Nuclear power cold source intercept net system operation state prediction method based on mixed model

Abstract

The application relates to the technical field of safety monitoring and prediction, and discloses a nuclear power cold source intercepting net system operation state prediction method based on a hybrid model, which comprises the following steps: acquiring and preprocessing multi-source time sequence data of a net body stress, hydrology and water quality; dividing the operation state into three categories of initial operation, filtering work and maintenance to be performed in an unsupervised manner through principal component analysis weighting and hierarchical K-means clustering, and constructing a labeled data set; then training an XGBoost model based on the data set, and combining SHAP analysis to construct an interpretable state classifier; meanwhile, a CNN-LSTM-Attention coupled model is established to perform time sequence prediction on key parameters; finally, the output of the time sequence prediction model is taken as a feature and input into the trained XGBoost classification model, so that joint prediction of the net body tension and the operation state at a future moment is realized. The problem of unlabeled initial data is solved, the space-time features are comprehensively considered, the prediction accuracy and interpretability are combined, and an intelligent solution is provided for predictive maintenance of the nuclear power cold source intercepting net.

Description

Technical Field

[0001] This application relates to the technical field of safety monitoring and prediction, and in particular to a method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model. Background Technology

[0002] Coastal nuclear power plants rely on large amounts of seawater for cooling, and their intakes are highly susceptible to blockage by marine organisms (such as algae and shellfish), floating debris, and other contaminants. In severe cases, this can force reactors to reduce power output or even shut down, posing a significant threat to nuclear safety and economic viability. The interception nets installed in the intake channels are a crucial first line of defense against such risks.

[0003] The operational status of the interception net, such as initial operation (after maintenance), normal filtration, and maintenance pending (severe clogging), is influenced by the complex dynamic coupling of multiple factors. These factors include: the stress distribution of the net itself (such as main cable tension and net strain), the hydrological conditions of the surrounding sea area (current velocity, tide level, wave height), and water quality parameters (temperature, dissolved oxygen, chlorophyll a concentration, etc.). Currently, most nuclear power plants still rely on manual periodic inspections or simple single-parameter threshold alarms to determine the status of the interception net. These methods have significant shortcomings: first, they are highly lagging and cannot provide early warnings; second, they have low accuracy and cannot distinguish between normal fluctuations and abnormal clogging; and third, they lack interpretability and cannot guide specific maintenance decisions (such as cleaning and replacement).

[0004] In recent years, IoT sensing systems related to nuclear power plant cold sources (such as network tension acquisition systems and hydrological and water quality monitoring systems) have been gradually built up and accumulated massive amounts of multi-source time-series data. However, existing technologies have not yet been able to effectively mine the deep coupling relationships between these data to form intelligent closed-loop diagnostic and predictive capabilities. The main technical bottlenecks are: First, there is a lack of labeled training datasets of operational status, and manual labeling is extremely costly and highly subjective; second, state evolution is a strongly time-dependent process, and traditional machine learning models struggle to effectively capture long-range spatiotemporal features; third, the model decision-making process is opaque, making it difficult for operation and maintenance personnel to trust its prediction results and take action accordingly. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model, employing the following technical solution: A method for predicting the operational status of a nuclear power plant cold source interception network system based on a hybrid model includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire multi-source time-series data containing various multi-source parameters, and perform cleaning, missing value imputation and normalization processing; the multi-source time-series data includes at least network stress parameters, hydrological parameters and water quality parameters; Step S2: Principal component analysis is used to perform feature weighting and dimensionality reduction on each multi-source parameter, and hierarchical K-means clustering algorithm is used to cluster the weighted data to construct a labeled dataset; Step S3: Using the multi-source parameters in the labeled dataset as features and the state label as the target, train an XGBoost classification model; use the SHAP method to perform interpretability analysis on the trained XGBoost model, quantify the contribution of each input feature to the state classification, and form an interpretable state classifier. Step S4: Divide the multi-source parameters to be predicted into several parameter groups according to their correlation, and establish an independent CNN-LSTM-Attention coupled prediction model for each group of parameters. Use historical time series data as input and future time data as labels to train each prediction model. Step S5: Connect the trained prediction models with the trained state classifier. Take the historical time series data sequence of length L before the current time t as input. First, each prediction model outputs the predicted multi-source parameter value for the future time t+1. Then, use the predicted parameter value as a feature input to the state classifier to output the predicted running state and its probability for the future time t+1, thus achieving joint prediction.

