Power plant production management method and system based on domestic distributed control system

Abstract

The present invention provides a power plant production management method and system based on a domestic distributed control system, which relates to the technical field of production management, including collecting historical production data of power plant equipment through a pre-built domestic distributed control system and pre-processing, dividing the pre-processed historical production data into different working condition modes and forming a production scenario library, and in the real-time production process, using a pre-trained working condition discrimination model to distinguish the current production working condition in real time; building a corresponding production scheduling optimization model, generating an initial scheduling strategy by solving the production scheduling optimization model, and iteratively optimizing the production scheduling strategy using a reinforcement learning algorithm to obtain a target scheduling strategy; converting the target scheduling strategy into a control instruction, and sending it to a controller of the domestic distributed control system, while displaying the execution progress of the scheduling plan and the response of the equipment in real time through a human-computer interaction interface, and allowing the dispatcher to manually intervene and correct the scheduling instruction through the interface.

Description

Technical Field

[0001] The present invention relates to production management technology, and in particular to a power plant production management method and system based on a domestic distributed control system. Background Art

[0002] Power plant production management is an important part of power system operation, which directly affects the safe, economical and efficient operation of power plants. Traditional power plant production management methods mainly rely on manual experience and centralized control systems, which have the problems of low management efficiency, low control accuracy and poor system reliability. In order to meet the development needs of modern power systems, an advanced power plant production management technology is urgently needed.

[0003] In recent years, distributed control systems (DCS) have been widely used in the field of industrial automation. DCS adopts distributed structure and digital communication technology, and has the advantages of high control accuracy, good system flexibility and high reliability. my country's independently developed domestic DCS system has made great progress in recent years, and has approached or reached the level of similar foreign products in terms of reliability, performance and functions. Domestic DCS systems have price advantages and localized service advantages, and are more suitable for the actual needs of domestic power plants. However, how to effectively integrate domestic DCS systems with power plant production management to achieve intelligent, refined and efficient management of power plant operations is still a strategy that needs to be solved urgently.

[0004] In view of this, the present invention proposes a power plant production management method based on a domestic distributed control system, aiming to give full play to the technical advantages of the domestic DCS system and improve the efficiency and level of power plant production management. The present invention introduces advanced management concepts and optimization algorithms, and systematically integrates and uniformly manages equipment monitoring, parameter optimization, fault diagnosis and other links in the power plant production process, realizing real-time monitoring, intelligent decision-making and rapid response of power plant production and operation. Summary of the invention

[0005] The embodiment of the present invention provides a power plant production management method and system based on a domestic distributed control system, which can solve the strategies in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A power plant production management method based on a domestic distributed control system is provided, comprising:

[0008] The historical production data of power plant equipment is collected and preprocessed through a pre-built domestic distributed control system. The pre-processed historical production data is divided into different operating modes to form a production scenario library. In the real-time production process, the latest production data is dynamically intercepted and its key features are extracted. The pre-trained operating condition discrimination model is used to discriminate the current production condition in real time to obtain the production condition recognition result.

[0009] Based on the production condition identification results and the pre-built power plant process knowledge base, a corresponding production scheduling optimization model is built. By solving the production scheduling optimization model, an initial scheduling strategy is generated, and a reinforcement learning algorithm is used to iteratively optimize the production scheduling strategy to obtain the target scheduling strategy.

[0010] The target scheduling strategy is converted into control instructions at the device level, which are sent from the industrial Internet platform to the controller of the domestic distributed control system through the industrial communication protocol. The controller automatically adjusts the operating status and parameters of the equipment according to the instructions. At the same time, the progress of the scheduling plan execution and the response of the equipment are displayed in real time through the human-computer interaction interface, and the dispatcher is allowed to manually intervene and correct the scheduling instructions through the interface.

[0011] In an optional embodiment,

[0012] The pre-trained working condition identification model is used to identify the current production working condition in real time, and the production working condition identification results include:

[0013] The power plant production system is abstracted into a heterogeneous graph network, where nodes represent equipment and edges represent logical connections and material flows between equipment.

[0014] For each device node, we collect multi-source heterogeneous data and static attribute data collected in real time, use attribute encoder to extract and encode the node attribute data, and obtain the low-dimensional attribute embedding vector of the device node.

[0015] For each edge connecting devices, the interaction information of power plant devices is collected, and the edge encoder is used to extract and encode the edge interaction data to obtain a low-dimensional embedding vector of the device interaction information;

[0016] Input the low-dimensional attribute embedding vector of the device node and the low-dimensional embedding vector of the device interaction information into the pre-trained graph neural network model to obtain the updated state embedding vector of the device node;

[0017] The updated state embedding vector of the equipment node is used as the input feature of the working condition discrimination model to generate the embedded representation of the production status graph. Cluster analysis and anomaly detection are performed on the embedded representation of the production status graph, combined with the prior knowledge of the scene and the pre-built anomaly discrimination rules, to obtain the final production condition recognition result.

[0018] In an optional embodiment,

[0019] Training a graph neural network model involves:

[0020] According to the multi-source heterogeneous data collected in real time, an adaptive learning rate is introduced to dynamically update the parameters of the node attribute encoder and edge encoder. The calculation formula is as follows:

[0021]

[0022] Among them, ω (t) represents the updated model parameters, ω (t-1) represents the model parameters before updating, t represents the current time, η (t) represents the adaptive learning rate at time t, m (t) represents the first-order momentum term at time t, v (t) represents the second-order momentum term at time t, ε represents the smoothing term of parameter update, λ represents the regularization coefficient, and R(·) represents the model parameter regularization term.

