Multi-modal convolution and full-connection hybrid network data driving method of digital twin nuclear reactor core

Abstract

The invention discloses a multi-mode convolution and full-connection hybrid network data driving method for a digital twin nuclear reactor core. The method comprises the following steps: 1, establishing a data driving library according to a large amount of data monitored by a nuclear reactor; 2, multi-modal features are extracted in parallel according to different parameters of reactor core monitoring; 3, carrying out dynamic fusion on the extracted multi-modal features; 4, building a multi-task mixed output system based on the extracted multi-modal fusion features, and obtaining a data driving model; and 5, according to measurement parameters transmitted in real time by the reactor core of the nuclear reactor, taking the measurement parameters as input of the data driving model, obtaining a correction value of the digital twin reactor core in the state, and assisting dynamic accurate mapping of the solid reactor core characteristics to the digital reactor core. According to the method disclosed by the invention, the construction of the data driving core algorithm of the digital twin nuclear reactor core is realized, dynamic accurate mapping of a large number of actually measured data of different types on the digital reactor core is assisted, an intelligent driving engine with physical interpretability is provided for the digital twin reactor core, and the real-time interaction capability of the digital twin reactor core is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twins of nuclear reactor cores, and specifically provides a multi-modal convolution and fully connected hybrid network data-driven method for digital twin nuclear reactor cores. Background Art

[0002] As the core energy-generating equipment in nuclear power plants, nuclear reactor cores feature high temperature, high pressure, and strong radioactivity. Their safety and economic efficiency directly determine the operational quality of nuclear power systems. However, the extreme environment within nuclear reactor cores (such as strong radiation, high temperature, and high pressure) severely limits sensor deployment, making direct measurement of key parameters such as core power distribution and fuel rod deformation difficult. Conventional operations and maintenance methods, constrained by limited measurement methods and expensive experimental costs, struggle to gain a deep understanding of the specific conditions within nuclear power plants, directly hindering the scientific and economical nature of operations and maintenance decisions. Therefore, constructing a digital twin of the nuclear reactor core—by integrating advanced modeling, big data, and artificial intelligence technologies, establishing a dynamic mapping mechanism between the physical and digital cores, and forming a new logical framework for nuclear reactor core industrial software—has become an inevitable choice for breaking through physical boundary constraints and achieving intelligent upgrades for nuclear reactors.

[0003] Currently, mainstream digital twin core solutions can be roughly divided into two categories. The first is solely data-driven: Using measured data, proxy models are constructed through methods such as data fitting or neural networks, replacing traditional numerical calculations to display or monitor key core parameters in real time. This approach has the advantages of facilitating dynamic management and operating speed. However, the displayed key parameters lack self-consistency and interpretability, making it difficult to develop reliable self-evolution capabilities. This makes it unsuitable for complex physical processes such as nuclear reactor cores, and it is impossible to provide reliable operation and maintenance guidance for the physical core based on the digital core. The second category relies on both mechanism-driven and data-driven approaches. First, the mechanism-driven approach involves building a reliable core model based on the core design and manufacturing parameters. Preliminary numerical calculations are performed using reliable theoretical models to simulate the core's internal physical fields. Second, the data-driven approach utilizes advanced algorithms such as artificial intelligence to refine the model-driven approach based on extensive field data from the nuclear power plant and within the constraints of the actual core conditions, thereby enabling the physical core to guide the digital core. This type of approach offers high credibility and strong interpretability, providing reliable guidance for actual core operations. However, it places greater demands on the accuracy and real-time nature of data-driven approaches. This demonstrates that data-driven approaches play a crucial role in the implementation of digital twin cores, directly impacting the interactivity between the physical and digital cores. Appropriate data-driven approaches can facilitate the realization of digital twins for reactor cores.

