A method and apparatus for transient frequency control of power systems based on boosting learning

Abstract

This invention discloses a method and apparatus for transient frequency control of power systems based on boosting learning. The method includes: obtaining a set of instructions for frequency control based on a preset combination of the input amount and timing of pre-selected active power resources; training a preset number of first models one by one based on historical data and the principle of boosting learning; after training, weighting and integrating the first models to obtain a second model; the first model includes a control effect calculation model; the historical data includes a preset number of frequency data before and after system faults and corresponding control instructions; periodically acquiring real-time frequency data of the system for a preset duration and combining it with the second model to determine whether instability is possible; if instability is possible, obtaining the minimum frequency value of each instruction within a preset future duration based on the real-time frequency data, the instruction set, and the second model; and selecting the instruction corresponding to the minimum frequency value closest to the preset value to control the system.

Description

Technical Field

[0001] This invention relates to the field of power system frequency assessment and control, and in particular to a method and apparatus for power system transient frequency control based on boosting learning. Background Technology

[0002] With the high penetration of renewable energy, traditional power systems are gradually evolving into modern power systems with significant uncertainty and limited predictability. As system inertia decreases significantly, power systems are highly susceptible to active power imbalances caused by large-scale faults. These imbalances can lead to rapid and severe frequency deviations and transient frequency instability across the entire system, ultimately resulting in large-scale power outages.

[0003] After a power system fault occurs, existing methods typically identify the fault type and collect information such as changes in system frequency before executing frequency control commands in conjunction with a control strategy table. This table records the frequency control commands to be taken when the frequency deviation or deviation time reaches the instability threshold under different fault types. However, control measures taken when the frequency reaches the instability threshold cannot completely prevent frequency instability accidents; their control effect is limited to shortening the occurrence time of frequency instability accidents and reducing the scope of their impact. Some existing transient frequency control methods can predict the constantly changing frequency trajectory of the faulted power system and dynamically adjust control commands in real time before the frequency reaches the instability standard, effectively reducing the probability of frequency instability accidents. However, real-time adjustment of control commands places strict demands on computational speed. To meet the requirements of rapid computation, existing predictive control methods are usually developed based on simplified power system physical models, and their reliability is fundamentally limited by the accuracy of the simplified model.

[0004] Therefore, a new technical solution is urgently needed to address the technical problem of how to regulate the transient frequency of the power system to avoid transient instability. Summary of the Invention

[0005] This invention provides a method and apparatus for controlling the transient frequency of a power system based on boosting learning, which solves the technical problem of how to control the transient frequency of a power system to avoid transient instability.

[0006] To achieve the above objectives, this invention provides a power system transient frequency control method based on boosting learning, comprising:

[0007] The instruction set for frequency regulation is obtained based on a preset combination of the input amount and timing of the pre-selected active power resources.

[0008] Based on historical data and the principle of enhancement learning, a preset number of first models are trained one by one; after training is completed, the first models are weighted and integrated to obtain the second model; the first model includes a control effect calculation model; the historical data includes a preset number of frequency data and corresponding control commands before and after system failure.

[0009] Periodically acquire real-time frequency data for a preset duration from the system and combine it with the second model to determine whether instability is possible;

[0010] If instability is possible, the minimum frequency of each instruction within a preset time period is obtained based on real-time frequency data, instruction set, and second model; the instruction corresponding to the minimum frequency closest to the preset value is selected to regulate the system.

[0011] Preferably, the instruction set for frequency regulation, obtained based on a preset combination of the amount and timing of the pre-selected active power resources, includes:

[0012] Pre-select from the power system before a power failure occurs. Active power resources include fast-response active power resources.

[0013] The input of each type of active power resource starts from 0%, and is divided into input steps with a preset percentage, until 100% input is achieved; the input timing is set after a fault. Within the time interval, with For step size, there are a total of The timing of power input is determined by pre-selecting and combining the input amount and timing of each type of active power resource as instructions, resulting in a set of instructions for frequency regulation. , , Indicates the total number of instructions. , Indicate the number of input types; and set Indicates in The instruction is to not allocate any active power resources at any time, and to let the system frequency evolve on its own.

