Generator tripping stability control strategy determination method and device, computer equipment and storage medium

Abstract

The invention relates to a generator tripping stability control strategy determination method and device, computer equipment and a storage medium, and relates to the technical field of power system control. The method comprises the following steps: determining training feature data of a power system; the training feature data is obtained by screening electrical quantity data of the power system in a preset time period; based on the training feature data corresponding to the power system, training the security constraint model to obtain a trained security constraint model; the security constraint model is obtained by modeling a generator tripping stability control strategy of the power system based on a Markov decision process; the security constraint model comprises a quintuple corresponding to a Markov decision process and a deep learning network; the quintuple comprises a state space, an action space, a constraint condition, a reward function and a discount coefficient; and inputting target electrical quantity data of the power system into the trained security constraint model to obtain a generator tripping stability control strategy corresponding to the target electrical quantity data. By adopting the method, the efficiency of generating the stability control strategy can be improved.

Description

Technical Field

[0001] The present application relates to the field of power system control technology, and in particular to a method, device, computer equipment, storage medium and computer program product for determining a generator switching stabilization control strategy. Background Art

[0002] With the access of a large number of new energy sources and a large number of power electronic devices to the power system, the power system has equipment diversity, circuit mechanism complexity and power uncertainty. In addition, the output generated by new energy is uncertain, which can change the transient stability characteristics of the power grid, resulting in insufficient safety and stability of the power system and the power grid.

[0003] In the related technology, the stabilization control strategy can be determined through physical models and real-time simulation methods, and the safety and stability of the power system can be improved through the stabilization control strategy. However, the physical model needs to be additionally improved to ensure the coordinated work of the system, which will increase the complexity of the system. There is a lack of theoretical guidance for the calculation of the machine cutting amount of the new energy power system in real-time simulation, resulting in redundant machine cutting stabilization control strategies in the process of generating the stabilization control strategy. Therefore, when generating the stabilization control strategy through the above-mentioned related technologies, there is a problem of insufficient efficiency in generating the stabilization control strategy. Summary of the invention

[0004] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium and computer program product for determining a cutting machine stabilization control strategy that can improve the efficiency of generating stabilization control strategies in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for determining a cutting machine stabilization control strategy. The method comprises:

[0006] Determining training characteristic data of the power system; the training characteristic data is obtained by screening the electrical quantity data of the power system in a preset time period;

[0007] Based on the training feature data corresponding to the power system, the safety constraint model is trained to obtain a trained safety constraint model; the safety constraint model is obtained by modeling the power system's machine switching stabilization control strategy based on the Markov decision process; the safety constraint model includes a quintuple and a deep learning network corresponding to the Markov decision process; the quintuple includes a state space, an action space, a constraint condition, a reward function, and a discount coefficient;

[0008] Inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain a generator tripping stabilization control strategy corresponding to the target electrical quantity data;

[0009] Among them, the state space is used to represent the electrical quantity data corresponding to multiple nodes corresponding to each generator in the power system at different time points; the action space is used to represent the power cutting instructions corresponding to each generator in the power system; and the reward function is used to represent the reward value caused by the power cutting instruction after the power system is cut off through the power cutting instruction in the action space.

[0010] In one embodiment, before training the safety constraint model based on the training feature data corresponding to the power system to obtain the trained safety constraint model, the method further includes:

[0011] Generate a state space based on electrical quantity data of each generator in the power system at different time points; the electrical quantity data includes generator rotor speed, generator power angle, rotor acceleration, bus voltage and bus current;

[0012] Based on the tripping instructions of each generator in the power system at different time points, an action space is generated; wherein the tripping instruction is used to determine whether the generator is to be tripped;

[0013] Determining a reward function based on a power angle stability reward corresponding to each generator in the power system and a power-off amount penalty of the generator after taking a power-off instruction corresponding to an action space;

[0014] At least one constraint condition is determined; the constraint condition includes a machine cutting action timing constraint, a machine cutting group selection constraint, and a machine cutting quantity constraint.

[0015] In one embodiment, determining a reward function based on a power angle stability reward corresponding to each generator in the power system and a power cut amount penalty after the generator takes a power cut instruction corresponding to the action space includes:

[0016] Determine the maximum power angle difference between the generators at the same time point, and determine the power angle stability reward corresponding to the maximum power angle difference according to the value of the maximum power angle difference;

[0017] The product value of the state space and action space corresponding to each time point is determined as the machine cutting penalty;

[0018] The difference between the power angle stability reward and the machine cutting amount penalty is determined as a reward function.

[0019] In one embodiment, determining at least one constraint condition includes:

[0020] Based on the time period corresponding to the electrical quantity data and based on the difference of the action space corresponding to adjacent time points, the timing constraint of the machine cutting action is determined; the timing constraint of the machine cutting action is used to limit the value range of the action space at different time points;

[0021] Determining the cutting machine group selection constraint based on the working angular acceleration of the generator; the cutting machine group selection constraint is used to limit the generator selected by the cutting machine;

[0022] The machine cutting amount constraint is determined based on the load power of the electric power system; the machine cutting amount constraint is used to limit the number of generators that issue machine cutting instructions.

[0023] In one embodiment, determining training characteristic data of the power system includes:

[0024] Obtaining electrical quantity data of the power system in a preset time period;

[0025] Based on the intra-class distances corresponding to the electrical quantity data of each type, determining the discrimination factor corresponding to the electrical quantity data; the intra-class distances are determined based on the distances between the electrical quantity data of each type at each time point; the discrimination factor is determined based on the intra-class distances corresponding to different time points;

[0026] Based on the discrimination factors corresponding to the electrical quantity data arranged in order of magnitude, determining the level difference between adjacent discrimination factors, and based on the level difference, determining the level correlation coefficients corresponding to the adjacent discrimination factors;

[0027] If the rank correlation coefficient is greater than a preset value, the electrical quantity data corresponding to the rank correlation coefficient is determined to be training feature data.

[0028] In one embodiment, the deep learning network includes an evaluation network and an action network; the safety constraint model is trained based on the training feature data corresponding to the power system to obtain a trained safety constraint model, including:

[0029] Determine an initial state value of the training feature data in the state space, an initial action instruction in the action space, and an initial reward value corresponding to the reward function;

[0030] Based on the action network, the evaluation network and the discount coefficient, the initial state value and the initial action instruction are updated, and the predicted reward value and the real reward value corresponding to the initial reward value are determined; the updating process satisfies the constraint condition;

[0031] The first parameter of the action network and the second parameter of the evaluation network are updated until the loss value between the predicted reward value and the actual reward value meets the training end condition, thereby obtaining a trained safety constraint model.

[0032] In one embodiment, inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain the generator tripping stabilization control strategy corresponding to the target electrical quantity data includes:

[0033] Inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain a target state space corresponding to the target electrical quantity data;

[0034] Determine the action instruction corresponding to the target state space as the machine cutting stabilization control strategy.

