GS simulation model construction and rapid prediction device and method capable of being updated online
By employing a data-driven approach, combining multi-channel data acquisition, feature preprocessing, and LSTM neural network modeling, the problem of untimely updates in traditional physical mechanism modeling on nuclear power simulation platforms is solved. This enables rapid and high-precision prediction of thermodynamic processes, making it suitable for modeling and predictive control of complex industrial systems.
Patent Information
- Application Number
- CN202511091243.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-11-07
AI Technical Summary
Traditional physical mechanism-based modeling methods are not timely in updating models when dealing with complex, highly coupled, and rapidly changing nuclear power simulation platforms (GS systems), making it difficult to adapt to system changes and affecting prediction accuracy and control performance.
A data-driven approach is adopted, which involves multi-channel data acquisition, Z-score and principal component analysis feature preprocessing, LSTM neural network modeling, and the introduction of an online correction mechanism to achieve fast and accurate thermal simulation modeling and prediction.
It achieves rapid, accurate, and robust prediction of the operating status of the GS system, has good online adaptability, reduces engineering deployment and maintenance costs, and improves the model's adaptability and generalization ability.
Smart Images
Figure CN120910464A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of industrial process modeling and predictive control, in particular to a thermal simulation modeling and prediction method suitable for a nuclear power simulation system (referred to as GS system) with multiple parameters, strong coupling and obvious time delay characteristics, which integrates data acquisition, feature preprocessing, neural network modeling and prediction mechanism, and can be widely applied to complex industrial system operation state analysis and simulation. BACKGROUND
[0002] The present application relates to the technical field of industrial process modeling and predictive control, in particular to an intelligent modeling method for high-complexity thermal fluid system, which is suitable for thermal process modeling and multi-time step state prediction tasks in a nuclear power simulation platform (hereinafter referred to as GS system) with multivariable coupling, significant time delay and strong dynamic nonlinearity.
[0003] With the development of nuclear energy and process industry intelligence, traditional modeling methods based on physical mechanism face great challenges in dealing with high system coupling strength, fast running state change and large parameter uncertainty: on the one hand, modeling under complex conditions needs to rely on a large number of fine physical formula derivation and engineering experience, which is long in cycle and poor in robustness; on the other hand, the model cannot be updated in time to adapt to the changes in structure and parameters during long-term operation of the system, affecting the prediction accuracy and control effect.
[0004] The GS system is a simulation platform for evaluating and controlling algorithms of thermal parameters of nuclear power plants, which has a complex structure involving multiple subsystems such as pumps, valves and heat exchangers. The signals of various measuring points are strongly coupled in physics, and the system has dynamic time-varying, large thermal inertia and significant nonlinear characteristics. These factors pose higher requirements on the modeling accuracy and real-time response capability of the thermal process. SUMMARY
[0005] In order to solve the above problems, the present application provides a GS thermal simulation modeling and prediction device and method which integrates data-driven modeling, feature processing and synchronous prediction and can be updated online.
[0006] The device and method have the following functions:
[0007] 1. Full-parameter multi-channel data acquisition;
[0008] 2. Z-score and principal component analysis feature preprocessing;
[0009] 3. LSTM neural network modeling;
[0010] 4. Online correction mechanism;
[0011] 5. Real-time interaction and prediction visualization of upper computer.
[0012] To achieve the above object, the first technical solution adopted by the present application is:
[0013] The present application provides a kind of online updateable GS simulation model construction and fast prediction device, and the device mainly includes the following modules:
[0014] (1) data acquisition and pretreatment module:
[0015] The data acquisition module is connected with PLC system through Snap7 protocol, and real-time acquisition of various thermal parameter data in GS system, including flow, pressure, temperature, electric power etc.;And have breakpoint continuation and redundant storage function, ensure data integrity and timeliness.
[0016] The data pretreatment module: a series of pretreatment operations are performed on the collected data, including Z-score outlier identification, missing value interpolation, principal component analysis (PCA) dimensionality reduction processing and correlation analysis between variables;This module can effectively improve the robustness and representation ability of modeling features.
