Automatic control method and system for tripping of low-load auxiliary fan under deep peak regulation working condition

Abstract

The invention relates to the technical field of power system control, and discloses an automatic control method and system for tripping of a low-load auxiliary fan under a deep peak regulation working condition, and the method comprises the steps: constructing a fuzzy cognitive map of operation influence factors of the auxiliary fan, and predicting the tripping risk of the auxiliary fan; establishing a multi-dimensional fairness index system including instant load fairness, accumulative operation time fairness and life consumption rate fairness; determining a load distribution scheme through a multi-agent negotiation game; guiding the system to evolve towards the Pareto optimal solution direction through an adaptive evolution strategy; introducing a reputation mechanism to evaluate the historical performance of each subject, and continuously optimizing a load distribution strategy through historical decision effect feedback; according to the method, the dynamic balance between the fairness of auxiliary fan load distribution and the system stability is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power system control, and more specifically, to an automatic control method and system for low-load auxiliary fan tripping under deep peak shaving conditions. Background Art

[0002] With the continuous increase in the proportion of renewable energy, the power system has put forward higher requirements for the peak shaving ability of thermal power units. Under deep peak shaving conditions, the auxiliary fans (hereinafter referred to as auxiliary fans) of thermal power units need to operate under low-load conditions to meet the peak shaving requirements of the power grid. However, due to significant differences in the equipment types, aging degrees, and maintenance states of each auxiliary fan, the traditional load average distribution strategy cannot adapt to the heterogeneous characteristics between devices.

[0003] Currently, the mainstream control methods for auxiliary fans mainly include the average load distribution method, the priority control method, and the simple state perception control method. The average load distribution method ignores the differences between devices, resulting in new devices bearing too high a load for a long time and old devices being overprotected, affecting the long-term reliability and economy of the overall system; although the priority control method takes into account device differences, the fixed priority strategy is difficult to adapt to the dynamically changing operating environment; although the state perception control method can adjust the load distribution according to the current state, due to the lack of the ability to model complex causal relationships, it cannot comprehensively consider the problem of uneven load distribution caused by multi-factor interaction.

[0004] The existing control systems also have the following technical problems: lack of the ability to model complex causal relationships between auxiliary fans and unable to predict the tripping risk under low-load conditions; lack of a dynamic equilibrium mechanism that comprehensively considers short-term and long-term fairness; lack of global information when each auxiliary fan makes independent decisions, and it is easy to generate sub-optimal or even incorrect load distribution decisions. Therefore, an intelligent control method for auxiliary fans that can solve the above problems is needed. Summary of the Invention

[0005] The present invention provides an automatic control method and system for low-load auxiliary fan tripping under deep peak shaving conditions, which solves the technical problems of uneven load distribution of auxiliary fans, inaccurate prediction of tripping risk, and lack of global information in decision-making in related technologies.

[0006] The present invention provides an automatic control method for low-load auxiliary fan tripping under deep peak shaving conditions, including: Constructing a fuzzy cognitive map of the influencing factors of auxiliary fan operation to predict the tripping risk of the auxiliary fan; Based on the tripping risk prediction results, establishing a multi-dimensional fairness index system including instant load fairness, cumulative operation time fairness, and life consumption rate fairness to provide fairness constraints for auxiliary fan load distribution; According to the multi-dimensional fairness index system, each auxiliary air blower is constructed as an intelligent agent with autonomous decision-making ability, and the load distribution plan is determined through multi-agent negotiation and game; Based on the results of the negotiation game, a temporary alliance of equipment with similar performance within the auxiliary air blower group is formed, and the system is guided to evolve towards the Pareto optimal solution through an adaptive evolutionary strategy; For the alliance evolution process, a reputation mechanism is introduced to evaluate the historical performance of each agent, and the load distribution strategy is continuously optimized through the feedback of historical decision-making effects.

[0007] In a preferred embodiment, the steps of constructing the fuzzy cognitive map of the influencing factors of the auxiliary air blower operation include: Collect data on the influencing factors of the auxiliary air blower operation, including peak shaving depth, inlet air temperature, vibration parameters, load level, operation duration, and equipment aging degree; Construct a set of nodes for the fuzzy cognitive map, and use each influencing factor as a map node; Based on historical operation data and expert knowledge, determine the fuzzy weights between nodes and establish a weight matrix; Predict potential tripping risks by iteratively calculating the activation values of each node.

[0008] In a preferred embodiment, the update of the activation value of the node in the fuzzy cognitive map uses the hyperbolic tangent function as the activation function, and a tripping risk threshold is set to output a tripping risk warning signal.