[0035] In a second aspect, the present application also provides a device for determining a cutting machine stabilization control strategy. The device comprises:

[0036] A data determination module, used to determine training characteristic data of the power system; the training characteristic data is obtained by screening the electrical quantity data of the power system in a preset time period;

[0037] A model training module is used to train the safety constraint model based on the training feature data corresponding to the power system to obtain a trained safety constraint model; the safety constraint model is obtained by modeling the power system's machine-cutting stabilization control strategy based on the Markov decision process; the safety constraint model includes a quintuple and a deep learning network corresponding to the Markov decision process; the quintuple includes a state space, an action space, a constraint condition, a reward function and a discount coefficient; wherein the state space is used to represent the electrical quantity data corresponding to multiple nodes corresponding to each generator in the power system at different time points; the action space is used to represent the machine-cutting instructions corresponding to each generator in the power system; the reward function is used to represent the reward value caused by the machine-cutting instruction after the power system is cut off through the machine-cutting instruction in the action space;

[0038] The strategy determination module is used to input the target electrical quantity data of the power system into the trained safety constraint model to obtain the generator tripping stabilization control strategy corresponding to the target electrical quantity data.

[0039] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in the first aspect are implemented.

[0040] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0041] In a fifth aspect, the present application further provides a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0042] The above-mentioned method, device, computer equipment, storage medium and computer program product for determining the machine-cutting stabilization control strategy obtain training feature data for training the safety constraint model by screening the electrical quantity data of the power system in a preset time period. And the safety constraint model is trained by the training feature data to obtain a trained safety constraint model. The target electrical quantity data of the power system is input into the trained safety constraint model to obtain the machine-cutting stabilization control strategy corresponding to the target electrical quantity data. Among them, the safety constraint model is obtained by modeling the machine-cutting stabilization control strategy of the power system based on the Markov decision process, and the Markov decision process can be represented by a five-tuple consisting of a state space, an action space, a constraint condition, a reward function and a discount coefficient. Based on this, the electrical quantity data of each generator in the power system at different time points can be represented by the state space, the machine-cutting instructions that can be executed by each generator can be represented by the action space, and the reward value caused by the machine-cutting instruction in the action space can be represented by the reward function. Thus, the relationship between the electrical quantity data, the power system and the machine-cutting instruction is constrained by the five-tuple, and the safety constraint model is constructed by the five-tuple. By training the safety constraint model, the safety constraint model can output the optimal generator cutting and stabilization control strategy under different state spaces. After building the trained safety constraint model, only the target electrical quantity data of the power system needs to be input to quickly obtain the generator cutting and stabilization control strategy corresponding to the target electrical quantity data, thereby improving the efficiency of generating the stabilization control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] FIG1 is an application environment diagram of a method for determining a cutting machine stabilization control strategy in one embodiment;

[0045] FIG2 is a schematic flow chart of a method for determining a cutting machine stabilization control strategy in one embodiment;

[0046] FIG3 is a flow chart of a method for determining a cutting machine stabilization control strategy in another embodiment;

[0047] FIG4 is a flow chart of a DDPG learning algorithm strategy update in one embodiment;

[0048] FIG5 is a topological diagram of an IEEE39 node in one embodiment;

[0049] FIG6 is a schematic diagram of decision reward function value comparison in one embodiment;

[0050] FIG7 is a diagram showing power angle fluctuation and speed fluctuation of an uncut machine in one embodiment;

[0051] FIG8 is a diagram showing power angle fluctuation and speed fluctuation after cutting off the machine in one embodiment;

[0052] FIG9 is a structural block diagram of a device for determining a stabilization control strategy for cutting machines in one embodiment;

[0053] FIG. 10 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] The method for determining the machine-cutting stabilization control strategy provided in the embodiment of the present application can be applied in the application environment shown in FIG1 . The power system 102 communicates with the server 104 via a communication network. Each generator node in the power system 102 can send electrical quantity data to the power system in real time. The server 104 can store and filter the electrical quantity data, and obtain training feature data for training therefrom. The server 104 can also obtain target electrical quantity data from the power system 102, and obtain the corresponding machine-cutting stabilization control strategy by inputting the target electrical quantity data into the trained safety constraint model.

[0056] The server 104 can first train the pre-built safety constraint model by training feature data. The safety constraint model includes a quintuple and a deep learning network. The quintuple in the safety constraint model is obtained by the server modeling the power system's machine-cutting strategy based on the Markov decision process, and the deep learning network in the safety constraint model is determined based on the quintuple. When training the safety constraint model, the server converts the training feature data into data corresponding to the state space and action space in the quintuple, and trains the deep learning network through the constraints, reward functions, and discount coefficients in the quintuple to obtain a trained safety constraint model.

[0057] The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. Among them, the power system 102 can be, but is not limited to, various electronic power devices, generators and other power grid equipment.

[0058] In an exemplary embodiment, as shown in FIG2 , a method for determining a machine switching stabilization control strategy is provided, and the method is applied to the server in FIG1 as an example for explanation, including the following steps S202 to S206 . Among them:

[0059] Step S202, determining training characteristic data of the power system.

[0060] The training feature data is obtained by screening the electrical quantity data of the power system in a preset time period.

[0061] Specifically, the server can continuously receive the electrical quantity data sent by each generator in the power system within a preset time period. The server can sample different types of electrical quantity data, and obtain the electrical quantity data corresponding to each time point according to different sampling methods for each type of electrical quantity data within the preset time period, and use it as training feature data. In one example, multiple valid time points in the preset time period can be determined, and the electrical quantity data corresponding to the valid time points can be sampled and merged in chronological order to obtain training feature data. For example, for type A electrical quantity data, January, April, June, and July are selected as valid time points in the past year, and training feature data are obtained by sampling. For type B electrical quantity data, working days can be selected as valid time points in the past year, and training feature data can be obtained by sampling.

[0062] Step S204: training the security constraint model based on the training feature data corresponding to the power system to obtain a trained security constraint model.

[0063] Among them, the safety constraint model is obtained by modeling the power system's machine-cutting stabilization control strategy based on the Markov decision process; the safety constraint model contains the quintuple and deep learning network corresponding to the Markov decision process; the quintuple includes state space, action space, constraint conditions, reward function and discount coefficient; the state space is used to represent the electrical quantity data corresponding to multiple nodes corresponding to each generator in the power system at different time points; the action space is used to represent the machine-cutting instructions corresponding to each generator in the power system; the reward function is used to represent the reward value caused by the machine-cutting instruction after the power system is cut off through the machine-cutting instruction in the action space.

