters would result in significant performance overhead and resource consumption. Therefore, a hybrid storage mechanism combining on-chain and off-chain storage is adopted in the storage architecture to achieve an optimal balance between data security and system performance.

[0101] At the on-chain storage layer, the system carefully selects key data for on-chain storage, including the unique identifier ID of the incident, the core incident parameter of the breach size, and the hash values ​​corresponding to the weight parameters of the LSTM model. By storing only the hash values ​​instead of the complete parameters, the traceability of the model data is ensured, and the on-chain data load is significantly reduced.

[0102] In the off-chain storage layer, each node uses a well-structured and highly versatile JSON file format to record the nine core parameters of the LSTM model. This format supports efficient data management and cross-platform sharing, facilitating parameter synchronization and verification across multiple power plants.

[0103] This combined on-chain and off-chain storage approach has significant advantages. On the one hand, it effectively saves on-chain storage resources, reducing storage costs and performance overhead; on the other hand, by verifying the consistency between the hash value stored on-chain and the actual parameters stored off-chain, it can further ensure the security and integrity of the data.

[0104] (4) Model update and closed-loop iteration

[0105] To achieve reliable management of accident prediction results and continuous model evolution, this embodiment constructs a collaborative closed-loop mechanism based on an LSTM model and blockchain. After the model completes accident prediction, the system extracts nine key parameters, calculates their SHA-256 hash values, and uploads these hash values ​​along with the corresponding accident number and breach area to the blockchain, achieving secure sharing and tamper-proof evidence storage of the prediction results. Specific parameter files are securely stored offline by the system. During the file generation stage, the system performs AES-256 symmetric encryption on the parameter content and saves the encrypted file in a locally controlled directory. The encryption key and initialization vector (IV) are generated by the system security module and protected through the key management module or a controlled interface to prevent plaintext storage locally. To prevent data leakage or tampering, the system combines operating system permission control and access log recording mechanisms to strictly limit the scope of file access. When it is necessary to verify the authenticity of parameter files or retrain the model, the system can obtain the corresponding key and IV through the key management module, decrypt the encrypted file, recover the original parameter content, recalculate the SHA-256 hash value of the file, and compare it with the digest stored on the blockchain. This ensures the consistency between the offline file and the on-chain record, achieving reliable verification of the prediction results. Through this mechanism, the prediction results can be securely stored off-chain and tamper-proof and verifiable on-chain, constructing a reliable closed loop of "verifiable on-chain and manageable off-chain".

[0106] In subsequent use, the nuclear power plant's intelligent prediction nodes can verify the uniqueness and integrity of prediction results through the blockchain interface, ensuring model trustworthiness. Regarding uniqueness, the smart contract stipulates that each accident number can only correspond to one valid prediction summary. When a new model generates results and uploads them to the blockchain, the system automatically updates the corresponding hash and retains historical version information, thereby preventing duplicate registrations or result conflicts and achieving end-to-end unique traceability of prediction records. Regarding integrity, after receiving the prediction results, the node retrieves the stored hash value and timestamp information from the blockchain based on the accident number and recalculates the hash value of the locally stored parameter file for comparison. If the two match, it indicates that the file content has not been tampered with, and the verification is successful. Through the above mechanisms, the system ensures the unique identifier of the prediction results and the consistency of the data content, achieving trusted verification and secure sharing of prediction results. Simultaneously, the system supports parameter version replacement and recording: when new data arrives or prediction performance degrades, model retraining can be initiated based on historical records, and the new round of prediction parameters can be re-hashed and uploaded to the blockchain, replacing the old identifier and retaining version history. Through a closed-loop process of prediction—sharing—verification—updating, the system achieves adaptive learning and continuous trust assurance for accidents in multiple scenarios.

