Liquid level prediction method, system, equipment and medium

Abstract

The invention provides a liquid level prediction method, system and equipment and a medium, and solves the problems that an existing liquid level prediction model is poor in generalization ability and unstable in prediction precision under various working conditions with huge differences such as normal, swinging and crevasse accidents. The method comprises the following steps: constructing a working condition sensing module, automatically identifying and quantifying a current operation working condition, and generating a working condition feature vector; and the working condition information is injected into the prediction model through a FiLM feature modulation mechanism, so that adaptive feature modulation and differential prediction in a multi-working-condition scene are realized. Constructing a teacher model, capturing a long-range global dependency relationship among sensor data by using Transform, capturing a local time sequence mode of the data by using LSTM, and fusing a working condition feature vector as priori knowledge into the model to realize working condition self-adaptive accurate prediction; the knowledge distillation technology is adopted, knowledge of the teacher model is migrated to the student model, and it is ensured that high-precision and real-time liquid level prediction can still be achieved in an on-site control system with limited computing resources.

Description

Technical Field

[0001] This invention belongs to the field of condition monitoring and safety analysis technology for nuclear power, thermal power and chemical industry, and relates to a method and system for predicting the liquid level of a steam generator based on deep learning, operating condition perception and knowledge distillation. Background Technology

[0002] The steam generator is a critical piece of equipment between the primary and secondary loops of a nuclear power plant, and its liquid level is a core parameter for ensuring the safe operation of the reactor. An excessively high liquid level can lead to water carryover in the steam, damaging the turbine; an excessively low liquid level can cause the heating tubes to be exposed above the water surface, leading to overheating or even rupture and triggering a serious accident. Therefore, accurate and reliable prediction of the steam generator's liquid level is crucial for proactive control and accident mitigation.

[0003] Existing liquid level prediction methods have the following limitations: 1. Insufficient model generalization ability: Traditional mechanistic models or data-driven models (such as ARIMA and LSTM) usually perform well under specific working conditions (such as normal steady state), but once the working conditions change drastically (such as ship swaying or pipeline rupture), the model performance will drop sharply and cannot adapt to the complex dynamic characteristics of multiple working conditions.

[0004] 2. Ignoring operating condition information: Most methods train data from different operating conditions together, making it difficult for the model to learn the liquid level change patterns under different physical mechanisms, resulting in prediction lag or inaccuracy when operating conditions change.

[0005] 3. The contradiction between computational complexity and real-time performance: High-precision deep learning models (such as Transformer) usually have large computational loads and many parameters, making it difficult to implement real-time inference on embedded devices in nuclear power plant control systems, thus limiting their engineering application value.

[0006] Therefore, there is an urgent need for a steam generator liquid level prediction method that can adapt to various operating conditions and has both high accuracy and high real-time performance. Summary of the Invention

[0007] This invention proposes a liquid level prediction method and system to solve the problems of poor generalization ability and unstable prediction accuracy of existing liquid level prediction models under various significantly different working conditions such as normal operation, swaying, and breakage accidents.

[0008] The technical solution of the present invention to solve the above problems is: In a first aspect, the present invention proposes a liquid level prediction method, which is characterized by including the following steps: S1: Collect historical operating data (parameters such as liquid level, inlet and outlet flow rate, pressure and temperature) of the steam generator under various operating conditions to build a training dataset; S2: Construct and train a working condition awareness module that can identify the current working condition type based on input data and output the corresponding working condition feature vector; S3: Construct a hybrid neural network based on Transformer and LSTM as the teacher model. This model takes the historical operation data sequence as input and receives the operating condition feature vector from S2. It adaptively modulates the encoded features of the hidden layer through the FiLM feature modulation mechanism, thereby achieving accurate prediction of future liquid level changes. S4: Train the teacher model using the training dataset to minimize the error between the predicted liquid level value and the actual value of the target data. S5: Build a student model with a simpler structure and lower computational cost; S6: Using knowledge distillation technology, the trained teacher model is used as the "teacher" to guide the training of the student model, so that the output of the student model not only fits the real liquid level label, but also fits the "soft label" output of the teacher model. S7: In practical applications, the real-time collected operating data is input into the trained operating condition perception module and student model to obtain the final steam generator liquid level prediction result.