[0064] Specifically, the server converts the training feature data into feature data corresponding to the state space and action space in the five-tuple, and determines the reward value corresponding to the feature data through the reward function, and continuously adjusts the parameters in the deep learning network through the reward value corresponding to each set of training feature data, thereby training the safety constraint model. When the parameters in the deep learning network or the reward value corresponding to the feature data meet the preset training end conditions, it is determined that the trained safety constraint model is obtained.

[0065] Step S206 , inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain a generator tripping stabilization control strategy corresponding to the target electrical quantity data.

[0066] The target electrical quantity data may be electrical quantity data corresponding to a time period formed by a current time point and a historical time point. For example, the target electrical quantity data may be electrical quantity data corresponding to each generator in the power system in the past hour.

[0067] Specifically, the server receives the real-time electrical quantity data of each generator in the power system in real time and stores the real-time electrical quantity data in chronological order. When the electrical quantity data stored in the server meets the time step corresponding to the target electrical quantity data, the target electrical quantity data is determined and input into the trained safety constraint model. The safety constraint model can analyze the target electrical quantity data and output the machine-cutting instruction with the highest reward value, thereby obtaining the machine-cutting stability control strategy.

[0068] In the above-mentioned method for determining the machine-cutting stabilization control strategy, the electrical quantity data of the power system in a preset time period is screened to obtain training feature data for training the safety constraint model. And the safety constraint model is trained by the training feature data to obtain a trained safety constraint model. The target electrical quantity data of the power system is input into the trained safety constraint model to obtain the machine-cutting stabilization control strategy corresponding to the target electrical quantity data. Among them, the safety constraint model is obtained by modeling the machine-cutting stabilization control strategy of the power system based on the Markov decision process, and the Markov decision process can be represented by a five-tuple consisting of a state space, an action space, a constraint condition, a reward function and a discount coefficient. Based on this, the electrical quantity data of each generator in the power system at different time points can be represented by the state space, the machine-cutting instructions that can be executed by each generator can be represented by the action space, and the reward value caused by the machine-cutting instruction in the action space can be represented by the reward function. Thus, the relationship between the electrical quantity data, the power system and the machine-cutting instruction is constrained by the five-tuple, and the safety constraint model is constructed by the five-tuple. By training the safety constraint model, the safety constraint model can output the optimal machine-cutting stabilization control strategy under different state spaces. After building a trained safety constraint model, you only need to input the target electrical quantity data of the power system to quickly obtain the generator cutting stabilization control strategy corresponding to the target electrical quantity data, thereby improving the efficiency of generating the stabilization control strategy.

[0069] In an exemplary embodiment, before the step of “training the security constraint model based on the training feature data corresponding to the power system to obtain a trained security constraint model”, the method further includes:

[0070] Based on the electrical quantity data of each generator in the power system at different time points, a state space is generated. Based on the machine-cutting instructions of each generator in the power system at different time points, an action space is generated. Based on the power angle stability reward corresponding to each generator in the power system and the machine-cutting quantity penalty of the generator after taking the machine-cutting instruction corresponding to the action space, a reward function is determined; and at least one constraint condition is determined.

[0071] The electrical quantity data include generator rotor speed, generator power angle, rotor acceleration, bus voltage and bus current. The generator cutting instruction is used to determine whether each generator performs a generator cutting operation. The constraints include the timing constraint of the generator cutting action, the selection constraint of the generator cutting unit and the generator cutting quantity constraint.

[0072] Specifically, the server needs to define the meaning of each element in the quintuple in advance. Based on this, the server can pre-create a state space and configure the data structure in the state space so that the data structure can receive electrical quantity data at different time points. The data structure can carry different types of electrical quantity data and electrical quantity data at different time points. In one example, the data structure can be a vector, and the generator power angle, generator rotor speed, rotor acceleration, bus voltage and bus current can be used as a group of vectors. And for different generators, the above vectors can be used as basic units at different time points to obtain multiple groups of vectors corresponding to different generators at different time points. The state space is used to store the above vectors.

[0073] The server can pre-create an action space. For different time points, the power system can store the machine-cutting instructions executed at the time point and use them as data in the action space. The server can pre-configure the data structure of the action space. The action space is used to store the machine-cutting instructions for each generator at different times. The instructions can be machine-cutting or machine-non-cutting.

[0074] The server can predefine the reward function and determine the reward function by configuring the power angle stability reward and the machine cutting penalty. The power angle stability reward is used to determine whether the generator is stable after the power system executes the machine cutting command, and obtain the corresponding reward value. The machine cutting penalty is used to determine the penalty value corresponding to each generator that executes the machine cutting command after the power system executes the machine cutting command. The server can obtain the reward value corresponding to the reward function by combining the reward value and the penalty value. Therefore, the server can obtain the reward function through the calculation strategy of the power angle stability reward and the calculation strategy of the machine cutting penalty.

[0075] The server may pre-configure constraints, which are used to constrain whether each generator in the power system can execute the power cutting instruction. For example, the constraints may include power cutting action timing constraints, power cutting unit selection constraints, and power cutting quantity constraints. Power cutting action timing constraints are used to constrain the power cutting instruction from the time dimension. Power cutting unit selection constraints are used to constrain the power cutting instruction from the dimension of the generator. Power cutting quantity constraints are used to constrain the power cutting quantity corresponding to each generator in the entire power system.

[0076] In this embodiment, by establishing the state space, action space, reward function and constraint conditions in the quintuple through the server, the basic machine-cutting strategy in the power system can be converted into a quintuple corresponding to the Markov decision process, which can increase the dynamic association between the electrical quantity data and the machine-cutting instruction, and facilitate the subsequent training of the safety constraint model through the quintuple.

[0077] In an exemplary embodiment, the specific implementation process of the step of "determining the reward function based on the power angle stability reward corresponding to each generator in the power system and the power-off amount penalty of the generator after taking the power-off instruction corresponding to the action space" includes:

[0078] Determine the maximum power angle difference between generators at the same time point, and determine the power angle stability reward corresponding to the maximum power angle difference based on the value of the maximum power angle difference; determine the product value of the state space and action space corresponding to each time point as the machine cutting penalty; determine the difference between the power angle stability reward and the machine cutting penalty as the reward function.

[0079] Specifically, for the same time point, the server traverses the power angle difference between the two generators in the power system and determines the maximum power angle difference therefrom. If the maximum power angle difference is less than a preset value, the power angle stability reward is determined to be a first reward value. If the maximum power angle difference is greater than or equal to the preset value, the power angle stability reward is determined to be a second reward value. In one example, the preset value may be 180 degrees. The first reward value may be 1, and the second reward value may be -1. The server may determine all power angle stability rewards at each time point.

[0080] The server can multiply the state space and action space corresponding to each time point to obtain the product value. Since the value corresponding to no machine cutting in the action space is 0, the product value of no machine cutting is 0. The product value can represent the machine cutting capacity and serve as the machine cutting penalty. Based on this, the server can perform a weighted subtraction of all power angle stability rewards and all machine cutting penalties, and use the weighted subtraction formula as the reward function.