[0107] Compared to traditional centralized sharing mechanisms and existing full-scale on-chain solutions, this embodiment proposes a secure synchronization mechanism that integrates consortium blockchain and off-chain storage to address the cross-organizational sharing requirements of deep learning model prediction weight parameters in nuclear power plant Loss of Coolant Accident (LOCA) scenarios. This mechanism combines the Practical Byzantine Fault-Tolerant (PBFT) consensus algorithm with the FISCO BCOS consortium blockchain platform to construct a secure, efficient, and verifiable prediction model parameter sharing system with the following significant advantages:

[0108] 1) Construct a reliable prediction parameter sharing architecture across nuclear power plants.

[0109] This embodiment uses an LSTM model to train key accident parameters, extracts key weight data, generates hash digests, and stores them on the blockchain, enabling verifiable and secure sharing of model weights. While ensuring model privacy, it achieves synchronous consistency verification among multiple nodes.

[0110] 2) The hybrid architecture of "on-chain evidence storage + off-chain storage" is adopted to effectively improve data synchronization efficiency and system scalability.

[0111] The weight parameter file is stored locally or on a trusted server, and only its hash digest and metadata are uploaded to the blockchain. This avoids the performance bottleneck caused by uploading big data to the blockchain and realizes the "light chain, heavy storage" design concept in data sharing.

[0112] 3) A secure, transparent and efficient weight parameter sharing mechanism is proposed, which takes into account both data privacy protection and system consistency verification.

[0113] This embodiment employs a hybrid architecture of "on-chain evidence storage + off-chain storage," uploading the hash values ​​of key weight parameters of the deep learning model to the blockchain to avoid privacy leaks and storage burdens caused by directly uploading complete parameters to the blockchain. During data synchronization, the system performs consistency checks between the on-chain hash and the off-chain weight parameters to ensure that the parameters shared across nuclear power plants are authentic, valid, and untampered with. Furthermore, this mechanism boasts excellent scalability and flexibility, making it suitable for critical application scenarios such as nuclear power accident emergency response, accident-assisted decision-making, and cross-organizational collaboration, effectively enhancing the system's security, transparency, and credibility.

[0114] Specific application examples

[0115] In a typical implementation environment, this application example is deployed on the VMware® Workstation 17 Pro virtualization platform, running Ubuntu 18.04.1 LTS. The blockchain platform uses FISCO BCOS 2.x, and the development tool is IntelliJ IDEA, which supports SSH remote connections. For the deep learning component, the PaddlePaddle 2.x framework is used, running in a Python 3.8 environment. Hardware-wise, the system configuration includes an NVIDIA Tesla V100 GPU or a graphics processing unit with equivalent computing power, and 32 GB of memory. This environment ensures the stable operation of the blockchain platform and the training and inference of the deep learning model, thus supporting the smooth execution of processes such as incident prediction parameter extraction, on-chain evidence storage, cross-node sharing and verification in this application example.

[0116] In the above environment, this application example can be implemented by following these steps:

[0117] Step 1: Accident Data Acquisition and Preprocessing. This dataset was obtained by simulating a Loss of Coolant Accident (LOCA) using an industrial-grade nuclear power plant (NPP) simulation platform. The simulation conditions were set as a loss of coolant accident under 100% reactor power operation, with breach sizes of 0.9 cm² and 2.0 cm² to cover two typical accident scenarios: small and medium breaches. The raw time-series data collected during the simulation included multiple operating parameters such as pressurizer pressure, coolant flow rate, and core temperature. After acquisition, the system preprocessed the data, including outlier removal, missing value imputation, time step alignment, and normalization, and uniformly converted it into a standardized time-series input format that can be recognized by deep learning frameworks to ensure data integrity and training effectiveness.