[0009] In a preferred embodiment, the multi-dimensional fairness index system calculates the instant load fairness index through the Gini coefficient, calculates the cumulative operation time fairness index through the cumulative operation time difference, and calculates the life consumption rate fairness index through the life consumption rate difference.

[0010] In a preferred embodiment, the multi-dimensional fairness index system also calculates the marginal contribution of each auxiliary air blower to the system stability based on the Shapley value, and introduces a time discount mechanism to balance the short-term and long-term fairness requirements, and constructs a comprehensive fairness index.

[0011] In a preferred embodiment, the steps of determining the load distribution plan through multi-agent negotiation and game include: Establish a multi-agent system model for the auxiliary air blower, and each auxiliary air blower is used as an intelligent agent with autonomous decision-making ability; Construct an inter-agent communication protocol to achieve efficient sharing of the operation information of the auxiliary air blower; Construct a load distribution strategy based on evolutionary game theory; Use the Nash bargaining solution in non-cooperative games to handle the load distribution conflicts of the auxiliary air blower group; Design a multi-agent decision fusion algorithm to calculate the final load distribution plan based on the negotiation results and trust levels of each auxiliary fan.

[0012] In a preferred embodiment, the inter-agent communication protocol includes status information messages, decision intention messages, negotiation request messages, and negotiation response messages. The communication timing includes a status broadcast phase, an intention declaration phase, a conflict detection phase, a negotiation resolution phase, and an execution confirmation phase.

[0013] In a preferred embodiment, the steps for forming a temporary coalition of devices with similar performance within the auxiliary fan group include: Construct an auxiliary fan similarity evaluation model to calculate the similarity between devices; Construct an auxiliary fan coalition partition algorithm based on the similarity matrix, using the spectral clustering method; Construct a collaborative decision-making model within the coalition; Construct an adaptive evolution strategy and construct a load distribution game model between coalitions; Adopt a Pareto solution set management strategy based on the dominance relationship to achieve convergence to the Pareto optimal solution.

[0014] In a preferred embodiment, the steps for the reputation mechanism to evaluate the historical performance of each agent include: Construct an auxiliary fan agent reputation evaluation model; Construct a decision weight adjustment algorithm with reputation awareness; Construct an adaptive optimization model for the load distribution strategy based on reinforcement learning; Construct a multi-agent experience replay pool mechanism; Construct a historical decision effect feedback evaluation model to optimize the strategy parameters through historical data analysis.

[0015] In a preferred embodiment, a low-load auxiliary fan tripping automatic control system for deep peak shaving conditions is used to execute a low-load auxiliary fan tripping automatic control method for deep peak shaving conditions, including: A fuzzy cognitive map modeling module for constructing a fuzzy cognitive map of the influencing factors of auxiliary fan operation and predicting the tripping risk of the auxiliary fan; A multi-dimensional fairness index construction module for establishing a multi-dimensional fairness index system including immediate load fairness, cumulative operation time fairness, and life consumption rate fairness; A multi-agent negotiation game module for constructing each auxiliary fan as an intelligent agent with autonomous decision-making ability and determining the load distribution plan through multi-agent negotiation game; A dynamic coalition formation module for realizing the formation of a temporary coalition of devices with similar performance within the auxiliary fan group and guiding the system to evolve towards the Pareto optimal solution through an adaptive evolution strategy; A reputation mechanism and feedback optimization module, which is used to evaluate the historical performance of each entity and continuously optimize the load distribution strategy through the feedback of historical decision-making effects.

[0016] The beneficial effects of the present invention are as follows: Through the five innovative steps of combining fuzzy cognitive map modeling, multi-dimensional fairness index construction, multi-agent negotiation game mechanism, dynamic coalition formation and adaptive evolution, and reputation mechanism and feedback optimization, the present invention realizes the intelligent control of the auxiliary fan group under the deep peak shaving condition, solves the technical problems that the traditional control method cannot adapt to the heterogeneous characteristics of equipment, is difficult to predict the tripping risk, and lacks global decision-making, and achieves technical effects such as improving the overall stability of the system, reducing the tripping risk, enhancing the peak shaving ability, extending the equipment life, and enhancing the adaptability. Description of the Drawings

[0017] Figure 1 It is a flowchart of an automatic control method for tripping of low-load auxiliary fans under the deep peak shaving condition of the present invention. Detailed Embodiments

[0018] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed can be changed without departing from the protection scope of the content of this specification. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0019] In at least one embodiment of the present invention, an automatic control method for tripping of low-load auxiliary fans under the deep peak shaving condition is disclosed, as Figure 1 shown, including the following steps: Step 1, construct a fuzzy cognitive map of the influencing factors of the auxiliary fan operation to predict the tripping risk of the auxiliary fan; Specifically, it includes the following sub-steps: Step 1.1, collect data on the influencing factors of the auxiliary fan operation, including multi-dimensional data parameters such as the peak shaving depth, inlet air temperature, vibration parameters, load level, operation duration, and equipment aging degree.