[0014] Preferably, a predetermined number of first models are trained one by one based on historical data and the principle of boosting learning, including:

[0015] Acquire historical data and perform data preprocessing based on the historical data to obtain the training set;

[0016] A predetermined number of first models are constructed. Each first model comprises an input layer, a convolutional layer, a pooling layer, a fully connected layer, a regression layer, and an output layer connected in sequence. The input layer receives the input data matrix. The convolutional layer automatically extracts and learns hierarchical local features from the input data using a predetermined number of convolutional kernels to obtain a set of feature maps, where each convolutional kernel is trained to respond to specific features. The pooling layer downsamples the feature maps to obtain a set of flattened one-dimensional vectors. The fully connected layer and the regression layer calculate the minimum transient frequency of the power system within a predetermined time period after the execution of the control command based on the one-dimensional vectors.

[0017] Based on the training set and the principle of boosting learning, a predetermined number of first models are trained one by one.

[0018] Preferably, the training set is obtained by preprocessing historical data, including:

[0019] Assume the historical data includes g operational data points, specifically g frequency data points before and after system failures, along with corresponding control commands; each operational data point includes:

[0020] The lowest value of the system frequency during power system operation is obtained from the operating data. ;

[0021] The duration is obtained from the runtime data. Time after power system failure Inside Frequency data from each substation were integrated into... OK The column matrix is ​​then normalized to obtain the frequency data matrix. ,in ;

[0022] Based on operational data, control commands and the time from the occurrence of a fault to the execution of the control command are obtained. Resulting in 2 lines Column instruction data matrix The data in each column of the first row represents the active power injection value of the control command at the emergency response substation, and the values ​​in each column of the second row are... ;

[0023] frequency data matrix and instruction data matrix To integrate and obtain OK Column input data matrix ;by To validate the results, the data will and By performing pairing and combination, training data samples are obtained. ;

[0024] Obtain training data samples of g running data. , consisting of g training data samples Forming a training set .

[0025] Preferably, a predetermined number of first models are trained sequentially based on the training set and the principle of boosting learning, including:

[0026] Assume it exists h The first model, defining the first... h The first model is ;

[0027] Training the first model At that time, the training set The weight of each data sample is initialized to 1. ,in i This represents the index of the data sample. There are a total of g data samples. Sample weight set According to the sample weight set and training set right Perform training and calculate the normalized loss for each sample. :

[0028] ;

[0029] The normalized loss of each sample is integrated to obtain Normalized loss set ;

[0030] according to calculate Model error and network weight :

[0031] ;

[0032] according to and Update training set The weights of the data samples are used as the basis for training the second first model. Sample weight set , The weights of each data sample include:

[0033] ;

[0034] Based on the sample weight set and training set right Perform training and calculate the normalized loss for each sample. Model error and network weight and repeatedly update the training set. The weights of the data samples, until completion. h Training of the first model.

[0035] Preferably, after training, the second model is obtained by weighted integration of the first models, including:

[0036] After training is complete, the h first models are weighted and integrated to obtain the second model. :

[0037] ;

[0038] in, Indicates the first h The network weights of the first model.

[0039] Preferably, periodically acquiring real-time frequency data for a preset system duration and combining it with the second model to determine whether instability is possible includes:

[0040] The system's frequency status is monitored in real time, and the current system frequency status is periodically obtained. Within the system during the time period The frequency data of each substation is used to obtain real-time frequency data.

[0041] Integrate real-time frequency data into OK The column matrix is ​​then normalized to obtain the frequency data matrix. P ;Will P and After integration, the data is input into the second model, which outputs the minimum value of the first frequency within a preset future time period. According to the minimum value of the first frequency Determine if the system is likely to become unstable:

[0042] when Greater than the preset frequency stability threshold At that time, it was determined that the system had no risk of instability;

[0043] when Less than or equal to the preset frequency stability threshold At that time, it was determined that the system was at risk of instability.

[0044] Preferably, the minimum frequency of each instruction within a preset future time period is obtained based on real-time frequency data, the instruction set, and the second model, including:

[0045] Will PWith instruction set To integrate and obtain k One input data; combined with the second model k Each input data is processed to obtain an execution instruction set. The minimum frequency values ​​within a preset time period after each instruction are integrated to obtain a set of minimum frequency values. .