[0006] Optionally, step S2 includes: First, water quality parameters are clustered. Then, hydrological parameters are clustered again within each water quality cluster. Finally, network stress parameters are clustered a third time within each water quality-hydrological cluster. Based on the stress distribution characteristics and time distribution characteristics of each cluster in the clustering results, each cluster is mapped to an initial running state, a filtering working state, or a state awaiting maintenance.

[0007] Optionally, the process of mapping each cluster specifically includes: The ratio of the stress parameter of the network body to the synchronous hydrological baseline parameter in each cluster is calculated and defined as the stress ratio; Based on the historical maintenance record time points, clusters that are within a specified hour of the most recent maintenance record time point and whose average stress ratio is less than 0.6 times the global average are marked as the initial operating state; Clusters whose stress ratio quartiles on the box plot exceed 1.5 times the global mean, or whose stress distributions, after statistical testing, show a significant difference from the filtered working state clusters and have a higher mean, are marked as being in a state requiring maintenance. The remaining clusters are marked as being in filter working status.

[0008] Optionally, the CNN-LSTM-Attention coupled prediction model uses CNN layers to extract local spatial coupling features of parameters within the same time step, uses LSTM layers to capture long-term dependencies of parameters over time, and uses Attention layers to dynamically assign weights to features at different historical time steps. The CNN layer contains at least one convolutional layer and a pooling layer. The convolutional kernel size is 3 and the stride is 1. The pooling layer uses max pooling and the pooling window size is 2. The LSTM layer contains at least one LSTM unit, and the number of units is 64 or 128. The Attention layer employs the Bahdanau attention mechanism, which performs a weighted summation of the hidden states at each time step output by the LSTM layer.

[0009] Optionally, the stress parameters of the net body include the tension of the main cable, the tension of the reinforcing ribs, and the strain of the net; the hydrological parameters include flow velocity, flow direction, wave height, and tidal level; and the water quality parameters include temperature, salinity, dissolved oxygen, and chlorophyll a concentration.

[0010] Optionally, the historical data window length L is set to 24 hours, and the sampling interval is 1 hour.

[0011] Optionally, in step S1, the missing value imputation uses a sliding window with a window size of 5, and uses the mean of the non-missing data in the window to imput the missing value at the center position; if there are fewer than 3 valid data in the window, linear interpolation is used for imputation.

[0012] Optionally, in step S4, the parameters to be predicted are divided into mechanical parameter group, hydrological parameter group and water quality parameter group, and an independent CNN-LSTM-Attention coupled prediction model is established for each group.

[0013] Optionally, in step S3, when training the XGBoost classification model, the class imbalance problem can be addressed by using synthetic minority class oversampling technology or by setting the scale_pos_weight parameter.