[0020] Step 1.2, construct a set of nodes of the fuzzy cognitive map, and use each influencing factor as a map node.

[0021] The fuzzy cognitive map can be expressed as: ; wherein, represents the fuzzy cognitive map, represents the set of map nodes, , , Fuzzify the , , th factor nodes in the fuzzy cognitive map, denotes the total number of nodes in the fuzzy cognitive map, denotes the weight matrix, denotes a row column matrix, where each element denotes the influence weight from node to node . denotes the range of influence weight values of node on node . A positive value indicates a positive influence, a negative value indicates a negative influence, and 0 indicates no influence.

[0022] Step 1.3: Based on historical operation data and expert knowledge, determine the fuzzy weights between nodes and establish the weight matrix .

[0023] The weight value is calculated by the following formula: ; where denotes the influence weight from node to node , denotes the correlation coefficient between node and node in the th historical data sample, denotes the observed value of node in the th historical data sample, denotes the number of historical data samples, denotes the sum of correlation coefficients for all historical samples.

[0024] Step 1.4: Predict the potential tripping risk by iteratively calculating the activation values of each node.

[0025] The update formula for the node activation value is: ; where denotes the activation value of node in the th iteration, denotes the activation value of node in the th iteration, denotes the node The activation value in the th iteration, represents the influence weight of node on node . represents the total number of nodes in the fuzzy cognitive map, represents the activation function, using the hyperbolic tangent function: ; Among them, represents the natural constant to the power, represents the natural constant to the power, The function maps the input value to the interval to keep the node activation value within a reasonable range.

[0026] Step 1.5, set the tripping risk threshold of the auxiliary fan . When the activation value of the tripping risk node exceeds the threshold, output a tripping risk warning signal: ; Among them, represents the tripping risk warning signal, represents the activation value of the tripping risk node in the th iteration, represents the preset tripping risk threshold, represents that there is a tripping risk and preventive measures need to be taken, represents that there is no tripping risk and the system can operate normally.

[0027] Step 2, based on the tripping risk prediction result, establish a multi-dimensional fairness index system including instant load fairness, cumulative operation time fairness, and life consumption rate fairness, providing fairness constraints for the auxiliary fan load distribution; Specifically, it includes the following sub-steps: Step 2.1, construct an instant load fairness index to evaluate the balance degree of the load distribution of each auxiliary fan at the current moment.

[0028] The instant load fairness index is calculated by the Gini coefficient: ; Among them, represents the instant load fairness index, represents the total number of auxiliary fans, , respectively represent the rd and the The current load level of the auxiliary fan represents the th and the th auxiliary fan absolute difference in load levels, represents the average load level of all auxiliary fans, represents the double summation of all pairs of auxiliary fans.

[0029] Step 2.2, construct the cumulative running time fairness index to evaluate the balance degree of the historical load accumulation of each auxiliary fan.

[0030] The cumulative running time fairness index is calculated as follows: ; where, represents the cumulative running time fairness index, represents the total number of auxiliary fans, 、 respectively represent the th and the th auxiliary fan cumulative running time, represents the th and the th auxiliary fan cumulative running time absolute difference - represents the average cumulative running time of all auxiliary fans.

[0031] Step 2.3, construct the life consumption rate fairness index, considering the health status and expected life of different auxiliary fans.

[0032] The life consumption rate fairness index is calculated as follows: ; where, represents the life consumption rate fairness index, represents the life consumption rate of the th auxiliary fan, represents the life consumption rate of the th auxiliary fan, represents the th and the th auxiliary fan life consumption rate absolute difference, represents the th auxiliary fan expected total life, represents the average life consumption rate of all auxiliary fans.

[0033] Step 2.4, calculate the marginal contribution of each auxiliary fan to the system stability based on the Shapley value.

[0034] For any subset of auxiliary fans , define the characteristic function Represents the contribution value of the system stability of this subset. The Shapley value of the auxiliary fan is calculated as follows: ; Among them, represents the Shapley value of the th auxiliary fan, represents any subset of auxiliary fans that does not include the auxiliary fan , represents the set of all auxiliary fans after removing the auxiliary fan , represents the number of auxiliary fans in the set , represents the total number of auxiliary fans, represents factorial, represents factorial, represents factorial, represents the contribution value of the system stability after adding the auxiliary fan to the subset , represents the contribution value of the system stability of the subset , represents the marginal contribution of the auxiliary fan to the subset .