[0023] In an optional embodiment,

[0024] Based on the production condition identification results and the pre-built power plant process knowledge base, a corresponding production scheduling optimization model is built. By solving the production scheduling optimization model, the initial scheduling strategy is generated, including:

[0025] Building a power plant process knowledge base based on historical data of power plant production and operation, wherein the power plant process knowledge base includes empirical rules and optimization strategies for power plant production and operation, and building a production scheduling optimization model based on the power plant process knowledge base;

[0026] Using the empirical rules and optimization strategies in the power plant process knowledge base, the strategy search algorithm is used to solve the production scheduling optimization model, the first scheduling plan is generated as the search starting point, and the neighborhood structure and move operation are defined to generate candidate scheduling plans. According to the predetermined acceptance criteria, it is decided whether to accept the candidate scheduling plan. The strategy search algorithm is repeatedly executed until the preset termination condition is reached, and the target scheduling plan is finally output;

[0027] The target scheduling plan is adjusted using sensitivity analysis and parameter optimization methods, and the adjusted scheduling plan is re-substituted into the production scheduling optimization model to generate the final initial scheduling strategy for power plant production scheduling.

[0028] In an optional embodiment,

[0029] The strategy search algorithm used to solve the production scheduling optimization model includes:

[0030] The power plant production scheduling strategy is decomposed into unit combination sub-strategy and load allocation sub-strategy, and the search space of each sub-strategy is divided according to unit characteristics and load characteristics to obtain multiple independent sub-spaces;

[0031] In each subspace of the unit combination sub-strategy, the optimal unit combination scheme is searched through the local search algorithm, and the optimal schemes of each subspace are compared, and the optimal solution is selected as the target strategy of the unit combination sub-strategy;

[0032] In each subspace of the load distribution sub-strategy, the optimal load distribution scheme is searched through the local search algorithm, and the optimal schemes of each subspace are compared, and the optimal solution is selected as the target strategy of the load distribution sub-strategy;

[0033] The main strategy is introduced to coordinate the target strategies of the unit combination sub-strategy and the load distribution sub-strategy. The quality of the current strategy is evaluated according to the target strategies of each sub-strategy, and the evaluation information is fed back to the main strategy. The main strategy adjusts the optimization direction and constraints according to the feedback information and generates new strategy coordination information.

[0034] Establish an information interaction mechanism between sub-strategies. The unit combination sub-strategy transmits the unit combination plan to the load allocation sub-strategy. The load allocation sub-strategy optimizes the load allocation based on the unit combination plan and feeds back the result to the unit combination sub-strategy.

[0035] The mathematical model of collaborative optimization sub-strategies is constructed by combining the strategy coordination information generated by the main strategy and the information interaction between sub-strategies. The distributed optimization algorithm is used to solve it and the optimal solution of the production scheduling optimization model is obtained.

[0036] In an optional embodiment,

[0037] The reinforcement learning algorithm is used to iteratively optimize the production scheduling strategy, and the target scheduling strategy includes:

[0038] According to the continuous control variables in the power plant production scheduling strategy, the state space and action space are constructed. The state space includes equipment load and parameter variables, and the action space is the continuous control instructions.

[0039] A policy generation network and a policy evaluation network are introduced. Sample data are collected by interacting with the environment and stored in an experience replay pool. A batch of samples are randomly drawn from the experience replay pool. The policy evaluation network is used to calculate the current estimated value and the target value. The parameters of the policy evaluation network are updated based on the temporal difference error to obtain an updated policy evaluation network.

[0040] The updated policy evaluation network is used to evaluate the actions generated by the policy generation network, and the policy gradient method is used to update the parameters of the policy generation network. The trained policy generation network is used as the target scheduling strategy.

[0041] In an optional embodiment,

[0042] The policy evaluation network is used to calculate the current estimated value and the target value, and the parameters of the policy evaluation network are updated based on the temporal difference error. The updated policy evaluation network includes:

[0043] The discount factor is introduced to calculate the target value, and the loss function of the policy evaluation network is constructed in combination with the current estimated value. The policy evaluation network parameters are updated by minimizing the loss function. The loss function calculation formula is as follows:

[0044]

[0045] Among them, L(θ Q ) represents the loss function of the strategy evaluation network, N represents the number of samples, r i represents the immediate reward, γ represents the discount factor, Q' represents the target strategy evaluation network, s i+1 represents the state at time i+1, μ' represents the target strategy generation network, θ μ′ represents the network parameters corresponding to the target strategy generation network, θ Q′ represents the network parameters corresponding to the target policy evaluation network, Q(·) represents the estimated value of the target policy evaluation network, and s i represents the state at time i, a i represents the action taken at time i, θ Q Represents the policy evaluation network parameters.

[0046] According to a second aspect of the embodiments of the present invention,

[0047] Provide a power plant production management system based on a domestic distributed control system, including:

[0048] The first unit is used to collect and preprocess the historical production data of power plant equipment through a pre-built domestic distributed control system, divide the pre-processed historical production data into different working mode modes and form a production scenario library. In the real-time production process, the latest production data is dynamically intercepted and its key features are extracted. The pre-trained working condition discrimination model is used to distinguish the current production working condition in real time to obtain the production working condition recognition result;

[0049] The second unit is used to build a corresponding production scheduling optimization model based on the production condition identification results and the pre-built power plant process knowledge base, generate an initial scheduling strategy by solving the production scheduling optimization model, and use a reinforcement learning algorithm to iteratively optimize the production scheduling strategy to obtain a target scheduling strategy;

[0050] The third unit is used to convert the target scheduling strategy into control instructions at the device level, which are sent from the industrial Internet platform to the controller of the domestic distributed control system through the industrial communication protocol. The controller automatically adjusts the operating status and parameters of the equipment according to the instructions, and displays the progress of the scheduling plan execution and the response of the equipment in real time through the human-computer interaction interface, and allows the dispatcher to manually intervene and correct the scheduling instructions through the interface.

