A nuclear power plant distributed control system and control method

Abstract

The application discloses a nuclear power plant distributed control system and a control method, and relates to the technical field of digital control of nuclear power stations, and comprises the following steps: a global optimization layer generates a cross-unit general control strategy library based on a migration learning algorithm according to a decision cycle of the global optimization layer; a coordination control layer dynamically allocates corresponding secondary loop subsystem control weights based on a multi-agent attention network according to a decision cycle of the coordination control layer when each secondary loop subsystem is operated by using the corresponding migratable control strategy of the unit equipment; and an execution control layer acquires a primary loop temperature, a primary loop pressure and a coolant flow rate according to a decision cycle of the control layer when the dynamically allocated secondary loop subsystem control weights are operated, calculates a safety constraint matrix and solves a quadratic programming problem with a relaxation factor through a safety manifold projection module, obtains modified secondary loop subsystem control weights, and converts the modified secondary loop subsystem control weights into control instructions and issues the control instructions to a DCS field control station.

Description

Technical Field

[0001] This application relates to the field of digital control technology for nuclear power plants, and in particular to a distributed control system and control method for nuclear power plants. Background Technology

[0002] The distributed control system (DCS) of a nuclear power plant is the core technology carrier for ensuring the safe and stable operation of nuclear power units and achieving economical power generation and long-term operation and maintenance of equipment.

[0003] The stable operation of a nuclear power plant requires the DCS system to simultaneously achieve multiple objectives: maximizing power generation efficiency, which means increasing reactor power output as much as possible while ensuring safety, optimizing the energy conversion process, and reducing energy loss; extending equipment lifespan, which requires precise control of equipment operating parameters to avoid accelerated equipment aging due to factors such as overstress and overheating; and maintaining safety margins, which requires ensuring sufficient safety redundancy in the system under various operating conditions to prevent accidents. However, traditional proportional-integral-derivative (PID) control strategies mainly rely on linear adjustment based on system deviation. When faced with the interdependent objectives of power generation efficiency, equipment lifespan, and safety margins, it is difficult to dynamically balance them based on real-time operating conditions. For example, during sudden load changes, pursuing increased power generation efficiency may lead to equipment overload, shortening equipment lifespan, and reducing safety margins. Summary of the Invention

[0004] The purpose of this application is to provide a distributed control system and control method for nuclear power plants, which can solve the problems of multi-objective conflict, time-varying characteristics, and delayed response in emergency conditions.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] In a first aspect, this application provides a distributed control system for a nuclear power plant, comprising: a global optimization layer, a coordination control layer, and an execution control layer; the global optimization layer is used to generate a cross-unit general control strategy library based on a transfer learning algorithm according to the decision cycle of the global optimization layer; wherein, the transferable strategies in the cross-unit general control strategy library can minimize fuel costs and maximize equipment maintenance cycles; the coordination control layer is used to dynamically allocate corresponding secondary loop subsystem control weights based on a multi-agent attention network according to the decision cycle of the coordination control layer when using the transferable control strategies of the unit equipment where each secondary loop subsystem is located; the execution control layer is used to obtain the primary loop temperature, primary loop pressure, and coolant flow rate according to the decision cycle of the control layer when using the dynamically allocated secondary loop subsystem control weights, calculate the safety constraint matrix based on the primary loop temperature, primary loop pressure, and coolant flow rate through a safety manifold projection module, solve a quadratic programming problem with relaxation factors, obtain the corrected secondary loop subsystem control weights, and convert the corrected secondary loop subsystem control weights into control commands and issue them to the DCS field control station.

[0007] Secondly, this application provides a control method for a distributed control system of a nuclear power plant. The control method is applied to the distributed control system of a nuclear power plant, which includes a global optimization layer, a coordination control layer, and an execution control layer. The control method includes: the global optimization layer generating a cross-unit general control strategy library based on a transfer learning algorithm according to its decision cycle; wherein the transferable strategies in the cross-unit general control strategy library can minimize fuel costs and maximize equipment maintenance cycles; when the coordination control layer uses the transferable control strategies of the unit equipment where each secondary loop subsystem is located, it dynamically allocates the corresponding secondary loop subsystem control weights based on a multi-agent attention network according to the decision cycle of the coordination control layer; when the execution control layer uses the dynamically allocated secondary loop subsystem control weights, it obtains the primary loop temperature, primary loop pressure, and coolant flow rate according to the decision cycle of the control layer, calculates the safety constraint matrix based on the primary loop temperature, primary loop pressure, and coolant flow rate using a safety manifold projection module, solves a quadratic programming problem with relaxation factors, obtains the corrected secondary loop subsystem control weights, and converts the corrected secondary loop subsystem control weights into control commands and sends them to the DCS field control station.

[0008] This application provides a distributed control system and method for nuclear power plants. By constructing a three-level control architecture comprising a global optimization layer, a coordination control layer, and an execution control layer, each layer collaborates to effectively handle multi-objective conflicts in nuclear power plant operation. The global optimization layer aims to minimize fuel cycle costs and maximize equipment maintenance cycles. It generates a cross-unit general control strategy library through transfer learning to optimize resource utilization and equipment maintenance at a macro level. The coordination control layer uses a multi-agent attention network to dynamically allocate control weights for the secondary loop subsystems to improve power generation efficiency while balancing equipment lifespan. The execution control layer maintains a safety margin through a safety manifold projection module. Therefore, this disclosure can balance multiple objectives such as power generation efficiency, equipment lifespan, and safety margin according to actual operating conditions, avoiding the limitations of traditional PID control strategies in multi-objective conflicts. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of a nuclear power plant DCS control system according to an exemplary embodiment;

[0011] Figure 2 A flowchart illustrating the control method of a nuclear power plant DCS control system provided in an embodiment of this application. Figure 1 ;

[0012] Figure 3 A flowchart illustrating the control method of a nuclear power plant DCS control system provided in an embodiment of this application. Figure 2 . Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] Figure 1 This is a schematic diagram of a nuclear power plant DCS control system according to an exemplary embodiment, such as... Figure 1 As shown, the DCS control system includes: a global optimization layer, a coordination control layer, and an execution control layer;

[0016] The global optimization layer is used to generate a cross-unit general control strategy library based on the transfer learning algorithm according to the decision cycle of the global optimization layer; among them, the transferable strategies in the cross-unit general control strategy library can minimize fuel costs and maximize equipment maintenance cycle.

[0017] Migrative policies are policies that can be shared across all units.

[0018] The transferable control policies are a component of the cross-unit common control policy library. This library stores these transferable policies for use by the coordination control layer and the execution control layer. The generation process of the cross-unit common control policy library involves learning common patterns from multi-unit data, thereby avoiding the need to design policies individually for each unit and improving efficiency and consistency.

[0019] Specifically, the global optimization layer does not generate an independent control policy for each unit and store them all in the policy library. Instead, it generates a general, shareable, and transferable policy through transfer learning algorithms, reducing policy redundancy.

[0020] A nuclear power plant unit refers to a relatively independent and complete power generation unit. It is a core system combination that can independently complete the entire process of nuclear reaction → heat transfer → power generation. It typically includes one nuclear reactor, supporting primary loop subsystems, secondary loop subsystems, and corresponding power generation equipment. A nuclear power plant may have multiple such independent units (for example, a nuclear power plant has Unit 2 and Unit 3). A control strategy that can be adapted to different units is needed to avoid designing a separate strategy for each unit. At the same time, a general strategy can be optimized by using historical data from different units (such as fuel burnup and equipment fouling rate) to achieve the goals of cost savings and extended lifespan.

[0021] The decision cycle for the global optimization layer can be 1-24 hours.

[0022] The general control strategy mainly revolves around two core objectives: minimizing fuel costs and maximizing equipment maintenance cycles. Specifically, it includes:

[0023] Power control strategies for different unit fuel burnup characteristics (such as adjusting reactor power output according to the burnup curve to reduce uranium resource consumption and reprocessing costs).

[0024] Optimization strategies for equipment operating parameters to adapt to changes in the scaling rate of steam generators (such as adjusting water flow rate to reduce scaling and extending maintenance intervals).

[0025] A cross-unit common resource allocation and maintenance planning framework (such as developing a transferable equipment maintenance cycle model based on the common aging patterns of different units).

[0026] The transfer learning process of the global optimization layer involves a deep neural network model. When integrating data from different units, feature extraction technology is used to filter out key features that are highly relevant to operation optimization. These features include not only direct data such as fuel consumption curves and steam generator fouling rates, but also implicit information such as the changing trends of unit operating conditions and seasonal load fluctuation patterns. The generated cross-unit general strategy is not static, but is continuously updated through an online learning mechanism. When new unit operating data is added or the existing unit operating mode changes significantly, the global optimization layer automatically retrains the model to ensure that the general strategy always adapts to the actual situation. In addition, in order to more accurately balance fuel costs and equipment maintenance cycles, a multi-objective optimization algorithm can be used. This algorithm can find a series of optimal strategy combinations that satisfy different preferences in a complex solution space, enabling nuclear power plants to achieve efficient operation under different economic and safety requirements (the goal is to combine multiple objectives into a single objective function by adjusting the weights, for example, by using the weighted sum method or the ε-constraint method).

[0027] In one embodiment, in generating a cross-unit universal control strategy library based on a transfer learning algorithm, the global optimization layer is specifically used to perform the following steps A1-A3:

[0028] A1. Extract key operating characteristics of different units from historical databases. The key operating characteristics include at least one of the following data: fuel consumption curves and steam generator scaling rate data.

[0029] A2. Transform key operational features into meta-strategy feature vectors;

[0030] A3. Through the model-agnostic meta-learning algorithm, the meta-policy feature vector is iteratively optimized to generate a transferable control strategy that can be adapted to different units. In the iterative process of the model-agnostic meta-learning algorithm, a comprehensive loss function that integrates fuel cost and maintenance cycle is adopted.