[0009] Furthermore, the condition sensing module in step S2 above is based on HDBSCAN clustering and TransformerEncoder encoding, and its specific implementation is as follows: Statistical features (including mean, standard deviation, median, range of variation, skewness, kurtosis, etc.), frequency domain features (fast Fourier transform results, autocorrelation coefficient, etc.), and structural features (PCA principal components) of the working condition time series are extracted and standardized to obtain the working condition feature dataset. The HDBSCAN density hierarchical clustering algorithm is used to cluster the feature dataset, automatically remove noise, and generate an anchor-positive-negative triplet sample set. The original working condition time series is processed using Transformer Encoder, and latent features are extracted through a multi-layer attention mechanism to output a high-dimensional working condition embedding. A loss function is constructed based on the contrastive learning criterion to maximize the embedding similarity of samples within the same cluster and minimize the embedding similarity of samples across clusters, and the encoder parameters are iteratively optimized. Save the encoder parameters after training convergence, input any working condition time series into the encoder, and output a low-dimensional working condition embedding for subsequent prediction models or multi-working condition analysis.

[0010] Furthermore, the teacher model in step S3 above includes: The encoder part consists of multiple layers of stacked Transformer-LSTM, and each Transformer-LSTM layer integrates a multi-head attention mechanism and an LSTM module. This part takes the input multivariate time series as the processing object, captures the global dependencies between variables through multi-head attention, enhances the learning of temporal features through LSTM, and finally outputs a high-dimensional feature representation rich in contextual information. Fusion and modulation section: The high-dimensional feature representation of the encoder output is modulated using FiLM technology, and the condition information on which the modulation process depends is the working condition coding. Through channel-level linear scaling and offset, the features are adapted to the data distribution of different working conditions. Output section: A Dense fully connected layer is set up; this layer receives the FiLM modulated features and outputs the final liquid level value through linear transformation.

[0011] Furthermore, the student model in step S5 above includes: The student model is constructed using a single-layer Transformer-LSTM architecture, which significantly reduces the number of model parameters and computational complexity, facilitating subsequent online prediction, real-time adjustment, and deployment.

[0012] Furthermore, the loss function of the knowledge distillation process in step S6 above consists of two parts: Student loss: the mean square error between the student model's predicted value and the actual liquid level value; Distillation loss: KL divergence between the output distributions of the student model and the teacher model; The total loss function is: , where α is a hyperparameter balancing the weights of the two terms.

[0013] Furthermore, the various operating conditions in step S1 above include at least the normal steady-state operating condition, the swaying operating condition caused by ocean conditions or external disturbances, and the breach accident operating condition caused by pipeline rupture.

[0014] Secondly, based on the above-mentioned liquid level prediction method, the present invention also proposes a liquid level prediction system, which is characterized by including a data acquisition and preprocessing module, a working condition perception module, a teacher model module, a student model module, a knowledge distillation training module, and an output module. The data acquisition and preprocessing module is used to collect historical operating data (parameters such as liquid level, inlet and outlet flow rate, pressure and temperature) under various working conditions, clean and normalize the data, and construct a training dataset. The working condition perception module can identify the current working condition type based on the input data and output the corresponding working condition feature vector. The teacher model module takes historical operating data sequences as input and receives operating condition feature vectors from the operating condition perception module. It adaptively modulates the hidden layer encoded features through the FiLM feature modulation mechanism, thereby achieving accurate prediction of future liquid level changes. The knowledge distillation training module uses the trained teacher model as the "teacher" to guide the training of the student model, so that the output of the student model not only fits the real liquid level label, but also fits the "soft label" output of the teacher model. The output module is used to output the liquid level prediction value of the student model.

[0015] Furthermore, the aforementioned teacher model module includes: The encoder unit consists of multiple layers of stacked Transformer-LSTM, with each Transformer-LSTM layer integrating a multi-head attention mechanism and an LSTM module. This part takes the input multivariate time series as the processing object, captures the global dependencies between variables through multi-head attention, enhances the learning of temporal features through LSTM, and finally outputs a high-dimensional feature representation rich in contextual information. Fusion and Modulation Unit: The high-dimensional feature representation of the encoder output is modulated using FiLM technology, and the condition information on which the modulation process depends is the working condition code. Through channel-level linear scaling and offset, the features are adapted to the data distribution of different working conditions. Output unit: Receives the FiLM modulated features, performs a linear transformation through the Dense layer, and finally outputs the liquid level value.

