A real-time online intelligent prediction method for high-dimensional parameters of a nuclear reactor

Abstract

The application provides a nuclear reactor high-dimensional parameter real-time online intelligent prediction method, and belongs to the technical field of nuclear reactor operation monitoring and intelligent prediction, and particularly relates to a technology for unified transient monitoring and real-time prediction of massive, high-dimensional and dynamic time sequence parameters generated in the operation process of a reactor by using a deep learning model, which comprises feature extraction and time sequence learning by a Transformer and a long short-term memory (LSTM) neural network, and the generalization ability of the prediction model under different initial conditions, system transients and external intervention is realized by combining a deep operator network (DeepONet), so as to improve the intelligent operation and maintenance level of the nuclear reactor and the decision support efficiency of the operator.

Description

Technical Field

[0001] This invention belongs to the field of nuclear reactor operation monitoring and intelligent prediction technology, specifically relating to a real-time online intelligent prediction method for high-dimensional parameters of nuclear reactors. Background Technology

[0002] As a large and complex energy system, the safe, stable, and efficient operation of a nuclear reactor relies on the continuous monitoring and accurate prediction of massive amounts of operating parameters. During operation, tens of thousands of sensors are deployed within the reactor system, collecting real-time, multi-physics, high-dimensional time-series data, including temperature, pressure, flow rate, neutron flux, and coolant chemical parameters. These parameters collectively constitute a complete mapping of the reactor's operating state. When the system deviates from normal operating conditions due to planned operations, equipment failures, or external disturbances, all relevant parameters undergo complex dynamic transient processes. Therefore, achieving unified, real-time, and dynamic prediction of the entire system's high-dimensional parameters is a core technology for improving the intelligent operation and maintenance level of nuclear reactors, enhancing state awareness, and assisting operators in making forward-looking decisions. It has significant practical application value for ensuring nuclear safety, optimizing operational performance, and extending equipment lifespan.

[0003] However, the current field of nuclear reactor operation monitoring and prediction faces a series of severe challenges and technical limitations. First, the data is highly dimensional and tightly coupled. The number of reactor parameters is enormous, and the physicochemical processes are highly nonlinearly coupled. Traditional monitoring methods based on single-parameter or low-dimensional models struggle to capture the overall dynamic behavior of the system and the complex relationships between parameters, leading to biased and delayed identification and prediction of abnormal conditions or transient processes. Second, real-time performance and uniformity are insufficient. Existing technologies mostly focus on independent monitoring of a few key safety parameters or offline simulations based on simplified physical models, lacking effective means to simultaneously process tens of thousands of parameters and perform online unified rolling predictions. This limits the operator's ability to grasp the overall transient evolution trend of the system in real time. Third, the model generalization ability is weak. Reactor operation faces different initial states, diverse transient processes (such as rod lifting, power reduction, and accident conditions), and different operator intervention strategies. Traditional data-driven models or physical models with fixed parameters struggle to adapt to such a wide range of changing boundary conditions and external actions, resulting in a significant decrease in the reliability of prediction results when scenarios shift. Finally, the extraction of temporal dynamic features is not deep enough. Reactor parameter variations exhibit strong long-range dependence and multi-timescale characteristics. Existing methods based on shallow machine learning or conventional recurrent neural networks have limitations in capturing complex long-term time-series patterns and transient features in high-dimensional parameter sequences, making it difficult to meet the requirements for high-precision prediction.