[0031] In one embodiment, in the aspect of iteratively optimizing the meta-policy feature vector through a model-agnostic meta-learning algorithm to generate a transferable control policy adaptable to different units, the global optimization layer is specifically used for:

[0032] Based on the model-agnostic meta-learning algorithm, the meta-policy feature vector is iteratively optimized through a parameter update formula, which is expressed as follows:

[0033] ;

[0034] in, This represents the parameters in the meta-policy feature vector at time t+1 after the update. Let represent the parameters in the meta-policy eigenvector at time t. Indicates the learning rate. The gradient operator for parameter θ This represents the control task of the i-th unit. Let represent the probability density function of the control task for the i-th unit. Represents the comprehensive loss function. This represents a control strategy model based on the parameter θ in the meta-policy feature vector.

[0035] Specifically, fuel consumption curves and steam generator scaling rate data of different units are extracted from historical databases to construct meta-strategy feature vectors.

[0036] The fuel consumption curves and steam generator scaling rate data in the historical database are not simply numerical records, but contain rich operational information. When extracting data, data mining techniques, such as association rule mining, are used to discover potential relationships between operating parameters of different units. These relationships can be incorporated as additional features into the meta-strategy feature vector. Simultaneously, considering the time-series characteristics of the data, time-series analysis methods are used to predict the trends of fuel consumption curves and steam generator scaling rate data. The prediction results are also incorporated into the meta-strategy feature vector, making the feature vector more forward-looking. To improve data quality and usability, historical data is cleaned and preprocessed to remove outliers and noisy data. Data interpolation and smoothing techniques are used to fill in missing data, ensuring data integrity and continuity.

[0037] A transferable control policy is generated using the Model-Agnostic Meta-Learning (MAML) algorithm, satisfying the parameter update formula:

[0038] ;

[0039] in, This represents the parameters in the meta-policy eigenvector at time t+1 after the update, obtained by adjusting the gradient of the parameters in the meta-policy eigenvector at time t. Let represent the parameters in the meta-policy eigenvector at time t. The learning rate controls the step size for parameter updates and determines the magnitude of each gradient descent. Indicates parameters The gradient operator represents the gradient with respect to the parameters. Find the partial derivative. This represents the control task of the i-th unit. Let represent the probability density function of the control task of the i-th unit, which is also the overall probability distribution of the task space. Represents the comprehensive loss function. Represents the parameters in the meta-policy feature vector. The control strategy model, the parameters here These are all the parameters in the meta-policy feature vector.

[0040] The control task refers to the control strategy optimization scenario of a nuclear power unit under specific operating conditions with the goal of "minimizing fuel costs and maximizing equipment maintenance cycle". It is specifically defined by the core operating state parameters of the unit, including fuel consumption and steam generator scaling rate.

[0041] To accelerate the convergence speed of the MAML algorithm, an adaptive learning rate adjustment strategy, such as Adagrad or Adadelta, can be employed when applying the MAML algorithm. This strategy dynamically adjusts the learning rate based on gradient changes during parameter updates. Simultaneously, considering the diversity of task distribution across different units, task clustering technology is introduced to group similar unit tasks into one category. Meta-learning is then performed on each category of tasks to improve the algorithm's adaptability to different tasks (specifically including: collecting task data: extracting relevant data for each unit's tasks from historical databases, including fuel consumption curves, steam generator scaling rates, and corresponding operating parameters to form an initial task dataset; extracting task features: converting each unit's task into a task feature vector, where features include key indicators such as fuel consumption depth, equipment aging degree, and load fluctuation patterns, which can be referenced from the meta-strategy feature vector construction method; task clustering: using clustering algorithms such as K-means, and clustering all unit tasks based on the similarity of task feature vectors). Clustering is performed to group tasks with similar features into the same cluster. Clustering results are validated using metrics such as silhouette coefficients to assess the effectiveness, ensuring high similarity among tasks within the same cluster and significant differences between tasks across clusters. If necessary, clustering parameters are adjusted and re-clustering is performed. Cluster-specific meta-learning is then applied to each clustered task cluster, using the MAML algorithm to iteratively update meta-policy parameters based on task data within that cluster, generating transferable control policies suitable for that type of task. Furthermore, fuel costs and equipment maintenance cycles in the comprehensive loss function are standardized to eliminate the impact of dimensional differences on algorithm training, ensuring the algorithm treats different optimization objectives more fairly. During training, early stopping and cross-validation techniques are employed to avoid overfitting and improve the model's generalization ability.

[0042] In the above embodiments, the meta-strategy feature vector includes fuel consumption curve related features and steam generator scaling rate related features.

[0043] Characteristics of the fuel consumption curve include:

[0044] Burn-up depth: A core indicator reflecting the degree of nuclear fuel consumption, usually measured in megawatt-days per ton of uranium. It may be an instantaneous value at a specific point in time, or an average, cumulative value, or trend (slope) over an operating cycle.

[0045] Reactivity change trend: The change in the remaining reactivity of the reactor as burnup increases.

[0046] Historical power characteristics: The distribution of power levels during historical operation (such as base load operating time, load tracking amplitude and frequency) significantly affects the shape of the fuel consumption curve.

[0047] Boron concentration change curve: The history of boron concentration adjustment introduced to compensate for changes in reactivity.

[0048] Characteristics related to the scaling rate of steam generators include:

[0049] Scaling rate: The rate at which scale or deposits accumulate on the heat transfer tubes of a steam generator, directly affecting heat transfer efficiency. It may be a statistical value (such as the average daily scaling rate) or a parameter derived from a time series model.

[0050] Increment in thermal resistance: The change in heat transfer resistance caused by scaling is a direct reflection of the scaling rate.

[0051] Historical water quality indicators: long-term trends of water chemical parameters closely related to scaling (such as pH, conductivity, and impurity ion concentration).

[0052] Cleaning and maintenance history: the operating time since the last chemical or mechanical cleaning, serving as an indirect indicator of the degree of scaling.

[0053] Standardizing the fuel costs and equipment maintenance cycles in the comprehensive loss function includes:

[0054] Suppose we have historical data from N generator units to determine standardized parameters.

[0055] Step 1: Data collection and preparation.

[0056] Extract N samples from the historical database, each containing raw indicator values ​​that need to be standardized.

[0057] : Original fuel cost of the nth unit (unit: RMB 10,000 / MWh).

[0058] : The reciprocal of the original maintenance cycle for the nth unit (unit: 1 / month). (Using the reciprocal, its physical meaning is "maintenance frequency per unit time," and the smaller the value, the better, consistent with the target direction of fuel cost).

[0059] Step 2: Calculate the sample statistic for each indicator.

[0060] Regarding fuel cost metrics:

[0061] Calculate the mean of fuel costs for all historical samples;

[0062] Calculate the standard deviation of fuel costs for all historical samples.

[0063] For the inverse maintenance cycle indicator:

[0064] This is used to calculate the mean of the inverse of the maintenance period of all historical samples;

[0065] , is used to calculate the standard deviation of the inverse of the maintenance cycle of all historical samples.

[0066] Step 3: Define the standardized formula.

[0067] For any new unit (or new data point), its original index values ​​are standardized to a similar order of magnitude range.

[0068] The standardized value of the fuel cost for the nth unit: .

[0069] The standardized value of the reciprocal of the maintenance cycle of the nth unit: .

[0070] Step 4: Substitute the standardized values ​​into the loss function.

[0071] The expression for the comprehensive loss function is:

[0072] ;

[0073] in, , , , which is a preset fixed weighting coefficient; fuel costs include uranium resource prices and reprocessing costs. This represents the parameters in the meta-strategy feature vector after adaptation for any given generator set. This represents the parameter in the global meta-policy feature vector. This term measures the difference between the two (L2 norm squared), which is used to prevent the policy from being over-adapted to a single unit and losing cross-unit mobility, thus ensuring the universality of the generated policy.

[0074] Global, general meta-policy parameters After performing a few gradient updates on the historical operating data of a specific generating unit, the resulting customized policy parameters (i.e., the parameters in the meta-policy feature vector adapted for any generating unit in the above embodiments) specifically include:

[0075] Step 1: Task sampling.

[0076] One unit is randomly selected from among many units. It represents a specific adaptation task.

[0077] Step 2: Calculate the adaptation loss.

[0078] Using this unit Historical operating data (note: this uses the unit's own data) is input into the current control strategy model. In this context, the performance of the calculation strategy on the unit, i.e., the adaptation loss, is considered. This loss function is the same as the comprehensive loss function used in meta-training, but the data used in the calculation comes only from the flight crew. .

[0079] Step 3: Perform gradient descent (adaptation step).

[0080] The parameters are updated through 1-5 steps of gradient descent to obtain the adapted parameters:

[0081] ;

[0082] in: It is the inner loop learning rate, used to control the step size for adaptation. It is a loss function For parameters gradient (based on unit) (Data computation).

[0083] The comprehensive loss function here does not have traditional predicted values ​​and sample labels. It is a policy performance evaluation function that scores the policy by simulating its performance in a real environment and finds the optimal policy by optimizing this score.

[0084] Weighting coefficients in the overall loss function , , Although currently set to fixed values, these can be dynamically adjusted through reinforcement learning algorithms in practical applications. For example, algorithms such as Deep Q-Network (DQN) can be used to input environmental information, including: primary loop core parameters (core temperature, primary loop pressure, coolant flow); secondary loop subsystem status: sensor readings of temperature, pressure, and flow of feedwater pumps, turbines, condensers, etc., as well as equipment health indices (reflecting equipment aging, wear, etc.); fuel and equipment related parameters: fuel burnup curves (reflecting the degree of nuclear fuel consumption), steam generator scaling rate (reflecting equipment operating status), etc., reflecting the time-varying characteristics of the system; and operating condition indicators: such as load fluctuation patterns, power output levels, etc., reflecting the real-time operating needs of the nuclear power plant. With the goal of minimizing the comprehensive loss function value, different weight combinations are continuously tried to learn the optimal weight coefficients. The uranium resource price in fuel cost calculation is affected by various factors such as market supply and demand, and therefore exhibits significant volatility. Therefore, a uranium resource price prediction model can be established (using time series analysis, machine learning, or multi-factor regression models), combined with macroeconomic data and other information, to make short-term and long-term predictions of uranium resource prices, thus making fuel cost calculations more accurate. The calculation of reprocessing costs includes not only direct processing costs but also the impact of environmental requirements, technological advancements, and other factors on reprocessing processes and costs. This is achieved by establishing a cost model (specifically a dynamic cost hybrid model based on machine learning, integrating data-driven and physical constraints, dynamically updating fuel (uranium price, reprocessing) and maintenance costs, supporting online learning and threshold-triggered adjustments) for dynamic updates. Furthermore, to more accurately reflect the impact of maintenance cycles on nuclear power plant operation, the reciprocal of the maintenance cycle can be replaced with a maintenance cost function based on equipment reliability theory (maintenance cost = fixed maintenance cost + k × λ(t) / L(t), where λ(t) is the equipment failure rate (calculated based on a reliability model such as the Weibull distribution), L(t) is the remaining life, and k is the variable cost coefficient). This function comprehensively considers factors such as equipment failure probability and maintenance costs, making the overall loss function more consistent with reality.