[0016] Thirdly, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-mentioned liquid level prediction method.

[0017] Fourthly, the present invention proposes a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the above-described liquid level prediction method.

[0018] Advantages of this invention: 1. Strong adaptability to multiple operating conditions: By explicitly introducing an operating condition perception module, the model can dynamically adjust its prediction strategy according to the current operating state, responding quickly and with high prediction accuracy when operating conditions change. 2. High prediction accuracy: The Transformer-LSTM hybrid architecture fully leverages the advantages of both networks, enabling it to model global spatiotemporal correlations and capture local temporal dynamics, thus achieving a more accurate characterization of the liquid level system, which is characterized by strong nonlinearity and multivariable coupling. 3. Balancing accuracy and efficiency: Through knowledge distillation technology, knowledge from large models is transferred to smaller models, so that the final deployed model maintains high accuracy while significantly reducing computational overhead, meeting the stringent real-time requirements of industrial sites. 4. Good robustness: Especially under extreme conditions such as breach accidents, the model can provide more reliable advance predictions based on the learned accident evolution patterns, giving operators valuable time to take intervention measures. Attached Figure Description

[0019] Figure 1 This is the overall system architecture diagram of the liquid level prediction method proposed in this invention; Figure 2 This is a structural diagram of the working condition sensing module; Figure 3 This is a detailed structural diagram of the Transformer-LSTM teacher model; Figure 4 It is a FiLM characteristic modulation; Figure 5 This is a flowchart of the knowledge distillation training process. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.

[0021] See Figure 1 This invention proposes a liquid level prediction method, comprising the following steps: S1: Collect historical operating data (parameters such as liquid level, inlet and outlet flow rate, pressure and temperature) of the steam generator under various operating conditions to build a training dataset.

[0022] Specifically, the various operating conditions include at least the normal steady-state operating condition, the swaying operating condition caused by marine conditions or external disturbances, and the rupture accident operating condition caused by pipeline rupture.

[0023] S2: Build and train a working condition awareness module that can identify the current working condition type based on the input data and output the corresponding working condition feature vector.

[0024] Specifically, see Figure 2The working condition perception module is based on HDBSCAN clustering and Transformer Encoder encoding, and its specific implementation is as follows: Statistical features (including mean, standard deviation, median, range of variation, skewness, kurtosis, etc.), frequency domain features (fast Fourier transform results, autocorrelation coefficient, etc.), and structural features (PCA principal components) of the working condition time series are extracted and standardized to obtain the working condition feature dataset. The HDBSCAN density hierarchical clustering algorithm is used to cluster the feature dataset, automatically remove noise, and generate an anchor-positive-negative triplet sample set. The original working condition time series is processed using Transformer Encoder, and latent features are extracted through a multi-layer attention mechanism to output a high-dimensional working condition embedding. A loss function is constructed based on the contrastive learning criterion to maximize the embedding similarity of samples within the same cluster and minimize the embedding similarity of samples across clusters, and the encoder parameters are iteratively optimized. Save the encoder parameters after training convergence, input any working condition time series into the encoder, and output a low-dimensional working condition embedding for subsequent prediction models or multi-working condition analysis.

[0025] S3: Construct a hybrid neural network based on Transformer and LSTM as the teacher model. This model takes the historical operating data sequence as input and receives the operating condition feature vector from S2. It adaptively modulates the encoded features of the hidden layer through the FiLM feature modulation mechanism, thereby achieving accurate prediction of future liquid level changes.

[0026] Specifically, see Figure 3 The teacher model consists of three parts: The encoder part consists of multiple layers of stacked Transformer-LSTM, and each Transformer-LSTM layer integrates a multi-head attention mechanism and an LSTM module. This part takes the input multivariate time series as the processing object, captures the global dependencies between variables through multi-head attention, enhances the learning of temporal features through LSTM, and finally outputs a high-dimensional feature representation rich in contextual information. Fusion and modulation section: The high-dimensional feature representation of the encoder output is modulated using FiLM technology, and the condition information on which the modulation process depends is the working condition coding. Through channel-level linear scaling and offset, the features are adapted to the data distribution of different working conditions. Output section: A Dense fully connected layer is set up; this layer receives the FiLM modulated features and outputs the final liquid level value through linear transformation.

[0027] S4: Train the teacher model using the training dataset to minimize the error between the predicted liquid level value and the actual value of the target data. S5: Build a student model with a simpler structure and less computation.