[0051] According to a third aspect of the embodiments of the present invention,

[0052] An electronic device is provided, comprising:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] According to a fourth aspect of the embodiments of the present invention,

[0057] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0058] In this embodiment, by using the pre-collected and pre-processed historical production data, combined with the production scenario library and the working condition discrimination model, the current production conditions can be identified in real time, providing data support for production scheduling optimization, and improving the scientificity and accuracy of decision-making. Based on the production condition identification results and the power plant process knowledge base, a production scheduling optimization model is constructed, and the scheduling strategy is iteratively optimized using a reinforcement learning algorithm, which can generate the optimal scheduling plan to improve production efficiency and economy. The optimized scheduling strategy is converted into a control instruction at the equipment level, and is issued and executed through the industrial Internet platform and the distributed control system to achieve automatic adjustment of the equipment status and parameters, and improve the automation level of the production process. Through the human-computer interaction interface, the dispatcher can monitor the execution progress of the scheduling plan and the response of the equipment in real time, and can manually intervene and correct the scheduling instructions to achieve intelligent scheduling of human-machine collaboration. Through real-time production condition identification and intelligent scheduling optimization, the production strategy can be quickly adjusted according to changes in production conditions, and the flexibility and adaptability of the production process can be improved. Through the power plant production management method based on the domestic distributed control system, the safety and controllability advantages of the domestic system are fully utilized, and a variety of advanced information technologies are integrated to form a set of intelligent management solutions covering the entire life cycle of power plant production, improve power plant production efficiency, and reduce operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1It is a flow chart of a power plant production management method based on a domestic distributed control system according to an embodiment of the present invention;

[0060] Figure 2 It is a structural schematic diagram of a power plant production management system based on a domestic distributed control system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a power plant production management method based on a domestic distributed control system according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S101. The historical production data of power plant equipment is collected and preprocessed through a pre-built domestic distributed control system. The pre-processed historical production data is divided into different operating modes to form a production scenario library. In the real-time production process, the latest production data is dynamically intercepted and its key features are extracted. The pre-trained operating condition discrimination model is used to discriminate the current production condition in real time to obtain the production condition recognition result.

[0065] Among them, the domestic distributed control system architecture is composed of the field equipment layer, control layer and management layer from bottom to top. The field equipment layer is connected to intelligent sensors and actuators, and equipment access and management are achieved through domestic intelligent gateways. The control layer is mainly based on domestic PLCs, and communicates with the field equipment layer and management layer through industrial communication protocols. The management layer builds power plant production management applications based on the domestic industrial Internet platform.

[0066] In an optional embodiment,

[0067] The pre-trained working condition identification model is used to identify the current production working condition in real time, and the production working condition identification results include:

[0068] The power plant production system is abstracted into a heterogeneous graph network, where nodes represent equipment and edges represent logical connections and material flows between equipment.

[0069] For each device node, we collect multi-source heterogeneous data and static attribute data collected in real time, use attribute encoder to extract and encode the node attribute data, and obtain the low-dimensional attribute embedding vector of the device node.

[0070] For each edge connecting devices, the interaction information of power plant devices is collected, and the edge encoder is used to extract and encode the edge interaction data to obtain a low-dimensional embedding vector of the device interaction information;

[0071] Input the low-dimensional attribute embedding vector of the device node and the low-dimensional embedding vector of the device interaction information into the pre-trained graph neural network model to obtain the updated state embedding vector of the device node;

[0072] The updated state embedding vector of the equipment node is used as the input feature of the working condition discrimination model to generate the embedded representation of the production status graph. Cluster analysis and anomaly detection are performed on the embedded representation of the production status graph, combined with the prior knowledge of the scene and the pre-built anomaly discrimination rules, to obtain the final production condition recognition result.

[0073] For example, the power plant production system is first abstracted into a heterogeneous graph network. The production status graph is constructed with the equipment as nodes and the logical connections and material flows between the equipment as edges. Different node types are set for different types of equipment, such as boilers, steam turbines, generators, etc. Different edge types are set for different types of connections and interactions, such as process connections, electrical connections, thermal connections, etc. The topological structure and interaction mode of the power plant production system are finely portrayed through the modeling method of heterogeneous graphs.

[0074] For each device node, we collect multi-source heterogeneous data collected in real time, such as temperature, pressure, flow, vibration and other sensor data, as well as static attribute data such as material properties, geometric parameters, and health status of the equipment. We use attribute encoders, such as multi-layer perceptrons and convolutional neural networks, to extract and encode node attribute data and obtain low-dimensional attribute embedding vectors of device nodes.

[0075] For each edge connecting devices, data reflecting the interaction information such as material flow and energy transfer between devices, such as pipeline flow, heat load, etc., are collected. Edge encoders, such as attention mechanism and gating mechanism, are used to extract and encode the interaction data of the edge to obtain a low-dimensional embedding vector of the interaction between devices. At the same time, considering the directionality and heterogeneity of the edge, edge type embedding is introduced to distinguish the semantic information of different types of interactions.

[0076] Taking the attribute embedding of the device node and the edge embedding of the interaction between devices as input, the state representation of the device node is updated through the message passing and aggregation operation of the graph neural network. Using graph neural network variants such as graph attention network and graph convolution network, through the weighted aggregation of the state of neighbor nodes, combined with the node's own attributes and the interaction information of the edges, the updated state embedding vector of the device node is obtained.

[0077] With the dynamic evolution of the production process, the attributes of device nodes and the interaction intensity between devices are also changing dynamically. Through real-time data collection and online learning, the parameters of node attribute encoders and edge encoders are dynamically updated to adapt to changes in device status and interaction modes. At the same time, combined with contextual information such as external environmental factors and changes in production tasks, the hyperparameters and aggregation methods of the graph neural network are adjusted to maintain the adaptability and robustness of the state graph model.

[0078] The state embedding vector of the equipment node is used as the input feature of the scene recognition model, and the embedding representation of the entire production state graph is generated by using graph-level pooling operations, such as global average pooling, attention pooling, etc. Different production scenarios and abnormal patterns are identified through clustering analysis and anomaly detection of the embedded representation, and the final production condition recognition result is obtained.