[0046] Preferably, selecting the instruction corresponding to the minimum frequency value closest to the preset value to regulate the system includes:

[0047] Set preset value :

[0048] like This indicates that the system frequency will fluctuate significantly. In this case, the primary objective of the control command is to reduce the system frequency fluctuation. Set to a preset value closer to B within the range of A to B;

[0049] like At this point, the primary objective of the control commands is to prevent frequency instability. Set to a preset value close to A within the range of A to B; where A and B are both preset values;

[0050] Known preset value In the case of the minimum frequency set Select the value closest to the preset value. The instruction corresponding to the minimum frequency value controls the system.

[0051] The present invention also provides a power system transient frequency control device based on boosting learning, which is used in the method of the present invention. The device includes a first module, a second module, a third module and a fourth module.

[0052] The first module is used to obtain a set of instructions for frequency regulation based on a preset combination of the input amount and timing of pre-selected active power resources.

[0053] The second module is used to train a preset number of first models one by one based on historical data and the principle of boosting learning. After training is completed, the first models are weighted and integrated to obtain the second model. The first model includes a control effect calculation model. The historical data includes a preset number of frequency data before and after system failure and corresponding control commands.

[0054] The third module is used to periodically acquire real-time frequency data for a preset duration of the system and combine it with the second model to determine whether instability is possible.

[0055] The fourth module is used to obtain the minimum frequency of each instruction within a preset time period in the event of potential instability, based on real-time frequency data, instruction set, and the second model; and to select the instruction corresponding to the minimum frequency closest to the preset value to regulate the system.

[0056] The present invention has the following beneficial effects:

[0057] This invention presents a power system transient frequency control method based on boosting learning. Taking a data-driven approach, it eliminates reliance on physical models. After a power system fault, it directly calculates the minimum future transient frequency under pre-set control commands using a data-driven model, rather than judging the control effect of corresponding commands based on frequency trajectories. This substitution of the computational object significantly reduces computational load. Multiple first models are trained using boosting learning based on historical data, allowing each model to focus on different feature points. By weighted integration of all first models, a second model is obtained. This second model can deeply explore the coupling relationship between power system frequency evolution and various frequency control commands, enabling more accurate calculations. This provides a solid theoretical foundation for high accuracy of the calculation results and the effectiveness of control commands based on the calculation results, thereby achieving highly reliable and adaptive control commands for different power system operating environments. By executing the control command whose transient minimum value is closest to the expected control target, the system frequency is controlled to near the target value before a power system frequency instability accident occurs, ensuring the safe and stable operation of the power system.

[0058] The power system transient frequency control device based on boosting learning of the present invention, used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0059] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0060] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0061] Figure 1 This is a schematic diagram of the method flow of a preferred embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the first model of a preferred embodiment of the present invention. Detailed Implementation

[0063] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0064] See Figure 1 In a preferred embodiment of the present invention, a method for transient frequency control of a power system based on boosting learning is provided, comprising:

[0065] S1. Obtain a set of instructions for frequency regulation based on a preset combination of the input amount and timing of the pre-selected active power resources.

[0066] S1 specifically includes:

[0067] Pre-select from the power system before a power failure occurs. Active power resources include fast-response active power resources, specifically including but not limited to DC power support, electrochemical energy storage power stations, and backup generators.

[0068] The input of each type of active power resource starts from 0%, and is divided into input steps with a preset percentage, until 100% input is achieved; the input timing is set after a fault. Within the time interval, with For step size, there are a total of The timing of power input is determined by pre-selecting and combining the input amount and timing of each type of active power resource as instructions, resulting in a set of instructions for frequency regulation. , , Indicates the total number of instructions. , Indicate the number of input types; and set Indicates in The instruction is to not allocate any active power resources at any time, and to let the system frequency evolve on its own.

[0069] In a preferred embodiment of the present invention, the input amount of each type of active power resource starts from 0%, and is divided into increments of 10%, until 100% is input, resulting in 10 input amounts, i.e. a =10. Considering the time consumed by data collection, instruction decision-making, and instruction execution, totaling 1 second, we set... =1.0s, =1.5s, =0.02s, =25. Three types of active power resources are pre-selected from the power system: DC power support, electrochemical energy storage power stations, and standby generators. x =3; therefore, we get k =750.

[0070] S2. Based on historical data and the principle of boosting learning, train a preset number of first models one by one; after training, weight and integrate the first models to obtain the second model.

[0071] The first model includes a control effect calculation model; historical data includes frequency data and corresponding control commands before and after a system failure, based on a preset number of bars.