[0085] The coordination control layer is used to dynamically allocate the corresponding control weights of the secondary loop subsystems based on the multi-agent attention network according to the decision cycle of the coordination control layer when the unit equipment where each secondary loop subsystem is located is running using the transferable control strategy.

[0086] The secondary loop is the energy conversion system in a nuclear power plant that does not directly contact the nuclear reactor and is responsible for converting the heat from the primary loop into electrical energy. It is a crucial link in power generation. Its core logic is as follows: without contact with radioactive materials, it receives heat transferred from the primary loop through a steam generator (the heat exchange equipment between the primary and secondary loops) and heats the water in the secondary loop into steam. The steam drives the turbine to rotate, which in turn drives the generator to generate electricity. After power generation, the steam is cooled back into water by a condenser and then pumped back to the steam generator for reuse. The coordination and control layer dynamically allocates control weights to the pumps, turbines, and condensers to ensure more efficient coordination among these devices (e.g., adjusting the feedwater pump flow and turbine speed based on real-time steam volume), thus ensuring power generation efficiency while avoiding equipment overload.

[0087] The decision cycle of the coordination and control layer is 1-10 minutes. A multi-agent attention network is used to dynamically allocate the control weights of the second-loop subsystem. The second-loop subsystem includes a water pump, a steam turbine, and a condenser.

[0088] During operation, multi-agent attention networks not only adjust weights by exchanging state information among agents (corresponding to different secondary loop subsystems), but also utilize reward mechanisms from reinforcement learning. For example, when the system operates stably and various performance indicators reach the optimization target, agents are given positive rewards; conversely, they are penalized, thereby guiding agents to continuously optimize weight allocation strategies. Considering the complexity of the nuclear power plant operating environment, the coupling relationship between secondary loop subsystems is not fixed and may be affected by various factors such as equipment aging and water quality changes. Therefore, multi-agent attention networks can introduce adaptive mechanisms to monitor changes in the coupling strength between subsystems in real time and automatically adjust the parameters of attention calculation to more accurately capture the dynamic relationships between subsystems. At the same time, to improve the robustness of the system, adversarial training techniques can be added during network training to simulate various interference scenarios, enabling the network to maintain good control performance when facing uncertainty.

[0089] In one embodiment, regarding the dynamic allocation of control weights for corresponding two-loop subsystems based on a multi-agent attention network, the coordination control layer is specifically configured to perform the following steps B1-B5:

[0090] B1. Collect real-time operating data of the secondary loop subsystem, wherein the real-time operating data includes at least one of the following: sensor readings of temperature, pressure, flow rate, etc., and equipment health index of the secondary loop subsystem.

[0091] B2. The encoder module converts the real-time operating data of the two-loop subsystem into feature vectors.

[0092] B3. A four-head attention mechanism combined with feature vectors is used to calculate the coupling weight matrix between the two-loop subsystems.

[0093] B4. Obtain the score value of each real-time running data in each two-loop subsystem based on the coupling weight matrix.

[0094] B5. Obtain the control weights of each secondary loop subsystem based on the score values ​​of each real-time operating data.

[0095] Specifically, the encoder module converts the states of each subsystem into feature vectors. The encoder module contains a 3-layer fully connected neural network;

[0096] In designing the three-layer fully connected neural network of the encoder module, different combinations of activation functions, such as ReLU and LeakyReLU, can be used to improve the network's ability to extract nonlinear features. To further enhance the network's expressive power, residual connection structures can be added to the network to alleviate the gradient vanishing problem, enabling the network to learn deeper features. When training the encoder module, mini-batch gradient descent combined with momentum optimization techniques is used to accelerate training speed and improve convergence stability. Meanwhile, to prevent overfitting, in addition to regularization techniques, Dropout techniques can be used to randomly discard some neurons, reducing co-adaptation among neurons. Furthermore, considering the characteristics of state data from different subsystems, the input data can be normalized, such as using standardized normalization or max-min normalization, to distribute the data within a suitable range and improve network training efficiency.

[0097] A four-head attention mechanism is used to calculate the coupling weight matrix between subsystems. Where Q and K are the query and key matrices, and d = 64;

[0098] Output normalized control weights ,in, The score for each subsystem;

[0099] Specifically, the subsystem status data is obtained by collecting real-time operating data of the secondary loop subsystems (feed pump, steam turbine, condenser), including sensor readings of temperature, pressure, flow rate, and equipment health index, forming input data with a dimension of 256.

[0100] Feature extraction: The encoder module converts the subsystem state data into a 128-dimensional feature vector s. i This vector contains key features of the subsystem's operating state.

[0101] Calculate the score: Based on a 4-head attention mechanism of a multi-agent attention network, combined with feature vector s i Calculate the coupling relationships between subsystems (e.g., through the interaction of query matrix Q and key matrix K) to generate a score z reflecting the importance of each subsystem.i This is for the subsequent normalization of control weights w i This provides the foundation for calculations.

[0102] Each attention in the four-head attention mechanism can focus on different aspects of the relationships between subsystems. For example, one head might focus on flow-related coupling; another on temperature-related coupling; and yet another on the dynamic coupling between pressure parameters. Pressure-related couplings include significant correlations between pressure parameters such as feedwater pump outlet pressure, turbine inlet pressure, and condenser vacuum in the secondary loop (e.g., feedwater pump pressure directly affects turbine operating pressure). The third head can focus on health-state-based relationships between subsystems, including health-state-related couplings: input data includes equipment health indices, and the health status of subsystems (e.g., feedwater pump wear, turbine blade aging) can mutually influence operational coordination (e.g., a decrease in the efficiency of one piece of equipment may lead to abnormal loads on other equipment). To more effectively utilize the results of the multi-head attention mechanism, attention fusion technology can be employed to integrate the multiple heads... The results of attention calculation are weighted and fused, and weights are assigned according to the importance of different heads (the importance and weight allocation of each head need to be combined with the real-time operating conditions of the nuclear power plant; the importance of different heads changes with the operating scenario (e.g., when the load fluctuates, heads related to flow / pressure may be more important; when the equipment is aging, heads related to health status may be more important), and there is no fixed division of "absolutely important" or "absolutely unimportant"); the weight allocation method is as follows: when training the encoder module based on the dual-delay deep deterministic policy gradient (TD3) algorithm, the network automatically learns the weights of each head through the objective function (including the squared term of Q-value error). During the training process, the weights of heads that contribute more to the system control effect (such as subsystem coordination stability and operating efficiency) will be dynamically amplified through the gradient optimization mechanism; conversely, the weights of heads that contribute less will be reduced. The final weights need to meet the normalization requirements (similar to control weight w). i (Normalization logic) ensures that the sum of the weights of each head is reasonable and adapts to the coupling requirements of the subsystem under real-time operating conditions; when calculating the query and key matrices Q and K, the historical operating data and current state information of the subsystem can be combined to extract features through convolutional neural networks or recurrent neural networks, making the obtained query and key matrices more representative; in addition, in order to improve the computational efficiency of the attention mechanism, sparse attention calculation methods can be used to reduce unnecessary computation, especially when there are many subsystems, which can significantly improve the running speed of the system;

[0103] The sparse attention computation method is specifically as follows:

[0104] Determine the sparsity criteria: Based on the physical relationships between subsystems (e.g., flow-pressure coupling between the feedwater pump and the turbine, and steam-vacuum coupling between the turbine and the condenser), pre-define core coupling pairs (e.g., feedwater pump-turbine, turbine-condenser) as attention pairs that must be considered. Set dynamic filtering thresholds: Analyze the distribution of coupling weights between subsystems using historical operating data, and set weight thresholds (e.g., retain only coupling weights). The top 30% of pairs), filtering out non-critical attention pairs with low relevance (such as the weak indirect correlation between the water pump and the condenser).

[0105] Generate a sparse mask matrix: Based on preset core association pairs and dynamic thresholds, generate a binary mask matrix M ( This indicates that the attention pair is retained. (Indicates to ignore), only for The pairs are then used for subsequent calculations;

[0106] Sparse attention computation: When computing the interaction between the query matrix Q and the key matrix K, only the values ​​in the mask matrix are retained. Skip the elements Element-wise calculations reduce the amount of matrix operations;

[0107] Dynamically update the sparsity strategy: Every 1-10 minutes (coordination and control layer decision cycle), the threshold or core associated pairs are readjusted based on the real-time status of the subsystem (such as changes in equipment health index and parameter fluctuations) to ensure that sparsity does not affect the capture of key coupling relationships.

[0108] The training method for the encoder module includes:

[0109] The input data has 256 dimensions and includes sensor readings such as temperature, pressure, and flow rate, as well as equipment health indices.