[0035] Step 2.5, introduce a time discount mechanism to balance the short-term and long-term fairness requirements, and construct a comprehensive fairness index: ; Among them, represents the comprehensive fairness index at time , represents the current time, represents the immediate load fairness index at time , represents the cumulative operation time fairness index at time , represents the life consumption rate fairness index at time , represents the initial weight coefficient of the immediate load fairness index, represents the initial weight coefficient of the cumulative operation time fairness index, represents the weight coefficient of the life consumption rate fairness index, represents the time discount coefficient, represents an exponential function that decays with time, represents a function that grows with time, And It means that the sum of the three weight coefficients is 1 and they are all non - negative.

[0036] Step 3: According to the multi - dimensional fairness index system, each auxiliary air blower is constructed as an intelligent agent with autonomous decision - making ability, and the load distribution plan is determined through multi - agent negotiation and game. Specifically, it includes the following sub - steps: Step 3.1: Establish a multi - agent system model for auxiliary air blowers. Each auxiliary air blower is regarded as an intelligent agent with autonomous decision - making ability and is defined as: ; Among them, represents the intelligent agent corresponding to the th auxiliary air blower, represents the state space, represents the action space, represents the utility function.

[0037] Step 3.2: Construct an inter - agent communication protocol to achieve efficient sharing of the operation information of auxiliary air blowers.

[0038] The communication protocol includes: Message type: : State information message, used to transmit the current operation state of the auxiliary air blower; : Decision intention message, used to transmit the load adjustment intention that the auxiliary air blower plans to take; : Negotiation request message, used to initiate load distribution negotiation; : Negotiation response message, used to respond to the negotiation request; Message format: ; Among them, represents the unique identifier of the sending auxiliary air blower; represents the unique identifier of the receiving auxiliary air blower; represents the message type, which can be state, intention, request or response; represents the message content, including specific data information; represents the timestamp of the message sending; Communication timing: State broadcast stage → Intention declaration stage → Conflict detection stage → Negotiation resolution stage → Execution confirmation stage.

[0039] Step 3.3: Construct a load distribution strategy based on evolutionary game theory, and set the strategy set where 、 、 respectively represent the , , types of load distribution strategies, represents the total number of load distribution strategies.

[0040] main body selection strategy The profit function is defined as: ; where represents the comprehensive profit of the main body adopting the strategy , represents the load performance profit, represents the tripping risk cost, represents the fairness profit, , , respectively represent the weight coefficients of the performance profit, tripping risk cost, and fairness profit.

[0041] Step 3.4, implement the replicator dynamic equation to describe the evolution process of the auxiliary fan strategy selection.

[0042] main body selection strategy The probability evolution equation is: ; where represents the probability that the main body selects the strategy at time , represents the rate of change of the probability with time, represents the profit of the main body adopting the strategy , represents the average profit of the main body at time .

[0043] Step 3.5, construct a multi-round negotiation mechanism to solve the load distribution conflict of the auxiliary fan group.

[0044] The negotiation process adopts the Nash bargaining solution in non-cooperative game: ; where represents the optimal load distribution plan, represents the feasible solution space, represents the utility of the main body under the plan , Represents the main body The reserved utility of Represents the total number of auxiliary fans participating in the negotiation Represents the product of the utility gains of all auxiliary fans Represents the maximum value

[0045] Step 3.6, introduce a dynamic adjustment mechanism for negotiation weights based on trust. The trust calculation formula is: ; Among them, Represents the main body At time The trust of Represents the main body The trust at the next moment Represents the weight coefficient of historical trust Represents the main body At the Degree of cooperation in the Represents the number of historical interactions under investigation Represents the main body In the recent Average degree of cooperation in interactions

[0046] Step 3.7, design a multi-agent decision fusion algorithm. According to the negotiation results and trust of each auxiliary fan, calculate the final load distribution plan: ; Among them, Represents the final load value assigned to the auxiliary fan after fusion Represents the main body Suggested load value to be assigned to the auxiliary fan Represents the main body At time The trust of Represents the total number of auxiliary fans participating in the decision-making Represents all agents' weighted sum of load suggestions for the auxiliary fan Represents the sum of the trust of all agents

[0047] Step 4, based on the negotiation game results, form a temporary alliance of equipment with similar performance within the auxiliary fan group, and guide the system to evolve towards the Pareto optimal solution through an adaptive evolutionary strategy; Specifically, it includes the following sub-steps: Step 4.1, construct a similarity evaluation model for auxiliary fans and calculate the similarity between devices: ;​​ Among them, represents the auxiliary fan and the auxiliary fan similarity, represents the auxiliary fan the th performance parameter, represents the auxiliary fan the th performance parameter, represents the parameter dimension, represents the th parameter weight coefficient, represents the standard deviation parameter.