[0017] (2) deep modeling and online updating module: long short-term memory neural network (LSTM) is used to build thermal agent model, which supports multi-variable, multi-time step output structure;
[0018] (3) multi-step prediction module: single model multi-output structure is used for synchronous prediction to avoid error accumulation problem in traditional rolling prediction, and improve long-term prediction accuracy.
[0019] (4) prediction visualization and interaction module: integrate host computer human-computer interaction interface to realize visual display of model prediction process, error alarm, parameter adjustment and manual intervention operation.
[0020] The above modules run cooperatively to form an end-to-end data-driven thermal simulation modeling and prediction system, which can realize fast, accurate and high robustness prediction of GS system running state, and has good online adaptation and industrial deployment capability.
[0021] To achieve the above object, the second technical solution adopted by the present application is:
[0022] The present application provides a kind of online updateable GS simulation model construction and fast prediction method, and the method above online updateable GS simulation model construction and fast prediction device, the method includes the following steps:
[0023] (1) data acquisition and pretreatment:
[0024] The data acquisition step is connected with PLC system through Snap7 protocol, and real-time acquisition of various thermal parameter data in GS system;And have breakpoint continuation and redundant storage function;
[0025] The data preprocessing step: a series of preprocessing operations are performed on the collected data, including Z-score outlier identification, missing value interpolation, principal component analysis PCA dimension reduction processing, and correlation analysis between variables.
[0026] (2) Deep modeling and online updating: a long short-term memory neural network (LSTM) is used to construct a thermal agent model, supporting a multi-variable, multi-time step output structure.
[0027] (4) Prediction visualization and interaction: an upper computer human-computer interaction interface is integrated to realize visualization display, error alarm, parameter adjustment and manual intervention operation of the model prediction process.
[0028] As an embodiment of the present application, the data collection frequency in step (1) is 1 Hz.
[0029] As an embodiment of the present application, the Z-score outlier identification in step (1) data preprocessing is based on the Z-score standardization method, a 3σ threshold is set to identify and remove system jump interference points; the missing value interpolation uses maximum likelihood estimation in the training phase and historical sliding window mean interpolation in the online phase; the principal component analysis PCA dimension reduction processing uses the PCA method to extract the main characteristic factors; and the correlation analysis between variables uses the Pearson coefficient matrix to analyze the correlation between variables.
[0030] As an embodiment of the present application, the long short-term memory neural network (LSTM) in step (2) has 2 layers, each layer has 128 units.
[0031] As an embodiment of the present application, step (2) combines the entropy value method and the error feedback mechanism to realize periodic updating and replacement of model parameters.
[0032] As an embodiment of the present application, the periodic updating and replacement refers to periodic evaluation of RMSE / IEE and drift based on entropy value; when the drift is low, the model is fine-tuned, and when the drift is high, the asynchronous retraining and hot replacement are performed.
[0033] As an embodiment of the present application, the multi-step prediction in step (3) uses a single model multi-output structure to perform synchronous prediction for 5 seconds, 10 seconds and 20 seconds.
[0034] As an embodiment of the present application, in step (3), the RMSE of the key parameters is kept less than 5%.
[0035] The advantages and beneficial technical effects of the present application are as follows:
[0036] (I) End-to-end integrated platform
[0037] The application integrates multi-source data acquisition (Snap7-PLC), time sequence storage, data preprocessing (outlier rejection, interpolation, PCA dimension reduction, correlation screening), deep modeling (double-layer LSTM), online correction (entropy method + error feedback), multi-step prediction and visualization (Qt interface) functions, and the modular interface is unified, deployed "ready to use", which greatly simplifies the engineering online and operation process.
[0038] (2) High-fidelity working condition simulation capability
[0039] By multi-point sampling of the GS system pump, valve, heat exchanger and other subsystems, and combining with historical and experimental data fusion analysis, the application can reproduce the system dynamic working condition in real time on the software platform; virtual disturbance can be applied to different loads and thermal parameters to realize true machine level simulation and online calibration.