[0028] Specifically, the student model is constructed using a single-layer Transformer-LSTM architecture, which can significantly reduce the number of model parameters and computational complexity, facilitating subsequent online prediction, real-time adjustment, and deployment.

[0029] S6: See also Figure 5 The knowledge distillation technique is used to guide the training of student models by using a trained teacher model as the "teacher". This allows the output of the student model to not only fit the real liquid level label, but also fit the "soft label" output of the teacher model.

[0030] Specifically, the knowledge distillation process in step S6 has a loss function consisting of two parts: Student loss: the mean square error between the student model's predicted value and the actual liquid level value; Distillation loss: KL divergence between the output distributions of the student model and the teacher model; The total loss function is: , where α is a hyperparameter balancing the weights of the two terms.

[0031] S7: In practical applications, the real-time collected operating data is input into the trained operating condition perception module and student model to obtain the final steam generator liquid level prediction result.

[0032] Based on the above liquid level prediction method, this invention proposes a module including a data acquisition and preprocessing module, a working condition sensing module, a teacher model module, a student model module, a knowledge distillation training module, and an output module.

[0033] The data acquisition and preprocessing module is used to collect historical operating data (parameters such as liquid level, inlet and outlet flow rate, pressure and temperature) under various working conditions, clean and normalize the data, and construct a training dataset. The operating condition perception module processes operating condition time series data and outputs low-dimensional operating condition features to support subsequent prediction models or multi-operating condition analysis. Its working process is as follows: first, it extracts and standardizes the statistical / frequency domain features of the operating condition time series; then, it uses HDBSCAN clustering to generate contrastive learning triples, extracts latent features through a Transformer Encoder, and optimizes parameters using contrastive learning; finally, it outputs a low-dimensional operating condition embedding.

[0034] The teacher model module is a Transformer-LSTM hybrid prediction master model, which serves as the core of high-precision liquid level prediction, realizing multivariate time series feature extraction, working condition information fusion, and final liquid level numerical output. The structure of the teacher model module includes: a. Temporal encoder unit: It consists of multiple layers of Transformer-LSTM stacked together. Each layer integrates a multi-head attention mechanism and an LSTM module. It processes the input multivariate time series, captures the global dependencies between variables through multi-head attention, and combines LSTM to enhance the learning of temporal features. Finally, it outputs a high-dimensional feature representation rich in contextual information. b. Operating condition information fusion and modulation unit: The high-dimensional feature representation output by the Transformer encoder is modulated using FiLM technology (the modulation condition information is the operating condition code). Through channel-level linear scaling and offset, the features are adapted to the data distribution of different operating conditions. c. Output unit: Receives the FiLM modulated features, performs a linear transformation through the Dense layer, and finally outputs the liquid level value.

[0035] The knowledge distillation training module compresses the "knowledge" of the teacher model into a smaller student model and acts as the "teacher" to guide the training of the student model, so that the output of the student model not only fits the real liquid level label, but also fits the "soft label" output of the teacher model.

[0036] The training process for the knowledge distillation training module is as follows: a. First, the complex teacher model is fully trained using complete training data to achieve extremely high prediction accuracy.

[0037] b. Then, construct a simpler student model (single-layer Transformer-LSTM).

[0038] c. During the distillation stage, the student model not only learns from real liquid level data (hard labels), but also from the probabilistic prediction distribution output by the teacher model (soft labels). The soft labels contain rich patterns of liquid level changes under different operating conditions learned by the teacher model, enabling the student model to learn similar generalization capabilities with fewer parameters.

[0039] The output module is used to output the liquid level prediction value of the student model.

[0040] Example: The following uses the steam generator of a marine nuclear power plant as an example to predict the liquid level using the above method: S1: Data Acquisition and Preprocessing Historical data were collected under various operating conditions, including normal operation, oscillation, and main steam pipe rupture accidents. Variables included: liquid level, feedwater flow rate, steam flow rate, average temperature of the primary coolant, and pressure. The data were cleaned, normalized, and constructed into a time series sample.