[0014] In summary, this application includes at least the following beneficial technical effects: This application designs a PCA weighted and hierarchical K-means clustering algorithm to automatically and objectively construct a physically meaningful state label dataset from completely unlabeled measured data. By cascading and coupling a CNN-LSTM-Attention temporal prediction model with an XGBoost state classification model, it can not only accurately predict specific parameters such as network stress at future moments but also output the overall future operating state, providing a direct and clear decision-making basis for predictive maintenance. The XGBoost model itself has excellent classification performance, and combined with SHAP global interpretability analysis, it makes the model's decision logic transparent, enhancing the confidence of maintenance personnel in the prediction results. Simultaneously, the CNN-LSTM-Attention model, through spatial feature extraction, temporal dependent memory, and key information focusing mechanisms, can effectively handle multi-source heterogeneous temporal data, ensuring the accuracy and robustness of the predictions. The complete data processing, clustering labeling, model training and coupling framework provided in this application has good versatility and portability. When applied to different nuclear power plants or different network types, the training process of this method can be re-executed using local historical data to obtain a prediction system adapted to the scenario, which has high engineering application value. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the hybrid prediction model construction method of the present invention; Figure 2 This is a schematic diagram of the process of constructing a state label dataset based on unsupervised clustering in step S2 of the present invention; Figure 3 This is a schematic diagram of the process of constructing an interpretable state classification model based on XGBoost in step S3 of the present invention; Figure 4 This is a schematic diagram illustrating the construction process of the time-series prediction model based on the key parameters of CNN-LSTM-Attention in step S4 of the present invention. Figure 5 This is a flowchart of the joint prediction system (time series prediction + state classification) in step S5 of the present invention. Detailed Implementation

[0016] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0017] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0018] This application discloses a method for predicting the operational status of a nuclear power plant cold source interception network system based on a hybrid model, including the following steps: Step S1, Data Acquisition and Preprocessing: Acquire multi-source time-series data containing various multi-source parameters, and perform cleaning, missing value imputation and normalization processing; the multi-source time-series data includes at least network stress parameters, hydrological parameters and water quality parameters; Step S2: Principal component analysis is used to perform feature weighting and dimensionality reduction on each multi-source parameter, and hierarchical K-means clustering algorithm is used to cluster the weighted data to construct a labeled dataset; Step S3: Using the multi-source parameters in the labeled dataset as features and the state label as the target, train an XGBoost classification model; use the SHAP method to perform interpretability analysis on the trained XGBoost model, quantify the contribution of each input feature to the state classification, and form an interpretable state classifier. Step S4: Divide the multi-source parameters to be predicted into several parameter groups according to their correlation, and establish an independent CNN-LSTM-Attention coupled prediction model for each group of parameters. Use historical time series data as input and future time data as labels to train each prediction model. Step S5: Connect the trained prediction models with the trained state classifier. Take the historical time series data sequence of length L before the current time t as input. First, each prediction model outputs the predicted multi-source parameter value for the future time t+1. Then, use the predicted parameter value as a feature input to the state classifier to output the predicted running state and its probability for the future time t+1, thus achieving joint prediction.

[0019] This embodiment uses the interception net of the intake channel of a coastal nuclear power plant as an application object to verify the proposed method for predicting the operating status of a nuclear power cold source interception net system based on a hybrid model. The system is equipped with net tension sensors, an acoustic Doppler current profiler (ADCP), and a multi-parameter water quality analyzer, enabling real-time data acquisition. This embodiment uses historical data from January 1, 2023 to December 31, 2023, with a sampling frequency of 1 time / hour, obtaining a total of 8760 samples at consecutive time points.

[0020] like Figure 1 As shown, step S1: Data acquisition and preprocessing: The three types of multi-source time series data acquired specifically include: Net stress parameters: main cable tension (unit: kN), stiffener tension (unit: kN), net strain (dimensionless, expressed as micro-strain με); Hydrological parameters: surface velocity (unit: m / s), tidal level (unit: m), significant wave height (unit: m); Water quality parameters: seawater temperature (unit: °C), dissolved oxygen concentration (unit: mg / L), chlorophyll a concentration (unit: μg / L); First, the raw data is cleaned to remove null values ​​and outliers (such as negative tensile force) caused by sensor malfunctions, communication interruptions, etc. For the cleaned missing values, a sliding window of size 5 is used to fill in the missing values ​​at the center using the mean of the non-missing data within the window; if there are fewer than 3 valid data points in the window, linear interpolation is used for filling.