[0004] In summary, the nuclear reactor field urgently needs breakthroughs in intelligent prediction technologies capable of simultaneously handling massive high-dimensional inputs, achieving online real-time rolling predictions, and possessing strong cross-scenario generalization capabilities. Existing technologies have significant limitations in handling high-dimensional unified predictions, complex time-series dynamic learning, and coping with changing operating conditions, thus hindering the further development of intelligent operation and maintenance and rapid prediction and early warning technologies for nuclear reactors. This invention aims to provide an innovative solution to address the aforementioned challenges. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an online real-time prediction method for high-dimensional parameters of nuclear reactor systems. This invention aims to solve the problem that current technologies struggle to perform unified online transient monitoring and real-time prediction of tens of thousands of high-dimensional, strongly coupled parameters during reactor operation. By integrating Transformer and LSTM neural networks, it achieves efficient extraction and learning of deep features and long-range dependencies in massive time-series data; and by introducing DeepONet, the prediction model can generalize to different initial states, system transients, and external intervention scenarios. Ultimately, it achieves dynamic, high-precision prediction synchronized with the actual operating state of the reactor, thereby comprehensively improving the intelligent operation and maintenance level of nuclear reactors, enhancing the real-time perception capability of system status, and providing operators with forward-looking decision support to ensure nuclear safety and optimize operational performance.

[0006] To achieve the aforementioned objectives, this invention provides a real-time online intelligent prediction method for high-dimensional parameters of a nuclear reactor system, comprising: acquiring high-dimensional time-series parameter data of the nuclear reactor system in real time; performing feature extraction and time-series learning on the high-dimensional time-series parameter data using a neural network model that integrates Transformer and Long Short-Term Memory; mapping the learned features to a dynamic response space under different initial conditions, system transients, and external interventions by combining a deep operator network; and performing online rolling prediction based on the mapping to output the predicted values ​​of the high-dimensional parameters of the system over a future period.

[0007] Preferably, the high-dimensional time-series parameter data includes temperature, pressure, flow rate, neutron flux, and coolant chemical parameters from tens of thousands of sensors.

[0008] Preferably, the Transformer neural network is used to process high-dimensional parameters in parallel and captures the global spatial coupling and nonlinear correlation between parameters through a self-attention mechanism.

[0009] Preferably, the Long Short-Term Memory (LSTM) neural network is used to learn long-range dependencies and dynamic evolution features across multiple time scales in the parameter sequence.

[0010] Preferably, the deep operator network is used to decouple and learn the generalized mapping relationship between the system's transient process and abstract conditions, wherein the abstract conditions include the initial state, boundary conditions, system actions, and operator intervention strategies.

[0011] Preferably, the deep operator network includes a branch network and a backbone network. The branch network encodes abstract conditions, and the backbone network encodes spatiotemporal coordinates. The two are combined to generate prediction results.

[0012] Preferably, the online rolling prediction is based on the latest received measured data, and the predicted trajectory is updated in a fixed time window to achieve synchronization with the reactor operating status.

[0013] Preferably, the method further includes visualizing the prediction results to assist the operator in system status perception and operational decision-making.

[0014] Preferably, the neural network model is obtained by offline training using a combination of historical running data and high-fidelity simulation data.

[0015] Preferably, the training data covers a variety of normal operating conditions, planned transients, fault conditions, and accident sequence scenarios.

[0016] Preferably, the prediction method is deployed on the real-time computing platform of the nuclear power plant and interacts with the distributed control system (DCS) for data exchange.

[0017] Preferably, the predictive capability achieved by the method is generalized to different reactor designs, different power levels, and different fuel cycle stages.

[0018] Preferably, another aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the online real-time prediction method for high-dimensional parameters of a nuclear reactor system.

[0019] Preferably, another aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the online real-time prediction method for high-dimensional parameters of a nuclear reactor system.

[0020] The present invention has the following beneficial effects: This invention effectively addresses several key technical bottlenecks that have long existed in the field of nuclear reactor operation monitoring and prediction. First, addressing the challenge of unified real-time prediction of high-dimensional parameters, a prediction model capable of simultaneously processing tens of thousands of parameters in time-series data streams is constructed by integrating Transformer and LSTM network architectures. The powerful parallel computing and global attention mechanism of Transformer efficiently capture the complex spatial coupling and nonlinear correlations among massive parameters; while LSTM focuses on learning the long-range dependencies and multi-timescale features in the dynamic evolution of parameters. Second, by introducing DeepONet, the specific transient processes are decoupled from abstract initial / boundary conditions, system actions, and external interventions, enabling the model to possess powerful cross-scenario generalization capabilities. This means that the trained model can accurately predict the dynamic response of parameters under different initial states, different operations (such as rod lifting and power reduction), and even different accident condition sequences without retraining or fine-tuning, fundamentally solving the problem of traditional data-driven models failing due to scenario changes. Finally, by designing an online rolling prediction framework, this method achieves real-time synchronization with the actual operating status of the reactor within seconds. It can continuously update the predicted trajectory based on the latest measured data, transforming the previous offline and lagging analysis into online and forward-looking perception, significantly improving the real-time performance of prediction and the accuracy of dynamic tracking.