[0040] (3) Rapid and high-precision modeling and prediction
[0041] Distributed PCA is used for dimension reduction and decomposition of system submodules, and a double-layer LSTM is used to construct a multi-output synchronous structure, which generates 5s / 10s / 20s three time step predictions at a time, greatly reducing the rolling error accumulation. The measured RMSE is less than 5%, which meets the engineering level rapid response demand.
[0042] (4) Self-adaptive online correction mechanism
[0043] The double-layer update strategy based on entropy method and error feedback is created: online fine tuning when the error drifts slightly, asynchronous retraining and hot switching when it falls out of steady state, without interrupting the prediction service, ensuring that the model always maintains high consistency with the physical bench.
[0044] (5) Modularized software and hardware co-design
[0045] Software and hardware are decoupled and layered, and the interface is standardized, supporting PLC field bus, time sequence database, Dockerized deep learning service and cross-platform Qt host computer interconnection; the communication adopts a custom frame protocol, meeting the real-time and reliability requirements of industrial sites.
[0046] (6) Engineering easy maintenance
[0047] The host computer interface visually displays real-time data, prediction curves, error alarms and parameter configurations; system logs, model versions and performance indicators are recorded throughout the link, which is convenient for fault tracing and scheme iteration, significantly reducing operation and maintenance costs and improving system stability.
[0048] In summary, the present application proposes a thermal modeling and prediction scheme based on a data-driven method, which integrates multi-source sensor data acquisition, Z-score outlier detection and PCA dimensionality reduction preprocessing, LSTM deep neural network model construction, and error feedback driven online model correction mechanism, to realize high-precision, high-stability modeling and fast prediction of the thermal process of the GS system. This method does not require knowledge of complex physical mechanisms, has good adaptability and generalization ability, and is suitable for process modeling, digital twin construction and intelligent prediction control scenarios of complex industrial systems. BRIEF DESCRIPTION OF DRAWINGS
[0049] Figure 1 is a system function structure diagram of the present application.
[0050] Figure 2 is a data acquisition schematic diagram.
[0051] Figure 3 is a data preprocessing flowchart.
[0052] Figure 4 is a modeling and online updating flowchart.
[0053] Figure 5 is a multi-step prediction structure diagram. DETAILED DESCRIPTION
[0054] In order to better illustrate the embodiments of the present application, the following will be described in detail in combination with the GS system modeling and prediction experiment process:
[0055] Acquisition process: By configuring the data area and sampling period of S7-1200 / 1500 series PLC, using Snap7 interface to collect more than 10 kinds of variables in the system in real time, including the temperature, pressure, flow, heater outlet temperature, pump power of primary / secondary main pipeline, etc., the sampling frequency is set to 1Hz, and the storage period covers multiple states such as working condition switching and steady-state operation.
[0056] Preprocessing process:
[0057] Outlier detection: based on the Z-score standardization method, set the 3σ threshold, identify and eliminate the system jump interference points;
[0058] Missing value filling: maximum likelihood estimation is used for interpolation in the training stage, and historical sliding window mean interpolation is used in the online stage;
[0059] Dimensionality reduction: PCA method is used to extract main feature factors, reduce redundancy, and avoid "dimension disaster";
[0060] Correlation analysis: analyze the correlation between variables through Pearson coefficient matrix to improve the expression ability of input data structure.
[0061] Model training: A double-layer LSTM network structure is constructed, with 128 hidden layer nodes, and a sliding window structure is used to model the time series data; the output layer is designed as a multi-output node to simultaneously predict multiple parameters at t+5, t+10 and t+20 seconds in the future; the MSE loss function and Adam optimizer are used for training, and the RMSE, MAE and other indicators are used to evaluate the performance of the validation set.
[0062] Online updating mechanism: The system calculates the prediction error of the current model every fixed period (e.g., 10 minutes); if the RMSE or the extreme error IEE exceeds the threshold (e.g., 5%), the entropy method is used to determine the degree of model drift; according to the drift level, the model parameters are adjusted slightly (fine tuning) or replaced directly with a newly trained model (hot start retraining); all updating processes are completed online to avoid system downtime.