[0079] In an optional embodiment,

[0080] Training a graph neural network model involves:

[0081] According to the multi-source heterogeneous data collected in real time, an adaptive learning rate is introduced to dynamically update the parameters of the node attribute encoder and edge encoder. The calculation formula is as follows:

[0082]

[0083] Among them, ω (t) represents the updated model parameters, ω (t-1) represents the model parameters before updating, t represents the current time, η (t) represents the adaptive learning rate at time t, m (t) represents the first-order momentum term at time t, v (t) represents the second-order momentum term at time t, ε represents the smoothing term of parameter update, λ represents the regularization coefficient, and R(·) represents the model parameter regularization term.

[0084] In this embodiment, it is possible to integrate multi-source heterogeneous data from power plant equipment, including real-time acquisition data, static attribute data, and equipment interaction information, etc., and perform feature extraction and encoding through attribute encoders and edge encoders to fully tap the value of data. By training the graph neural network model, the device attribute embedding vector and the device interaction information embedding vector can be automatically learned, and the embedded representation of the production status graph can be generated end-to-end without complex manual feature engineering. Based on the embedded representation of the production status graph, combined with prior knowledge and abnormality discrimination rules, accurate identification of production conditions and timely detection of abnormal conditions can be achieved, providing support for scheduling decisions. By integrating advanced technologies such as heterogeneous data fusion, graph structure modeling, and deep representation learning, the intelligence level and accuracy of power plant production condition identification can be effectively improved, providing solid data support for production scheduling optimization. At the same time, it has certain interpretability and visualization capabilities, which is helpful for optimization analysis and decision-making assistance of production processes.

[0085] S102. Construct a corresponding production scheduling optimization model based on the production condition identification results and the pre-built power plant process knowledge base, generate an initial scheduling strategy by solving the production scheduling optimization model, and use a reinforcement learning algorithm to iteratively optimize the production scheduling strategy to obtain a target scheduling strategy.

[0086] In an optional embodiment,

[0087] Based on the production condition identification results and the pre-built power plant process knowledge base, a corresponding production scheduling optimization model is built. By solving the production scheduling optimization model, the initial scheduling strategy is generated, including:

[0088] Building a power plant process knowledge base based on historical data of power plant production and operation, wherein the power plant process knowledge base includes empirical rules and optimization strategies for power plant production and operation, and building a production scheduling optimization model based on the power plant process knowledge base;

[0089] Using the empirical rules and optimization strategies in the power plant process knowledge base, the strategy search algorithm is used to solve the production scheduling optimization model, the first scheduling plan is generated as the search starting point, and the neighborhood structure and move operation are defined to generate candidate scheduling plans. According to the predetermined acceptance criteria, it is decided whether to accept the candidate scheduling plan. The strategy search algorithm is repeatedly executed until the preset termination condition is reached, and the target scheduling plan is finally output;

[0090] The target scheduling plan is adjusted using sensitivity analysis and parameter optimization methods, and the adjusted scheduling plan is re-substituted into the production scheduling optimization model to generate the final initial scheduling strategy for power plant production scheduling.

[0091] For example, we first collect historical data on power plant production and operation, including unit parameters, load curves, fuel costs, etc., and organize the empirical rules and optimization strategies for power plant production scheduling to form a structured knowledge representation. We then use knowledge representation tools such as ontology and semantic web to establish a power plant process knowledge base.

[0092] Define decision variables, including unit start-stop state variables, unit output variables, unit start-up time variables and unit shutdown time variables; define the optimization objective function, that is, minimize the production cost, including fuel cost and start-stop cost; define constraints, including load balance constraints, spare capacity constraints, unit ramp rate constraints, minimum start-stop time constraints and logical constraints, etc., so as to establish a production scheduling optimization model.

[0093] Using the empirical rules and optimization strategies in the power plant process knowledge base, a feasible initial scheduling plan is generated as the search starting point, and the neighborhood structure and mobile operations are defined, such as changing the start and stop status of the unit, adjusting the unit output, etc., to generate new candidate scheduling plans, and an evaluation function is designed to comprehensively consider the objective function value and constraint violation degree of the candidate scheduling plan to evaluate the quality of the candidate solution. By selecting appropriate acceptance criteria, such as better solution acceptance, simulated annealing acceptance, etc., it is decided whether to accept the candidate solution as the new current solution, and the termination conditions, such as the maximum number of iterations, convergence of the objective function value, etc., are set to determine whether to end the search process, and the optimal scheduling plan obtained by the search is output as the preliminary scheduling strategy.

[0094] By analyzing the advantages and disadvantages of the preliminary scheduling strategy and identifying the room for improvement, the key parameters in the preliminary scheduling strategy are adjusted using sensitivity analysis, parameter optimization and other methods. The adjusted scheduling strategy is re-substituted into the mixed integer programming model for solution, the optimization effect is evaluated, and based on the optimization results, the optimal scheduling strategy is selected as the improved preliminary scheduling strategy.

[0095] Convert the improved preliminary dispatching strategy into executable dispatching instructions and operation plans, deploy the dispatching strategy in the power plant production management system, guide the actual production operation, monitor the execution of the dispatching strategy, collect actual operation data, evaluate the actual effect of the dispatching strategy, and dynamically adjust and optimize the dispatching strategy based on actual operation feedback to achieve closed-loop optimization control.

[0096] The improved preliminary scheduling strategy is used as the starting point of the deep reinforcement learning algorithm to accelerate the convergence of strategy training, and deep reinforcement learning models such as Actor-strategy evaluation model and DQN model are constructed to further optimize the scheduling strategy. The deep reinforcement learning model is trained through simulation and actual production data to adapt to the dynamic changes of power plant production and operation. The trained and optimized deep reinforcement learning model is applied to the real-time scheduling decision-making of the power plant to achieve adaptive optimization of the scheduling strategy.