[0021] The implementation of this invention will generate significant social and economic benefits, powerfully promoting the safe development and intelligent upgrading of the nuclear energy industry. In terms of social benefits, its primary contribution is a significant enhancement of the nuclear safety barrier. This method provides operators with decision support tools, enabling them to more quickly and comprehensively reveal parameter evolution trends and potential risk paths in the early stages of abnormal or accidental conditions. This buys operators valuable intervention time, assisting them in making more scientific and precise control decisions, thereby significantly reducing the risk of human error from a technical perspective, improving the nuclear power plant's ability to prevent and mitigate serious accidents, and strengthening the public safety defense line. In terms of economic benefits, this method will profoundly transform the operation and maintenance model of nuclear power plants. Firstly, by implementing predictive maintenance, it can accurately assess equipment performance degradation trends, optimize maintenance plans and spare parts inventory, shifting from traditional periodic or reactive maintenance to condition-based precision maintenance, significantly reducing unplanned downtime, improving unit availability and operating efficiency, and directly generating power generation revenue. Secondly, high-precision real-time prediction provides a data foundation for operation optimization, supporting more flexible and economical load tracking and scheduling strategies, improving fuel utilization efficiency, and reducing operating costs. Third, this method, as a core intelligent engine, can be integrated into digital twin power plant systems, driving the entire nuclear energy industry towards an intelligent, less-manned operation and maintenance model, reducing long-term operating costs, and enhancing the international competitiveness of my country's advanced nuclear energy technologies. Therefore, this invention is not only a technological innovation but also an important tool for ensuring national energy security and promoting the high-quality development of the nuclear energy industry. Attached Figure Description

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments. The accompanying drawings, which constitute a part of this application, are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0023] Figure 1 Flowchart of the method of this invention.

[0024] Figure 2 A schematic diagram of the intelligent prediction neural network framework proposed in this invention.

[0025] Figure 3 A schematic diagram illustrating the online rolling prediction effect of this invention. Detailed Implementation

[0026] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0027] Example 1: A schematic diagram of the embodiment of the present invention is shown below. Figure 1 As shown.

[0028] Step 1: Construct a high-dimensional parameter time series database and preprocessing module for nuclear reactors.

[0029] This step forms the data foundation for implementing the prediction method. First, key parameters closely related to system safety and operational status, including but not limited to core power, coolant temperature and pressure, flow rate, control rod position, and neutron flux density, are collected in real-time from the target nuclear reactor's distributed control system, protection system, and various sensor networks, forming a raw high-dimensional time-series data stream. Next, a data preprocessing module is established to clean the raw data, removing missing and outlier values ​​caused by sensor malfunctions or communication interruptions. Then, timestamp alignment and synchronization are performed on the multi-source heterogeneous data to ensure data consistency. Finally, preliminary feature grouping and normalization are performed based on the physical correlations between parameters, forming a standardized, high-quality high-dimensional parameter time-series dataset, providing input for subsequent model training and online learning.

[0030] Step 2: Design a temporal feature extraction network architecture that integrates Transformer and LSTM.