[0081] In this embodiment, the power angle difference and the machine cutting amount are used as functions of the two control rewards, namely the power angle stability reward and the machine cutting amount penalty, and the two functions are subtracted to obtain the final reward function, which can improve the accuracy of the reward function in calculating the reward value.

[0082] In an exemplary embodiment, the specific implementation process of the step of “determining at least one constraint condition” includes:

[0083] Based on the time period corresponding to the electrical quantity data and the difference in action space corresponding to adjacent time points, the timing constraints of the machine cutting action are determined; based on the angular acceleration of the generator, the selection constraints of the machine cutting unit are determined; based on the load power of the power system, the machine cutting quantity constraints are determined.

[0084] Among them, the cutting action timing constraint is used to limit the value range of the action space at different time points; the cutting unit selection constraint is used to limit the generator selected for cutting; and the cutting quantity constraint is used to limit the number of generators that issue cutting instructions.

[0085] Specifically, the server can use the time period corresponding to the electrical quantity data as the range of executable machine cutting instructions. If the range corresponds to the time period, the server cannot generate the corresponding machine cutting instructions. For the same generator, the server can determine that there should be differences in the machine cutting instructions corresponding to the generator at different time points. And the value corresponding to the machine cutting instruction at the latter time point should be greater than or equal to the value corresponding to the machine cutting instruction at the previous time point. This is equivalent to that after the generator is cut off at the previous time point, the generator cannot be reversed at the next time point, that is, the generator cannot be turned on. The generator was not cut off at the previous time point, but it can be cut off at the next time point. Both of the above constraints can be used as timing constraints for the machine cutting action.

[0086] Based on the experience gained in engineering practice, the server can configure and select to start cutting off the generator set with the largest power angular acceleration, giving priority to cutting off the generator set with excess power. In addition, when the power angular acceleration of the generator set is less than a preset value, the generator set may not be cut off. The server can also limit the amount of generator sets to be cut off based on the load power of the power system. In one example, the combined power of all generators in the power system should be greater than or equal to the load power of the power system. When the combined power of all generators is less than the power compliance of the power system, the power system becomes unstable. Therefore, this situation needs to be avoided.

[0087] In this embodiment, by multiple constraints on the timing constraints of the machine cutting action, the selection constraints of the machine cutting group and the machine cutting amount constraints, it is possible to ensure from multiple aspects that the power system can still operate stably after executing the machine cutting instruction, thereby improving the stability of the power system.

[0088] In an exemplary embodiment, the specific implementation process of the step of “determining training characteristic data of the power system” includes:

[0089] Obtain electrical quantity data of the power system in a preset time period; determine the discrimination factor corresponding to the electrical quantity data based on the intra-class distance corresponding to each type of electrical quantity data; determine the grade difference between adjacent discrimination factors based on the discrimination factors corresponding to each electrical quantity data arranged in order of size, and determine the grade correlation coefficient corresponding to the adjacent discrimination factors based on the grade difference; if the grade correlation coefficient is greater than a preset value, determine the electrical quantity data corresponding to the grade correlation coefficient as training feature data.

[0090] Among them, the intra-class distance is determined based on the distance between each type of electrical quantity data at each time point; the discrimination factor is determined based on the intra-class distance corresponding to different time points. The discrimination factor is used to indicate whether there is a large difference between two time points. If there is a large difference, it means that the electrical quantity data corresponding to the two time points are quite different.

[0091] Specifically, the server can determine the discrimination factor corresponding to each electrical quantity data through a feature evaluation method. The server can divide the electrical quantity data for the electrical quantity data within a preset time period based on the data type of the electrical quantity data and multiple time points within the preset time period to obtain the divided electrical quantity data. For each data type of electrical quantity data, at the same time point, the server can traverse the electrical quantity data of two different samples, and for a set of sample data traversed, the distance value between the sample data can be calculated to obtain the total distance value of all the multiple sets of sample data. The server can obtain the total distance value corresponding to each time point and the total distance value corresponding to each data type in the preset time period through the above method, and determine the total distance value corresponding to each data type in the preset time period as the intra-class distance.

[0092] For multiple time points within a preset time period, the server can determine the intra-class distances corresponding to the first time point and the second time point, respectively, and sum the intra-class distances to obtain a distance sum value. Based on this, the server can divide the total distance value corresponding to the preset time period by the distance sum value to obtain the discrimination factor between the first time point and the second time point. Based on this, the server can obtain the discrimination factor between two different time points in the preset time period for each type of electrical quantity data.

[0093] The server can determine the correlation coefficient between the discrimination factors by the method of the Spearman correlation coefficient. Specifically, the server can sort all the discrimination factors within the preset time period in the order from small to large to obtain a sorting sequence. For each discrimination factor in the sorting sequence, the level corresponding to each discrimination factor is determined. Based on this, the server can determine the level difference between two different discrimination factors based on the level. The server can obtain the level correlation coefficient between the two different discrimination factors through the level difference and the amount of data corresponding to the discrimination factor.

[0094] In one example, the calculation formula of the rank correlation coefficient can be:

[0095]

[0096] Where is the rank difference of each pair of data points; n is the amount of data corresponding to the discrimination factor.

[0097] Based on this, if the rank correlation coefficient is greater than a preset value configured in advance, the electrical quantity data corresponding to the rank correlation coefficient is determined to be training feature data.

[0098] In this embodiment, by using feature evaluation and Spearman rank correlation coefficient, electrical quantity data with discrimination is screened from electrical quantity data within a preset time period as training feature data, which can improve the accuracy of obtaining the training feature data.

[0099] In an exemplary embodiment, the deep learning network includes an evaluation network and an action network; the specific implementation process of the step of "training the safety constraint model based on the training feature data corresponding to the power system to obtain the trained safety constraint model" includes:

[0100] Determine the initial state value of the training feature data in the state space, the initial action instruction in the action space, and the initial reward value corresponding to the reward function; based on the action network, the evaluation network and the discount coefficient, update the initial state value and the initial action instruction, and determine the predicted reward value and the real reward value corresponding to the initial reward value; the updating process satisfies the constraint conditions; update the first parameter of the action network and the second parameter of the evaluation network until the loss value between the predicted reward value and the real reward value meets the training end condition, and obtain a trained safety constraint model.

[0101] Specifically, the server inputs the training feature data into the state space to obtain the initial state value in the state space. The server inputs the machine cutting operation data corresponding to the training feature data into the action space to obtain the initial action instruction in the action space. The server calculates the initial state value and the initial action instruction based on the reward function for each time point in the preset time period to obtain the initial reward value corresponding to each time point.