[0110] For the temperature, pressure, flow rate, and equipment health index in the input data, feature engineering techniques are used for further processing. For example, wavelet transform is applied to the temperature data to extract features of its different frequency components, better reflecting the details and trends of temperature changes; statistical analysis is performed on the pressure data to calculate its mean, variance, peak value, and other statistical characteristics as additional input features. The equipment health index can be calculated using deep learning methods, such as methods based on convolutional neural networks (CNN) or recurrent neural networks (RNN), combining multi-source information such as historical operating data, fault records, and maintenance logs to more accurately assess the health status of the equipment. During data acquisition, considering the differences in sensor installation locations and measurement accuracies, sensor fusion technology is employed to fuse data from multiple sensors. This fusion process includes: data preprocessing (synchronizing data from each sensor in time, removing outliers, and labeling data confidence levels based on installation location differences); dynamic weight allocation (assigning weights to different sensors based on their accuracy level and historical data reliability); fusion calculation (using Kalman filtering or weighted averaging algorithms, combined with weights, to fuse data from similar parameters (such as temperature sensors at different locations) to reduce random errors); and consistency verification (comparing the fusion results with the predicted values ​​from the thermo-hydraulic equations, triggering secondary fusion of redundant sensor data when deviations exceed limits to ensure reliable output). This improves data accuracy and reliability. Simultaneously, to ensure data consistency and compatibility, data from different types of sensors are standardized and normalized to ensure they have the same dimensions and data range.

[0111] The network parameters are optimized using the dual-delay deep deterministic policy gradient (TD3) algorithm, with the objective function being:

[0112] ;

[0113] in, As a discount factor, These are the target network parameters.

[0114] This represents the objective function, which measures the parameters of the Q-network. The optimization error needs to be minimized through gradient descent to improve the network's accuracy in estimating action value.

[0115] It represents the mathematical expectation of the system state s and the control action a, and is used to smooth the error of different samples and avoid the influence of individual extreme data on parameter optimization;

[0116] Indicates the current Q network (parameter is) The output represents the "state-action value" estimate of executing control action a (such as adjusting the speed of the feedwater pump or the opening of the turbine valve) under the current state s (such as the temperature, pressure and other parameters of the secondary loop of a nuclear power plant);

[0117] y is the target value, which serves as the learning objective of the Q-network and consists of two parts:

[0118] r: Immediate reward, such as positive feedback such as reduced parameter fluctuations and improved stability brought about by action a in the coordinated control of nuclear power plant subsystems;

[0119] The "double Q mechanism" of TD3 takes the minimum value of the value estimates of the two target Q networks for "execution of target action a' (optimized control instruction in the new state) under the next state s' (new running parameters after execution of action a)" to reduce Q-value overestimation bias and improve training stability.

[0120] This represents the squared error between the current Q-value and the target value, ensuring that the error is differentiable, which facilitates updating the parameters through gradient descent. .

[0121] When optimizing encoder module network parameters using the TD3 algorithm, noise, such as Gaussian noise or Ornstein-Uhlenbeck noise, can be added during action selection to enhance the algorithm's exploration capability. This allows the agent (in nuclear power plant coordinated control scenarios, an intelligent control unit that uses reinforcement learning algorithms (such as TD3) to perceive system operating states (e.g., subsystem temperature, pressure, etc.) and optimize control strategies by exploring the control action space (e.g., adjusting feedwater pump speed, turbine valve opening, etc.)) to explore new action spaces to some extent, avoiding getting trapped in local optima. Simultaneously, considering the dynamic changes in the nuclear power plant operating environment, online learning and incremental learning techniques are employed to enable the network to update parameters promptly based on new observation data, adapting to environmental changes. Regarding the discount factor in the objective function... The algorithm can be dynamically adjusted according to different operating stages and task characteristics of a nuclear power plant. For example, during the unit startup phase, due to significant changes in system state, the discount factor can be appropriately reduced to focus more on short-term rewards and accelerate the network's learning speed. During the stable operation phase, the discount factor can be appropriately increased to focus more on long-term rewards, improving system stability and optimization effectiveness. Furthermore, to improve the algorithm's computational efficiency, distributed training technology is employed, distributing training tasks across multiple computing nodes for parallel execution, thus shortening training time.

[0122] In particular, considering the dynamic changes in the operating environment of nuclear power plants, online learning and incremental learning technologies are adopted to enable the network to update parameters in a timely manner based on new observation data and adapt to environmental changes. Specifically, this includes:

[0123] Real-time data access and preprocessing:

[0124] The DCS system collects key parameters of the primary / secondary loop (such as core temperature, pressure, equipment health index, etc.) at a sampling rate of ≥100Hz. After noise is removed by sensor fusion technology (such as Kalman filtering), an input sample set for incremental learning is generated to ensure the real-time performance and reliability of the data.

[0125] Incremental parameter update mechanism:

[0126] Based on the online gradient descent algorithm, network parameters (such as the Q-network weights of the TD3 algorithm) are evaluated. Dynamic adjustment: Each time a new data sample is added, only that sample is used to calculate the loss function (such as the objective function). The gradient is used to update parameters through mini-batch gradient updates, avoiding computational delays caused by retraining on full data. Update verification is performed under safety constraints.

[0127] After each parameter update, the control strategy corresponding to the new parameters is quickly simulated and verified to meet the safety constraints through thermal-hydraulic equations. If there is a safety risk, the parameter backtracking mechanism is triggered, and the historical optimal parameters are temporarily replaced. The system is then re-optimized after new data is accumulated.

[0128] The learning rate is adaptively adjusted based on the intensity of environmental changes (such as load fluctuation amplitude and equipment parameter drift rate): a small learning rate is used when the system is stable to maintain strategy stability; when significant changes are detected (such as sudden equipment failure), the learning rate is increased to accelerate parameter convergence and quickly adapt to new operating conditions.

[0129] Regular lightweight model refactoring:

[0130] Every 24 hours, the network is reconstructed in a lightweight manner based on accumulated incremental data. The knowledge distillation technique is used to retain core parameters while eliminating redundant weights, preventing the model size from expanding and affecting real-time performance, and ensuring computational efficiency and adaptability in long-term operation.

[0131] The execution control layer is used to obtain the primary loop temperature, primary loop pressure, and coolant flow rate according to the decision cycle of the control layer when the control weights of the dynamically allocated two-loop subsystem are used. Through the safety manifold projection module, the safety constraint matrix is ​​calculated based on the primary loop temperature, primary loop pressure, and coolant flow rate, and a quadratic programming problem with relaxation factors is solved to obtain the corrected control weights. The corrected control weights are then converted into control commands and sent to the DCS field control station.

[0132] The primary coolant loop is the core safety system in a nuclear power plant that directly contacts the reactor core and is responsible for removing heat from the core; it is also a critical barrier to nuclear safety. Its core function is to absorb the enormous heat generated by the nuclear reaction by allowing coolant to flow through the reactor core, while simultaneously strictly isolating radioactive materials (the coolant may be radioactive, but the primary coolant loop system is sealed to prevent radioactive leakage). The execution control layer must monitor the temperature, pressure, and flow rate of the primary coolant loop in real time. Because these parameters are directly related to nuclear safety, exceeding the limits for primary coolant loop parameters (such as excessively high temperature or abnormal pressure) could trigger safety risks. Therefore, safety constraints need to be calculated using a safety manifold projection module to ensure that control commands do not exceed safety limits.

[0133] The decision cycle of the execution control layer is less than 1 second, and a safety manifold projection module based on a primary loop thermal-hydraulic model is embedded.

[0134] The safety manifold projection module is based on a primary loop thermal-hydraulic model. This model is a highly complex physical model that, in addition to considering basic parameters such as temperature, pressure, and flow rate, incorporates detailed descriptions of the reactor core's physical properties and coolant phase transition processes. In actual operation, to improve the model's real-time performance and accuracy, model reduction techniques can be employed to simplify the model structure and reduce computational load while retaining key physical characteristics. When calculating the safety constraint matrix, the safety manifold projection module considers the uncertainty of measurement data and uses data fusion techniques such as Kalman filtering to process sensor data, improving data reliability and accuracy. Furthermore, the parameters of the safety manifold projection module can be customized for different nuclear power plant design and operating conditions, ensuring effective system safety operation under various circumstances.

[0135] In one embodiment, in the aspect of calculating the safety constraint matrix based on the primary loop temperature, primary loop pressure, and coolant flow rate using the safety manifold projection module and solving a quadratic programming problem with a relaxation factor to obtain the corrected control weights, the execution control layer is specifically used for:

[0136] The control safety manifold projection module will acquire the primary loop temperature. Primary circuit pressure Substitute the values ​​into the thermo-hydraulic equation to calculate the safety constraint matrix; the thermo-hydraulic equation is expressed as: ;in, Represents the safety constraint matrix. This represents the control weights of the two-loop subsystem dynamically allocated by the coordination control layer through a multi-agent attention network;

[0137] The quadratic programming problem aims to minimize the L2 norm squared deviation between the original control weights and the modified control weights output by the coordinated control layer. As a constraint, and = ,in, This represents the safety margin threshold that changes over time (also known as the dynamic safety margin threshold). Represents the safety constraint matrix. This indicates the corrected control weights. represents the basic safety margin, and k represents the margin decay coefficient.

[0138] Real-time acquisition of primary loop temperature ,pressure Coolant flow rate Substitute into the thermo-hydraulic equation to calculate the safety constraint matrix. :

[0139] ;

[0140] in, This represents the input under the control weights of a two-loop subsystem dynamically allocated using a multi-agent attention network.

[0141] Coolant flow rate is a key input parameter for calculating the safety constraint matrix. The execution control layer will acquire the primary loop temperature, pressure, and coolant flow rate in real time, all of which are input data for the thermo-hydraulic equation.

[0142] The safety constraint matrix formula is the Jacobian matrix of the thermal-hydraulic equation with respect to the control input u. The explicit expression lists the partial derivatives of core temperature and primary loop pressure with respect to u (focusing on the core safety indicators and the direct sensitivity of u). However, the coolant flow rate will affect temperature (e.g., changing the core heat transfer efficiency) and pressure (e.g., affecting the primary loop pressure distribution). When calculating the above partial derivatives, their effects are implicitly incorporated through the thermal-hydraulic equation.

[0143] In deriving the partial derivatives of temperature and pressure with respect to the control input u, the influence of coolant flow rate must be incorporated through the thermo-hydraulic equations, and its role is implicit in the calculation of the partial derivatives.