[0031] This step aims to construct a core feature extractor capable of efficiently capturing complex spatiotemporal dependencies among high-dimensional parameters. A hierarchical hybrid neural network model is designed: the bottom layer employs multiple parallel LSTM unit groups to process short sequences of high-dimensional parameters in different physical groups, capturing the short-term dynamics and nonlinear evolution patterns within each parameter group. The upper layer introduces a Transformer encoder layer, taking the abstract feature sequences extracted by each LSTM group as input. Utilizing the Transformer's multi-head self-attention mechanism, the association weights of all parameter features at different time steps are globally calculated, thus explicitly modeling the strong coupling relationships and long-range dependencies among tens of thousands of parameters. This architecture achieves collaborative learning of local temporal dynamics and global interaction relationships, laying the structural foundation for deep feature extraction.

[0032] Step 3: Integrate DeepONet to build a generalized and enhanced predictive operator network.

[0033] The core of this step lies in improving the model's generalized prediction ability for unknown operating conditions. The DeepONet architecture is integrated on top of the spatiotemporal feature extraction network built in the second step. The context vector, which condenses historical sequence information and is output by the feature extraction network, is used as input to the "branch network" of DeepONet to encode the current system state and historical trajectory. Simultaneously, future prediction time points are used as input to the "backbone network." The outputs of the two networks interact (dot product or more complex operations) to form a "neural operator." This design enables the model not only to learn the mapping from historical sequences to future values ​​but also to learn a generalized prediction operator that can adapt to different function inputs (i.e., different initial states, system transients, and external intervention scenarios), thus maintaining robust prediction performance even when faced with initial conditions or external interventions not covered by the training data.

[0034] The schematic diagram of the intelligent prediction neural network framework proposed in steps 2 and 3 is shown below. Figure 2 As shown.

[0035] Step 4: Offline and transfer learning training of the model based on historical running data.

[0036] This step utilizes accumulated historical data to fully train the model. Using the preprocessed historical time-series dataset constructed in the first step, a large number of "historical sequence-future sequence" sample pairs are constructed using a rolling time window approach. A phased training strategy is adopted: First, the hybrid feature extraction network is pre-trained on a subset of data containing various typical operating conditions (such as steady state, load increase, load decrease, and common transients) to learn general spatiotemporal feature representations. Then, the pre-trained weights are loaded, and the complete model integrating DeepONet is trained end-to-end on a more complete dataset. The optimization objective is to minimize the mean squared error between predicted parameter values ​​and true values, as well as other related physical constraint losses. For newly built reactors or situations where data is scarce, transfer learning can be used to fine-tune parameters using models trained on similar reactor types or simulation systems.

[0037] Step 5: Deploy the online real-time rolling prediction engine and interface module.

[0038] This step involves deploying the trained model to a real-time computing server or high-performance edge computing device at the reactor site. An online prediction engine is developed, continuously receiving the latest high-dimensional parameter data stream from the data acquisition system. Internally, the engine implements a dynamic buffer that continuously stores the latest time-series data at fixed time intervals required by the model (e.g., data from the past hour, sampled at 1-second intervals). At each prediction cycle (e.g., per second), the engine automatically calls the preprocessing module to process the data in the buffer and inputs it into the loaded deep learning model. The model outputs a rolling prediction sequence of all high-dimensional parameters for a future time period (e.g., the next 10 minutes) in real time. Simultaneously, a standardized data interface is developed to push the prediction results in real time to the human-machine interface, database, and subsequent early warning module. Figure 3 A schematic diagram illustrating the online rolling prediction of an example parameter (regulator pressure) is shown.

[0039] Step 6: Develop a multi-level prediction result visualization and real-time synchronous comparison system.

[0040] This step aims to effectively present the prediction results to the operator. A dedicated human-machine interface is developed, and a multi-level visualization scheme is designed: At the overview level, the predicted trajectories of key parameters (such as core outlet temperature and primary circuit pressure) are dynamically plotted synchronously with their real-time measurements using overlay trend curves, intuitively displaying the prediction accuracy and future trends. At the detail level, operators are allowed to drill down and view detailed comparisons between predictions and actual measurements for any group or individual parameter. The system must ensure that the prediction curves are updated dynamically as real-time data arrives, and highlight parameters whose deviations from the predicted values ​​exceed a preset threshold.