[0063] Prediction effect: The experimental results show that under variable operating conditions (e.g., 240s and 1150s) and steady-state operating conditions (500s), the RMSE of the water pressure and flow rate at the end of the No. 1 main pipe is controlled within 5%; the model has good prediction consistency and time response ability for multiple variables and multi-step thermal parameters.
[0064] Host computer interaction: A visual interface is provided to display real-time prediction curves and error trends; model switching, threshold setting, manual updating and other operations are supported to improve the convenience of operation and maintenance personnel.
[0065] As shown in Figure 1 , the system of the present application mainly comprises a data acquisition and preprocessing module 1, a deep modeling and online updating module 2, a multi-step prediction module 3 and a host computer visual interaction module 4, and each module is interconnected through a unified data interface to form a closed-loop system of "acquisition - modeling - prediction - updating - display". The system runs in the GS system of the nuclear power plant simulation platform, and the data variables collected include more than 10 thermal indicators such as temperature, flow rate, pressure and power, which have high coupling and time-varying characteristics.
[0066] (1) GS system data acquisition and preprocessing module
[0067] As shown in Figure 2 , the data acquisition module is deployed on the GS system test bench. Based on the Snap7 and S7-1500 series PLC hardware platform, it is connected to the host computer through Ethernet and Snap7 communication protocol. This module obtains key physical parameter signals such as flow rate, pressure, temperature and voltage from nodes such as the first main pipe, the second main pipe, the pump outlet and the heater in the system.
[0068] The data collection period is usually set to 1 Hz, combined with ring buffer and time series database, and supports configuration item adjustment to adapt to different dynamic characteristics under steady state / variable working conditions. To ensure data consistency, stability and reliable collection, the module has a sampling synchronization control mechanism to uniformly schedule the sampling rate of different types of sensors, and embeds a timestamp correction strategy.
[0069] After the data enters the collection buffer, it will be automatically output with high-quality modeling features through the following preprocessing process:
[0070] 1. Outlier detection: Z-score method is used to identify data points outside ±3σ, and to eliminate implicit outliers within constraints ( Figure 3 as shown);
[0071] 2. Missing value interpolation: Combine historical sliding window mean method and maximum likelihood estimation method to interpolate online and training phase data in stages;
[0072] 3. Principal component dimensionality reduction: Through PCA analysis, the principal component with the largest explainable variance in the original variable is extracted to improve data representation;
[0073] 4. Variable selection: Based on Pearson correlation analysis, select strong correlation input features for target variables (such as main pump flow, main pipe pressure) to reduce model dimension and improve training efficiency.
[0074] (2) Thermal modeling and online updating module
[0075] Distributed PCA-LSTM modeling architecture: After system decomposition, the subsystems are respectively reduced in dimension and independently modeled, and then a double-layer LSTM is used for multi-step (5s / 10s / 20s) synchronous prediction to achieve fast and high-precision simulation.
[0076] The core of this module is the LSTM neural network modeling subsystem. As Figure 4 shown, this model uses a double-layer LSTM structure for time series modeling, with input being historical feature variables within the past 30 seconds and output being predicted values of target variables (such as main pipe outlet pressure, flow, etc.) at future time steps.
[0077] The network configuration is as follows:
[0078] • Input dimension: feature dimension after PCA dimensionality reduction (e.g. 12 dimensions);
[0079] • LSTM hidden layer: 2 layers, each with 128 units;
[0080] • Output layer: supports simultaneous output of 5s, 10s, and 20s future steps;
[0081] • Optimizer: Adam, learning rate 0.001;
[0082] • Loss function: Mean Squared Error (MSE);
[0083] To enhance the system's adaptability during dynamic operation, the module introduces an online model updating mechanism, with the following strategy:
[0084] • The system calculates the RMSE and IEE of the current model every 5 minutes, periodically evaluating RMSE / IEE;
[0085] • If the error exceeds the preset threshold (such as 5%), the model drift is evaluated by entropy method;
[0086] • When the drift coefficient exceeds the tolerance range, the system will activate the update process:
[0087] - Light update: use the latest 30 minutes of data, fine-tune the current model when the drift is low;
[0088] - Hot reconstruction: re-initialize the new model training and replace the old model parameters (high drift when asynchronous retraining and hot replacement), ensuring real-time accuracy.