[0102] The server can adjust the initial action instructions in the action space through the action network, and evaluate the adjusted initial action instructions through the evaluation network, reward function and discount coefficient to obtain the predicted reward value. At this time, the initial reward value can be regarded as the real reward value. The server can update the first parameter in the action network and the second parameter in the evaluation network by the loss value between the predicted reward value and the initial reward value. After the update, the training feature data at the next time point is traversed to re-determine the initial reward value and the predicted reward value. The server then repeats the above process of updating the first parameter and the second parameter until the loss value between the predicted reward value and the real reward value is less than the preset value, and a trained safety constraint model is obtained.

[0103] In this embodiment, by updating the first parameter and the second parameter through the action network and the evaluation network, convergence can be completed quickly to obtain a trained safety constraint model, which can improve the efficiency of training the safety constraint model.

[0104] In an exemplary embodiment, the specific implementation process of the step of "inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain the generator tripping stabilization control strategy corresponding to the target electrical quantity data" includes:

[0105] The target electrical quantity data of the power system is input into the trained safety constraint model to obtain the target state space corresponding to the target electrical quantity data; the action instructions corresponding to the target state space are determined as the machine tripping stabilization control strategy.

[0106] Specifically, the server can input the target electrical quantity data into the trained safety constraint model. The safety constraint model will determine the target state space corresponding to the target electrical quantity data through the deep learning network, and can also obtain the target action space corresponding to the target state space. The target action space contains the machine cutting instructions corresponding to the target electrical quantity data. The server can obtain the machine cutting instructions of each generator from the target action space, and merge the action instructions as a machine cutting stabilization control strategy. The server can send a machine cutting instruction to each generator in the power system according to the machine cutting stabilization control strategy, thereby completing the machine cutting operation.

[0107] In this embodiment, the target state space corresponding to the target electrical quantity data can be determined through the safety constraint model, and the corresponding target action space can be determined through the target state space, thereby obtaining the machine cutting instruction in the target action space, and the corresponding machine cutting stabilization control strategy can be quickly obtained through the safety constraint model.

[0108] As shown in FIG3 , the specific implementation process of the above-mentioned method for determining the machine cutting stabilization control strategy is described in detail below in combination with a specific embodiment, including the following steps:

[0109] Step S1, determining the power system emergency control tripping action strategy and the safety constraint model involved.

[0110] Specifically, an embodiment of the emergency control cutting action strategy of the power system is as follows. Taking a single-machine infinite system as an example, the transient process of the power grid is explained. The generator power angle speed equation of the system is:

[0111]

[0112] In the formula, is the generator rotor speed rad / s, , are the prime mover power and generator electromagnetic power respectively, is the generator power angle; M is the generator moment of inertia.

[0113] When an accident occurs, the system In the short term, The decrease in the power angle causes the generator rotor to accelerate, resulting in problems with the power angle stability. In order to analyze the effect of emergency control measures, the stabilization control measures of cutting the machine and shedding the load can be added to the model. Cutting the machine and shedding the load correspond to reducing the power angle. and increase , these two methods can reduce the system power difference, thereby enhancing the power angle stability margin. In order to ensure that the power system emergency control cutting action strategy is applicable to various environments, the embodiment of the present application adopts cutting measures in emergency control to avoid the limitations of too many specific conditions.

[0114] Step S2: the power system stability control security constraint model is converted into a Markov decision process.

[0115] The Markov decision process mainly includes: state space, action space, reward function, and safety constraints. In order to use reinforcement learning methods to solve the emergency control problem of transient stability switching of power grids, it is necessary to transform the safety constraint model of the power grid into a Markov model containing constraints. The Markov model consists of a five-tuple (S, a, P, R, ), namely, the model state space, action space, constraints, reward function and discount factor. In the power system stability control problem, the elements of the quintuple are defined as follows:

[0116] State space: In the stabilization server, the setting of the state vector needs to reflect the transient process characteristics of the power grid and the power angle characteristics of the generator rotor. The state vector is set to be a combination of multiple types of electrical quantities, which is specifically expressed as:

[0117]

[0118] In the formula, G is the number of generators, N is the number of buses; T is the training time period; I is the bus current; a is the rotor acceleration; V is the bus voltage; the superscript represents the distribution network node (generator) number, and the subscript represents the moment (time point) number, such as represents the power angle of generator i at time t.

[0119] Action space: The emergency control action taken in the embodiment of the present application is a machine cutting operation. The server can set the controllable object to be G generators in the power grid, and define the emergency control action as:

[0120]

[0121] In the formula, It represents the shutdown command of the i-th generator at time t, and its value is 1 or 0, indicating shutdown or no shutdown respectively.

[0122] Reward function: In order to objectively evaluate the machine cutting action, the embodiment of the present application divides the reward function into two parts: one is the power angle stability control effect, and the other is the rationality of the number of machine cutting units. As shown in the following formula:

[0123]

[0124] In the formula, It is the power angle stability reward; when the maximum power angle difference of the generator is less than 180°, the value is 1; when the power angle is unstable, the value is -1; Penalize the amount of power cuts to minimize the cost of grid stabilization and control actions, while reducing the adverse impact on the grid. and is the reward function coefficient.

[0125] Safety constraints (SL): In order to ensure the safety and reliability of actions in emergency control, the embodiment of the present application configures safety constraints and uses them as inviolable specifications for regulating the server. Safety constraints include machine switching action timing constraints, machine switching unit selection constraints, and machine switching quantity constraints.

[0126] Among them, the timing constraint of the machine tripping action can set the fault removal time to ensure that the machine tripping operation must be performed after the fault removal time, and once the unit completes the machine tripping, the subsequent grid connection operation will no longer be considered. The specific formula is:

[0127]

[0128]

[0129] Selection constraint for cutting off units: When a major disturbance occurs in the power grid, the load on the power grid is usually reduced due to the action of relay protection, resulting in excess electromagnetic power in the system. In engineering practice, it is usually chosen to start cutting off the unit with the largest power angular acceleration, that is, to give priority to cutting off the unit with the most serious power excess. Based on this, the embodiment of the present application formulates an action constraint: when the power angular acceleration of the unit is less than a certain threshold When , the machine cutting operation is not performed. The formula is as follows:

[0130]

[0131] Machine shedding constraint: After a fault occurs, if the machine shedding action is excessive, it may lead to another extreme situation, that is, the possibility of insufficient power supply increases. Therefore, when performing power angle stabilization control, it is necessary to consider whether the power on both ends of the supply and demand is balanced. The formula is as follows:

[0132]

[0133] In the formula, For the The power of the generator; is the system load; is the load power supply coefficient, and the embodiment of the present application can The value is 0.98.

[0134] Step S3, feature evaluation and Spearman rank correlation coefficient method are used to screen out feature data.