[0004] At present, most of the existing data-driven implementations are based on the first type of digital twin core solution, which mainly focuses on the prediction of core key parameters. It only reflects the restoration of key parameters of the physical core by the digital twin core, and cannot fundamentally establish a two-way accurate mapping and synchronous presentation of the physical core and the digital twin core. Although the second type of digital twin solution is expected to combine mechanism-driven and retain the systematic and self-consistent nature of numerical calculation, in the numerical driving part, (1) it is impossible to use the actual measured data of the nuclear power plant in real time. How to transfer the data characteristics generated by the operation of the physical core to the numerical calculation in real time is still a difficult problem that needs to be solved; (2) the data generated in real time by the nuclear power plant is complex, and conventional methods are difficult to comprehensively consider the various forms from scalar data to high-dimensional data containing spatiotemporal characteristics; (3) the extreme environment of the core causes inherent deviations between the sensor measurement results and the real physical state (such as temperature and power measurement uncertainty), and factors such as sensor aging and radiation damage may cause data anomalies or failures. The existing methods lack the ability to quantitatively evaluate uncertainty in real time and identify abnormal data. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a multimodal convolution and fully connected hybrid network data-driven method for a digital twin nuclear reactor core, build a multimodal convolution and fully connected hybrid neural network architecture for different forms of measured data of nuclear power plants, integrate multiple measured parameters, realize dynamic and accurate mapping of physical core characteristics to digital cores, and enhance the real-time interactive capabilities of digital twin cores.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The data-driven approach of a multimodal convolutional and fully connected hybrid network for a digital twin nuclear reactor core consists of the following steps:

[0008] Step 1: Establish a data-driven library based on a large amount of data monitored by the nuclear reactor: the data-driven library includes at least two of the total nuclear power of the nuclear reactor, i.e., the rate at which nuclear reactions release energy; the axial power offset within the reactor, i.e., the difference between the upper power and the lower power; the relative power of the core; the position of each group of control rods within the core; the primary coolant flow rate; the temperature of the primary coolant at the core inlet and outlet; the boron concentration of the primary coolant; and the response current of each detector within the core.

[0009] Step 2: Extract multimodal features in parallel for different parameters in the data-driven library: Use a fully connected network to process scalar parameters, deploy convolutional networks in parallel to extract high-dimensional features, and use a channel-incrementing model to achieve nonlinear projection of low-dimensional features into high-dimensional space, enhancing feature expression capabilities.

[0010] Step 3: Dynamically fuse the extracted multimodal features: Establish a learnable dynamic weight matrix in the parameterized fusion layer, and optimize feature contributions through a dual mechanism of initial value presetting and adaptive adjustment. In terms of multimodal collaboration, feature decoupling technology is used to separate common or individual features to achieve effective cross-modal information complementarity.

[0011] Step 4: Build a multi-task hybrid output system based on the extracted multimodal fusion features to obtain a data-driven model: Design a multi-task output architecture consisting of a feature fusion layer and a deep fully connected layer cascade. Use dimensionality reduction mapping to convert the fusion features into predicted values ​​for various parameters of the digital twin core, thereby achieving parameter regression.

[0012] Step 5: Based on the measurement parameters transmitted in real time by the nuclear reactor core, they are input into the data-driven model constructed in the above steps, and multiple measured parameters are integrated to obtain the correction value of the digital twin core in this state, which helps to dynamically and accurately map the physical core characteristics to the digital core, ensuring that the digital core can timely and accurately reflect the state changes of the physical core.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] 1) A multimodal convolution and fully connected hybrid network data-driven approach for digital twin nuclear reactor cores was proposed. A data-driven library was built based on a large amount of measured data from the core. This solved the problem that traditional numerical calculations could not immediately utilize measured data, and achieved real-time dynamic and accurate mapping of physical core characteristics to digital cores. This is an effective supplement to the data-driven module in the digital twin core solution and reflects the real-time dynamic interaction characteristics of the digital twin core.

[0015] 2) By building a multi-task output architecture with multiple branches, dynamic fusion, and deep mapping, and achieving high-precision state regression through a feature dimension reduction design, the data-driven part of the digital twin nuclear reactor core has enhanced its strong adaptability to different types of data, high generalization ability, and multi-feature extraction capabilities;

[0016] 3) The method of the present invention reduces the error when the measured core parameters are transferred to the digital core. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1As shown in the figure, the multimodal convolution and fully connected hybrid network data-driven method for the digital twin nuclear reactor core of the present invention selects the core relative power and the two-dimensional response current of each detector inside the core as the measured core parameters. The correction of parameters such as the rod position of each group of control rods in the core of the digital core, the temperature of the primary coolant at the core inlet, and the boron concentration content of the primary coolant are illustrated by example. The specific steps are as follows:

[0020] Step 1: Build a data-driven library based on the large amount of data monitored by the nuclear reactor. The selected parameters include the core relative power, the two-dimensional response current of each detector inside the core, the position of each group of control rods in the core, the temperature of the primary coolant at the core inlet, and the boron concentration of the primary coolant.

[0021] Step 2: Parallel extraction of multimodal features for different core monitoring parameters: A fully connected network is used to process the scalar parameter core relative power, and a convolutional network is deployed in parallel to extract the two-dimensional response current of each detector within the 160-dimensional core. A channel-incrementing model is used to achieve nonlinear projection of low-dimensional features into a high-dimensional space, enhancing feature expression capabilities.