[0048] Step 4.2, construct an auxiliary fan alliance division algorithm based on the similarity matrix, using the spectral clustering method: Construct the similarity matrix , where represents the total number of auxiliary fans; Calculate the Laplacian matrix , where is the degree matrix; Calculate the first th smallest eigenvalue corresponding eigenvectors, forming the eigenmatrix ; Perform K-means clustering on the row vectors of the eigenmatrix to obtain the auxiliary fan alliance division result.

[0049] Step 4.3, implement the load distribution algorithm within the alliance based on game theory. Within each alliance, set the total load target of the alliance , and the load distribution formula for the auxiliary fan within the alliance is: ; Among them, represents the load allocated to the auxiliary fan , represents the load bearing capacity of the auxiliary fan , represents the life consumption rate of the auxiliary fan , represents the set of auxiliary fans within the current alliance, represents the adjustment coefficient, represents the total weighted capacity of all auxiliary fans within the alliance.

[0050] Step 4.4, implement the load negotiation mechanism between alliances based on the genetic algorithm, including the following steps: Encoding: Encode the load distribution scheme of each alliance into a chromosome , where represents the distribution ratio of the total load, and

[0051] Initialization: Randomly generate an initial population, where each individual represents a load distribution scheme among coalitions; Evaluation: Calculate the fitness of each individual. The fitness function comprehensively considers system stability, load balance, and fairness; Selection: Use the tournament selection algorithm to select excellent individuals to enter the next generation; Crossover: Perform crossover operations on the selected individual pairs to generate new load distribution schemes; Mutation: Perform mutation operations on some individuals to increase population diversity; Update: Update the population, replacing individuals with lower fitness until the termination condition is reached.

[0052] Step 5: For the coalition evolution process, introduce a reputation mechanism to evaluate the historical performance of each entity, and continuously optimize the load distribution strategy through the feedback of historical decision-making effects.

[0053] Specifically, it includes the following sub-steps: Step 5.1: Construct a reputation evaluation model for the auxiliary fan entity. The reputation value calculation formula is: ; where represents the reputation value of the auxiliary fan at time , represents the reputation value of the auxiliary fan at the previous time, represents the decay coefficient of historical reputation, represents the performance score of the auxiliary fan at time , which is comprehensively calculated through the following indicators: ; where represents the stability index, represents the cooperation index, represents the fairness index, , , respectively represent the weight coefficients of the stability, cooperation, and fairness indexes.

[0054] Step 5.2: Construct a decision weight adjustment algorithm based on reputation perception. The weight calculation formula of the auxiliary fan in collective decision-making is: ; where Indicates the auxiliary fan At time The decision-making weight of Indicates the total number of auxiliary fans in the system Indicates the sensitivity parameter of reputation impact Indicates the weighted sum of the reputation values of all auxiliary fans 、 Respectively indicate the auxiliary fan And the auxiliary fan At time The reputation value of

[0055] Step 5.3, construct an adaptive optimization model for the load distribution strategy based on reinforcement learning

[0056] Define the Q-value function for each auxiliary fan agent : ; Wherein Indicates the auxiliary fan Taking action In the state The expected return of Indicates the expected value under the policy Of Indicates the auxiliary fan At time The return obtained by Indicates the current system state Indicates the currently taken action

[0057] Step 5.4, implement the Q-value update algorithm based on temporal difference learning ; Wherein Indicates the auxiliary fan Taking action In the state The Q-value of Indicates the learning rate Indicates the immediate return Indicates the discount factor Indicates the next state The maximum Q-value of Indicates the temporal difference error

[0058] Step 5.5, construct a multi-agent experience replay pool mechanism to improve learning efficiency Maintain an experience replay pool for each agent To store the experience sample tuple Wherein Indicates the current state Indicates the action taken Indicates the immediate reward obtained Indicates the next state to which the transfer is made

[0059] During the training process, a batch of samples is randomly sampled from the experience replay pool for Q-value update to reduce the correlation between samples Introduce a priority experience replay mechanism, allocate sample priorities according to the temporal difference error, the greater the error, the higher the probability of being sampled, and accelerate the learning of key experiences