[0021] Subsequently, all parameters are normalized using Min-Max to eliminate the influence of dimensions and map them to the [0,1] interval.

[0022] like Figure 2 As shown, step S2: Constructing a state label dataset based on unsupervised clustering; This step aims to automatically construct a state label dataset from 8760 unlabeled samples.

[0023] First, principal component analysis (PCA) was performed to reduce the dimensionality of the water quality parameter group (temperature, dissolved oxygen, chlorophyll a), the hydrological parameter group (flow velocity, tidal level, wave height), and the net structure stress parameter group (main cable tension, stiffener tension, net strain). Taking the water quality parameter group as an example, the eigenvalues ​​and eigenvectors of its covariance matrix were calculated. The top k principal components with a cumulative variance contribution rate of 85% were selected, and the original three-dimensional water quality parameters were weighted and fused into a new low-dimensional eigenvector using the loading matrix. The hydrological parameter group and the net structure stress parameter group were processed in the same way.

[0024] Then, hierarchical K-means clustering algorithm is used to cluster the weighted data. The clustering process is as follows: First-level clustering: K-means clustering is performed on the weighted feature vectors of water quality for all samples. The optimal number of clusters is determined to be 3 by combining the elbow rule (calculating the sum of squared errors within the cluster under different K values ​​and selecting the elbow point) and the silhouette coefficient (selecting the K value that maximizes the average silhouette coefficient), resulting in three water quality clusters: C_w1, C_w2, and C_w3.

[0025] The second layer of clustering: Within each water quality cluster, K-means clustering was performed on the hydrological weighted feature vectors of the samples. Based on the elbow rule and silhouette coefficient evaluation, the optimal number of clusters within each water quality cluster was 2, resulting in six water quality-hydrological composite clusters: C_w1h1, C_w1h2, C_w2h1, C_w2h2, C_w3h1, and C_w3h2.

[0026] The third layer of clustering: Within each water quality-hydrology composite cluster, K-means clustering was performed on the network stress-weighted eigenvectors of the samples. Evaluation showed that the optimal number of clusters within each composite cluster was 2, resulting in a total of twelve final fine-grained clusters.

[0027] After clustering is completed, each final cluster is mapped to a running status label according to objective quantification rules. The specific rules are as follows: Obtain the accurate record time of each net bag maintenance (cleaning or replacement).

[0028] For each final cluster, the "stress ratio" of all samples within the cluster is calculated and defined as: ; in, For a very small positive number (e.g.) To prevent division by zero errors, the main cable tension is used as a representative of the stress.

[0029] Initial operating status: If more than 80% of the sample points in a cluster are located within 72 hours after any maintenance record, and the mean stress ratio of the cluster is less than 0.6 times the global mean stress ratio of all samples, then the cluster is marked as "initial operating status".

[0030] "Maintenance pending": If the quartile (Q3) of the stress ratio of a cluster on the box plot exceeds 1.5 times the global mean of the stress ratio of all samples, the cluster is marked as "Maintenance pending".

[0031] Filtering status: Clusters that do not meet any of the above conditions are marked as "filtering status".

[0032] Using the above automated objective rules, state labels were assigned to all 8760 samples. Statistical analysis revealed approximately 1800 initial running state samples, approximately 5400 filtered working state samples, and approximately 1560 samples awaiting maintenance, initially showing class imbalance. The normalized multi-source parameter data was then merged with the assigned state labels to form a complete state classification dataset for supervised learning.

[0033] like Figure 3 As shown, step S3: Construction of an interpretable state classification model based on XGBoost: The dataset constructed in step S2 is divided chronologically, with the first 80% (approximately 7008 samples) used as the training set and the last 20% (approximately 1752 samples) used as the test set to avoid future information leakage.

