A power system frequency stability control method based on a double-channel deep neural network

Abstract

The application discloses a power system frequency stability control method based on a double-channel deep neural network, which comprises the following steps: determining a possible fault setting set in a receiving end system; carrying out time domain simulation on the receiving end system; in the allowed range of normal operation of the receiving end system, combining various possible frequency stability control measures in the receiving end system to form a series of feasible control action sets; in actual operation of the power system, if a fault occurs to cause active power shortage, characteristic quantity data of key nodes are collected in real time, each action quantity data in the control action set is combined with the characteristic quantity data to be sent into a pre-trained regression neural network to predict the frequency of the receiving end system, and reliable control actions are screened out according to the frequency as a standard; all reliable control actions are subjected to secondary screening to obtain optimal frequency stability control actions under the current fault condition of the receiving end system. The frequency stability control actions output by the application can effectively restore the frequency of the receiving end system.

Description

Technical Field

[0001] The present invention relates to the technical field of power system stability assessment and control, and in particular to a power system frequency stability control method based on a dual-channel deep neural network. Background Art

[0002] Frequency is a key indicator of system active power balance. When the active power generated by each generating device in the system exceeds the active power consumed by the load, the system frequency rises; when the active power generated by each generating device in the system is less than the active power consumed by the load, the system frequency drops. Frequency issues encountered in receiving-end systems with high penetration of renewable energy in UHV AC and UHV DC grids are often of the latter type. To address frequency drops, existing stability control measures for improving system frequency can be categorized into two categories: "increase power" and "throttle power." Increase power increases the system's active power output by increasing generator output, activating backup generators, and deploying energy storage plants. These measures are fast and economical, but the active power they provide is small and often cannot fully compensate for power shortfalls in the event of a fault. Throttling achieves system active power balance by selectively removing non-essential loads with low power supply priority. These measures can quickly remove a large number of loads, quickly restoring system frequency stability, but the economic cost of interrupting load power is high, so these measures are generally given the lowest priority among various control measures. Actual frequency stability control plans typically incorporate both increase power and throttling measures. In recent years, deep learning, as an emerging machine learning method, has been introduced into the development of power system frequency stability control schemes. By deeply mining and learning the mapping relationship between the system frequency response characteristics and control actions after a receiving system encounters a fault, frequency stability control actions are adaptively derived. However, the operating conditions of actual receiving systems are complex and diverse, and the training data required for deep learning often cannot cover all possible fault conditions. Limited by the constraints of training data, the generalization performance of existing deep learning methods needs to be improved. When the system encounters a new fault not included in the training data, the frequency stability control actions output by such methods may not be able to effectively restore the system frequency. Summary of the Invention

[0003] In view of this, the present invention provides a power system frequency stability control method based on a dual-channel deep neural network. It comprehensively considers all feasible frequency stability control measures in the receiving system, pre-sets all available frequency stability control actions in the receiving system, and after a fault occurs in the receiving system, performs multiple rounds of screening on these actions to obtain the optimal frequency stability control action. Regardless of any type of fault that occurs in the system, the optimal frequency stability control action that takes into account both reliability and economy can always control the system frequency to a safe range and ensure reliable power supply to the system. Compared with traditional methods, this method does not directly output frequency stability control actions, but instead evaluates the frequency control effects of a series of pre-set frequency stability control actions, and screens out the most appropriate frequency stability control action based on the performance. Regardless of whether the fault encountered by the system has appeared in the training data, the most appropriate frequency stability control action among all preset frequency control actions can be given.

[0004] The present invention discloses a method for controlling power system frequency stability based on a dual-channel deep neural network, which includes:

[0005] Step 1: Determine the possible fault setting set in the receiving system under the UHV AC and UHV DC feeding grid structure;

[0006] Step 2: Build a receiving system framework in the transient time-domain simulation software to perform time-domain simulation of the receiving system under fault conditions, and screen out the fault settings that cause frequency instability in the receiving system. Integrate the available frequency control measures in the power system, apply frequency stability control actions to the frequency instability samples, and conduct a second round of simulations. Continuously optimize the frequency control actions until the system simulation frequency reaches a safe and stable range under all fault conditions in the fault setting set. During this process, monitor and record the key node feature quantities, frequency stability control actions, and system frequency data in the time-domain simulation to obtain a model training dataset.