[0072] In S2, a predetermined number of first models are trained one by one based on historical data and the principle of boosting learning. Specifically, this includes:

[0073] Acquire historical data and perform data preprocessing based on the historical data to obtain the training set, specifically including:

[0074] Assuming the historical data includes g operational data points, i.e., g frequency data before and after a system failure and corresponding control commands; in a preferred embodiment of this invention, g = 4392; each operational data point includes:

[0075] The lowest value of the system frequency during power system operation is obtained from the operating data. .

[0076] The duration is obtained from the runtime data. Time after power system failure Inside Frequency data from substations equipped with detection devices are integrated into... OK The column matrix is ​​then normalized to obtain the frequency data matrix. ,in In a preferred embodiment of the present invention, =0.001s, =300, =15, =0.3s.

[0077] Based on operational data, control commands and the time from the occurrence of a fault to the execution of the control command are obtained. Resulting in 2 lines Column instruction data matrix The data in each column of the first row represents the active power injection value of the control command at the emergency response substation, and the values ​​in each column of the second row are... .

[0078] frequency data matrix and instruction data matrix To integrate and obtain OK Column input data matrix ;by To validate the results, the data will and By performing pairing and combination, training data samples are obtained. .

[0079] Obtain training data samples of g running data. , consisting of g training data samples Forming a training set .

[0080] In a preferred embodiment of the present invention, historical data can be replaced with simulation data.

[0081] A predetermined number of first-order models were built using Python 3.10 and the PyTorch framework; see [link / reference]. Figure 2 The first model comprises an input layer, a convolutional layer, a pooling layer, a fully connected layer, a regression layer, and an output layer connected in sequence. The input layer receives the input data matrix. The convolutional layer automatically extracts and learns hierarchical local features from the input data using a preset number of convolutional kernels to obtain a set of feature maps, where each convolutional kernel is trained to respond to specific features. The pooling layer downsamples the feature maps to reduce their dimensionality while retaining the most salient information, resulting in a set of flattened one-dimensional vectors. The fully connected layer and the regression layer calculate the minimum transient frequency of the power system within a preset time period after the execution of the control command based on the one-dimensional vector.

[0082] In a preferred embodiment of the present invention, the first model can predict the minimum system frequency under a preset control command intervention, and then select the optimal control command from a series of control commands based on the prediction result; rather than simply predicting the minimum system frequency and using the minimum system frequency information to guide the construction of the control command. This model can achieve an efficient combination of prediction and control steps, effectively ensuring the effectiveness of the selected optimal control command.

[0083] Based on the training set and the principle of boosting learning, train a predetermined number of first models one by one:

[0084] Assume it exists h The first model, defining the first... h The first model is .

[0085] Training the first model At that time, the training set The weight of each data sample is initialized to 1. ,in i This represents the index of the data sample. There are a total of g data samples. Sample weight set According to the sample weight set and training set right Perform training and calculate the normalized loss for each sample. :

[0086] ;

[0087] The normalized loss of each sample is integrated to obtain Normalized loss set ;

[0088] according to calculate Model error and network weight :

[0089] ;

[0090] according to and Update training set The weights of the data samples are used as the basis for training the second first model. Sample weight set , The weights of each data sample include:

[0091] ;

[0092] In a preferred embodiment of the present invention, the normalization loss corresponding to the sample with poor prediction performance is... Large, close to 1 Approaching 0, resulting in The closer the value is to 1, the more unchanged the updated sample weights remain; the normalization loss corresponds to the sample with the best prediction performance. Small, Close to 1 close to Typically, a well-trained model has a high prediction error. Less than 0.5, at this time If the value is less than 1, the updated weight will be smaller. This update method enables the learning of different features with emphasis.

[0093] Based on the sample weight set and training set right Perform training and calculate the normalized loss for each sample. Model error and network weight and repeatedly update the training set. The weights of the data samples, until completion. h Training of the first model.

[0094] In a preferred embodiment of the present invention, increasing the number of first models increases the computational accuracy of the second model, but also increases model complexity and decision-making time. Therefore, considering both timeliness and model performance, the second model is obtained by weighted aggregation of eight first models. During the training process of each first model, the Optuna algorithm is used for parameter optimization to obtain the second model with optimal performance.

[0095] In a preferred embodiment of the present invention, by training each first model independently and updating the weights of the samples before training, each independent first model can focus on different features of the training dataset during the training process.