[0022] Step 3: Dynamically fuse the extracted multimodal features: Establish a learnable dynamic weight matrix in the parameterized fusion layer, and optimize the feature contribution through a dual mechanism of initial value preset and adaptive adjustment:

[0023] F=λ·F CNN +(1-λ)F DNN #(1)

[0024] Among them, the initial value of weight coefficient λ is 0.2, F CNN Represents the characteristic of the scalar parameter core relative power, F DNN represents the 160-dimensional detector current signal response characteristic, and F represents the scalar parameter characteristic F CNN Feature contribution after fusion with the 160-dimensional detector current signal response feature;

[0025] In terms of multimodal collaboration, feature decoupling technology is used to separate common or individual features to achieve effective complementarity of cross-modal information;

[0026] Step 4: Build a multi-task hybrid output system based on the extracted multimodal fusion features to obtain a data-driven model: Design a multi-task output architecture consisting of a feature fusion layer and a deep fully connected layer cascade. Use dimensionality reduction mapping to convert the fusion features into predicted values ​​for various parameters of the digital twin core, thereby achieving parameter regression.

[0027] Step 5: Based on the measurement parameters transmitted in real time by the nuclear reactor core, they are input into the data-driven model constructed in the above steps, and multiple measured parameters are integrated to obtain the correction value of the digital twin core in this state, which helps to dynamically and accurately map the physical core characteristics to the digital core, ensuring that the digital core can timely and accurately reflect the state changes of the physical core.

[0028] The method of the present invention can be used to realize data-driven digital twin nuclear reactor cores, establish a data-driven library for the complex data detected by nuclear power plants, build a multimodal and fully connected hybrid neural network architecture, integrate multiple measured parameters, realize dynamic and accurate mapping of physical core characteristics to digital cores, and enhance the real-time interactive capabilities of digital twin cores.

[0029] Table 1. Error changes of measured data under the data-driven method

[0030]

[0031] As shown in Table 1, the present invention's method, when implementing a digital twin core solution, uses data-driven approaches based on various forms of measured data, such as core relative power and the two-dimensional response currents of detectors within the core. This method reduced the rod position error of each group of control rods within the core from -3.20846 cm to -0.80525 cm, the temperature error of the primary coolant at the core inlet from 0.535 K to 0.076 K, and the boron concentration error of the primary coolant from -0.661 ppm to -0.098 ppm. These results demonstrate the effectiveness of the present method.

Claims

1. A data-driven approach using a multimodal convolutional and fully connected hybrid network for digital twin nuclear reactor cores, characterized by: Here are the steps: Step 1: Establish a data-driven library based on a large amount of data monitored by the nuclear reactor: the data-driven library includes at least two of the total nuclear power of the nuclear reactor, i.e., the rate at which nuclear reactions release energy; the axial power offset within the reactor, i.e., the difference between the upper power and the lower power; the relative power of the core; the position of each group of control rods within the core; the primary coolant flow rate; the temperature of the primary coolant at the core inlet and outlet; the boron concentration of the primary coolant; and the response current of each detector within the core. Step 2: Extract multimodal features in parallel for different parameters in the data-driven library: A fully connected network is used to process scalar parameters, and convolutional networks are deployed in parallel to extract high-dimensional features. The channel-incrementing mode is used to achieve nonlinear projection of low-dimensional features into high-dimensional space, thereby enhancing feature expression capabilities. Step 3: Dynamically fuse the extracted multimodal features: Establish a learnable dynamic weight matrix in the parameterized fusion layer, and optimize feature contributions through a dual mechanism of initial value presetting and adaptive adjustment. In terms of multimodal collaboration, feature decoupling technology is used to separate common or individual features to achieve effective cross-modal information complementarity. Step 4: Build a multi-task hybrid output system based on the extracted multimodal fusion features to obtain a data-driven model: Design a multi-task output architecture consisting of a feature fusion layer and a deep fully connected layer cascade. Use dimensionality reduction mapping to convert the fusion features into predicted values ​​for various parameters of the digital twin core, thereby achieving parameter regression. Step 5: Based on the measurement parameters transmitted in real time by the nuclear reactor core, they are input into the data-driven model constructed in the above steps, and multiple measured parameters are integrated to obtain the correction value of the digital twin core in this state, which helps to dynamically and accurately map the physical core characteristics to the digital core, ensuring that the digital core can timely and accurately reflect the state changes of the physical core.

Images