[0007] Step 3: Use the long short-term memory recurrent neural network and convolutional neural network to build a dual-channel deep neural network frequency regression prediction model. Use the data in the training dataset to optimize and train the dual-channel deep neural network frequency regression prediction model. The trained prediction model takes node feature quantities and frequency stability control action data as input and outputs system frequency data.

[0008] Step 4: Debug the model hyperparameters to obtain a dual-channel deep neural network frequency regression prediction model with minimal error and the fastest convergence speed. The trained model can predict the system frequency value under control action intervention and determine the reliability of the frequency stability control action. At the same time, within the allowable range of normal operation of the receiving system, a series of feasible control actions are formed by combining various possible frequency stability control measures in the system.

[0009] Step 5: In the actual operation of the power system, if a fault occurs to cause a lack of active power, the characteristic data of the key nodes are collected in real time, and each action data in the control action set is combined with the characteristic data to be sent into the pre-trained regression neural network to predict the system frequency. Based on the frequency as the standard, reliable control actions are screened out;

[0010] Step 6: Under the constraint condition of the target function, all reliable control actions are secondarily screened to obtain the optimal frequency stability control action under the current fault condition of the system.

[0011] Further, the step 1 comprises:

[0012] The simulation fault setting of the receiving end system under the UHV AC and UHV DC grid structure is determined, which covers a series of possible typical faults, including the DC single-pole blocking fault and the DC bipolar blocking fault that may occur in the UHV DC line, the off-grid fault of the new energy generator set in the system, and the simulation of different typical operation modes of the receiving end system. The fault condition is composed of multiple faults superimposed on different power system operation states. All possible fault states are arranged and combined to form a simulation fault setting set S, and the simulation fault setting is numbered.

[0013] Further, the step 2 comprises:

[0014] The receiving end system is sequentially subjected to time-domain simulation under the first round of fault conditions according to the order of numbering, and the fault number corresponding to the system frequency instability in the simulation process is screened out. At the same time, the available frequency stability control measures in the receiving end system are determined, the fault number corresponding to the fault is screened out according to the set strategy, and the second round of time-domain simulation is performed to obtain the node characteristic data, system frequency stability control action data and system frequency data in the time-domain simulation.

[0015] Further, the frequency stability control measures include the input of pumped storage M1, the input of centralized energy storage M2, and the load shedding control M3.

[0016] The set strategy comprises:

[0017] The moment of fault occurrence to the first preset time after the fault is determined as the transient observation time window. After the preset time, M1 acts, and after the second preset time, M2 acts. After the third preset time after the fault, M3 acts. Among them, M1 and M2 are input at one time, and M3 is increased gradually until the system frequency is stable or the load shedding upper limit is reached.

[0018] Further, the step 2 further comprises:

[0019] Record the monitoring node characteristic quantities and frequency stability control action data in the second round of time domain simulation, and extract R transient frequency evolution cases from them. For the i-th case, determine the period after the fault occurs as the observation time window ΔT, and collect the active power data P of h nodes in the receiving system within the observation time window. i , voltage data V i and frequency data F i And the lowest frequency value Y that appears in the system during the simulation i , the sampling time interval is Δt; where the monitoring node characteristic quantities include active power data, voltage data and frequency data; 1≤i≤R;

[0020] P i 、V i 、F i They are sorted into multiple q-row and h-column feature quantity matrices, and the frequency stability control action data is processed to obtain the action quantity matrix A with a row and h columns i , a is the maximum control action round allowed by the system, and the R integrated feature data (P i ,V i ,F i ), frequency stability control action data A i The input data set X and R frequency data Y are combined i The two are combined into the output dataset Y, and together they constitute the model training dataset U, where U = {(X, Y)}, q = ΔT / Δt.