[0097] In this embodiment, by constructing the empirical rules and optimization strategies of the long-term production and operation of the power plant as a process knowledge base, and integrating these domain knowledge into the production scheduling optimization model, the practicality and interpretability of the model are improved. By defining the neighborhood structure and moving operations to generate candidate solutions, and combining the acceptance criteria for heuristic search, the solution space can be explored more efficiently and the amount of calculation of the enumeration search can be reduced. By taking the first feasible solution as the starting point through the strategy search algorithm, and gradually improving the scheduling scheme until the preset termination condition is reached, the global optimal or suboptimal production scheduling strategy can be obtained to avoid falling into the local optimum. Through sensitivity analysis and parameter optimization, the degree of influence of each decision variable on the objective function can be evaluated, and the initial scheduling strategy can be further adjusted and optimized. By organically combining artificial experience knowledge with the optimization algorithm, not only can the global optimal or suboptimal initial scheduling strategy be generated, but also it has strong interpretability and credibility, which helps to improve the intelligence level and decision-making quality of power plant production scheduling.

[0098] In an optional embodiment,

[0099] The strategy search algorithm used to solve the production scheduling optimization model includes:

[0100] The power plant production scheduling strategy is decomposed into unit combination sub-strategy and load allocation sub-strategy, and the search space of each sub-strategy is divided according to unit characteristics and load characteristics to obtain multiple independent sub-spaces;

[0101] In each subspace of the unit combination sub-strategy, the optimal unit combination scheme is searched through the local search algorithm, and the optimal schemes of each subspace are compared, and the optimal solution is selected as the target strategy of the unit combination sub-strategy;

[0102] In each subspace of the load distribution sub-strategy, the optimal load distribution scheme is searched through the local search algorithm, and the optimal schemes of each subspace are compared, and the optimal solution is selected as the target strategy of the load distribution sub-strategy;

[0103] The main strategy is introduced to coordinate the target strategies of the unit combination sub-strategy and the load distribution sub-strategy. The quality of the current strategy is evaluated according to the target strategies of each sub-strategy, and the evaluation information is fed back to the main strategy. The main strategy adjusts the optimization direction and constraints according to the feedback information and generates new strategy coordination information.

[0104] Establish an information interaction mechanism between sub-strategies. The unit combination sub-strategy transmits the unit combination plan to the load allocation sub-strategy. The load allocation sub-strategy optimizes the load allocation based on the unit combination plan and feeds back the result to the unit combination sub-strategy.

[0105] The mathematical model of collaborative optimization sub-strategies is constructed by combining the strategy coordination information generated by the main strategy and the information interaction between sub-strategies. The distributed optimization algorithm is used to solve it and the optimal solution of the production scheduling optimization model is obtained.

[0106] For example, according to the characteristics of the power plant production scheduling strategy, it is decomposed into two parts: a unit combination sub-strategy and a load distribution sub-strategy.

[0107] The unit combination sub-strategy mainly determines the start-stop combination mode of each type of unit in the power plant. It generates a feasible unit combination plan by considering the type, capacity, parameters and other characteristics of the unit, as well as the start-stop constraints and minimum operating time constraints between units.

[0108] Given a unit combination plan, the load allocation sub-strategy mainly determines the output allocation method of each enabled unit. The optimal output of each unit is determined by considering the upper and lower limits of the unit's output, the ramp rate constraint, the load balance constraint of the power grid, and the spare capacity constraint.

[0109] For the unit combination sub-strategy, the search space is divided into multiple sub-spaces according to the unit type, capacity and other characteristics. For the load allocation sub-strategy, the search space is divided into multiple sub-spaces according to the load characteristics, unit parameters and so on.

[0110] Then, a suitable neighborhood structure is defined in each subspace, such as the neighborhood of the start-stop change of a single unit, the neighborhood of exchanging the start-stop status of two units, etc. Local search is performed in the divided subspace, and the current solution is continuously improved through state transfer in the neighborhood until the local optimum is reached.

[0111] The main strategy is introduced to coordinate the solutions of the sub-strategies. The main strategy adjusts the optimization direction and constraints according to the feedback information of the sub-strategies, and passes the updated information to the sub-strategies. The goal of the main strategy is to minimize the total cost of power plant production scheduling, including unit start-up and shutdown costs and power generation costs, and then adjust the optimization direction and constraints according to the feedback information of the unit combination sub-strategy and load allocation sub-strategy.

[0112] Specifically, if the target value of the unit combination sub-strategy is large, it means that the start-up and shutdown cost of the current unit combination plan is high, and the unit combination plan needs to be adjusted. The main strategy can guide the unit combination sub-strategy to search for a plan with lower start-up and shutdown costs by adding penalty items or constraints.

[0113] If the target value of the load allocation sub-strategy is large, it means that the power generation cost of the current load allocation scheme is high and the load allocation scheme needs to be adjusted. The main strategy can guide the load allocation sub-strategy to search for a scheme with lower power generation cost by adding penalty items or constraints.

[0114] If the solutions of the unit combination sub-strategy and the load distribution sub-strategy are near the boundary of the feasible region, it means that the current solution may not meet certain constraints and the constraints need to be relaxed or tightened. The main strategy can relax or tighten the feasible region of the sub-strategy by adjusting the right-hand side of the constraints.

[0115] The unit combination sub-strategy passes the target unit combination plan to the load allocation sub-strategy as a prerequisite for load allocation. Given the unit combination plan, the load allocation sub-strategy solves the optimal output of each unit and feeds the optimal output plan back to the unit combination sub-strategy to evaluate the pros and cons of the current unit combination plan.

[0116] Combining the coordination of the main strategy and the information interaction between sub-strategies, a collaborative optimization sub-strategy can be obtained. Under the guidance of the main strategy, the collaborative optimization sub-strategy continuously optimizes the unit combination plan and load distribution plan through iterative interaction between sub-strategies, and finally obtains the global optimal solution of the power plant production scheduling strategy.

[0117] Among them, the collaborative optimization sub-strategy can be solved by distributed optimization algorithms, such as the alternating direction method of multipliers (ADMM) and the sub-gradient method.