[0041] S2: Construct a training condition awareness module 1. Collect operating condition time series data and extract two types of core features: first, statistical features (including mean, standard deviation, median, range of variation, skewness, kurtosis, etc.); second, frequency domain features (fast Fourier transform results, autocorrelation coefficient, etc.); and third, structural features (PCA principal components). Standardize the extracted features to eliminate dimensional differences and obtain the operating condition feature dataset. 2. The HDBSCAN density hierarchical clustering algorithm is used to cluster the standardized working condition feature dataset; the algorithm automatically analyzes the local density differences of the samples, removes noisy data, and generates the anchor-positive-negative triplet sample set required for comparative learning. 3. Construct a Transformer Encoder as the core of feature encoding, input the original working condition time series into the encoder; capture the global temporal dependency through a multi-layer attention mechanism, extract the implicit working condition features in the data, and output a high-dimensional working condition embedding. 4. Construct a loss function based on the contrastive learning criterion, using the triplet sample set as training data, and achieve dual optimization through the loss function: maximize the embedding similarity of samples in the same cluster (anchor and positive), and minimize the embedding similarity of samples across clusters (anchor and negative); iteratively optimize the Transformer Encoder parameters until the model converges. 5. Save the Transformer Encoder parameters after training convergence to form a trained working condition perception module; input any working condition time series into this module, and the encoder outputs a low-dimensional working condition embedding, which is used as a working condition feature vector for subsequent prediction models or multi-working condition analysis.

[0042] S3: Build and train the teacher model 1. Encoder section It consists of multiple layers of stacked Transformer-LSTM, with each Transformer-LSTM layer integrating a multi-head attention mechanism and an LSTM module. This part takes the input multivariate time series as the processing object, efficiently captures the global dependencies between variables through multi-head attention, and enhances the learning of temporal features through LSTM. The final output is a high-dimensional feature representation with the dimension of [sequence length, feature dimension]. 2. Fusion and Modulation Section For a high-dimensional feature representation of the Transformer-LSTM output, see [link to documentation]. Figure 4 The modulation is performed using FiLM technology. The condition information on which the modulation process depends is the operating condition coding. Through channel-level linear scaling and offset, the features are adapted to the data distribution of different operating conditions. 3. Output section A single Dense layer is set up to output the predicted liquid level value at a specified future time step through linear transformation; the mean squared error is used as the loss function during the model training phase; the model is trained using the Adam optimizer with the mean squared error (MSE) as the loss function.

[0043] S4: Build a student model and train it using knowledge distillation. The student model uses a single-layer Transformer Encoder network. During the knowledge distillation training phase, the loss function is set as follows:

[0044] S5: Output prediction results The trained student model and the condition sensing module are integrated into the target control system. The system receives sensor data in real time, the condition sensing module first determines the condition, and then sends the data and condition vector to the student model to output the liquid level prediction value in real time for display or advanced controller.

[0045] In summary, this invention proposes a steam generator liquid level prediction method based on operating condition perception mechanism and knowledge distillation. This method can be applied to condition monitoring and fault prediction technologies in nuclear power, thermal power, and chemical industries, addressing the problems of poor generalization ability and unstable prediction accuracy of existing liquid level prediction models under various significantly different operating conditions, such as normal operation, swaying, and breach accidents. This invention significantly improves the accuracy and robustness of steam generator liquid level prediction under various complex operating conditions, providing key technical support for the safe operation and advanced control of nuclear power plants.

[0046] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0047] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0048] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0049] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0050] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions. These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams.

[0051] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0052] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can still adjust the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Therefore, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A liquid level prediction method, characterized in that, Includes the following steps: S1: Collect historical operating data under various working conditions to construct a training dataset; S2: Construct and train the working condition perception module, which can identify the current working condition type based on the input data and output the corresponding working condition feature vector; S3: Construct a hybrid neural network based on Transformer and LSTM as the teacher model. This model takes the historical operation data sequence as input and receives the operating condition feature vector from step S2. It adaptively modulates the encoded features of the hidden layer through the FiLM feature modulation mechanism, thereby achieving accurate prediction of future liquid level changes. S4: Train the teacher model using the training dataset to minimize the error between the predicted liquid level value and the actual value of the target data. S5: Construct a student model; S6: Using knowledge distillation technology, the trained teacher model guides the training of the student model, so that the output of the student model not only fits the real liquid level label, but also fits the soft label output of the teacher model. S7: Input the real-time collected operating data into the trained operating condition perception module and student model to obtain the final steam generator liquid level prediction result.