[0034] Given the class imbalance in the dataset (the number of filtered working state samples is significantly greater than the other two classes), a synthetic minority oversampling technique (SMOTE) is used to address this issue on the training set. Specifically, for the initial running state and the state to be maintained, which are minority classes, new samples are randomly generated along the lines connecting their K=5 nearest neighbors, thus balancing the number of samples in the three classes in the training set.

[0035] An XGBoost classification model was trained using all nine normalized multi-source parameters (main cable tension, reinforcing rib tension, netting strain, current velocity, tidal level, wave height, temperature, dissolved oxygen, and chlorophyll a) as input features and the state labels generated in step S2 as prediction targets. Model training employed a grid search to optimize hyperparameters, with search ranges including: learning rate (0.01, 0.05, 0.1), maximum tree depth (3, 5, 7), and number of estimators (100, 200, 300). The optimal hyperparameter combination was selected through cross-validation: learning rate 0.05, maximum tree depth 5, and number of estimators 200. Simultaneously, a `scale_pos_weight` parameter was set during training, weighting the samples according to the inverse proportion of the number of samples in each class to further mitigate the impact of class imbalance.

[0036] To enhance the interpretability of the model, the SHAP method is used for global interpretability analysis after training. Specifically, the SHAP value for each feature of each sample in the test set is calculated. The SHAP value, derived from game theory based on Shapley values, fairly distributes the model's predictions to each input feature by simulating the marginal contribution of feature subsets. For classification problems, the SHAP value of a feature for a sample represents the change in the probability that the presence of that feature makes the model predict a certain state compared to the baseline probability (the average prediction probability of all samples). By summing the SHAP values ​​of all samples, feature importance bar charts (taking the mean of the absolute values ​​of the SHAP values ​​of each feature) and partial dependency plots (PDPs) can be created, visually demonstrating how each parameter affects the state judgment result.

[0037] like Figure 4 As shown, step S4: Constructing a temporal prediction model based on the key parameters of CNN-LSTM-Attention: This step aims to predict the input parameter values ​​at future moments, providing dynamic input for the state prediction in step S5. First, based on the calculation and physical meaning of the Pearson correlation coefficient, the nine parameters to be predicted are divided into three groups, and a separate CNN-LSTM-Attention prediction model is built for each group: Mechanical parameters: main cable tension, reinforcing rib tension, and mesh strain.

[0038] Hydrological parameters: flow velocity, tidal level, wave height.

[0039] Water quality parameters: temperature, dissolved oxygen, chlorophyll a.

[0040] Each prediction model takes as input historical values ​​of the within-group parameters over the past L=24 hours (24 consecutive time steps) and outputs predicted values ​​of the within-group parameters for the next hour (1 time step). For example, the mechanical parameter group model has an input shape of (24,3) and an output shape of (1,3). The first 80% (in chronological order) of the dataset from step S2 is used as training data, and the last 20% is used as test data.

[0041] The constructed CNN-LSTM-Attention model structure is as follows (taking the mechanical parameter group as an example; the other two groups have the same structure but different feature dimensions): Input layer: Shape (24,3).

[0042] CNN layers: used to extract local spatial coupling features between three mechanical parameters within the same time step. They include, in sequence: One-dimensional convolutional layer (Conv1D): Uses 64 convolutional kernels of size 3 with a stride of 1, and the activation function is ReLU. The input and output are padded with "same" to keep the time step constant.

[0043] MaxPooling1D: The pooling window size is 2, and the stride is 2, used for dimensionality reduction. After this layer, the time step length becomes 12.

[0044] One-dimensional convolutional layer (Conv1D): uses 128 convolutional kernels of size 3, stride of 1, and activation function ReLU.

[0045] MaxPooling1D: The pooling window size is 2, and the stride is 2. After this layer, the time step length becomes 6.

[0046] The feature map output by the CNN layer has a shape of (6, 128).