[0118] In this embodiment, the complex production scheduling optimization strategy is decomposed into unit combination sub-strategy and load distribution sub-strategy, and the search space of each sub-strategy is divided according to the unit and load characteristics, which reduces the complexity of the problem and improves the efficiency of solution. In each subspace, a local search algorithm is used to search for the optimal sub-strategy in parallel, and then coordinated through the main strategy, so as to realize the parallelization of the search process and the collaborative optimization between sub-strategies. The mathematical model of the collaborative optimization sub-strategy is solved by a distributed optimization algorithm, which has good scalability, can adapt to the growth of the problem scale, and supports solving in a distributed computing environment. By dividing the search space and adjusting the optimization direction of the main strategy, the search space can be effectively reduced, the algorithm can be prevented from falling into the local optimum, and the robustness of the solution can be improved. By efficiently solving large-scale and complex production scheduling optimization problems, generating high-quality scheduling plans, and improving the production efficiency and economic benefits of power plants, it has important theoretical significance and application value.

[0119] In an optional embodiment,

[0120] The reinforcement learning algorithm is used to iteratively optimize the production scheduling strategy, and the target scheduling strategy includes:

[0121] According to the continuous control variables in the power plant production scheduling strategy, the state space and action space are constructed. The state space includes equipment load and parameter variables, and the action space is the continuous control instructions.

[0122] A policy generation network and a policy evaluation network are introduced. Sample data are collected by interacting with the environment and stored in an experience replay pool. A batch of samples are randomly drawn from the experience replay pool. The policy evaluation network is used to calculate the current estimated value and the target value. The parameters of the policy evaluation network are updated based on the temporal difference error to obtain an updated policy evaluation network.

[0123] The updated policy evaluation network is used to evaluate the actions generated by the policy generation network, and the policy gradient method is used to update the parameters of the policy generation network. The trained policy generation network is used as the target scheduling strategy.

[0124] Exemplarily, the production condition identification results, equipment parameters, scheduling constraints and other information are first encoded into a state vector. Optionally, a strategy generation network is used to generate continuous scheduling actions, such as equipment start and stop, load adjustment, parameter optimization, etc. The strategy generation network usually adopts a multi-layer feedforward neural network or a recurrent neural network structure, and outputs a deterministic strategy based on the current state vector, that is, the optimal action to be taken under a given state.

[0125] According to the production scheduling optimization goal, a reward function is designed to evaluate the pros and cons of scheduling actions. The reward function needs to comprehensively consider multiple optimization goals such as production efficiency, energy consumption cost, equipment health, etc., reasonably set the reward value and penalty value, and learn the optimal scheduling strategy.

[0126] The policy evaluation network is used to estimate the Q value of the state-action pair and judge the long-term benefits of the scheduling strategy. The policy evaluation network usually adopts a multi-layer feedforward neural network structure. According to the current state and action, it outputs the corresponding Q value estimate. The training goal of the policy evaluation network is to minimize the TD error, that is, the difference between the current Q value estimate and the target Q value.

[0127] Sample data is collected by interacting with the environment and stored in the experience replay pool. The sample data includes information such as the current state, the action taken, the immediate reward obtained, the next state, etc. A batch of samples are randomly drawn from the experience replay pool, and the current Q value and target Q value are calculated using the policy evaluation network, and the parameters of the policy evaluation network are updated based on the temporal difference error.

[0128] The updated policy evaluation network is used to evaluate the actions generated by the policy generation network, and the policy gradient method is used to update the parameters of the policy generation network to optimize it towards a better policy.

[0129] Repeat the above process of sample collection and network update until the strategy generation network converges to the optimal strategy or reaches the preset number of training rounds.

[0130] The trained strategy generation network is used as the optimized production scheduling strategy. According to the current state of the power plant, the strategy generation network is used to generate continuous control instructions, optimize equipment load and parameter settings, and improve production efficiency and safety.

[0131] In an optional embodiment,

[0132] The policy evaluation network is used to calculate the current estimated value and the target value, and the parameters of the policy evaluation network are updated based on the temporal difference error. The updated policy evaluation network includes:

[0133] The discount factor is introduced to calculate the target value, and the loss function of the policy evaluation network is constructed in combination with the current estimated value. The policy evaluation network parameters are updated by minimizing the loss function. The loss function calculation formula is as follows:

[0134]

[0135] Among them, L(θ Q ) represents the loss function of the strategy evaluation network, N represents the number of samples, r i represents the immediate reward, γ represents the discount factor, Q' represents the target strategy evaluation network, s i+1 represents the state at time i+1, μ' represents the target strategy generation network, θ μ′ represents the network parameters corresponding to the target strategy generation network, θ Q′ represents the network parameters corresponding to the target policy evaluation network, Q(·) represents the estimated value of the target policy evaluation network, and s i represents the state at time i, a i represents the action taken at time i, θ Q Represents the policy evaluation network parameters.

[0136] In this embodiment, by constructing a state space and an action space and interacting with a simulation environment, the reinforcement learning algorithm can autonomously explore and learn to obtain the optimal production scheduling strategy without manually designing complex rules or optimization models. By adopting a deep reinforcement learning framework, it is possible to process continuous state spaces and action spaces, meeting the needs of continuous control variables such as equipment loads and parameters in power plant production scheduling. The introduction of the experience replay pool mechanism can efficiently utilize historical interaction data, avoid a large amount of repeated exploration, and improve the efficiency of sample use and the convergence speed of strategy learning. The separate design of the strategy generation network and the strategy evaluation network is conducive to improving the stability of strategy learning, avoiding the problem of training instability caused by excessive strategy variance, and realizing autonomous learning to obtain the optimal scheduling strategy, while being able to handle complex continuous control problems.

[0137] S103. Convert the target scheduling strategy into control instructions at the device level, and send them from the industrial Internet platform to the controller of the domestic distributed control system through the industrial communication protocol. The controller automatically adjusts the operating status and parameters of the equipment according to the instructions, and displays the progress of the scheduling plan execution and the response of the equipment in real time through the human-computer interaction interface, allowing the dispatcher to manually intervene and correct the scheduling instructions through the interface.