[0118] Step Two: Model Training and Parameter Extraction. This application example utilizes the PaddlePaddle 2.x framework to train an LSTM model on a Tesla V100 GPU to predict key accident parameters such as pressurizer pressure. Training data comes from nuclear power plant accident simulations or historical operation records, covering sensor monitoring data under different breach conditions, including pressurizer inlet and outlet pressures, coolant flow rate, breach area and location number, coolant temperature, system pump speed, and environmental parameters. Each parameter is sampled at a fixed time step to form a time series input, used to capture the dynamic changes of the system. The training label is the measured value of the pressurizer pressure at a future time. After training, the model can output pressure predictions for several future time steps, forming a continuous pressure curve, while providing corresponding timestamps and prediction error information. The results show that the predicted curve is highly consistent with the measured values ​​overall, and the predictions of key inflection points and peak positions are accurate. The maximum error for the small breach condition is approximately 0.2696 × 10⁻⁶.7 Pa, accounting for 17.8% of the peak pressure, with a maximum error of approximately 0.1615 × 10⁻⁶ Pa under large-break conditions. 7 Pa, ranging from 9.7% to 10.7%, indicates that the model possesses strong temporal feature capture capabilities and prediction stability under various operating conditions. After training, the system automatically extracts nine core weight parameters, including input gate weights, forget gate weights, output gate weights, unit state weights, input gate biases, forget gate biases, output gate biases, unit state biases, and the weights of the final fully connected output layer. These parameters can be used for secure off-chain storage and on-chain hash notarization, enabling reliable management of prediction results and supporting continuous training and updates of the model as new data arrives.

[0119] Step 3: Parameter Notarization, Off-Chain Storage, and On-Chain Verification. After extracting the nine core weight parameters of the LSTM model in Step 2, the system first performs a SHA-256 hash operation on the parameter set to generate a globally unique digest value, used to identify the uniqueness and completeness of each prediction result. Subsequently, the system completes on-chain notarization through a smart contract on the FISCO BCOS 2.x platform, writing the incident number, breach area, and parameter hash value into the blockchain ledger to achieve immutable records. The smart contract includes data validity verification, anomaly rollback mechanisms, and event logging (such as IncidentReported, IncidentUpdated, etc.) to ensure that the notarization operation has complete traceability and timestamp information. Considering the limited on-chain storage resources, the complete parameter file is stored locally on each prediction node in JSON format for easy access and cross-node synchronization. During parameter sharing and verification, the client reads the on-chain hash value through the Java SDK and compares it with the calculation result of the local JSON file to achieve on-chain verification and consistency verification. If verification fails, the system will automatically trigger an alarm and initiate a rollback mechanism. This will restore the on-chain record to the most recent successful version and roll back the local JSON parameter file to the last complete and validated backup, thus ensuring data consistency, integrity, and real-time performance in a multi-node environment. This combined on-chain and off-chain storage and verification strategy not only guarantees the tamper-proof and traceability of prediction parameters but also addresses system performance and efficient cross-node synchronization needs. Furthermore, the rollback mechanism enhances the system's fault tolerance and data security in abnormal situations.

[0120] Step Four: Parameter Update and Closed-Loop Iteration. When model prediction performance deteriorates or new accident data appears, the system automatically triggers a retraining mechanism to incrementally or fully train the LSTM model, generating new nine core weight parameters and their corresponding SHA-256 hash values. The new hash values ​​and related information are updated to the blockchain via a smart contract, while historical versions are retained to form a complete version chain, enabling traceable management of the model parameter update process. Each prediction node can query the latest version of the parameter hash and corresponding version number through the blockchain interface, and download the JSON parameter file directly from the designated storage node accordingly, achieving reliable synchronization of off-chain data. Nodes recalculate the hash value of the parameter file locally and compare it with the on-chain evidence. After successful verification, the local model is updated, ensuring the consistency and trustworthiness of parameters across multiple nodes. During synchronization or verification, if an anomaly occurs such as a pull failure or hash mismatch, the system automatically triggers a rollback mechanism to restore the local parameters to the most recent valid version and issues an alarm, ensuring data security and integrity. Through this closed-loop iterative process of prediction, sharing, verification, and updating, the system can continuously optimize model performance, adapt to different accident conditions, and achieve dynamic and reliable management of accident prediction parameters and multi-scenario adaptive learning.

[0121] Step 5: Evaluate System Performance. To verify the system's stability and efficiency, this embodiment conducted performance tests in a consortium blockchain environment. The tests covered three aspects: response time, TPS (transactions per second), and synchronization accuracy. Response time was used to calculate the average latency from data request to blockchain feedback; TPS was measured by multi-threaded concurrent requests to assess the system's processing throughput; and synchronization accuracy was used to evaluate the consistency of multiple nodes during parameter synchronization. Experimental results show that, while ensuring data trustworthiness and tamper-proofing, this system can achieve millisecond-level response, high transaction throughput, and near 100% synchronization accuracy, thus fully meeting the stringent requirements for high security and real-time performance in nuclear power emergency scenarios.