2. The liquid level prediction method according to claim 1, characterized in that: The condition sensing module in step S2 is based on HDBSCAN clustering and Transformer Encoder encoding, and its specific implementation is as follows: Statistical, frequency, and structural features of the operating condition time series are extracted, and the features are standardized to obtain the operating condition feature dataset. The HDBSCAN density hierarchical clustering algorithm is used to cluster the feature dataset, automatically remove noise, and generate an anchor-positive-negative triplet sample set. The original working condition time series is processed using Transformer Encoder, and latent features are extracted through a multi-layer attention mechanism to output a high-dimensional working condition embedding. A loss function is constructed based on the contrastive learning criterion to maximize the embedding similarity of samples within the same cluster and minimize the embedding similarity of samples across clusters, and the encoder parameters are iteratively optimized. Save the encoder parameters after training convergence, input any working condition time series into the encoder, and output a low-dimensional working condition embedding for subsequent prediction models or multi-working condition analysis.

3. The liquid level prediction method according to claim 2, characterized in that: The teacher model in step S3 includes: The encoder part consists of multiple layers of stacked Transformer-LSTM, and each Transformer-LSTM layer integrates a multi-head attention mechanism and an LSTM module. This part takes the input multivariate time series as the processing object, captures the global dependencies between variables through multi-head attention, enhances the learning of temporal features through LSTM, and finally outputs a high-dimensional feature representation rich in contextual information. Fusion and modulation section: The high-dimensional feature representation of the encoder output is modulated using FiLM technology, and the condition information on which the modulation process depends is the working condition coding. Through channel-level linear scaling and offset, the features are adapted to the data distribution of different working conditions. Output section: A Dense fully connected layer is set up; this layer receives the FiLM modulated features and outputs the final liquid level value through linear transformation.

4. The liquid level prediction method according to claim 3, characterized in that: The student model in step S5 includes: It is built using a single-layer Transformer-LSTM architecture, which significantly reduces the number of model parameters and computational complexity, making it easier for subsequent online prediction, real-time adjustment and deployment.

5. The liquid level prediction method according to claim 4, characterized in that: The knowledge distillation process in step S6 has a loss function consisting of two parts: Student loss: the mean square error between the student model's predicted value and the actual liquid level value; Distillation loss: KL divergence between the output distributions of the student model and the teacher model; The total loss function is: , where α is a hyperparameter balancing the weights of the two terms.

6. The liquid level prediction method according to claim 1, characterized in that: The various operating conditions in step S1 include at least the normal steady-state operating condition, the swaying operating condition caused by ocean conditions or external disturbances, and the breach accident operating condition caused by pipeline rupture.

7. A liquid level prediction system, characterized in that: It includes a data acquisition and preprocessing module, a working condition perception module, a teacher model module, a student model module, a knowledge distillation training module, and an output module; The data acquisition and preprocessing module is used to collect historical operating data under various working conditions, clean and normalize the data, and construct a training dataset. The working condition perception module can identify the current working condition type based on the input data and output the corresponding working condition feature vector. The teacher model module takes historical operating data sequences as input and receives operating condition feature vectors from the operating condition perception module. It adaptively modulates the hidden layer encoded features through the FiLM feature modulation mechanism, thereby achieving accurate prediction of future liquid level changes. The knowledge distillation training module guides the training of student models through a pre-trained teacher model, so that the output of the student model not only fits the real liquid level label, but also fits the soft label output of the teacher model. The output module is used to output the liquid level prediction value of the student model.

8. The liquid level prediction system according to claim 7, characterized in that: The teacher model module includes an encoder unit, a fusion and modulation unit, and an output unit; The encoder unit is composed of multiple layers of Transformer-LSTM stacked together, and each Transformer-LSTM layer integrates a multi-head attention mechanism and an LSTM module. This part takes the input multivariate time series as the processing object, captures the global dependency between variables through multi-head attention, enhances the learning of temporal features through LSTM, and finally outputs a high-dimensional feature representation rich in contextual information. The fusion and modulation unit is used to represent the high-dimensional features of the encoder output. It uses FiLM technology for modulation, and the condition information on which the modulation process depends is the working condition coding. Through channel-level linear scaling and offset, the features are adapted to the data distribution of different working conditions. The output unit is used to receive the FiLM modulated features, perform linear transformation through the Dense layer, and finally output the liquid level value.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the liquid level prediction method according to claims 1-6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of the liquid level prediction method as described in claims 1-6.

Images