[0138] The target scheduling strategy includes the output curve of each unit, the equipment combination commissioning plan, etc. The scheduling execution module converts these high-level decisions into control instructions at the equipment level, such as AGC load instructions, DCS start and stop instructions, etc. The instructions are sent from the industrial Internet platform to the controller of the distributed control system through the industrial communication protocol, and the controller automatically adjusts the operating status and parameters of the equipment according to the instructions. At the same time, the scheduling execution module displays the progress of the scheduling execution and the response of the equipment to the dispatcher in real time through the human-computer interaction interface, which facilitates the dispatcher to supervise the execution process of the instructions. According to the actual situation, the dispatcher can manually intervene and correct the scheduling instructions through the human-computer interaction interface to ensure the safety and controllability of the scheduling execution.

[0139] The human-computer interaction module uses visualization technology to provide dispatchers with a friendly and intelligent human-computer interaction interface, which is convenient for dispatchers to monitor the dispatch process in real time and assist in dispatch decision-making. The human-computer interaction interface intuitively displays dispatch information such as unit output, equipment operating conditions, and energy consumption indicators through graphical methods such as curve charts, bar charts, and pie charts. Dispatchers can view the progress of dispatch tasks, optimize the effect evaluation, etc. in real time through the interface to grasp the overall production status.

[0140] In this embodiment, the optimized target scheduling strategy is converted into specific equipment control instructions, which are automatically issued and executed through the industrial Internet platform and distributed control system, realizing a seamless link between scheduling decisions and control execution, and improving the automation level of the production process. Through the human-computer interaction interface, the dispatcher can monitor the execution progress and response of the scheduling plan at the equipment level in real time, fully grasp the dynamics of production operations, and provide a basis for decision-making adjustments. By closely combining advanced optimization decision-making technology with automated control execution, human-machine collaboration, and system integration technologies, efficient connection between intelligent scheduling and automated control is achieved, which improves the intelligence and transparency of the production process, and helps to comprehensively improve the production efficiency and economic benefits of power plants.

[0141] Figure 2 FIG. 1 is a schematic diagram of the structure of a power plant production management system based on a domestic distributed control system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0142] The first unit is used to collect and preprocess the historical production data of power plant equipment through a pre-built domestic distributed control system, divide the pre-processed historical production data into different working mode modes and form a production scenario library. In the real-time production process, the latest production data is dynamically intercepted and its key features are extracted. The pre-trained working condition discrimination model is used to distinguish the current production working condition in real time to obtain the production working condition recognition result;

[0143] The second unit is used to build a corresponding production scheduling optimization model based on the production condition identification results and the pre-built power plant process knowledge base, generate an initial scheduling strategy by solving the production scheduling optimization model, and use a reinforcement learning algorithm to iteratively optimize the production scheduling strategy to obtain a target scheduling strategy;

[0144] The third unit is used to convert the target scheduling strategy into control instructions at the device level, which are sent from the industrial Internet platform to the controller of the domestic distributed control system through the industrial communication protocol. The controller automatically adjusts the operating status and parameters of the equipment according to the instructions, and displays the progress of the scheduling plan execution and the response of the equipment in real time through the human-computer interaction interface, and allows the dispatcher to manually intervene and correct the scheduling instructions through the interface.

[0145] According to a third aspect of the embodiments of the present invention,

[0146] An electronic device is provided, comprising:

[0147] processor;

[0148] a memory for storing processor-executable instructions;

[0149] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0150] According to a fourth aspect of the embodiments of the present invention,

[0151] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0152] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power plant production management method based on a domestic distributed control system, characterized in that: include: The historical production data of power plant equipment is collected and preprocessed through a pre-built domestic distributed control system. The pre-processed historical production data is divided into different operating modes to form a production scenario library. In the real-time production process, the latest production data is dynamically intercepted and its key features are extracted. The pre-trained operating condition discrimination model is used to discriminate the current production condition in real time to obtain the production condition recognition result. Based on the production condition identification results and the pre-built power plant process knowledge base, a corresponding production scheduling optimization model is built. By solving the production scheduling optimization model, an initial scheduling strategy is generated, and a reinforcement learning algorithm is used to iteratively optimize the production scheduling strategy to obtain the target scheduling strategy. The target scheduling strategy is converted into control instructions at the equipment level, which are sent from the industrial Internet platform to the controller of the domestic distributed control system through the industrial communication protocol. The controller automatically adjusts the operating status and parameters of the equipment according to the instructions, and displays the progress of the scheduling plan execution and the response of the equipment in real time through the human-computer interaction interface, allowing the dispatcher to manually intervene and correct the scheduling instructions through the interface; Based on the production condition identification results and the pre-built power plant process knowledge base, a corresponding production scheduling optimization model is built. By solving the production scheduling optimization model, the initial scheduling strategy is generated, including: Building a power plant process knowledge base based on historical data of power plant production and operation, wherein the power plant process knowledge base includes empirical rules and optimization strategies for power plant production and operation, and building a production scheduling optimization model based on the power plant process knowledge base; Using the empirical rules and optimization strategies in the power plant process knowledge base, the strategy search algorithm is used to solve the production scheduling optimization model, the first scheduling plan is generated as the search starting point, and the neighborhood structure and move operation are defined to generate candidate scheduling plans. According to the predetermined acceptance criteria, it is decided whether to accept the candidate scheduling plan. The strategy search algorithm is repeatedly executed until the preset termination condition is reached, and the target scheduling plan is finally output; The target scheduling plan is adjusted using sensitivity analysis and parameter optimization methods, and the adjusted scheduling plan is re-substituted into the production scheduling optimization model to generate the final initial scheduling strategy for power plant production scheduling; The strategy search algorithm used to solve the production scheduling optimization model includes: The power plant production scheduling strategy is decomposed into unit combination sub-strategy and load allocation sub-strategy, and the search space of each sub-strategy is divided according to unit characteristics and load characteristics to obtain multiple independent sub-spaces; In each subspace of the unit combination sub-strategy, the optimal unit combination scheme is searched through the local search algorithm, and the optimal schemes of each subspace are compared, and the optimal solution is selected as the target strategy of the unit combination sub-strategy; In each subspace of the load distribution sub-strategy, the optimal load distribution scheme is searched through the local search algorithm, and the optimal schemes of each subspace are compared, and the optimal solution is selected as the target strategy of the load distribution sub-strategy; The main strategy is introduced to coordinate the target strategies of the unit combination sub-strategy and the load distribution sub-strategy. The quality of the current strategy is evaluated according to the target strategies of each sub-strategy, and the evaluation information is fed back to the main strategy. The main strategy adjusts the optimization direction and constraints according to the feedback information and generates new strategy coordination information. Establish an information interaction mechanism between sub-strategies. The unit combination sub-strategy transmits the unit combination plan to the load allocation sub-strategy. The load allocation sub-strategy optimizes the load allocation based on the unit combination plan and feeds back the result to the unit combination sub-strategy. The mathematical model of collaborative optimization sub-strategies is constructed by combining the strategy coordination information generated by the main strategy and the information interaction between sub-strategies. The distributed optimization algorithm is used to solve it and the optimal solution of the production scheduling optimization model is obtained.