[0047] LSTM layer: Used to capture long-term dependencies of parameters as they evolve over time. It consists of two LSTM layers: The first LSTM layer contains 128 units and returns a complete output sequence (return_sequences=True). The input shape is (6,128), and the output shape is (6,128).

[0048] The second LSTM layer contains 64 units and returns a complete output sequence (return_sequences=True). The input shape is (6, 128), and the output shape is (6, 64).

[0049] Attention layer: Dynamically assigns weights to features from different historical time steps (6 steps in total) of the LSTM output. The calculation process is as follows: For each time step Calculate attention weights ,in For the first The hidden state of the LSTM at each time step This is the hidden state at the last time step. , , It is a learnable parameter matrix / vector.

[0050] The weights are normalized using the Softmax function: .

[0051] Calculate the context vector as a weighted sum: The output context vector has a shape of (64,).

[0052] Output layer: A fully connected (Dense) layer with the number of neurons equal to the dimension of the predicted target (3 in this case, representing the main cable tension, reinforcing rib tension, and mesh strain over the next hour). The activation function is linear activation. The output shape is (3,).

[0053] The model training used mean squared error (MSE) as the loss function, Adam as the optimizer, an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. An early stopping strategy (patience=10) was employed to prevent overfitting. Each parameter group model independently completed the above training process, resulting in three trained time series predictors.

[0054] like Figure 5 As shown, step S5: Construction of the joint prediction system for network state: The three CNN-LSTM-Attention prediction models trained in step S4 are concatenated with the XGBoost classifier trained in step S3 to construct an end-to-end joint prediction system. The workflow is as follows: Input: The system receives a 24-hour historical multi-source time series data sequence up to the current time t, i.e., all 9 parameter values ​​of the 24 time steps [t-23,t-22,...,t].

[0055] Parameter prediction: The data from these 24 time steps are fed into three pre-trained prediction models respectively: Historical mechanical parameters (main cable tension, reinforcing rib tension, and mesh strain) are fed into the CNN-LSTM-Attention model of mechanical parameters to obtain the predicted mechanical parameter values ​​at the future time t+1.

[0056] Historical hydrological parameters (flow velocity, tidal level, wave height) are fed into the hydrological parameter group model to obtain the predicted hydrological parameter values ​​for t+1.

[0057] Historical water quality parameters (temperature, dissolved oxygen, chlorophyll a) are fed into the water quality parameter group model to obtain the predicted water quality parameter values ​​for t+1.

[0058] These three predicted values ​​are combined to form a complete 9-dimensional predicted parameter vector at time t+1.

[0059] State prediction: The 9-dimensional prediction parameter vector obtained in step 2 at time t+1 is used as a feature and input into the XGBoost state classification model trained in step S3.

[0060] Output: The XGBoost model outputs the predicted running status (initial running, filtering, or pending maintenance) at time t+1 and its corresponding probability value (e.g., pending maintenance: 0.85, filtering: 0.10, initial running: 0.05). The system also outputs the predicted values ​​of nine specific parameters at time t+1.

[0061] To verify the effectiveness of this invention, an evaluation was conducted on a test set (the last 1752 samples). The joint prediction system was applied to each time point in the test set (except for the last point), inputting real data from the past 24 hours to predict the state and parameters for the next hour. The predicted state was compared with the true state (the state label obtained in step S2). The results showed that the overall accuracy of the XGBoost state classifier reached 94.2%, and the recall rate for the "to be maintained" state reached 91.5%, significantly outperforming the static classification model that only uses the parameters of the current time. Meanwhile, the mean absolute percentage error (MAPE) of the three CNN-LSTM-Attention predictors on the test set were 4.2% for the mechanical parameter group, 5.8% for the hydrological parameter group, and 6.5% for the water quality parameter group, indicating high accuracy in time-series prediction.