[0060] Step 5.6, construct a multi-agent collaborative exploration strategy to balance exploration and exploitation ; Among them Indicates the agent In the state Select the action Probability of Indicates selecting the action with the largest Q-value Indicates the moment Exploration rate of Indicates the minimum value of the exploration rate Indicates the maximum value of the exploration rate Indicates the decay coefficient

[0061] Step 5.7, construct a historical decision effect feedback evaluation model to optimize the policy parameters through historical data analysis Periodically collect system operation data and construct an evaluation index set Among them 、 、 Respectively represent the 、 、 th evaluation index Indicates the total number of evaluation indexes, including Tripping rate: The frequency of tripping in the system per unit time Load balance degree: The degree of even distribution of the loads of each auxiliary fan Equipment life consumption rate: The speed of life consumption of each auxiliary fan System response time: The execution speed of the load adjustment instruction Energy efficiency: The system stability performance under unit energy consumption For each time window Calculate the change trend of the evaluation index ; Adjust the policy parameters based on the evaluation results, such as the fairness index weight, the reputation influence coefficient, the exploration rate, etc., to achieve the adaptive optimization of the policy

[0062] Application example of this embodiment: This embodiment is applied to the low-load auxiliary fan control under the deep peak shaving condition of a large thermal power plant. The power plant has 6 auxiliary fans with different models and service years, and frequent tripping events occur under the deep peak shaving condition, seriously affecting the safe operation of the power plant. The specific application scenario is as follows: The auxiliary fan group of the power plant is: 2 new GF-3000 models (in service for 1 year), 2 conventional GF-2000 models (in service for 5 years), and 2 old GF-1000 models (in service for 9 years).

[0063] Under the deep peak shaving condition, the generator set is required to operate at 40% of the rated load. At this time, the auxiliary fan cluster needs to operate in low load coordination to ensure the safe and stable operation of the boiler. The traditional control method adopts the load average distribution strategy, resulting in frequent tripping of old fans, while the load rate of new fans is very low, and the overall stability of the system is poor. By applying the fairness constraint multi-agent collaborative load balancing control method of the present invention, intelligent load distribution among auxiliary fans is achieved, significantly improving the system stability.

[0064] Implementation process of fuzzy cognitive map modeling: For the operating environment of the auxiliary fans in the power plant, first collect the key factors affecting tripping and establish a fuzzy cognitive map. The definition and initial activation values of the nodes in the fuzzy cognitive map are shown in Table 1. By analyzing the historical operation data of the past 3 years, 8 key nodes are identified: peak shaving depth, inlet air temperature, vibration intensity, load level, equipment aging degree, operation duration, environmental humidity, and dust concentration.

[0065] Table 1: Definition and initial activation values of nodes in the fuzzy cognitive map;

[0066] Based on historical data and expert experience, construct a fuzzy weight matrix between nodes. By iteratively calculating the node activation values (setting the number of iterations to 10 times), the tripping risk prediction results of each auxiliary fan are obtained. The prediction results are shown in Table 2: Table 2: Tripping risk prediction results of different auxiliary fans;

[0067] According to the setting of the tripping risk threshold (low risk: <0.5; medium risk: 0.5 - 0.7; high risk: >0.7), it is identified that two old auxiliary fans, F5 and F6, have high tripping risks, which is consistent with the high tripping frequency of these two devices in the actual operation records, verifying the effectiveness of the fuzzy cognitive map model.

[0068] Construction and application of multi-dimensional fairness indicators: Based on the above tripping risk prediction results, a multi-dimensional fairness index system is constructed to balance the load distribution of each auxiliary fan. The initial parameters of each auxiliary fan and the calculation of fairness indices are shown in Table 3. First, three fairness indices are calculated: Table 3: Calculation of initial parameters of each auxiliary fan and fairness indices;

[0069] Calculate the immediate load fairness index , the cumulative operating time fairness index , and the life consumption rate fairness index . Among them, the life consumption rate fairness index is the lowest, indicating that the existing load distribution strategy lacks sufficient consideration for the balance of equipment life consumption.

[0070] Calculate the marginal contribution of each auxiliary fan to system stability through the Shapley value. Based on the multi-dimensional fairness index and the Shapley value, through the time discount mechanism (set , , , ), the comprehensive fairness index is obtained, indicating that there is still much room for improvement in the current system fairness.