2. The method according to claim 1, characterized in that: The pre-trained working condition identification model is used to identify the current production working condition in real time, and the production working condition identification results include: The power plant production system is abstracted into a heterogeneous graph network, where nodes represent equipment and edges represent logical connections and material flows between equipment. For each device node, we collect multi-source heterogeneous data and static attribute data collected in real time, use attribute encoder to extract and encode the node attribute data, and obtain the low-dimensional attribute embedding vector of the device node. For each edge connecting devices, the interaction information of power plant devices is collected, and the edge encoder is used to extract and encode the edge interaction data to obtain a low-dimensional embedding vector of the device interaction information; Input the low-dimensional attribute embedding vector of the device node and the low-dimensional embedding vector of the device interaction information into the pre-trained graph neural network model to obtain the updated state embedding vector of the device node; The updated state embedding vector of the equipment node is used as the input feature of the working condition discrimination model to generate the embedded representation of the production status graph. Cluster analysis and anomaly detection are performed on the embedded representation of the production status graph, combined with the prior knowledge of the scene and the pre-built anomaly discrimination rules, to obtain the final production condition recognition result.

3. The method according to claim 2, characterized in that Training a graph neural network model involves: According to the multi-source heterogeneous data collected in real time, an adaptive learning rate is introduced to dynamically update the parameters of the node attribute encoder and edge encoder. The calculation formula is as follows: ； Among them, ω (t) represents the updated model parameters, ω (t-1) represents the model parameters before updating, t represents the current time, η (t) represents the adaptive learning rate at time t, m (t) represents the first-order momentum term at time t, v (t) represents the second-order momentum term at time t, ε represents the smoothing term of parameter update, λ represents the regularization coefficient, and R(·) represents the model parameter regularization term.

4. The method according to claim 1, characterized in that The reinforcement learning algorithm is used to iteratively optimize the production scheduling strategy, and the target scheduling strategy includes: According to the continuous control variables in the power plant production scheduling strategy, the state space and action space are constructed. The state space includes equipment load and parameter variables, and the action space is the continuous control instructions. A policy generation network and a policy evaluation network are introduced. Sample data are collected by interacting with the environment and stored in an experience replay pool. A batch of samples are randomly drawn from the experience replay pool. The policy evaluation network is used to calculate the current estimated value and the target value. The parameters of the policy evaluation network are updated based on the temporal difference error to obtain an updated policy evaluation network. The updated policy evaluation network is used to evaluate the actions generated by the policy generation network, and the policy gradient method is used to update the parameters of the policy generation network. The trained policy generation network is used as the target scheduling strategy.

5. The method according to claim 4, characterized in that The policy evaluation network is used to calculate the current estimated value and the target value, and the parameters of the policy evaluation network are updated based on the temporal difference error. The updated policy evaluation network includes: The discount factor is introduced to calculate the target value, and the loss function of the policy evaluation network is constructed in combination with the current estimated value. The policy evaluation network parameters are updated by minimizing the loss function. The loss function calculation formula is as follows: ； in, L(θ Q ) represents the loss function of the strategy evaluation network, N represents the number of samples, r i represents the immediate reward, γ represents the discount factor, Q' represents the target strategy evaluation network, s i+1 represents the state at time i+1, μ' represents the target strategy generation network, Indicates the network parameters corresponding to the target strategy generation network, represents the network parameters corresponding to the target policy evaluation network, Q(·) represents the estimated value of the target policy evaluation network, and s i represents the state at time i, a i It represents the action taken at time i. θ Q Represents the policy evaluation network parameters.

6. A power plant production management system based on a domestic distributed control system, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to collect and preprocess the historical production data of power plant equipment through a pre-built domestic distributed control system, divide the pre-processed historical production data into different working mode modes and form a production scenario library. In the real-time production process, the latest production data is dynamically intercepted and its key features are extracted. The pre-trained working condition discrimination model is used to distinguish the current production working condition in real time to obtain the production working condition recognition result; The second unit is used to build a corresponding production scheduling optimization model based on the production condition identification results and the pre-built power plant process knowledge base, generate an initial scheduling strategy by solving the production scheduling optimization model, and use a reinforcement learning algorithm to iteratively optimize the production scheduling strategy to obtain a target scheduling strategy; The third unit is used to convert the target scheduling strategy into control instructions at the device level, which are sent from the industrial Internet platform to the controller of the domestic distributed control system through the industrial communication protocol. The controller automatically adjusts the operating status and parameters of the equipment according to the instructions, and displays the progress of the scheduling plan execution and the response of the equipment in real time through the human-computer interaction interface, and allows the dispatcher to manually intervene and correct the scheduling instructions through the interface.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images