[0096] In S2, after training is completed, the weighted integration of the first models yields the second model, which specifically includes:

[0097] After training is complete, the h first models are weighted and integrated to obtain the second model. :

[0098] ;

[0099] in, Indicates the first h The network weights of the first model.

[0100] In a preferred embodiment of the present invention, by weighting and integrating all the first models, a strong data-driven model with better overall performance, namely the second model, is obtained.

[0101] S3. Periodically obtain real-time frequency data for a preset duration from the system and combine it with the second model to determine whether instability is possible.

[0102] S3 specifically includes:

[0103] The system's frequency status is monitored in real time, and the current system frequency status is periodically obtained. Within the system during the time period The frequency data of each substation is used to obtain real-time frequency data.

[0104] Integrate real-time frequency data into OK The column matrix is ​​then normalized to obtain the frequency data matrix. P ;Will P and After integration, the data is input into the second model, which outputs the minimum value of the first frequency within a preset future time period. According to the minimum value of the first frequency Determine if the system is likely to become unstable: each time The calculation interval is 0.05 seconds.

[0105] when Greater than the preset frequency stability threshold At that time, it was determined that the system had no risk of instability;

[0106] when Less than or equal to the preset frequency stability threshold At that time, it was determined that the system was at risk of instability.

[0107] In a preferred embodiment of the present invention, according to the guidelines for power system safety and stability, the steady-state frequency of the power system shall not be lower than 49.9 Hz; therefore, a preset frequency stability threshold is established. =49.9Hz.

[0108] S4. If instability is possible, the minimum frequency of each instruction within a preset time period is obtained based on real-time frequency data, instruction set, and second model; the instruction corresponding to the minimum frequency closest to the preset value is selected to regulate the system.

[0109] In S4, the minimum frequency value of each instruction within a preset future time period is obtained based on real-time frequency data, instruction set, and the second model, specifically including:

[0110] Will P With instruction set To integrate and obtain k One input data; combined with the second model k Each input data is processed to obtain an execution instruction set. The minimum frequency values ​​within a preset time period after each instruction are integrated to obtain a set of minimum frequency values. . This is the minimum frequency of the system within a preset time period after executing each instruction.

[0111] In S4, selecting the instruction corresponding to the minimum frequency value closest to the preset value to regulate the system specifically includes:

[0112] Set preset value :

[0113] like This indicates that the system frequency will fluctuate significantly. In this case, the primary objective of the control command is to reduce the system frequency fluctuation. Set to a preset value closer to B within the range of A to B;

[0114] like At this point, the primary objective of the control commands is to prevent frequency instability. Set to a preset value close to A within the range of A to B; where A and B are both preset values.

[0115] In a preferred embodiment of the present invention, A=49.5Hz and B=49.9Hz.

[0116] Known preset value In the case of the minimum frequency set Select the value closest to the preset value. The instruction corresponding to the minimum frequency value is used to regulate the system, thereby effectively suppressing the frequency decline trend and controlling the system frequency within a certain range. Nearby, to curb system frequency instability or reduce system frequency fluctuations.

[0117] This invention presents a power system transient frequency control method based on boosting learning. Taking a data-driven approach, it eliminates reliance on physical models. After a power system fault, it directly calculates the minimum future transient frequency under pre-set control commands using a data-driven model, rather than judging the control effect of corresponding commands based on frequency trajectories. This substitution of the computational object significantly reduces computational load. Multiple first models are trained using boosting learning based on historical data, allowing each model to focus on different feature points. By weighted integration of all first models, a second model is obtained. This second model can deeply explore the coupling relationship between power system frequency evolution and various frequency control commands, enabling more accurate calculations. This provides a solid theoretical foundation for high accuracy of the calculation results and the effectiveness of control commands based on the calculation results, thereby achieving highly reliable and adaptive control commands for different power system operating environments. By executing the control command whose transient minimum value is closest to the expected control target, the system frequency is controlled to near the target value before a power system frequency instability accident occurs, ensuring the safe and stable operation of the power system.