[0144] When acquiring real-time data on primary loop temperature, pressure, and coolant flow, high-precision sensors are used, combined with sensor redundancy technology to ensure data reliability. To compensate for sensor measurement errors, sensor calibration technology and error correction models are employed to process the data in real time. This real-time processing includes: real-time data acquisition and preprocessing: acquiring raw sensor data (such as temperature, pressure, flow, etc.) through a data acquisition unit (sampling rate ≥ 100Hz), synchronously recording data timestamps, and removing obvious outliers; and online calibration parameter retrieval: based on the sensor's factory calibration curve and periodically updated offline calibration correction parameters (such as zero drift and range error coefficients). The system performs real-time basic calibration on the raw data. Using data fusion techniques such as Kalman filtering, combined with measurements from redundant sensors, it performs noise filtering and dynamic correction on individual sensor data to reduce random errors. It also introduces error models related to equipment characteristics (such as the ambient temperature influence model for temperature sensors) and dynamically adjusts correction coefficients based on real-time operating conditions (such as equipment load and ambient temperature) to compensate for system errors. Data consistency verification: The corrected data is compared with the predicted values ​​from the thermal-hydraulic model. If the deviation is within the allowable range (e.g., ≤ preset threshold), valid data is output; if the deviation is too large, a sensor anomaly warning is triggered, and backup sensors are activated. According to the data); when calculating the safety constraint matrix, the thermal-hydraulic equations take into account uncertainties in the operation of the nuclear reactor, such as microphysical changes in nuclear fuel and turbulent effects of coolant flow. Uncertainty quantification methods are used to analyze the calculation results and provide confidence intervals for the safety constraint matrix. (The uncertainty analysis to obtain the confidence intervals for the safety constraint matrix specifically includes: 1. Identifying sources of uncertainty: measurement errors of input parameters, simplification errors of the thermal-hydraulic model, and time-varying characteristics of the system (such as core physical parameter drift caused by nuclear fuel burnup and changes in flow characteristics caused by equipment aging); 2. Quantifying the uncertainties of input parameters: [The text abruptly ends here, likely due to an incomplete translation or source material.] Uncertainty factors are probabilistically modeled. For sensor measurement errors, a probability distribution (e.g., normal distribution, with the mean being the measured value and the standard deviation being the error range corresponding to the sensor's accuracy level) is assigned to each input parameter based on historical calibration data. For model simplification errors, the probability distribution of the error term (e.g., uniform distribution or truncated normal distribution) is determined by comparing experimental data with model predictions. 3. Uncertainty propagation calculation: An efficient uncertainty quantification method (e.g., Latin hypercube sampling, balancing computational efficiency and accuracy) is used to sample the probability distribution of the input parameters, generating a large sample set. Each sample is substituted into the thermo-hydraulic equation to calculate the corresponding safety constraint matrix. Samples are used to realize the propagation of uncertainty from input to output; simultaneously, with the accumulation of nuclear power plant operation data, machine learning algorithms are used to perform online correction and optimization of the thermal-hydraulic equations (specifically, online correction and optimization includes: determining the correction object: targeting key parameters related to linear characteristics in the thermal-hydraulic equations (such as heat transfer coefficient, flow resistance coefficient, etc., which affect core temperature, pressure, and other outputs in the form of linear terms in the equations), and using them as linear coefficients to be corrected; real-time data acquisition and error calculation: acquiring the primary loop temperature in real time through a data acquisition unit (sampling rate ≥ 100Hz). ,pressure Coolant flow rate Measured data is collected; the predicted value of the current thermal hydraulic equation (calculated based on the original linear coefficients) is compared with the measured value to obtain the error value e = measured value - error value; linear coefficient update based on machine learning: online linear regression or incremental gradient descent algorithm is used, with the error e as the feedback signal, to dynamically adjust the linear coefficients in the equation; optimization verification and application: after each round of correction, it is verified whether the prediction error of the corrected equation is within the allowable range (e.g., ≤ preset threshold); if the accuracy requirements are met, the updated linear coefficients are applied to the thermal hydraulic equation for subsequent safety constraint matrix. The calculation is performed; if the conditions are not met, the equation is iteratively corrected based on newly acquired data to improve its accuracy and adaptability. In addition, in order to quickly calculate the safety constraint matrix, parallel computing techniques can be used, such as accelerating the calculation on a GPU, to meet the real-time requirements of the execution control layer.

[0145] Solving the quadratic programming problem with a relaxation factor, the specific mathematical form is as follows:

[0146] ;

[0147] = ;

[0148] The objective function is: ;

[0149] The optimization variable is "control weight adjustment amount". "and dominant relaxation factor" The goal is to minimize the total subsequent cost;

[0150] L2 norm squared measures the deviation between the "original weights of the coordination control layer" and the "corrected weights of the execution layer". The smaller the deviation, the better the power generation efficiency can be preserved.

[0151] : The original control weights output by the coordination control layer (generated based on a multi-agent attention network, with the core objective of improving power generation efficiency).

[0152] : The corrected control weights to be solved in the execution control layer (the safety weights that are eventually sent to the DCS field control station).

[0153] Relaxation factor penalty coefficient (preset value, usually 0.01~0.1), to avoid Excessive size causes safety constraints to fail;

[0154] : Dominant relaxation factor (≥0) allows for slight deviations in safety constraints, avoiding frequent system shutdowns due to sensor noise and parameter fluctuations;

[0155] Constraints: and ;

[0156] Safety constraint matrix;

[0157] Dynamic safety margin threshold;

[0158] K: Margin decay coefficient;

[0159] Basic safety margin;

[0160] In the above formula, the dynamic safety margin threshold Adjusted according to the variable characteristics during nuclear power plant operation.

[0161] To improve efficiency in solving quadratic programming problems with relaxation factors, an algorithm combining the interior-point method and the alternating direction multiplier method (ADMM) can be used. The interior-point method handles inequality constraints, while ADMM decomposes large-scale problems into multiple smaller subproblems for parallel solution. When determining the relaxation factor, its value is dynamically adjusted based on extensive simulation experiments, statistical analysis of actual operating data, and the results of nuclear power plant safety risk assessments. This ensures both system safety and improved control performance. The margin decay coefficient k is determined by considering not only burnup depth but also the type of nuclear fuel and reactor operating history, using a multi-factor model for accurate calculation. Furthermore, to address potential computational anomalies such as numerical instability or lack of solutions, an anomaly handling mechanism is designed to ensure reasonable control commands are provided to the system under any circumstances.

[0162] The execution control layer distributes the modified control weights of the secondary loop subsystem to the DCS field control station via the OPCUA protocol. The ultimate object controlled by the DCS field control station is mainly the equipment of the secondary loop subsystem (feed water pump, turbine valve, etc.), but the ultimate purpose of doing so is to control and stabilize the state of the primary loop.

[0163] When transmitting control commands via the OPCUA protocol, encryption techniques are employed to ensure data security, such as using the AES encryption algorithm. Digital signature technology is also used to ensure data integrity and source reliability. Considering the complexity of nuclear power plant network environments, potential network latency and packet loss issues are considered. Network congestion control and data retransmission mechanisms are implemented to ensure control commands reach the DCS field control station in a timely and accurate manner. The OPCUA protocol supports real-time data monitoring and feedback. While transmitting control commands, it can simultaneously transmit the execution status and equipment operation data from the field control station back to the control system, forming a closed-loop control system. To improve system compatibility and scalability, the OPCUA protocol can be integrated with other industrial communication protocols, such as Modbus and PROFIBUS, facilitating communication and interaction with equipment from different manufacturers.

[0164] The hardware implementation of the secure manifold projection module is as follows:

[0165] The Jacobian matrix is ​​calculated in real time using an FPGA chip, with a calculation frequency ≥1kHz;

[0166] The calculation of the Jacobian matrix specifically includes:

[0167] Define the calculation object: The Jacobian matrix is ​​constructed by taking the partial derivatives of the key output parameters and control input variables (such as coolant flow rate, control rod position, etc., and executing control commands u) in the thermal hydraulic equation.

[0168] Hardware parallel computing: The FPGA performs parallel computing on the partial derivative relationship between each output parameter and the control input through pre-configured hardware logic units.

[0169] Real-time operating condition adaptation: Based on FPGA high-speed data interfaces (such as real-time communication links with DCS systems), it can access the current control input in real time. And the sensor's measured parameters, substituted into the pre-compiled thermo-hydraulic equation partial derivative calculation formula, to ensure that... Complete derivative calculations under operating conditions.

[0170] High-frequency output guarantee: By optimizing the hardware logic layout (such as reducing data interaction latency and reusing intermediate calculation results), the calculation cycle of a single Jacobian matrix is ​​controlled within 1ms, meeting the real-time calculation frequency requirement of ≥1kHz, and providing high-frequency sensitivity data support for safety constraint matrix updates and control strategy optimization.