[0122] In summary, this application example, combining deep learning and blockchain technology in a virtualized environment, achieves trusted sharing and verification of accident prediction parameters in a LOCA (Location of Response and Compatibility) scenario. Through the complete process of steps one through six, the system not only efficiently completes accident data prediction and parameter storage but also ensures consistency and security during cross-node synchronization. Experimental results show that the system performs excellently in terms of response time, TPS (Transactions Per Second), and synchronization accuracy, meeting the stringent requirements of real-time performance and high security for nuclear power plant emergency response. Therefore, this application has significant practical value and potential for widespread application.

[0123] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0124] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0125] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0126] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0127] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0128] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0129] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0130] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for sharing parameters in a LOCA prediction model for a nuclear power plant loss-of-water accident, characterized in that, include: Obtain the set of model parameters of the LOCA prediction model trained by each intelligent prediction node, which is to be stored, as well as the training data, which includes: LOCA identifier and LOCA running data, and the LOCA running data includes LOCA core data; Perform a hash operation on the model parameter set to obtain the hash value to be stored; and, The model parameter set is encrypted and stored on a local storage node, and the hash value, LOCA identifier, and LOCA core data are uploaded to the consortium blockchain.

2. The sharing method as described in claim 1, characterized in that, The sharing method also includes: The intelligent prediction node queries the latest version number, latest hash value, and latest designated storage node of the LOCA prediction model and model parameter set through the alliance link port; The intelligent prediction node obtains the latest model parameter set from the latest designated storage node based on a secure link; and The intelligent prediction node performs verification based on the latest hash value and the latest model parameter set. When the verification is successful, the intelligent prediction node updates the LOCA prediction model using the latest model parameter set.

3. The sharing method as described in claim 1, characterized in that, The sharing method also includes: The intelligent prediction node uses the new LOCA running data to train the LOCA prediction model. The new LOCA running data corresponds to a new LOCA identifier, and the new LOCA identifier corresponds one-to-one with a new hash value.

4. The sharing method as described in claim 1, characterized in that, The model parameter set is encrypted and stored on a local storage node, specifically including: Generate the key and initialization vector required for encryption; and, Using the key and initialization vector, the model parameter set is subjected to AES-256 symmetric encryption, and the encrypted file is saved in the controlled directory of the local storage node.

5. The sharing method as described in claim 1, characterized in that, The LOCA prediction model is a Long Short-Term Memory (LSTM) network, and the model parameter set includes the weight matrices and bias terms of the input gate, forget gate, and output gate in the LSTM.

6. The sharing method as described in claim 1, characterized in that, The training data is either measured data or simulated data.

7. The sharing method as described in claim 1, characterized in that, The LOCA operating data includes: regulator pressure, coolant flow rate, rupture area and location number, coolant temperature, system pump speed, and environmental parameters. The core data of LOCA is the rupture area, the training label is the measured value of regulator pressure, and the hash operation uses the SHA-256 hash algorithm.

8. A system for sharing parameters of a LOCA prediction model, characterized in that, include: The application layer, interface layer, service layer, and storage layer are as follows: The service layer is used to obtain the set of model parameters of the LOCA prediction model trained by each intelligent prediction node, and the training data to be stored. The training data includes: LOCA identifier and LOCA running data, and the LOCA running data includes LOCA core data. The service layer is also used to perform a hash operation on the set of model parameters to obtain the hash value to be stored. The storage layer is used to encrypt and store the model parameter set on a local storage node, and to upload the hash value, LOCA identifier, and LOCA core data to the consortium blockchain.

9. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; and The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the shared method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising a computer program or instructions that, when executed on a computer, cause the computer to perform the sharing method as described in any one of claims 1-7.

Images