[0118] Compared to existing transient frequency control methods, the method of this invention has the following characteristics: it can directly utilize actual operating data or time-domain simulation data of the power system, effectively reducing the dependence on precise mathematical models of the power system; it uses a data-driven model to accurately calculate the future evolution trend of the power system, perceive risks in advance before frequency instability accidents occur, and execute control commands in advance; at the same time, its "offline training, online application" feature ensures the speed of online decision-making; it adopts boosting learning, focusing on learning different features of the input data through multiple independent models, and the strong data-driven model obtained through weighted aggregation can make full use of training data information, deeply understand the underlying operating logic of the power system, and improve the adaptability of the model in different power system operating environments; it can flexibly set frequency control targets according to the severity of system faults to better meet the different control needs of the power system.

[0119] In a preferred embodiment of the present invention, a power system transient frequency control device based on boosting learning is also provided for use with the method of the present invention. The device includes a first module, a second module, a third module, and a fourth module.

[0120] The first module is used to obtain a set of instructions for frequency regulation based on a preset combination of the input amount and timing of pre-selected active power resources.

[0121] The second module is used to train a preset number of first models one by one based on historical data and the principle of boosting learning. After training is completed, the first models are weighted and integrated to obtain the second model. The first model includes a control effect calculation model. The historical data includes a preset number of frequency data before and after system failure and corresponding control commands.

[0122] The third module is used to periodically acquire real-time frequency data for a preset duration of the system and combine it with the second model to determine whether instability is possible.

[0123] The fourth module is used to obtain the minimum frequency of each instruction within a preset time period in the event of potential instability, based on real-time frequency data, instruction set, and the second model; and to select the instruction corresponding to the minimum frequency closest to the preset value to regulate the system.

[0124] The power system transient frequency control device based on boosting learning of the present invention, used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for transient frequency control of power systems based on boosting learning, characterized in that, include: The instruction set for frequency regulation is obtained based on a preset combination of the amount and timing of the selected active power resources. Pre-selected from the power system before a power failure occurs. The active power resources include fast-response active power resources; the input of each active power resource starts from 0% and is divided into input steps with a preset percentage until 100% is input. Set the deployment timing after the fault. Within the time interval, with For step size, there are a total of The timing of power input is determined by a preset combination of input amounts and timings for each type of active power resource, resulting in the instruction set used for frequency regulation. , , Indicates the total number of instructions. , Indicate the number of input types; and set Indicates in The instruction to not allocate any active power resources at any time, allowing the system frequency to evolve automatically; Based on historical data and the principle of boosting learning, a predetermined number of first models are trained one by one; after training is completed, the first models are weighted and integrated to obtain a second model; the first model includes a control effect calculation model; the historical data includes a predetermined number of frequency data and corresponding control commands before and after system failure; Periodically acquire real-time frequency data for a preset duration from the system and combine it with the second model to determine whether instability is possible; If instability is possible, the minimum frequency of each instruction within a preset time period is obtained based on the real-time frequency data, the instruction set, and the second model; the instruction corresponding to the minimum frequency closest to the preset value is selected to regulate the system.

2. The power system transient frequency control method based on boosting learning according to claim 1, characterized in that, The step of training a predetermined number of first models one by one based on historical data and the principle of boosting learning includes: The historical data is acquired, and data preprocessing is performed based on the historical data to obtain the training set; A predetermined number of first models are constructed. Each first model comprises an input layer, a convolutional layer, a pooling layer, a fully connected layer, a regression layer, and an output layer connected sequentially. The input layer receives an input data matrix. The convolutional layer automatically extracts and learns hierarchical local features from the input data using a predetermined number of convolutional kernels to obtain a set of feature maps, where each convolutional kernel is trained to respond to specific features. The pooling layer downsamples the feature maps to obtain a set of flattened one-dimensional vectors. The fully connected layer and the regression layer calculate the minimum transient frequency of the power system within a predetermined time period after the execution of the control command based on the one-dimensional vectors. Based on the training set, a predetermined number of first models are trained one by one according to the principle of boosting learning.

3. The power system transient frequency control method based on boosting learning according to claim 2, characterized in that, Based on the historical data, data preprocessing is performed to obtain the training set, which includes: Assuming the historical data includes The running data includes The frequency data and corresponding control commands before and after the system failure are recorded; each piece of operational data includes: The lowest value of the system frequency during the operation of the power system is obtained based on the aforementioned operating data. ; The duration is obtained based on the aforementioned operational data. Time after power system failure Inside Frequency data from each substation, and integrated into OK The column matrix is ​​then normalized to obtain the frequency data matrix. ,in ; Based on the operational data, control commands and the time from the occurrence of a fault to the execution of the control commands are obtained. Resulting in 2 lines Column instruction data matrix The data in each column of the first row represents the active power injection value of the control command at the emergency response substation, and the values ​​in each column of the second row are... ; The frequency data matrix and instruction data matrix To integrate and obtain OK The input data matrix of the column ;by To validate the results, the data will and By performing pairing and combination, training data samples are obtained. ; Get Training data samples of the running data ,Depend on training data samples Forming a training set .