[0071] Implementation process of multi-agent negotiation game and dynamic coalition formation: Six auxiliary fans are constructed as six intelligent agents. Based on the fuzzy cognitive map model and the fairness index, the load distribution plan is determined through multi-agent negotiation game. The initial utility function and trust degree of the agents are shown in Table 4. Specifically: Table 4: Initial utility and trust degree parameters of agents;

[0072] Based on the calculation of the similarity of auxiliary fans, the system automatically forms two temporary coalitions: Coalition 1 consists of A1 and A2 (new auxiliary fan coalition); Coalition 2 consists of A3 and A4 (conventional auxiliary fan coalition); A5 and A6 (old and worn-out auxiliary fans) remain independent. The negotiation game results after coalition formation are shown in Table 5. Specifically: Table 5: Load distribution results after negotiation game;

[0073] This load distribution plan reflects the optimal adjustment under fairness constraints: new auxiliary fans undertake more loads, old and worn-out auxiliary fans reduce loads appropriately, while ensuring that the overall load demand remains unchanged, and the system comprehensive fairness index is improved to .

[0074] Implementation process of reputation mechanism feedback optimization: During the operation of the system, the performance of each entity is continuously evaluated through the reputation mechanism, and the decision weights are dynamically adjusted. With the implementation of the new load distribution scheme, the reputation values of each entity, especially the old and low-reputation auxiliary fan entities, have been significantly improved, reflecting the self-adaptive optimization ability of the system. The change in the reputation value leads to the corresponding adjustment of the decision weight, and the weight distribution becomes more balanced, further improving the overall fairness of the system.

[0075] During the policy optimization based on reinforcement learning, historical decision samples are accumulated through the experience replay mechanism, and the Q-value function is continuously optimized. During the 60-day operation period, the system completed approximately 2000 rounds of decision optimization, and the Q-value function converged well. The average temporal difference error decreased from the initial 0.82 to 0.15, indicating significant learning effects.

[0076] Verification of technical effects: After this implementation method is applied to the power plant, the operating stability and economy of the system have been significantly improved. By comparing the tripping frequency and equipment life consumption before and after the system application (as shown in Tables 6 and 7), the technical effects are verified: Table 6: Comparison of tripping frequencies before and after system application (3-month operation data);

[0077] The overall tripping rate of the system has decreased by 57.4%, which is consistent with the expected technical effects. Especially for the high-risk old auxiliary fans F5 and F6, the number of trips has decreased from 16 and 18 times to 7 and 8 times respectively, and the tripping risk has been effectively controlled.

[0078] Table 7: Changes in equipment life consumption rates before and after system application (predicted 30-year service life);

[0079] As can be seen from Table 7, although the new auxiliary fans undertake more loads and their life consumption rates increase slightly, the life consumption rates of the old auxiliary fans have decreased significantly, and the average life of the overall system has been extended by 4.8%. More importantly, the life consumption among various equipment is more balanced, and the maximum difference has been reduced from 1.67 service life cycles to 0.95 service life cycles, a decrease of 43.1%.

[0080] In addition, the system has also been significantly improved in terms of deep peak shaving ability. By optimizing the load distribution, the generating units can operate safely and stably at 38% of the rated load (a 5% increase compared to the original 40%), further enhancing the flexibility of the power grid for peak shaving and providing more space for the power system to absorb renewable energy.

[0081] The embodiments of the present invention have been described above. However, these embodiments are not limited to the specific implementation manners described above. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of these embodiments, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of these embodiments.

Claims

1. A method for automatically controlling the tripping of low-load auxiliary fans in deep peak-shaving conditions, characterized in that: The following steps are involved: Construct a fuzzy cognitive map of the factors affecting the operation of the auxiliary fan to predict the tripping risk of the auxiliary fan; Based on the tripping risk prediction results, a multi-dimensional fairness indicator system including instantaneous load fairness, cumulative operating time fairness and life consumption rate fairness is established to provide fairness constraints for auxiliary fan load distribution. According to the multi-dimensional fairness index system, each auxiliary wind turbine is constructed as an intelligent subject with autonomous decision-making ability, and the load distribution plan is determined through multi-agent negotiation game; Based on the negotiation game results, a temporary alliance of devices with similar performance within the auxiliary wind turbine group is formed, and the system is guided to evolve towards the Pareto optimal solution through an adaptive evolution strategy. In the process of alliance evolution, a reputation mechanism is introduced to evaluate the historical performance of each entity, and the load distribution strategy is continuously optimized through feedback on historical decision-making effects.

2. The automatic control method for tripping of low-load auxiliary fans in deep peak-shaving conditions according to claim 1 is characterized in that: The step of constructing a fuzzy cognitive map of factors affecting the operation of the auxiliary fan comprises: Collect data on factors affecting auxiliary fan operation, including peak load depth, air inlet temperature, vibration parameters, load level, operating time, and equipment aging degree; Construct a set of fuzzy cognitive graph nodes, and take each influencing factor as a graph node; Based on historical operation data and expert knowledge, the fuzzy weights between nodes are determined and the weight matrix is ​​established; The potential tripping risk is predicted by iteratively calculating the activation value of each node.