[0041] Step 7: Implement a closed loop for online adaptive model updates and performance monitoring.

[0042] This step ensures the continued accuracy and reliability of the prediction system during long-term operation. An online model performance monitoring module is established to continuously calculate the real-time performance of core prediction metrics, such as rolling prediction error and early warning accuracy. An online adaptive update mechanism for the model is designed: when the model's prediction performance is detected to be continuously degrading under specific new operating conditions, the system can automatically trigger an incremental learning process. While ensuring data security and isolation, the newly generated operating data is used to perform lightweight fine-tuning of the model. Simultaneously, the model is periodically (e.g., quarterly) comprehensively retrained and iterated using accumulated new data. The entire process forms an automated closed loop of "deployment-monitoring-update," enabling the prediction system to continuously evolve throughout the reactor's entire lifecycle, adapting to changes in system characteristics caused by equipment aging, fuel replacement, etc.

[0043] Example 2: This embodiment provides a real-time online intelligent prediction system for high-dimensional parameters of nuclear reactors, including: a data acquisition and preprocessing module, a hybrid neural network feature extraction module, a deep operator generalization prediction module, an online rolling prediction engine, a visualization monitoring module, and a model adaptive update module. The data acquisition and preprocessing module is used to acquire high-dimensional time series parameter data from the nuclear reactor distributed control system and high-fidelity simulation platform, and to perform outlier processing, timestamp synchronization and feature grouping normalization. The hybrid neural network feature extraction module is used to extract features from the preprocessed high-dimensional time-series parameter data. It includes a parallel LSTM unit group and a Transformer encoder layer. The parallel LSTM unit group is used to capture the short-range dependence and multi-time-scale dynamic features of different feature group parameters. The Transformer encoder layer is used to capture the global coupling and nonlinear correlation of cross-group parameters. The deep operator generalization prediction module includes a branch network and a backbone network. The branch network is used to encode the global feature vector output by the hybrid neural network feature extraction module, and the backbone network is used to encode the future prediction time coordinates. The outputs of the branch network and the backbone network interact and are mapped to the high-dimensional parameter space of the nuclear reactor to output the predicted parameter values. The online rolling prediction engine is used to trigger prediction tasks according to a preset period based on real-time data stored in a dynamic data buffer, and to reduce the cumulative prediction error through an online correction algorithm. The visualization monitoring module is used to display the comparison curve between the predicted and measured values ​​of parameters, and to issue an early warning when the parameter deviation exceeds the threshold. The model adaptive update module is used to monitor prediction performance metrics in real time and trigger an incremental learning process when performance degrades, thereby achieving online optimization of the model.

[0044] Furthermore, the high-dimensional time-series parameter data acquired by the data acquisition and preprocessing module includes at least one of the following: core fuel assembly temperature, primary coolant pressure and flow rate, neutron flux density distribution, steam generator water level and temperature, coolant boron concentration and pH, and control rod position. The sampling frequency of the high-dimensional time-series parameter data is 1 Hz, and the number of parameters is not less than 10,000.

[0045] Furthermore, the number of parallel LSTM unit groups is the same as the number of feature groups, each LSTM unit has a hidden layer dimension of 256, a dropout value of 0.2, and a time step of 60; the Transformer encoder layer has 3 layers, the multi-head attention head has 8 layers, the model dimension is 512, and the feedforward network dimension is 2048.

[0046] Furthermore, the branch network of the deep operator generalization prediction module includes three fully connected layers with dimensions of 512→256→128 respectively; the backbone network includes three fully connected layers with dimensions of 1→128→256 respectively; the outputs of the branch network and the backbone network interact through dot product operation.

[0047] Furthermore, the deep operator generalization prediction module introduces working condition label embedding during the training process. The working condition label is a one-hot encoded vector of normal working condition, planned transient, fault working condition, and accident sequence. The one-hot encoded vector is concatenated with the global feature vector and then input into the branch network.