[0135] Specifically, for feature evaluation, feature evaluation mainly evaluates the distinguishability of various electrical quantity data under different states. The principle of the feature evaluation method for screening feature data is as follows:

[0136] Extract feature data to form the state vector , each vector state Assume that There is a class of electrical quantities Sample, use express The time Class electrical quantity data samples, the distance metric is , intra-class distance The calculation formula is as follows:

[0137]

[0138] In the formula, and Respectively represent Class I electrical quantity data and samples.

[0139] The server can calculate First The mean value of the electrical quantity data , the specific formula is:

[0140]

[0141] In another example, t1 and t2 are time nodes within time t, and the time from t1 to t2 is calculated. The average distance between classes of electrical quantity data in all state categories , the specific formula is:

[0142]

[0143] In the formula, , Respectively In time, In time The average distance between classes and the distance within classes can be used as parameters for determining the discrimination factor in the future.

[0144] Determine the discrimination factor for:

[0145]

[0146] In the formula, , For the Within time, In time Average intra-class distance (or average inter-class distance) of electrical quantity data of each class; The size of Electrical quantity data for the In time, The larger the ratio is, the stronger the distinguishability is.

[0147] The server can obtain the discrimination factor corresponding to each electrical quantity data based on the above method. Based on this, the server can further process it through the Spearman rank correlation coefficient method:

[0148] For data types with the same variation characteristics, the present application embodiment uses the Spearman rank correlation coefficient to eliminate them. The Spearman rank correlation coefficient is a method for measuring the degree of correlation of variables, and the specific steps are as follows.

[0149] Step S31, data conversion to levels: For each variable, The values ​​are arranged from smallest to largest and a rank is assigned to each value. If there are duplicate values, their ranks are averaged.

[0150] Step S32, calculate the rank difference: for each pair of data points, calculate the rank difference between the two variables. The rank difference is the difference between the corresponding rankings of the two variables.

[0151] Step S33, calculate the Spearman rank correlation coefficient: measure the relationship between two variables by calculating the rank correlation coefficient of the rank difference. The value range of this coefficient is between -1 and 1, where 1 represents a complete positive correlation: an increase in one variable is always accompanied by an increase in the other variable. -1 represents a complete negative correlation: an increase in one variable is always accompanied by a decrease in the other variable. 0 represents no correlation: there is no linear relationship between the two variables.

[0152] The calculation formula of Spearman's rank correlation coefficient is:

[0153]

[0154] In the formula, is the rank difference of each pair of data points; n is the number of data points.

[0155] Step S4, using the stability control strategy of the deep deterministic gradient algorithm to train the safety constraint model, and implement the emergency power system tripping strategy. As shown in Figure 4, the training process of the DDPG algorithm is included.

[0156] The server can set the objective function. The specific formulas include:

[0157]

[0158] In the formula, is the mathematical expectation, is the ratio of the updated strategy to the strategy before the update.

[0159] The second item is clipped by the function Constrain to The convergence of the policy update process is improved within the interval. Among them, the policy update range is limited to avoid large fluctuations in the policy network in each update. This pruning strategy is an indirect optimization strategy that avoids directly using the Q value (reward value) and makes the update process more stable. When , the strategy training is moving in a positive direction, , the strategy develops in a worse direction; As the objective function, we optimize the server parameter set to conduct.

[0160] Training objective function, including:

[0161] Step S41, the reinforcement learning model of the stabilization strategy makes decisions based on the power system operation data. Its state space is a continuous space composed of continuous variables such as voltage and current. The DDPG algorithm can solve the problem of continuous action space. Therefore, the embodiment of the present application can use the DDPG algorithm for model training. The action reward value of the final decision is used The function value is used to evaluate the superiority of the strategy, and the Bellman equation is used to solve the Markov decision process. In the objective function of the policy network, the accumulation of the reward function is also the optimization goal, so the reward function determines the optimization direction of the objective function.

[0162] According to the state vector extracted in step S3 , determine the reward function.

[0163]

[0164] In the formula, Status Medium Action Rewards received; is the discount factor, which is one of the factors of change of network parameters; Take a state for the entire training process Medium Action The reward obtained; the size of the discount factor affects the importance of the objective function on long-term returns.

[0165] Step S42: The original DDPG algorithm adds an experience pool playback mechanism. After each interaction with the environment, the training samples are played back. , in the sample for The reward for the iteration, for To prevent overfitting in the later stages of training, the present embodiment introduces regularization to update the data in the experience pool, that is, to update the experience value. Specific formula:

[0166]

[0167] In the formula, is the regularization coefficient; is the reward weight. This encourages the model to use smaller weights, thereby reducing the complexity of the model and preventing overfitting.

[0168] Step S43, update the parameters of the evaluation network by minimizing the loss value (mean square error loss). This step is used to evaluate the difference between the predicted Q value of the evaluation network and the target Q value. By minimizing this loss function, the evaluation network can continuously approach the true Q value. The loss function when the evaluation network is updated is:

[0169]

[0170] In the formula, is the objective function value; It is the parameter set of the evaluation network, which refers to the weights and biases that need to be optimized in the evaluation network.

[0171] S44, the action target network is used to provide the strategy for the next state. Combined with the Q value function of the evaluation training network, the policy gradient of the action during parameter update can be obtained:

[0172]

[0173] In the formula, is the current action network parameter for gradient; is the gradient of the current evaluation network action; For Action The Q-value gradient of the evaluation network provides the gradient information of the value to help the action network select a better action; To estimate the state Execute an action Cumulative rewards when The state of Next, the action network outputs ; is the parameter set of the action network.

[0174] S45. DDPG uses soft updates to ensure that action parameters and evaluation parameters can be updated slowly, thereby improving the stability of learning. The specific formula is:

[0175]

[0176] In an exemplary embodiment, the simulation test of the embodiment of the present application is carried out in the IEEE39 node system, where a 10MW wind turbine is added to the generator node position 3, and G32 and G35 are replaced with wind turbine models of equal capacity to simulate the operation scenario of the power grid with a high proportion of new energy penetration. The proportion of new energy power generation in the example scenario is 10%~20%. The line topology diagram is shown in Figure 5, and the parameters such as the power grid operation load level, unit output, fault location and fault type are shown in Table 1.

[0177] Table 1

[0178]

[0179] In order to verify the emergency control decision-making effect of the method proposed in the embodiment of the present application, comparisons with existing research are made from two aspects: feature perception and domain knowledge fusion, as shown in Table 2.

[0180] Table 2

[0181]

[0182] In Table 2, comparison method 1 is a feature perception control group, comparison method 2 is a domain knowledge control group, and comparison method 3 is a basic algorithm control group. The basic enhancement algorithms of the above methods are all developed on the DDPG algorithm.

[0183] (1) Performance testing: The maximum iteration curve is set to , and the reward function and loss reduction curve changing trends of different algorithms during the training phase are compared as shown in Figure 6.