[0171] When using FPGA chips to calculate the Jacobian matrix in real time, the FPGA hardware architecture can be optimized to further improve the calculation frequency and reduce power consumption. For example, pipeline technology can be used to divide the calculation process into multiple stages, allowing data to be processed in parallel at different stages, thus improving computational efficiency. Hardware reuse technology can be employed (reused hardware includes: basic arithmetic units such as adders, multipliers, and floating-point units, which are repeatedly called when calculating different partial derivatives, avoiding the need to configure separate arithmetic hardware for each partial derivative term; data buffer and interface modules: register groups and FIFO buffers used to temporarily store real-time sensor data and intermediate calculation results, sharing data interaction channels in multiple sets of parallel partial derivative calculations; control logic modules: finite state machines responsible for timing scheduling and operation triggering, which adapt to different output parameters and control inputs in the partial derivative calculation process by dynamically switching input parameters, thus achieving control logic reuse), reducing the waste of hardware resources and lowering power consumption. Meanwhile, to improve the accuracy and stability of calculations, real-time monitoring and compensation of environmental parameters such as temperature and voltage are performed on the FPGA chip. (Real-time compensation specifically includes: real-time monitoring data access: integrating miniature temperature sensors (such as thermocouples and infrared sensors) and voltage monitoring circuits around the FPGA chip to obtain real-time temperature values ​​T_FPGA and supply voltage V_supply at high-frequency sampling (≥1kHz), and transmitting them to the on-chip control module via an internal high-speed data bus.) Compensation strategy triggering and execution: Temperature compensation: when the temperature exceeds the normal operating range (e.g., T_FPGA > 85 degrees Celsius), the core operating frequency is reduced by the dynamic clock management unit (e.g., from 100MHz to 80MHz) to reduce power consumption and heat generation; at the same time, the pre-stored temperature-delay calibration table is called to fine-tune the delay parameters of the timing logic units to avoid timing violations affecting the accuracy of real-time calculations such as the Jacobian matrix. Voltage Compensation: If voltage fluctuations exceed the threshold (e.g., V_supply deviates from the standard value by ±5%), the on-chip power management module (PMIC) is activated to dynamically adjust the supply voltage via the DC-DC converter, maintaining stable output. Voltage feedback calibration is enabled for sensitive arithmetic units (e.g., floating-point multipliers) to correct computational deviations caused by voltage fluctuations. Closed-Loop Verification and Iterative Optimization: After compensation, the built-in self-test module (BIST) verifies the FPGA's core functions (e.g., the output consistency of parallel computing units). If deviations still exist, the compensation parameters are further adjusted, forming a closed-loop mechanism of "monitoring-compensation-verification" to ensure stable operation under fluctuating environmental parameters and avoid computational errors caused by environmental changes. Considering the reconfigurable characteristics of FPGA chips, an online reconfiguration mechanism can be designed. When the system operating conditions change or a fault occurs, the FPGA's logic functions can be reconfigured in a timely manner to ensure that the Jacobian matrix calculation meets system requirements.In addition, to facilitate integration and communication with other hardware devices, standardized interface designs are adopted, such as high-speed serial interface (HSI) and universal asynchronous transceiver (UART), to improve system compatibility and scalability.

[0172] The online reconstruction mechanism includes:

[0173] Trigger monitoring: By comparing the current operating parameters (such as load and equipment status) with preset thresholds in real time, or by receiving fault diagnosis signals (such as sensor failure or excessive calculation deviation), the reconfiguration mechanism is triggered.

[0174] Rapid reconfiguration: The FPGA pre-stores multiple logic configuration files adapted to different operating conditions (such as high load and fault emergency modes). After triggering, the corresponding file is quickly loaded through the on-chip configuration interface to dynamically adjust the logic functions such as parallel computing units and data paths.

[0175] Verification effective: Immediately after reconstruction, the built-in self-test module verifies the logical functions, and the Jacobian matrix calculation results are compared with the output of the thermal-hydraulic model. Once the error meets the standard, the new configuration is officially enabled.

[0176] In the above embodiments, the quadratic programming solver integrates the interior point method algorithm, and the maximum solution time is ≤10ms;

[0177] When integrating the interior-point method algorithm into a quadratic programming solver, the algorithm can be optimized and parallelized to further shorten the maximum solution time. For example, sparse matrix storage technology and fast matrix operation algorithms can be employed (specific algorithms can be iterative fast algorithms based on the characteristics of sparse matrices, commonly including the conjugate gradient method (suitable for symmetric positive definite sparse matrices), GMRES (Generalized Minimum Residual Method, suitable for asymmetric sparse matrices), and fast factorization algorithms such as sparse LU decomposition and sparse Cholesky decomposition. These algorithms reduce invalid computations by utilizing the sparsity of matrices (a large number of zero elements), improving the efficiency of multiplication and inversion operations of sparse matrices such as Jacobian matrices, adapting to the real-time computing needs of nuclear power plants), reducing memory usage and computational load. The parallel computing capabilities of FPGAs or GPUs can be used to parallelize the iterative computation process in the interior-point method algorithm, improving the solution speed. In practical applications, to ensure the reliability and stability of the solver, extensive testing and verification work is carried out, including solving quadratic programming problems of different scales and complexities to check the accuracy and consistency of the solution results. Simultaneously, a fault detection and recovery mechanism is designed to promptly detect and implement corresponding recovery measures when the solver malfunctions, ensuring the normal operation of the system. Furthermore, to facilitate integration with other control algorithms, the quadratic programming solver can be designed with a modular structure, providing a unified interface and calling method for convenient use in different control scenarios.

[0178] The fault detection and recovery mechanisms include:

[0179] Multi-dimensional real-time monitoring: Hardware level: Real-time monitoring of FPGA chip temperature, voltage, computing unit load, and communication interface (such as OPCUA protocol) data packet loss rate, latency, etc. through sensors, and setting thresholds (such as triggering alarms when temperature > 90℃ or voltage fluctuation > ±10%).

[0180] Data level: Input data for solving quadratic programming problems (such as safety constraint matrices) The coefficients of the objective function are subjected to integrity checks (such as checksum comparison), and the output results are subjected to feasibility checks (verifying whether they meet the requirements). When the deviation exceeds the limit, it is marked as abnormal.

[0181] At the algorithm level: monitor the convergence status of the quadratic programming solver (e.g., the number of iterations exceeds the upper limit, the residual does not meet the target), and the parameter update stability of the online learning module (e.g., abnormal gradient fluctuations).

[0182] Fault Classification and Location:

[0183] By matching with a fault feature library (which contains typical features such as pre-stored hardware faults, data anomalies, and algorithm non-convergence), the source of the fault (such as FPGA computing unit faults, sensor data distortion, and solver parameter drift) can be quickly located.

[0184] In the above embodiments, the global optimization layer provides the coordination control layer with macro-level strategies (such as long-term optimization goals); the coordination control layer provides the execution control layer with subsystem control weights; the prediction error of the execution control layer (when it exceeds ±5%) will be fed back to prompt the upper layer to re-optimize.

[0185] The collaboration of each layer forms a closed loop, with non-independent output: the global macro framework is defined, the control weights are coordinated, the execution layer transforms them into specific control commands, and they are sent to the DCS field control station through the OPCUA protocol, which is the final result.

[0186] Fault recovery mechanism in this disclosure:

[0187] Hardware failure recovery:

[0188] Enable redundant FPGA modules (hot backup mode), quickly switch computing tasks through preset configuration files, and power off or isolate faulty modules so that they can be reconnected as backup nodes after repair.

[0189] Data failure recovery:

[0190] If the input data is abnormal, the system will automatically call up the historical valid data from the last 3 cycles for interpolation and completion, or switch to the backup sensor data channel; if the output result is not feasible, the system will temporarily activate the offline pre-calculated safety benchmark solution to ensure that the control commands are compliant.

[0191] Algorithm Fault Recovery:

[0192] When the quadratic programming solver fails to converge, a simplified solution algorithm (such as gradient descent) is temporarily used as a substitute; when the online learning parameters are abnormal, the algorithm is backtracked to the most recently validated parameter snapshot, incremental updates are stopped, and full data retraining and calibration are started.

[0193] Closed-loop verification and alarms:

[0194] After recovery, the control effect is verified by thermal-hydraulic model simulation. Once the indicators are confirmed to meet the standards, the normal operation process is restored. Simultaneously, audible and visual alarms are triggered and fault logs are recorded (including time, type, and recovery measures) to support later traceability and optimization.

[0195] Safety margin According to Section 5.2.2 of NUREG-0800, this is set at 90% of the design baseline value;

[0196] Safety margin Although it is set at 90% of the design baseline value according to Section 5.2.2 of NUREG-0800, it can be dynamically adjusted in actual operation based on the real-time risk assessment results of the nuclear power plant. For example, a risk assessment method based on Bayesian networks can be used to comprehensively consider various factors such as equipment aging, changes in the operating environment, and human error, and to assess the safety risk level of the nuclear power plant in real time.

[0197] In the above embodiments, the decision cycle of the global optimization layer is longer than the decision cycle of the coordination control layer; the decision cycle of the coordination control layer is longer than the decision cycle of the control layer.

[0198] For example, the decision cycle of the global optimization layer is 1-24 hours, the decision cycle of the coordination and control layer is 1-10 minutes, and the decision cycle of the control layer is less than 1 second.

[0199] There are a series of critical pain points in the distributed control system (DCS) architecture of today's nuclear power plants that urgently need to be addressed.

[0200] First, the problem of conflicting multiple objectives is prominent. The stable operation of a nuclear power plant requires the DCS system to simultaneously achieve multiple objectives: maximizing power generation efficiency means increasing reactor power output as much as possible while ensuring safety, optimizing the energy conversion process, and reducing energy loss; extending equipment lifespan requires precise control of equipment operating parameters to avoid accelerated equipment aging due to overstress, overheating, and other factors; and maintaining safety margins requires ensuring sufficient safety redundancy in the system under various operating conditions to prevent accidents. However, traditional proportional-integral-derivative (PID) control strategies mainly rely on linear adjustment based on system deviation. When faced with the interdependent objectives of power generation efficiency, equipment lifespan, and safety margins, it is difficult to dynamically balance them based on real-time operating conditions. For example, during sudden load changes, pursuing increased power generation efficiency may lead to equipment overload, shortening equipment lifespan, and reducing the safety margin.

[0201] Secondly, nuclear power plant systems exhibit significant time-varying characteristics. As nuclear fuel continues to burn up, its physical properties constantly change, leading to a gradual decrease in core reactivity. Simultaneously, long-term operating equipment ages due to wear, corrosion, and other factors. These changes cause significant drifts in key system parameters, such as coolant flow rate, temperature, and reactor power control parameters. For example, in a nuclear power unit that has been operating for many years, the actual flow rate of its main pump deviates from the design value by 10% due to impeller wear. If a fixed-parameter control strategy is still used, it will be impossible to accurately match the dynamic demands of the system, resulting in poor cooling performance, which in turn affects the reactor's thermal performance and severely degrades control performance.