[0048] Furthermore, the dynamic data buffer of the online rolling prediction engine stores the high-dimensional parameter data of the latest 3600 time steps, the prediction period is 1Hz, and a single prediction outputs a high-dimensional parameter prediction sequence for the next 600 time steps. The online correction algorithm fine-tunes the output weights of the backbone network based on the deviation between the predicted value and the measured value at the previous moment.

[0049] Furthermore, the visualization monitoring module includes an overview layer, a group details layer, and an early warning layer; the overview layer displays an overlay of the predicted trajectory and measured curve of key safety parameters; the group details layer displays the predicted value, measured value, and deviation percentage of parameters within any feature group; the parameter deviation threshold of the early warning layer is ±2%, and an audible and visual early warning is issued when the deviation exceeds the threshold for three consecutive time steps.

[0050] Furthermore, the performance metrics monitored by the model adaptive update module include root mean square error, mean absolute percentage error, and prediction accuracy. When the mean absolute percentage error is greater than 5% for 7 consecutive days, the incremental learning process is triggered, and the model is fine-tuned using a small batch and low learning rate method.

[0051] Furthermore, the system is deployed on a real-time computing platform equipped with an NVIDIA A100 GPU. The real-time computing platform and the distributed control system interact with each other via the OPC UA protocol, with a data transmission latency of no more than 100ms.

[0052] Example 3: This embodiment provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements any of the above-described methods for real-time online intelligent prediction of high-dimensional parameters of a nuclear reactor system.

[0053] Example 4: This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for real-time online intelligent prediction of high-dimensional parameters of a nuclear reactor system.

[0054] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

Claims

1. A real-time online intelligent prediction method for high-dimensional parameters of nuclear reactors, characterized in that, Includes the following steps: Acquire high-dimensional time-series parameter data of the nuclear reactor system in real time; use a neural network model that integrates Transformer and Long Short-Term Memory to extract features and learn the time series of the high-dimensional time-series parameter data; combine a deep operator network to map the learned features to the dynamic response space under different initial conditions, system transients and external interventions; perform online rolling prediction based on the mapping, and output the predicted values ​​of the high-dimensional parameters of the system in the future.

2. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The high-dimensional time-series parameter data includes temperature, pressure, flow rate, neutron flux, and coolant chemical parameters from tens of thousands of sensors.

3. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The Transformer neural network is used to process high-dimensional parameters in parallel and captures the global spatial coupling and nonlinear correlation between parameters through a self-attention mechanism.

4. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The long short-term memory neural network is used to learn long-range dependencies and dynamic evolution features across multiple time scales in parameter sequences.

5. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The deep operator network is used to decouple and learn the generalized mapping relationship between the system's transient process and abstract conditions, which include the initial state, boundary conditions, system actions, and operator intervention strategies.

6. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 5, characterized in that, The deep operator network includes a branch network and a backbone network. The branch network encodes abstract conditions, and the backbone network encodes spatiotemporal coordinates. The two are combined to generate prediction results.

7. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The online rolling prediction is based on the latest received measured data and updates the prediction trajectory in a fixed time window to achieve synchronization with the reactor operating status. The method also includes visualizing the prediction results to assist operators in system status perception and operational decision-making.

8. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The neural network model was obtained through offline training using a combination of historical running data and high-fidelity simulation data. The data used for model training covers a variety of normal operating conditions, planned transients, fault conditions, and accident sequence scenarios.

9. The method for real-time online intelligent prediction of high-dimensional parameters of nuclear reactors according to claim 1, characterized in that, The prediction method is deployed on the real-time computing platform of the nuclear power plant and interacts with the distributed control system (DCS) for data exchange. The predictive capability achieved by the method can be generalized to different reactor designs, different power levels, and different fuel cycle stages.

10. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the real-time online intelligent prediction method for high-dimensional parameters of a nuclear reactor as described in any one of claims 1 to 9.

Images