[0021] Furthermore, in step 3:

[0022] A dual-channel neural network model is constructed to perform regression learning on the model training data set. The framework of the dual-channel neural network model consists of a long short-term memory recursive neural network, a convolutional neural network, a feature vector weighted overlay layer, a fully connected layer, a regression layer and an output layer; the input end of the long short-term memory recursive neural network is used to input feature quantity data, and the output end is connected to the feature vector weighted overlay layer; the input end of the convolutional neural network is used to input action data, and the output end is connected to the feature vector weighted overlay layer; the feature vector weighted overlay layer is connected to the output layer through the fully connected layer and the regression layer in sequence.

[0023] Furthermore, in the channel of the long short-term memory recurrent neural network, the feature quantity data P with strong temporal correlation is i 、V i 、F iFor feature learning, the long short-term memory recurrent neural network is composed of multiple identical sub-units connected in series. Each sub-unit contains an input gate i, an output gate o, a forget gate f and a storage unit c structure. The forget gate f and the storage unit c can jointly determine whether the input is important enough to be remembered and whether it can be output by establishing a connection relationship between the previous moment and the current moment; the long short-term memory recurrent neural network finally outputs the node feature data (P i ,V i ,F i )’s feature vector T1, where randomly discarding neural units dropout is used to enhance the generalization performance of the model;

[0024] In the channel of the convolutional neural network, the action data A i The local feature data is obtained through the convolution layer, the pooling layer performs dimensionality reduction on the local feature data, and finally the feature vector T2 is output through the fully connected layer; T1 and T2 are weighted summed and fully connected layer, and the system frequency prediction value is regressed and output.

[0025] Furthermore, the step 4 includes:

[0026] Before a fault occurs, all available frequency stability control measures in the receiving system are integrated and put into operation in sequence according to the priority relationship in the time dimension. M1 and M2 should be put into operation as much as possible. According to the different load shedding amounts, a load shedding measure M3 is designed. i The construction method obtains l kinds of multi-measures mutually coordinated frequency stability control actions, which constitute the frequency stability control action set A={A1,A2,…,A i ,…,A l}stand-by.

[0027] Furthermore, the step 5 includes:

[0028] The first round screens out reliable actions in the frequency stability control action set; when a fault causes an active power shortage in the receiving system, the active power time series data, voltage time series data, and frequency time series data of the nodes in the receiving system are collected within the observation time window ΔT. The sampling time interval is Δt, and the lost information in the information transmission is supplemented with 0, and the characteristic matrix {P, V, F} is obtained; {P, V, F} is sequentially integrated with the frequency stability control data set matrix Ai to obtain l input data combinations {P, V, F, A i} is sent to the trained model to predict the system frequency. If the frequency prediction value y i The system frequency reaches the safety range, and the frequency stabilization control action A in the current input data combination i Recorded as reliable action; P i =[P 11 ...,P ij ,...,P ah] is the active power matrix, V i =[V 11 ...,V ij ,...,V ah ] is the voltage vector matrix, F i =[F 11 ...,F ij ,...,F ah ] and the frequency vector matrix.

[0029] Furthermore, the step 6 includes:

[0030] A second round of screening is performed on all reliable control actions obtained in step 5 to obtain the optimal frequency stabilization control action, with the following objective function as the screening condition:

[0031]

[0032] Frequency stabilization control action A i Load removal The total amount of load that can be removed by the receiving system Δp all The smaller it is, the more economical the frequency stability control action will be; A i Action round i accounts for the maximum allowed action round i all The smaller it is, the faster the frequency stability control action will respond;

[0033] Determine the weight coefficients r1 and r2 according to the operating requirements of the receiving system, and select the optimal frequency stability control action A of the receiving system under the current fault. best .