[0202] Finally, the lag in the response of existing safety systems under emergency conditions cannot be ignored. Currently, nuclear power plant safety systems largely rely on threshold-based triggering mechanisms, where corresponding protective actions are activated when monitored parameters exceed preset thresholds. However, in complex coupled fault scenarios, such as simultaneous cooling system pipe ruptures and pump failures, system parameters change in a complex and interconnected manner, making it difficult for a single threshold judgment to comprehensively and promptly detect potential faults. Due to the lack of foresight and control capabilities for such complex coupled faults, there is a significant delay from the occurrence of a fault to an effective response from the safety system, posing a significant potential risk to the safe operation of nuclear power plants and potentially leading to serious accidents.

[0203] In view of these problems, the hierarchical reinforcement learning-driven nuclear power plant DCS and related devices proposed in this disclosure effectively solve the problems of multi-objective conflict, time-varying characteristics, and delayed response in emergency conditions that exist in traditional nuclear power plant DCS control systems, and have many significant beneficial effects:

[0204] Optimized Multi-Objective Control: By constructing a three-tiered control architecture comprising a global optimization layer, a coordinated control layer, and an execution control layer, each layer collaborates to effectively address multi-objective conflicts in nuclear power plant operation. The global optimization layer aims to minimize fuel costs and maximize equipment maintenance cycles. Through transfer learning, it generates a cross-unit universal control strategy library, optimizing resource utilization and equipment maintenance at a macro-level. The coordinated control layer employs a multi-agent attention network to dynamically allocate control weights for the secondary loop subsystems. It can balance objectives such as power generation efficiency and equipment lifespan based on real-time operating conditions, avoiding the limitations of traditional PID control strategies in multi-objective conflicts. For example, during load surges, control weights can be rationally allocated to ensure power generation efficiency while minimizing negative impacts on equipment lifespan and safety margins.

[0205] Adapting to Time-Varying System Characteristics: Addressing the time-varying characteristics of nuclear power plant systems, the global optimization layer extracts data such as fuel burnup curves and steam generator fouling rates from historical databases to construct meta-policy feature vectors. A model-agnostic meta-learning (MAML) algorithm is then employed to generate transferable control strategies. These strategies can update parameters based on system changes such as nuclear fuel burnup and equipment aging, ensuring precise matching between the control strategy and dynamic system requirements. For example, in situations where main pump impeller wear causes flow deviations, the control strategy can be adjusted promptly to ensure cooling effectiveness and reactor thermal performance, effectively addressing drift in key system parameters and improving control performance.

[0206] Enhancing Emergency Response Capabilities: The execution control layer embeds a safety manifold projection module based on the primary loop thermal-hydraulic model. This module acquires parameters such as primary loop temperature, pressure, and coolant flow rate in real time to calculate the safety constraint matrix and solve a quadratic programming problem with relaxation factors. In complex coupled fault scenarios, this module can more comprehensively and promptly detect potential faults, significantly reducing the delay from fault occurrence to effective safety system response compared to traditional threshold-based triggering mechanisms. For example, when a cooling system pipe rupture and pump failure occur simultaneously, a rapid response can be initiated, reducing the risk to nuclear power plant safety and preventing serious accidents.

[0207] High-efficiency and reliable hardware implementation: The secure manifold projection module uses an FPGA chip to calculate the Jacobian matrix in real time at a frequency ≥1kHz. The quadratic programming solver integrates the interior-point method algorithm, with a maximum solution time ≤10ms, and the safety margin bbase is set to 90% of the design baseline value according to authoritative standards. This hardware implementation ensures the high efficiency and reliability of the secure manifold projection module, enabling it to complete complex calculations in a short time and providing strong support for the safe and stable operation of the system.

[0208] The system architecture is complete and practical: the supporting nuclear power plant DCS optimization control device includes a strategy server, a safety barrier module, a data acquisition unit, and an emergency rollback unit. The strategy server is equipped with a high-performance processor to run global optimization and coordination control layer algorithms. The safety barrier module uses a specific FPGA to implement safety manifold projection. The data acquisition unit collects data at a high sampling rate through a redundant protocol. The emergency rollback unit automatically switches to a preset PID controller when the prediction error exceeds the limit. The entire device architecture is complete, and all parts work collaboratively to ensure the effective implementation of optimized operation methods, improving the overall performance and stability of the nuclear power plant DCS system.

[0209] This disclosure also provides a nuclear power plant DCS optimization control device for implementing the above method, including:

[0210] Policy Server: Deployed on the DCS engineering station, equipped with an Intel Xeon Gold 6348 processor, running global optimization layer and coordination control layer algorithms;

[0211] The policy server can employ containerization technology, such as Docker, to encapsulate the global optimization layer and coordination control layer algorithms in separate containers, achieving environment isolation and fine-grained resource management. This facilitates the deployment, updating, and expansion of algorithms, improving system flexibility and maintainability. To cope with large-scale data processing and high-concurrency computing demands, a distributed computing architecture is adopted, distributing computing tasks across multiple server nodes for parallel processing, thereby improving computing efficiency. Simultaneously, a load balancing mechanism is established for the server cluster, dynamically allocating tasks based on the load of each node to ensure the stability of the overall system performance.

[0212] Safety barrier module: Safety manifold projection is implemented using Xilinx Virtex Ultra Scale+FPGA;

[0213] When implementing safety manifold projection using a Xilinx Virtex Ultra Scale+ FPGA, the safety barrier module leverages the FPGA's Dynamic Partial Reconfiguration (DPR) technology to adjust the allocation of internal logic resources in real time according to different operating conditions of the nuclear power plant, optimizing the computational performance of the safety manifold projection. For example, under special operating conditions such as unit startup and shutdown, FPGA resources are reconfigured to improve the calculation accuracy of key parameters. To enhance the security of the safety barrier module, encryption technology is used to encrypt the data transmitted to the FPGA to prevent data leakage and tampering. Simultaneously, a health monitoring system for the safety barrier module is established to monitor the FPGA's operating status in real time, such as temperature and power consumption, to promptly detect potential faults.

[0214] Data acquisition unit: Connects to field sensors via redundant Modbus / TCP protocol, with a sampling rate ≥100Hz;

[0215] When the data acquisition unit connects to field sensors via redundant Modbus / TCP protocols, it employs time synchronization technology to ensure that data collected by different sensors has accurate timestamps, facilitating subsequent data fusion and analysis. To improve the reliability of data acquisition, a fault detection and automatic switching mechanism is adopted. When the main sensor or main network link fails, it automatically switches to a backup sensor or backup network link to ensure the continuity of data acquisition. Simultaneously, the acquired data is compressed and preprocessed in real time to reduce data transmission volume and storage burden, thereby improving data processing efficiency. For example, wavelet compression algorithm is used to compress the data, and moving average filtering algorithm is used to denoise the data.

[0216] Emergency rollback unit: Automatically switches to the preset PID controller when the prediction error exceeds ±5%;

[0217] To achieve seamless switching when the emergency rollback unit switches to a preset PID controller, a pre-synchronization technique is employed. This involves adjusting the parameters of the preset PID controller before the switch to ensure its output closely approximates the current control system's output, reducing system fluctuations during the switch. Simultaneously, the prediction error calculation method is optimized using statistical analysis and machine learning-based approaches to more accurately assess prediction errors and avoid unnecessary switching due to misjudgments. A historical record and analysis system for the emergency rollback unit is established to record and analyze each switching event in detail, summarizing lessons learned and continuously optimizing the switching strategy and the parameters of the preset PID controller.

[0218] Figure 2 This is a flowchart illustrating a control method for a distributed control system in a nuclear power plant according to an exemplary embodiment, such as... Figure 2 As shown, this method is applied to the distributed control system in any of the above embodiments, the distributed control system comprising: a global optimization layer, a coordination control layer, and an execution control layer; the control method comprises:

[0219] 101. The global optimization layer generates a cross-unit general control strategy library based on the transfer learning algorithm according to the decision cycle of the global optimization layer; among them, the transferable control strategies in the cross-unit general control strategy library can minimize fuel costs and maximize equipment maintenance cycle;

[0220] 102. When the coordination control layer is running using the transferable control strategy of the unit equipment where each secondary loop subsystem is located, it dynamically allocates the corresponding control weight of the secondary loop subsystem based on the multi-agent attention network according to the decision cycle of the coordination control layer.

[0221] 103. When the execution control layer uses the dynamically allocated control weights of the two-loop subsystem, it obtains the primary loop temperature, primary loop pressure and coolant flow rate according to the decision cycle of the control layer. Through the safety manifold projection module, it calculates the safety constraint matrix based on the primary loop temperature, primary loop pressure and coolant flow rate and solves the quadratic programming problem with relaxation factors to obtain the corrected control weights of the two-loop subsystem. The corrected control weights of the two-loop subsystem are then converted into control commands and sent to the DCS field control station.

[0222] In one embodiment, a cross-unit universal control strategy library is generated based on a transfer learning algorithm, including:

[0223] Extract key operating characteristics of different units from historical databases. These key operating characteristics include at least one of the following data: fuel consumption curves and steam generator scaling rate data.

[0224] Transform key operational features into meta-policy feature vectors;

[0225] The meta-policy feature vector is iteratively optimized using a model-agnostic meta-learning algorithm to generate a transferable control strategy that can be adapted to different units. A comprehensive loss function that integrates fuel cost and maintenance cycle is adopted in the iterative process of the model-agnostic meta-learning algorithm.

[0226] In one embodiment, the step of iteratively optimizing the meta-policy feature vector using a model-agnostic meta-learning algorithm to generate a transferable control policy adaptable to different generator sets includes:

[0227] Based on the model-agnostic meta-learning algorithm, the meta-policy feature vector is iteratively optimized through a parameter update formula, which is expressed as follows:

[0228] ;

[0229] in, This represents the parameters in the meta-policy feature vector at time t+1 after the update. Let represent the parameters in the meta-policy eigenvector at time t. Indicates the learning rate. Indicates parameters gradient operator, This represents the control task of the i-th unit. Let represent the probability density function of the control task for the i-th unit. Represents the comprehensive loss function. Represents the parameters in the meta-policy feature vector. The control strategy model.

[0230] In one embodiment, the expression for the comprehensive loss function is:

[0231] ;

[0232] in, , and As a preset weighting factor, fuel costs include uranium resource prices and reprocessing costs. This represents the parameters in the meta-strategy feature vector after adaptation for any given generator set. This represents the parameters in the global meta-policy feature vector.