3. The automatic control method for tripping of low-load auxiliary fans in deep peak load regulation conditions according to claim 2 is characterized in that: The update of the node activation value in the fuzzy cognitive graph adopts the hyperbolic tangent function as the activation function, and sets the tripping risk threshold to output the tripping risk warning signal.

4. The automatic control method for tripping of low-load auxiliary fans in deep peak load regulation conditions according to claim 1 is characterized in that: The multi-dimensional fairness index system calculates the instantaneous load fairness index through the Gini coefficient, calculates the cumulative operating time fairness index through the cumulative operating time difference, and calculates the life consumption rate fairness index through the life consumption rate difference.

5. The method for automatic control of low-load auxiliary fan tripping in deep peak load regulation condition according to claim 4 is characterized in that: The multi-dimensional fairness index system also calculates the marginal contribution of each auxiliary fan to system stability based on the Shapley value, and introduces a time discount mechanism to balance short-term and long-term fairness requirements to construct a comprehensive fairness index.

6. The automatic control method for tripping of low-load auxiliary fans in deep peak load regulation conditions according to claim 1 is characterized in that: The steps of determining the load distribution scheme through multi-agent negotiation game include: Establish a multi-agent system model of auxiliary fans, with each auxiliary fan acting as an intelligent agent with autonomous decision-making capabilities; Build inter-subject communication protocols to achieve efficient sharing of auxiliary fan operation information; Construct a load distribution strategy based on evolutionary game theory; The Nash bargaining solution in non-cooperative game is used to deal with the load allocation conflict of auxiliary wind turbines. A multi-agent decision fusion algorithm is designed to calculate the final load distribution plan based on the negotiation results and trust levels of each auxiliary fan.

7. The automatic control method for tripping of low-load auxiliary fans in deep peak load regulation conditions according to claim 6 is characterized in that: The inter-subject communication protocol includes status information messages, decision intention messages, negotiation request messages, and negotiation response messages, and the communication sequence includes status broadcast phase, intention declaration phase, conflict detection phase, negotiation resolution phase, and execution confirmation phase.

8. The method for automatic control of low-load auxiliary fan tripping in deep peak load regulation condition according to claim 1 is characterized in that: The steps for forming a temporary alliance of devices with similar performance within the auxiliary wind turbine group include: Construct an auxiliary fan similarity evaluation model to calculate the similarity between devices; Based on the similarity matrix, the auxiliary fan alliance partition algorithm is constructed, and the spectral clustering method is adopted; Build a collaborative decision-making model within the alliance; Construct an adaptive evolution strategy and build a load distribution game model among alliances; A Pareto solution set management strategy based on dominance relationship is adopted to achieve convergence to the Pareto optimal solution.

9. The method for automatic control of tripping of auxiliary fans under low load in deep peak load regulation condition according to claim 1 is characterized in that: The steps of evaluating the historical performance of each subject by the reputation mechanism include: Construct a reputation evaluation model for auxiliary fan entities; Construct a reputation-aware decision weight adjustment algorithm; Construct an adaptive optimization model of load distribution strategy based on reinforcement learning; Construct a multi-agent experience replay pool mechanism; Build a historical decision-making effect feedback evaluation model and optimize strategy parameters through historical data analysis.

10. An automatic control system for tripping of low-load auxiliary fans under deep peak-shaving conditions, used to execute an automatic control method for tripping of low-load auxiliary fans under deep peak-shaving conditions as claimed in any one of claims 1 to 9, characterized in that: include: The fuzzy cognitive map modeling module is used to construct a fuzzy cognitive map of the factors affecting the operation of the auxiliary fan and predict the tripping risk of the auxiliary fan; A multi-dimensional fairness indicator construction module is used to establish a multi-dimensional fairness indicator system including instantaneous load fairness, cumulative operating time fairness and life consumption rate fairness; The multi-agent negotiation game module is used to build each auxiliary wind turbine into an intelligent entity with autonomous decision-making capabilities, and determine the load distribution plan through multi-agent negotiation games; Dynamic alliance formation module, which is used to realize the temporary alliance formation of devices with similar performance within the auxiliary wind turbine group, and guide the system to evolve towards the Pareto optimal solution through adaptive evolution strategy; The reputation mechanism and feedback optimization module are used to evaluate the historical performance of each entity and continuously optimize the load distribution strategy through feedback on historical decision-making effects.

Images