[0034] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0035] This invention uses a deep neural network to evaluate and screen all feasible frequency stability control actions in the receiving system, providing fast and economical multi-measure, coordinated frequency stability control actions for the receiving system in fault situations. This method analyzes frequency stability characteristics from the perspective of active power balance. Through simulation, it generates a large amount of characteristic quantities and frequency stability control action data for key nodes in the receiving system after a fault. A dual-channel neural network learns the characteristic quantity data and action data separately, extracts the data characteristics, performs weighted fusion and regression, and predicts the system frequency. This constructs a frequency stability analysis and control scheme suitable for receiving systems in ultra-high voltage AC and ultra-high voltage DC feed-in grid structures under the new energy environment. Unlike existing frequency stability control methods based on deep learning, this invention pre-defines a series of feasible frequency stability control actions. Using a model, it predicts the system frequency under the influence of frequency control actions. Reliable frequency stability control actions are screened in the first round. Taking into account the stable power supply requirements of the actual power system, a second round of screening is performed on these reliable actions to obtain the optimal frequency stability control action, ensuring the reliability and cost-effectiveness of the control action. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments described in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0037] Figure 1 A schematic flow chart of a method for controlling power system frequency stability using a dual-channel deep neural network according to an embodiment of the present invention;

[0038] Figure 2 This is a diagram showing the structure of the action data matrix in an embodiment of the present invention;

[0039] Figure 3 This is a structural diagram of a deep neural network regression model based on dual channels in an embodiment of the present invention;

[0040] Figure 4 This is a flowchart of the optimal frequency stabilization control action screening in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described with reference to the accompanying drawings and embodiments. The embodiments described are only a part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by those skilled in the art should fall within the scope of protection of the embodiments of the present invention.

[0042] The purpose of the present invention is to comprehensively utilize various feasible frequency stability control measures in the system when the receiving system encounters transient faults and frequency drops, efficiently formulate a set of reliable and economical frequency stability control schemes, and promptly provide a set of reliable frequency stability control actions to prevent system frequency instability or collapse accidents, and quickly and economically restore the system frequency to a safe and stable state. This scheme comprehensively considers the complex and diverse fault situations that may occur in the receiving system under the ultra-high voltage AC and ultra-high voltage DC feeding grid structure under the background of new energy, and uses the time domain simulation method to obtain fault data; through the encoding method, the characteristic quantity data of the monitoring node (500kV) and the frequency stability control action data in the system are converted into a data format that can be learned by the neural network, and the long short-term memory recursive neural network (LSTM) and the convolutional neural network (CNN) are used to perform feature learning on the two parts of the input data respectively, and the weighted combination is used to guide the improvement of the attention of the transient frequency stability model based on the neural network to the control action information during the prediction process, so that the model can accurately capture The impact of frequency control actions on system frequency under fault conditions is analyzed, and a highly accurate system frequency prediction value is output to obtain a deep neural network model with the best training effect; a series of available frequency stability control actions are set before the fault, and the characteristic data of the network monitoring nodes after the fault are fed into the pre-trained neural network model to predict the system frequency, thereby judging whether the system frequency control action under fault conditions is reliable, screening out reliable frequency stability control actions, and setting target constraints in combination with the actual economic operation needs of the power system to further screen and obtain the optimal frequency stability control action under the current fault. The two rounds of screening ensure the reliability and economy of the obtained transient frequency stability control action.

[0043] See also Figure 1 and Figure 4 The present invention provides an embodiment of a power system frequency stability control method based on a dual-channel deep neural network, which includes:

[0044] (1) Determine the simulation fault settings of the receiving system under the UHV AC and UHV DC feeding grid structure under the background of new energy, covering a series of possible typical faults. Including but not limited to the DC single-pole blocking fault, DC bipolar blocking fault, and the grid-off fault of the new energy generator set in the system, while simulating different typical operating modes of the receiving system. (In this embodiment, there is 1 UHV AC line, 61 groups of new energy generator sets, and 6 typical system operating modes). The fault situation is often composed of multiple typical faults superimposed on different power system operating states. All possible fault states are arranged and combined to form a simulation fault setting set S, and these simulation faults are numbered.