[0184] As can be seen from Figure 6, the basic Q learning training efficiency and convergence of comparison method 3 have no advantages over other methods. In contrast, the reward function value increase and training efficiency of the embodiment of the present application are significantly better than other methods. The reward function reaches 28.0 at 750 rounds, and after about 1000 rounds of iteration, the reward function gradually converges to 29.2; the convergence efficiency of comparison method 1 is slower than that of the embodiment of the present application. After 1400 rounds of iteration, the reward function value converges to 24.1, and the convergence value is only 82% of the embodiment of the present application; the final convergence of comparison method 2 is similar to comparison method 1. From the analysis of convergence speed and convergence, it is obviously at a disadvantage compared with the embodiment of the present application, and it tends to converge after 2500 rounds of iteration.

[0185] (2) Case analysis. In order to verify the effectiveness of the method proposed in the embodiment of the present application after a fault, representative scenarios outside the training set are selected for testing and verification, and the rationality of its decision is analyzed. The initial fault is a three-phase short circuit grounding fault at bus 2, which lasts for 0.2s. The power angle and speed fluctuation diagram when the server is not cut off are shown in Figure 7.

[0186] As shown in Figure 7, the power angle and speed of generators 1, 2, and 3 fluctuate after the fault. If no action is taken, the power angle will be unstable along the dotted line path. According to the server proposed in the embodiment of the present application, the machine cutting action is quickly performed after the fault is removed. G1 executes the machine cutting strategy, and the unit cuts off a total of 75MW. G2 executes the machine cutting strategy, and the unit cuts off a total of 70MW. G5 executes the machine cutting strategy, and the unit cuts off a total of 105MW. After the server machine cutting oscillation, recovery is shown in Figure 8.

[0187] As shown in Figure 8, the system successfully returns to normal after a period of oscillation, the power angle and speed fluctuate stably, and no more fluctuations and instabilities occur, thus ensuring the safety of the system.

[0188] Through the embodiments of the present application, the switching action strategies and related safety constraints of the power system under emergency control are first sorted out, and the stability control model of the power system is converted into a Markov decision process. Secondly, the most representative feature data are screened out using feature evaluation and Spearman rank correlation coefficient methods. Then, in order to improve the training efficiency of the stability control strategy server, a training framework based on the deep deterministic gradient algorithm DDPG is proposed. Finally, the effectiveness of this method is verified in the IEEE-39 node system.

[0189] The beneficial effect of the embodiments of the present application is that it can automatically adjust and generate a machine cutting stabilization control strategy according to the operating status of the system and the response to faults, and show better performance in terms of decision-making effect and efficiency.

[0190] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0191] Based on the same inventive concept, the embodiment of the present application also provides a device for determining a cutting machine stabilization control strategy for implementing the above-mentioned method for determining a cutting machine stabilization control strategy. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more embodiments of the device for determining a cutting machine stabilization control strategy provided below can refer to the limitations of the method for determining a cutting machine stabilization control strategy above, and will not be repeated here.

[0192] In an exemplary embodiment, as shown in FIG9 , a device 900 for determining a machine cutting stabilization control strategy is provided, comprising: a data determination module 901 , a model training module 902 and a strategy determination module 903 , wherein:

[0193] The data determination module 901 is used to determine the training characteristic data of the power system; the training characteristic data is obtained by screening the electrical quantity data of the power system in a preset time period;

[0194] Model training module 902 is used to train the safety constraint model based on the training feature data corresponding to the power system to obtain a trained safety constraint model; the safety constraint model is obtained by modeling the power system's machine-cutting stability control strategy based on the Markov decision process; the safety constraint model includes a five-tuple and a deep learning network corresponding to the Markov decision process; the five-tuple includes a state space, an action space, a constraint condition, a reward function and a discount coefficient; wherein the state space is used to represent the electrical quantity data corresponding to multiple nodes corresponding to each generator in the power system at different time points; the action space is used to represent the machine-cutting instructions corresponding to each generator in the power system; the reward function is used to represent the reward value caused by the machine-cutting instruction after the power system is cut off through the machine-cutting instruction in the action space;

[0195] The strategy determination module 903 is used to input the target electrical quantity data of the power system into the trained safety constraint model to obtain the generator tripping stabilization control strategy corresponding to the target electrical quantity data.

[0196] Furthermore, the device also includes a quintuple determination module, which is specifically used to: generate a state space based on electrical quantity data of each generator in the power system at different time points; the electrical quantity data include generator rotor speed, generator power angle, rotor acceleration, bus voltage and bus current; generate an action space based on the machine cutting instructions of each generator in the power system at different time points; wherein the machine cutting instruction is used to determine whether the generator performs a machine cutting operation; determine a reward function based on the power angle stability reward corresponding to each generator in the power system and the machine cutting amount penalty of the generator after taking the machine cutting instruction corresponding to the action space; determine at least one constraint condition; the constraint conditions include machine cutting action timing constraints, machine cutting unit selection constraints and machine cutting amount constraints.

[0197] Furthermore, the quintuple determination module is specifically used to: determine the maximum power angle difference between generators at the same time point, and determine the power angle stability reward corresponding to the maximum power angle difference based on the value of the maximum power angle difference; determine the product value of the state space and the action space corresponding to each time point as the machine cutting amount penalty; determine the difference between the power angle stability reward and the machine cutting amount penalty as the reward function.

[0198] Furthermore, the quintuple determination module is specifically used to: determine the machine cutting action timing constraints based on the time period corresponding to the electrical quantity data and the difference between the action spaces corresponding to adjacent time points; the machine cutting action timing constraints are used to limit the value range of the action space at different time points; determine the machine cutting unit selection constraints based on the working angle acceleration of the generator; the machine cutting unit selection constraints are used to limit the generators selected for machine cutting; determine the machine cutting quantity constraints based on the load power of the power system; the machine cutting quantity constraints are used to limit the number of generators that issue machine cutting instructions.

[0199] Furthermore, the data determination module 901 is specifically used to: obtain the electrical quantity data of the power system in a preset time period; determine the discrimination factor corresponding to the electrical quantity data based on the intra-class distance corresponding to each type of electrical quantity data; the intra-class distance is determined based on the distance between each type of electrical quantity data at each time point; the discrimination factor is determined based on the corresponding intra-class distances at different time points; based on the discrimination factors corresponding to each electrical quantity data arranged in order of size, determine the rank difference between adjacent discrimination factors, and based on the rank difference, determine the rank correlation coefficient corresponding to the adjacent discrimination factors; if the rank correlation coefficient is greater than a preset value, determine that the electrical quantity data corresponding to the rank correlation coefficient is the training feature data.