[0233] In one embodiment, the dynamic allocation of corresponding two-loop subsystem control weights based on a multi-agent attention network includes:

[0234] Collect real-time operating data of the secondary loop subsystem, wherein the real-time operating data includes at least one of the following: temperature, pressure, flow rate, and equipment health index of the secondary loop subsystem;

[0235] The encoder module converts the real-time operating data of the two-loop subsystem into feature vectors.

[0236] A four-head attention mechanism combined with feature vectors is used to calculate the coupling weight matrix between the two-loop subsystems;

[0237] The score value of each real-time running data in each two-loop subsystem is obtained based on the coupling weight matrix;

[0238] The control weights of each secondary loop subsystem are obtained based on the score values ​​of each real-time operating data.

[0239] In one embodiment, the step of calculating the safety constraint matrix based on the primary loop temperature, primary loop pressure, and coolant flow rate using the safety manifold projection module and solving a quadratic programming problem with a relaxation factor to obtain the corrected control weights for the second-loop subsystem includes:

[0240] The control safety manifold projection module substitutes the acquired primary loop temperature and primary loop pressure into the thermo-hydraulic equation to calculate the safety constraint matrix.

[0241] The quadratic programming problem aims to minimize the L2 norm squared deviation between the original control weights and the modified control weights output by the coordinated control layer. As a constraint, and = ,in, This represents the safety margin threshold that changes over time (also known as the dynamic safety margin threshold). Represents the safety constraint matrix. This indicates the modified control weights of the two-loop subsystem. represents the safety margin, and k represents the margin decay coefficient.

[0242] In one embodiment, the thermo-hydraulic equation is expressed as: ;

[0243] in, Represents the safety constraint matrix. Indicates the primary circuit temperature. Indicates the primary circuit pressure. This represents the control weights of the two-loop subsystem dynamically allocated by the coordination control layer through a multi-agent attention network.

[0244] In one embodiment, the decision cycle of the global optimization layer is longer than that of the coordination and control layer; the decision cycle of the coordination and control layer is longer than that of the control layer.

[0245] In one embodiment, the decision cycle of the global optimization layer is 1-24 hours, the decision cycle of the coordination and control layer is 1-10 minutes, and the decision cycle of the control layer is less than 1 second.

[0246] Figure 3 This is a flowchart illustrating a control method for a distributed control system in a nuclear power plant according to an exemplary embodiment, such as... Figure 3 As shown, this method is applied to the distributed control system in any of the above embodiments, the distributed control system comprising: a global optimization layer, a coordination control layer, and an execution control layer; the control method comprises:

[0247] Construct a three-level control architecture. The global layer operates by extracting data to construct meta-policy feature vectors. The MALL algorithm is used to generate transferable control policies. The control layer operation is coordinated. The encoder module converts real-time subsystem data into feature vectors. The coupling weight matrix between subsystems is calculated. Normalized control weights are output. The control layer operation is executed. First-loop parameters are obtained, and the safety constraint matrix is ​​calculated. A quadratic programming problem is solved. The prediction error is checked for ±5%. If it exceeds ±5%, the emergency rollback unit switches to the preset PID controller and continues the loop execution; if it does not exceed ±5%, the loop execution continues.

[0248] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0249] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A nuclear power plant distributed control system, characterized by, The distributed control system comprises a global optimization layer, a coordination control layer and an execution control layer; The global optimization layer is configured to generate a cross-unit general control strategy library based on a transfer learning algorithm according to a decision cycle of the global optimization layer; wherein the transferable control strategy in the cross-unit general control strategy library can minimize the fuel cost and maximize the equipment maintenance cycle; The coordination control layer is configured to dynamically allocate a corresponding two-loop subsystem control weight based on a multi-agent attention network according to a decision cycle of the coordination control layer when the transferable control strategy of the unit equipment where each two-loop subsystem is located is used to run, so as to improve the power generation efficiency; The execution control layer is configured to obtain a primary loop temperature, a primary loop pressure and a coolant flow rate according to a decision cycle of the control layer when the dynamically allocated two-loop subsystem control weight is used to run, calculate a safety constraint matrix and solve a quadratic programming problem with a relaxation factor based on the primary loop temperature, the primary loop pressure and the coolant flow rate through a safety manifold projection module, obtain a modified two-loop subsystem control weight, and convert the modified two-loop subsystem control weight into a control instruction and issue the control instruction to a DCS field control station; In the aspect of calculating the safety constraint matrix and solving the quadratic programming problem with the relaxation factor based on the primary loop temperature, the primary loop pressure and the coolant flow rate through the safety manifold projection module, the execution control layer is specifically configured to: control the safety manifold projection module to substitute the obtained primary loop temperature and primary loop pressure into a thermal hydraulic equation to calculate the safety constraint matrix; The quadratic programming problem aims to minimize the L2 norm square deviation of the original control weights and the modified control weights of the coordinated control layer output, with as constraint conditions, and = wherein, represents a safety margin threshold value that changes over time, represents a safety constraint matrix, represents a modified secondary loop subsystem control weight, represents a safety margin, and k represents a margin attenuation coefficient.

2. The system of claim 1, wherein, In the aspect of generating the cross-unit general control strategy library based on the transfer learning algorithm, the global optimization layer is specifically configured to: extract running key features of different units from a historical database, wherein the running key features at least include at least one of the following data: fuel consumption curve and steam generator fouling rate data; convert the running key features into a meta-strategy feature vector; generate a transferable control strategy that can adapt to different units by iteratively optimizing the meta-strategy feature vector through a model-agnostic meta-learning algorithm, and adopt a comprehensive loss function that fuses fuel cost and maintenance cycle in the iterative process of the model-agnostic meta-learning algorithm.

3. The system of claim 2, wherein, In the aspect of generating the transferable control strategy that can adapt to different units by iteratively optimizing the meta-strategy feature vector through the model-agnostic meta-learning algorithm, the global optimization layer is specifically configured to: iteratively optimize the meta-strategy feature vector based on the model-agnostic meta-learning algorithm through a parameter update formula, wherein the parameter update formula is expressed as: ； wherein, denotes a parameter in the meta-policy feature vector at time t+1 after update, denotes a parameter in the meta-policy feature vector at time t, denotes a learning rate, denotes a parameter of a gradient operator, denotes a control task of the i-th unit, denotes a probability density function of the control task of the i-th unit, denotes a comprehensive loss function, denotes a control policy model based on a parameter in the meta-policy feature vector .

4. The system of claim 3, wherein, the expression of the comprehensive loss function is: ； wherein, , and are preset weight coefficients, the fuel cost includes the uranium resource price and the reprocessing cost, denotes a parameter in the adapted meta-strategy feature vector for any unit, denotes a parameter in the global meta-strategy feature vector.

5. The system of claim 1, wherein, In the aspect of dynamically allocating the corresponding two-loop subsystem control weight based on the multi-agent attention network, the coordination control layer is specifically configured to: collect real-time running data of the two-loop subsystem, wherein the real-time running data at least includes at least one of the following data: temperature, pressure, flow rate and equipment health index of the two-loop subsystem; convert the real-time running data of the two-loop subsystem into a feature vector through an encoder module; calculate a coupling weight matrix among the two-loop subsystems by adopting a 4-head attention mechanism combined with the feature vector; According to the coupling weight matrix, score values of each real-time operation data in each two-loop subsystem are obtained; According to the score values of each real-time operation data, control weights of each two-loop subsystem are obtained.

6. The system of claim 1, wherein, The thermal-hydraulic equations are expressed as: ; wherein, denotes a security constraint matrix, denotes a primary loop temperature, denotes a primary loop pressure, denotes a secondary loop subsystem control weight dynamically assigned by the coordination control layer through the multi-agent attention network.

7. The system of claim 1, wherein, the decision-making period of the global optimization layer is greater than the decision-making period of the coordination control layer, and the decision-making period of the coordination control layer is greater than the decision-making period of the control layer.

8. The system of claim 7, wherein, the decision-making period of the global optimization layer is 1-24 hours, the decision-making period of the coordination control layer is 1-10 minutes, and the decision-making period of the control layer is less than 1 second.

9. A control method of a distributed control system of a nuclear power plant, characterized by, The control method is applied to a nuclear power plant distributed control system, and the nuclear power plant distributed control system comprises a global optimization layer, a coordination control layer, and an execution control layer. The global optimization layer generates a cross-unit general control strategy library based on a transfer learning algorithm according to a decision-making period of the global optimization layer, wherein a transferable control strategy in the cross-unit general control strategy library can minimize fuel cost and maximize equipment maintenance period. The coordination control layer dynamically allocates a corresponding two-loop subsystem control weight based on a multi-agent attention network according to a decision-making period of the coordination control layer when each two-loop subsystem is operated using a transferable control strategy of a unit device where the two-loop subsystem is located. The execution control layer obtains a primary loop temperature, a primary loop pressure, and a coolant flow rate according to a decision-making period of the control layer when each two-loop subsystem is operated using the dynamically allocated two-loop subsystem control weight, calculates a safety constraint matrix and solves a quadratic programming problem with a relaxation factor based on the primary loop temperature, the primary loop pressure, and the coolant flow rate through a safety manifold projection module, obtains a modified two-loop subsystem control weight, converts the modified two-loop subsystem control weight into a control instruction, and issues the control instruction to a DCS field control station. The calculation of the safety constraint matrix and the solving of the quadratic programming problem with the relaxation factor based on the primary loop temperature, the primary loop pressure, and the coolant flow rate to obtain the modified two-loop subsystem control weight comprises: The control safety manifold projection module substitutes the obtained primary loop temperature and primary loop pressure into a thermal hydraulic equation to calculate the safety constraint matrix. The quadratic programming problem aims to minimize the L2 norm square deviation of the original control weights and the modified control weights of the coordinated control layer output, with as constraint conditions, and = wherein, represents a safety margin threshold value that changes over time, represents a safety constraint matrix, represents a modified secondary loop subsystem control weight, represents a safety margin, and k represents a margin attenuation coefficient.

Images