[0089] (Three) Multi-step prediction module
[0090] As shown in Figure 5 , this module realizes multi-time step prediction of future thermal parameters. Unlike traditional rolling prediction structure, this system uses a "single model multi-output" structure, which outputs multiple time step prediction results at once, such as 5 seconds, 10 seconds, and 20 seconds. This effectively avoids error accumulation problems. For key parameters (pressure, flow, etc.), maintain RMSE < 5%.
[0091] The prediction target is configurable, including but not limited to:
[0092] • Primary supervisor end pressure / flow;
[0093] • Heater outlet temperature;
[0094] • Circulating pump outlet pressure, etc.
[0095] The prediction module will continuously receive input sequences from the data processing module, update the prediction results using a sliding window mechanism, and push them to the host computer for visualization and alarm judgment.
[0096] (Four) Host computer visualization interaction module
[0097] Modular hardware and software deployment and protocol: unified module interface (acquisition, preprocessing, modeling, prediction, visualization), customized data frame protocol and Qt host computer interaction interface, supporting engineering rapid integration and maintenance.
[0098] The host computer is developed using the Qt platform and has the following functional modules:
[0099] 1. Real-time monitoring interface: Displays the raw values of each physical quantity.
[0100] 2. Forecast Curve Interface: Displays historical and future forecast trends using a line chart format;
[0101] 3. Error assessment interface: Displays dynamic evaluation indicators such as RMSE and IEE;
[0102] 4. Parameter setting interface: Configure model update threshold, prediction step size, output channels, etc.;
[0103] 5. Anomaly Alarm Interface: Real-time early warning based on error thresholds, sudden change detection, and other mechanisms.
[0104] Data communication with the lower-level device is based on a custom frame structure, which includes: frame header, function code, device ID, data length, data body, CRC checksum, and frame tail. It supports command control and status feedback.
[0105] Compared with existing mainstream modeling / prediction methods, this invention has the following characteristics:
[0106] 1. End-to-end data-driven process
[0107] Existing physical mechanism models often rely on a large number of equipment parameters and empirical formulas, resulting in long model construction cycles and sensitivity to changes in system structure; while this invention acquires data in real time from multiple sources of sensors (…). Figure 2 Starting with a data acquisition diagram, the process involves automated preprocessing steps such as Z-score and PCA. Figure 4 The data preprocessing flowchart directly generates high-quality features without the need for in-depth mechanism derivation, significantly shortening the modeling preparation time and expanding the scope of application.
[0108] 2. Highly efficient dimensionality reduction and strong robustness
[0109] Traditional data-driven models often use all original variables for modeling, which are susceptible to noise and collinearity. This invention introduces principal component analysis (PCA) to compress the feature space and combines it with Pearson correlation to screen key variables. This retains the main information while avoiding redundancy, thus improving the training speed and stability of the model.
[0110] 3. Simultaneous prediction with multiple outputs to prevent error accumulation.
[0111] The existing LSTM or RNN multi-step prediction mostly adopts a rolling structure, and the iterative error is rapidly amplified with the step length; the scheme adopts a single model multi-output design, and outputs 5s, 10s, 20s and the like at one time, so as to effectively avoid the accumulation of rolling errors, and the 10-step and 20-step predictions can all keep the RMSE less than 5%.
[0112] 4. Online adaptive update, real-time response to system drift
[0113] The traditional data-driven model is generally fixed after offline training and cannot adapt to equipment degradation or working condition switching; the application constructs an online correction mechanism based on an entropy method + error feedback, monitors the RMSE / IEE index in real time, triggers a fine tuning (light update) or hot reconstruction once the threshold is exceeded, and ensures that the simulation model always fits the physical bench.
[0114] Light update: online fine tuning using the latest 30min data;
[0115] Hot reconstruction: asynchronous retraining, seamless replacement.
[0116] 5. Rapid deployment and industrial-grade visualization
[0117] The application integrates a "software and hardware integrated platform" of a Snap7-PLC interface, a Qt host computer interface, a time sequence database and a model service, can realize millisecond-level data transmission, has the functions of interactive friendly visualization monitoring, alarm and manual intervention, and is convenient for engineering deployment and maintenance.