[0062] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting the operational status of a nuclear power plant cold source interception network system based on a hybrid model, characterized in that, Includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire multi-source time-series data containing various multi-source parameters, and perform cleaning, missing value imputation and normalization processing; the multi-source time-series data includes at least network stress parameters, hydrological parameters and water quality parameters; Step S2: Principal component analysis is used to perform feature weighting and dimensionality reduction on each multi-source parameter, and hierarchical K-means clustering algorithm is used to cluster the weighted data to construct a labeled dataset; Step S3: Using the multi-source parameters in the labeled dataset as features and the state label as the target, train an XGBoost classification model; use the SHAP method to perform interpretability analysis on the trained XGBoost model, quantify the contribution of each input feature to the state classification, and form an interpretable state classifier. Step S4: Divide the multi-source parameters to be predicted into several parameter groups according to their correlation, and establish an independent CNN-LSTM-Attention coupled prediction model for each group of parameters. Use historical time series data as input and future time data as labels to train each prediction model. Step S5: Connect the trained prediction models with the trained state classifier. Take the historical time series data sequence of length L before the current time t as input. First, each prediction model outputs the predicted multi-source parameter value for the future time t+1. Then, use the predicted parameter value as a feature input to the state classifier to output the predicted running state and its probability for the future time t+1, thus achieving joint prediction.

2. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, Step S2 includes: First, water quality parameters are clustered. Then, hydrological parameters are clustered again within each water quality cluster. Finally, network stress parameters are clustered a third time within each water quality-hydrological cluster. Based on the stress distribution characteristics and time distribution characteristics of each cluster in the clustering results, each cluster is mapped to an initial running state, a filtering working state, or a state awaiting maintenance.

3. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 2, characterized in that, The process of mapping each cluster specifically includes: The ratio of the stress parameter of the network body to the synchronous hydrological baseline parameter in each cluster is calculated and defined as the stress ratio; Based on the historical maintenance record time points, clusters that are within a specified hour of the most recent maintenance record time point and whose average stress ratio is less than 0.6 times the global average are marked as the initial operating state; Clusters whose stress ratio quartiles on the box plot exceed 1.5 times the global mean, or whose stress distributions, after statistical testing, show a significant difference from the filtered working state clusters and have a higher mean, are marked as being in a state requiring maintenance. The remaining clusters are marked as being in filter working status.

4. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, The CNN-LSTM-Attention coupled prediction model uses CNN layers to extract local spatial coupling features of parameters within the same time step, LSTM layers to capture long-term dependencies of parameters over time, and Attention layers to dynamically assign weights to features at different historical time steps. The CNN layer contains at least one convolutional layer and a pooling layer. The convolutional kernel size is 3 and the stride is 1. The pooling layer uses max pooling and the pooling window size is 2. The LSTM layer contains at least one LSTM unit, and the number of units is 64 or 128. The Attention layer employs the Bahdanau attention mechanism, which performs a weighted summation of the hidden states at each time step output by the LSTM layer.

5. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, The stress parameters of the net body include the tension of the main cable, the tension of the reinforcing ribs, and the strain of the net; the hydrological parameters include flow velocity, flow direction, wave height, and tidal level; the water quality parameters include temperature, salinity, dissolved oxygen, and chlorophyll a concentration.

6. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, The historical data window length L is set to 24 hours, and the sampling interval is 1 hour.

7. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, The missing value imputation in step S1 uses a sliding window with a window size of 5, and uses the mean of the non-missing data in the window to imput the missing value at the center position; if there are fewer than 3 valid data in the window, linear interpolation is used for imputation.

8. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, In step S4, the parameters to be predicted are divided into mechanical parameter group, hydrological parameter group and water quality parameter group, and an independent CNN-LSTM-Attention coupled prediction model is established for each group.

9. The method for predicting the operating status of a nuclear power plant cold source interception network system based on a hybrid model according to claim 1, characterized in that, In step S3, when training the XGBoost classification model, the class imbalance problem is addressed by using synthetic minority class oversampling or setting the scale_pos_weight parameter.

Images