[0045] (2) Perform two rounds of time domain simulation. Perform the first round of time domain simulation on the receiving system in the order of the number, and select the fault numbers corresponding to the system frequency instability (in this implementation case, when the system frequency is lower than 49.8Hz, it is determined to be unstable) during the simulation process. At the same time, determine the frequency stability control measures available in the receiving system, and put pumped storage M1, centralized energy storage M2, and load shedding control M3 into operation. In order to obtain a universal and reliable frequency stability control strategy, the frequency stability control action corresponding to the fault number screened out is determined according to the following strategy and two rounds of time domain simulation are performed: the time from the time of fault occurrence to 0.3S after the fault is determined as the transient observation time window. After 0.3S, M1 acts, and then after 0.1S, M2 acts. After 0.5S after the fault, M3 acts. M1 and M2 are activated all at once (to ensure economic efficiency, the system's pumped hydro and centralized energy storage have limited active power reserves and are often unable to fully compensate for the system's active power shortfall under fault conditions, often requiring load shedding). M3's load shedding is incrementally increased until the system frequency stabilizes or reaches the load shedding limit. Frequency stability control measures are determined by dispatchers to obtain action data that stabilizes the system frequency and node feature data for model training. The optimal frequency stability control strategy is ultimately determined through two rounds of screening.

[0046] (3) Collect and organize the training data required for the deep learning model. Record the monitoring node characteristics (active power data, voltage data, frequency data) and frequency stability control action data in the second round of time domain simulation, and extract R transient frequency evolution cases from them. For the i-th (1≤i≤R) case, collect the active power data P of h 500kV nodes in the receiving system within the transient observation time window. i , voltage data V i and frequency data F i And the lowest value of system frequency Y during simulation i , the sampling time interval is Δt, and P i 、V i 、F i The frequency stabilization control action data is processed to obtain the fourth input matrix A with a row and h column. i ,a is the maximum control action round allowed by the system, and the construction method is as shown in the attached Figure 2 As shown: A i The values ​​in each column of the first and second rows are filled with the active power values ​​provided by the frequency stabilization control measures M1 and M2. i Each column from the 3rd row to the b+2th row is filled with the load shedding value of the corresponding node in the corresponding round (b is the number of load shedding operations). If a>b+2, the remaining rows are filled with 0. i ,V i,F i ), frequency stability control action data A i The input data set X is composed of R frequency data Y i The two are combined into the output data set Y, and the two together constitute the model training data set U, U = {(X, Y)}; Figure 2 shown.

[0047] (4) Construct a dual-channel neural network model to perform regression learning on the data set U obtained in step (3), see Figure 3 The model framework is composed of the long short-term memory recursive neural network (LSTM) and convolutional neural network (CNN) as the main input, feature vector weighted superposition layer and output regression layer. In the LSTM channel, the LSTM with the characteristics of long-term memory of important information and dynamic adjustment is used to process the feature quantity data P with strong temporal correlation. i ,V i ,F i Perform feature learning. LSTM is composed of multiple identical subunits connected in series. Each subunit contains an input gate i, an output gate o, a forget gate f, and a storage unit c structure. The forget gate f and the storage unit c can jointly determine whether the input is important enough to be remembered and whether it can be output by establishing a connection relationship between the previous moment and the current moment. In this way, the LSTM unit not only learns information from the past and current moments, but also passes the learned state to the subsequent moments, and finally outputs the node feature data (P i ,V i ,F i ) feature vector T1, in this process, random neural unit dropout is used to enhance the generalization performance of the model. In the CNN channel, the action data A i The convolutional layer obtains local feature data, which is then reduced in dimension by the pooling layer. These two steps significantly reduce the number of parameters and computational complexity, improving the model's generalization capabilities. Finally, the fully connected layer outputs the feature vector T2. T1 and T2 are then regressed through the weighted overlay layer and the fully connected layer to output the predicted system frequency value.