[0118] 6. Hybrid advantage of considering precision and efficiency
[0119] Compared with a pure mechanism or a pure data-driven method, the application combines the advantages of the two: without a large number of formula derivation, not "black box" type full variable input, but through scientific feature extraction + gated sequence model + online feedback, the high precision (RMSE < 5%) and high efficiency (minute-level retraining + real-time prediction) are considered.
Claims
1. An online-updatable GS simulation model construction and rapid prediction device, characterized in that: The device comprises the following modules: (1) Data acquisition and preprocessing module: The data acquisition module is connected with the PLC system through the Snap7 protocol, and real-time acquisition of various thermal parameters in the GS system is performed; and breakpoint continuation and redundant storage functions are provided; The data preprocessing module: performs a series of preprocessing operations on the collected data, including Z-score outlier identification, missing value interpolation, principal component analysis PCA dimensionality reduction processing, and correlation analysis between variables; (2) Deep modeling and online updating module: a long short-term memory neural network LSTM is used to construct a thermal proxy model, supporting a multi-variable, multi-time step output structure; (3) Multi-step prediction module: single model multi-output structure is used for synchronous prediction; (4) Prediction visualization and interaction module: integrates the upper computer human-computer interaction interface to realize the visual display of the model prediction process, error alarm, parameter adjustment and manual intervention operation.
2. An online-updatable GS simulation model construction and rapid prediction method, characterized in that: The method adopts the online updated GS simulation model construction and rapid prediction device of claim 1, and comprises the following steps: (1) Data acquisition and preprocessing: The data acquisition step is connected with the PLC system through the Snap7 protocol, and real-time acquisition of various thermal parameters in the GS system is performed; and breakpoint continuation and redundant storage functions are provided; The data preprocessing step: performs a series of preprocessing operations on the collected data, including Z-score outlier identification, missing value interpolation, principal component analysis PCA dimensionality reduction processing, and correlation analysis between variables; (2) Deep modeling and online updating: a long short-term memory neural network LSTM is used to construct a thermal proxy model, supporting a multi-variable, multi-time step output structure; (3) Multi-step prediction: single model multi-output structure is used for synchronous prediction; (4) Prediction visualization and interaction: integrates the upper computer human-computer interaction interface to realize the visual display of the model prediction process, error alarm, parameter adjustment and manual intervention operation.
3. The GS emulation model construction and fast prediction method with online update according to claim 2, characterized in that: The acquisition frequency of the data acquisition step in step (1) is 1 Hz.
4. The GS simulation model construction and fast prediction method with online update according to claim 2, characterized in that: In the Z-score outlier identification in the data preprocessing step (1), a 3σ threshold is set based on the Z-score standardization method to identify and eliminate system jump interference points; the maximum likelihood estimation is used for interpolation in the training phase, and the historical sliding window mean interpolation is used in the online phase; the PCA method is used to extract the main characteristic factors in the principal component analysis PCA dimensionality reduction processing; and the Pearson coefficient matrix is used to analyze the correlation between variables.
5. The GS emulation model construction and fast prediction method with online update according to claim 2, characterized in that: The long short-term memory neural network LSTM in step (2) is 2 layers, each layer having 128 units.
6. The GS emulation model construction and fast prediction method with online update according to claim 2, characterized in that: The step (2) combines the entropy method and error feedback mechanism to realize periodic updating and replacement of model parameters.
7. The GS emulation model construction and fast prediction method with online update according to claim 6, characterized in that: The periodic updating and replacement refers to periodic evaluation of RMSE / IEE and drift based on entropy; the model is fine-tuned when the drift is low, and the model is asynchronously retrained and replaced when the drift is high. 8.The GS simulation model construction and fast prediction method with online update according to claim 2, wherein: In step (3), the multi-step prediction uses a single model multi-output structure for synchronous prediction of 5 seconds, 10 seconds and 20 seconds.
9. The GS emulation model construction and fast prediction method according to claim 8, characterized in that: In step (3), the key parameters are kept with RMSE<5%.