4. The power system transient frequency control method based on boosting learning according to claim 3, characterized in that, Training the predetermined number of first models one by one based on the training set and the principle of boosting learning includes: Assume it exists The first model is defined as the first... The first model is ; Training the first model At that time, the training set The weight of each data sample is initialized to 1. ,in Indicates the sequence number of the data sample, total Data samples were obtained. Sample weight set According to the sample weight set and the training set right Perform training and calculate the normalized loss for each sample. : ； The normalized loss of each sample is integrated to obtain Normalized loss set ; according to calculate Model error and network weight : ； according to and Update the training set The weights of the data samples are used as the basis for training the second first model. Sample weight set , The weights of each data sample include: ； According to the sample weight set and the training set right Perform training and calculate the normalized loss for each sample. Model error and network weight And repeatedly update the training set. The weights of the data samples, until completion. The training of the first model.

5. The power system transient frequency control method based on boosting learning according to claim 4, characterized in that, The process of weightedly integrating the first models after training to obtain the second model includes: After training is complete, the h first models are weighted and integrated to obtain the second model. : ； in, Indicates the first The network weights of the first model.

6. The power system transient frequency control method based on boosting learning according to claim 5, characterized in that, The step of periodically acquiring real-time frequency data of the system for a preset duration and combining it with the second model to determine whether instability is possible includes: The frequency status of the system is monitored in real time, and the current state of the system is periodically obtained. Within the system during the time period The real-time frequency data is obtained by analyzing the frequency data of each substation. The real-time frequency data is integrated into OK The column matrix is ​​then normalized to obtain the frequency data matrix. ;Will and After integration, the data is input into the second model, which outputs the minimum value of the first frequency within a preset future time period. According to the minimum value of the first frequency Determine whether the system is likely to become unstable: when Greater than the preset frequency stability threshold At that time, it was determined that the system had no risk of instability; when Less than or equal to the preset frequency stability threshold At that time, it is determined that the system is at risk of instability.

7. The power system transient frequency control method based on boosting learning according to claim 6, characterized in that, Based on the real-time frequency data, the instruction set, and the second model, the minimum frequency value of each instruction within a preset future time period is obtained, including: Will With the instruction set To integrate and obtain Input data; based on the second model and combined with the... The input data is processed to obtain the instruction set to be executed. The minimum frequency values ​​within a preset time period after each instruction are integrated to obtain a set of minimum frequency values. .

8. The power system transient frequency control method based on boosting learning according to claim 7, characterized in that, The step of selecting the instruction corresponding to the minimum frequency value closest to the preset value to regulate the system includes: Set the preset value : like This indicates that the system frequency will fluctuate significantly. In this case, the primary objective of the control command is to reduce the system frequency fluctuation. Set to a preset value closer to B within the range of A to B; like At this point, the primary objective of the control commands is to prevent frequency instability. Set to a preset value close to A within the range of A to B; where A and B are both preset values; Given the preset value In the case of the set of minimum frequencies Select the value closest to the preset value. The instruction corresponding to the minimum frequency value controls the system.

9. A power system transient frequency control device based on boosting learning, used in the method described in any one of claims 1 to 8, characterized in that, The device includes a first module, a second module, a third module, and a fourth module; The first module is used to obtain a set of instructions for frequency regulation based on a preset combination of the input amount and timing of the pre-selected active power resources. The second module is used to train a preset number of first models one by one based on historical data and the principle of boosting learning; after training is completed, the first models are weighted and integrated to obtain the second model; the first model includes a control effect calculation model; the historical data includes a preset number of frequency data before and after system failure and corresponding control commands; The third module is used to periodically acquire real-time frequency data of the system for a preset duration and combine it with the second model to determine whether instability is possible. The fourth module is used to obtain the minimum frequency of each instruction within a preset time period in the event of potential instability, based on the real-time frequency data, the instruction set, and the second model; and to select the instruction corresponding to the minimum frequency value closest to the preset value to regulate the system.

Images