[0048] (5) Debug the model parameters, optimize the initial learning rate r, random number seed and data regularization method, and select the dual-channel deep neural network model with the fastest convergence speed and the smallest root mean square error (RMSE) for use. In this example, the machine learning part is written in Python and the model is built based on the PyTorch framework.

[0049] (6) Before the fault occurs, all available frequency stability control measures in the receiving system are integrated and put into operation in sequence in the time dimension according to the priority relationship. M1 and M2 should be put into operation as much as possible. According to the different load shedding amounts, a load shedding measure M3 is designed. According to step (3), A iConstruction method, there are l kinds of multi-measures coordinated frequency stability control actions, which constitute the frequency stability control action set A={A1,A2,…,A i ,…,A l}stand-by.

[0050] (7) The first round of screening of reliable actions in the frequency stability control data set. When the fault causes a serious active power shortage in the receiving system, the wide area measurement system collects the active power time series data, voltage time series data and frequency time series data of the 500kV node within the observation time window ΔT. The sampling time interval is Δt. The lost information in the information transmission is supplemented with 0, and the three-item feature quantity matrix {P, V, F} is obtained. {P, V, F} is combined with the action quantity matrix A in the frequency stability control data set in step (6) i Integrate sequentially to obtain l input data combinations {P, V, F, A i} is sent to the pre-trained model in step (6) to predict the system frequency. If the frequency prediction value y i The system frequency reaches the safety range, and the frequency stabilization control action A in the current input data combination i Denoted as a reliable action. There is only one unique feature matrix, and there are L action data. Combining the unique feature matrix with each action data yields L input data combinations, which are then fed into the model for prediction. In other words, among the L input data combinations, only the action data A differs.

[0051] (8) Perform a second round of screening on all reliable control actions obtained in step (7) to obtain the optimal frequency stabilization control action that is both economical and fast. A is the screening condition. i Load removal The total amount of load that can be removed by the receiving system Δp all The smaller it is, the more economical the action is; A i Action round i accounts for the maximum allowed action round i all The smaller it is, the faster the action will react. Determine the weight coefficients r1 and r2 according to the operating requirements of the receiving system, and select the optimal frequency stability control action A of the receiving system under the current fault. best .