[0200] Furthermore, the deep learning network includes an evaluation network and an action network, and the model training module 902 is specifically used to: determine the initial state value of the training feature data in the state space, the initial action instruction in the action space, and the initial reward value corresponding to the reward function; based on the action network, the evaluation network and the discount coefficient, the initial state value and the initial action instruction are updated, and the predicted reward value and the actual reward value corresponding to the initial reward value are determined; the updating process satisfies the constraint conditions; the first parameter of the action network and the second parameter of the evaluation network are updated until the loss value between the predicted reward value and the actual reward value meets the training end condition, thereby obtaining a trained safety constraint model.

[0201] Furthermore, the strategy determination module 903 is specifically used to: input the target electrical quantity data of the power system into the trained safety constraint model to obtain the target state space corresponding to the target electrical quantity data; determine the action instruction corresponding to the target state space as the machine cutting stabilization control strategy.

[0202] Each module in the above-mentioned device for determining the machine cutting stabilization control strategy can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0203] In an exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG10. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the electrical quantity data of each generator. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for determining a machine-cutting stabilization control strategy is implemented.

[0204] Those skilled in the art will understand that the structure shown in FIG. 10 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0205] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0206] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0207] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0208] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0209] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0210] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for determining a cutting machine stabilization control strategy, characterized in that: The method comprises: Determining training characteristic data of the power system; the training characteristic data is obtained by screening the electrical quantity data of the power system in a preset time period; Based on the training feature data corresponding to the power system, the safety constraint model is trained to obtain a trained safety constraint model; the safety constraint model is obtained by modeling the power system's machine switching stabilization control strategy based on the Markov decision process; the safety constraint model includes a quintuple and a deep learning network corresponding to the Markov decision process; the quintuple includes a state space, an action space, a constraint condition, a reward function, and a discount coefficient; Inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain a generator tripping stabilization control strategy corresponding to the target electrical quantity data; Among them, the state space is used to represent the electrical quantity data corresponding to multiple nodes corresponding to each generator in the power system at different time points; the action space is used to represent the power cutting instructions corresponding to each generator in the power system; and the reward function is used to represent the reward value caused by the power cutting instruction after the power system is cut off through the power cutting instruction in the action space.

2. The method according to claim 1, characterized in that Before training the safety constraint model based on the training feature data corresponding to the power system to obtain the trained safety constraint model, the method further includes: Generate a state space based on electrical quantity data of each generator in the power system at different time points; the electrical quantity data includes generator rotor speed, generator power angle, rotor acceleration, bus voltage and bus current; Based on the tripping instructions of each generator in the power system at different time points, an action space is generated; wherein the tripping instruction is used to determine whether the generator is to be tripped; Determining a reward function based on a power angle stability reward corresponding to each generator in the power system and a power-off amount penalty of the generator after taking a power-off instruction corresponding to an action space; At least one constraint condition is determined; the constraint condition includes a machine cutting action timing constraint, a machine cutting group selection constraint, and a machine cutting quantity constraint.

3. The method according to claim 2, characterized in that The determining of the reward function based on the power angle stability reward corresponding to each generator in the power system and the power-off amount penalty of the generator after taking the power-off instruction corresponding to the action space comprises: Determine the maximum power angle difference between the generators at the same time point, and determine the power angle stability reward corresponding to the maximum power angle difference according to the value of the maximum power angle difference; The product value of the state space and action space corresponding to each time point is determined as the machine cutting penalty; The difference between the power angle stability reward and the machine cutting amount penalty is determined as a reward function.

4. The method according to claim 2, characterized in that: The determining of at least one constraint condition comprises: Based on the time period corresponding to the electrical quantity data and based on the difference of the action space corresponding to adjacent time points, the timing constraint of the machine cutting action is determined; the timing constraint of the machine cutting action is used to limit the value range of the action space at different time points; Determining the cutting machine group selection constraint based on the working angular acceleration of the generator; the cutting machine group selection constraint is used to limit the generator selected by the cutting machine; The machine cutting amount constraint is determined based on the load power of the electric power system; the machine cutting amount constraint is used to limit the number of generators that issue machine cutting instructions.

5. The method according to claim 1, characterized in that The determining of training characteristic data of the power system comprises: Obtaining electrical quantity data of the power system in a preset time period; Based on the intra-class distances corresponding to the electrical quantity data of each type, determining the discrimination factor corresponding to the electrical quantity data; the intra-class distances are determined based on the distances between the electrical quantity data of each type at each time point; the discrimination factor is determined based on the intra-class distances corresponding to different time points; Based on the discrimination factors corresponding to the electrical quantity data arranged in order of magnitude, determining the level difference between adjacent discrimination factors, and based on the level difference, determining the level correlation coefficients corresponding to the adjacent discrimination factors; If the rank correlation coefficient is greater than a preset value, the electrical quantity data corresponding to the rank correlation coefficient is determined to be training feature data.

6. The method according to claim 1, characterized in that The deep learning network includes an evaluation network and an action network; the safety constraint model is trained based on the training feature data corresponding to the power system to obtain a trained safety constraint model, including: Determine an initial state value of the training feature data in the state space, an initial action instruction in the action space, and an initial reward value corresponding to the reward function; Based on the action network, the evaluation network and the discount coefficient, the initial state value and the initial action instruction are updated, and the predicted reward value and the real reward value corresponding to the initial reward value are determined; the updating process satisfies the constraint condition; The first parameter of the action network and the second parameter of the evaluation network are updated until the loss value between the predicted reward value and the actual reward value meets the training end condition, thereby obtaining a trained safety constraint model.

7. The method according to claim 6, characterized in that The step of inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain a generator tripping stabilization control strategy corresponding to the target electrical quantity data includes: Inputting the target electrical quantity data of the power system into the trained safety constraint model to obtain a target state space corresponding to the target electrical quantity data; Determine the action instruction corresponding to the target state space as the machine cutting stabilization control strategy.

8. A device for determining a cutting machine stabilization control strategy, characterized in that: The device comprises: A data determination module, used to determine training characteristic data of the power system; the training characteristic data is obtained by screening the electrical quantity data of the power system in a preset time period; A model training module is used to train the safety constraint model based on the training feature data corresponding to the power system to obtain a trained safety constraint model; the safety constraint model is obtained by modeling the power system's machine-cutting stabilization control strategy based on the Markov decision process; the safety constraint model includes a quintuple and a deep learning network corresponding to the Markov decision process; the quintuple includes a state space, an action space, a constraint condition, a reward function and a discount coefficient; wherein the state space is used to represent the electrical quantity data corresponding to multiple nodes corresponding to each generator in the power system at different time points; the action space is used to represent the machine-cutting instructions corresponding to each generator in the power system; the reward function is used to represent the reward value caused by the machine-cutting instruction after the power system is cut off through the machine-cutting instruction in the action space; The strategy determination module is used to input the target electrical quantity data of the power system into the trained safety constraint model to obtain the generator tripping stabilization control strategy corresponding to the target electrical quantity data.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images