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for power system frequency stability control based on a dual-channel deep neural network, characterized in that: include: Step 1: Determine the possible fault setting set in the receiving system under the UHV AC and UHV DC feeding grid structure; Step 2: Build a receiving system framework in the transient time-domain simulation software to perform time-domain simulation of the receiving system under fault conditions, and identify fault settings that cause frequency instability in the receiving system. Integrate available frequency control measures in the power system, apply frequency stability control actions to the frequency instability samples, and conduct a second round of simulations. Optimize the frequency control actions until the simulated frequency of the receiving system reaches a safe and stable range for all fault conditions in the fault setting set. During this process, monitor and record key node feature quantities, frequency stability control actions, and receiving system frequency data in the time-domain simulation to create a model training dataset. Step 3: Use the long short-term memory recurrent neural network and convolutional neural network to build a dual-channel deep neural network frequency regression prediction model. Use the data in the training dataset to optimize and train the dual-channel deep neural network frequency regression prediction model. The trained prediction model takes node feature quantities and frequency stability control action data as input and outputs the receiving system frequency data. Step 4: Debug the model hyperparameters to obtain a dual-channel deep neural network frequency regression prediction model with minimal error and the fastest convergence speed. The trained model predicts the frequency value of the receiving system under control action intervention to determine the reliability of the frequency stability control action. At the same time, within the allowable range of normal operation of the receiving system, a set of feasible control actions is formed by combining various possible frequency stability control measures in the receiving system. Step 5: During actual power system operation, if a fault occurs that causes an active power shortage, the characteristic data of key nodes are collected in real time. The action data of each control action set is combined with the characteristic data in sequence and fed into a pre-trained recurrent neural network to predict the receiving system frequency. This frequency is used as a criterion to select reliable control actions. Step 6: Using the objective function as a constraint, perform a secondary screening of all reliable control actions to obtain the optimal frequency stabilization control action under the current fault condition of the receiving system. The step 1 comprises: Determine the simulated fault settings for the receiving system under the UHV AC and UHV DC feed-in grid structure, encompassing a series of possible typical faults, including DC single-pole blocking faults and DC bipolar blocking faults that may occur in UHV DC lines, and grid-off faults of renewable energy generator sets in the receiving system, and simulate different typical operating modes of the receiving system; the fault situation is composed of multiple faults superimposed on different power system operating states, and all possible fault states are arranged and combined to form a simulated fault setting set S, and the simulated faults are numbered; The step 2 includes: Perform a first round of time domain simulation on the receiving system in fault conditions in order of numbering, and select the fault numbers corresponding to the frequency instability of the receiving system during the simulation; simultaneously determine the available frequency stability control measures in the receiving system, determine the frequency stability control actions for the faults corresponding to the selected fault numbers according to the set strategy, and perform a second round of time domain simulation to obtain node characteristic quantity data, receiving system frequency stability control action data, and receiving system frequency data in the time domain simulation; The step 6 comprises: A second round of screening is performed on all reliable control actions obtained in step 5 to obtain the optimal frequency stabilization control action, with the following objective function as the screening condition: Frequency stabilization control action A i Removal of load The total amount of load that can be removed by the receiving system Δp all The smaller it is, the more economical the frequency stability control action will be; A i Action round i accounts for the maximum allowed action round i all The smaller it is, the faster the frequency stability control action will respond; Determine the weight coefficients r1 and r2 according to the operating requirements of the receiving system, and select the optimal frequency stability control action A of the receiving system under the current fault. best .

2. The method according to claim 1, characterized in that The frequency stabilization control measures include the use of pumped storage M1, the use of centralized energy storage M2, and load shedding control M3; The policies that are set include: The transient observation time window is determined from the moment of fault occurrence to the first preset time after the fault. After the preset time, M1 is activated, and then after the second preset time, M2 is activated. After the third preset time after the fault, M3 is activated. Among them, M1 and M2 are activated at one time, and the load shedding amount of M3 increases gradually until the frequency of the receiving system is stable or the load shedding upper limit is reached.

3. The method according to claim 2, characterized in that The step 2 further comprises: Record the monitoring node characteristic quantities and frequency stability control action data in the second round of time domain simulation, and extract R transient frequency evolution cases from them. For the i-th case, determine the period after the fault occurs as the observation time window ΔT, and collect the active power data P of h nodes in the receiving system within the observation time window. i , voltage data V i and frequency data F i The lowest frequency value Y of the receiving system during the simulation process i , the sampling time interval is Δt; where the monitoring node characteristic quantities include active power data, voltage data and frequency data; 1≤i≤R; P i 、V i 、F i They are sorted into multiple q-row and h-column feature quantity matrices, and the frequency stability control action data is processed to obtain the action quantity matrix A with a row and h columns i , a is the maximum control action round allowed by the receiving system, and the R integrated feature data (P i ,V i ,F i ), frequency stability control action data A i The input data set X and R frequency data Y are combined i The two are combined into the output dataset Y, and together they constitute the model training dataset U, where U = {(X, Y)}, q = ΔT / Δt.

4. The method according to claim 1, wherein In step 3: A dual-channel neural network model is constructed to perform regression learning on the model training data set. The framework of the dual-channel neural network model consists of a long short-term memory recursive neural network, a convolutional neural network, a feature vector weighted overlay layer, a fully connected layer, a regression layer and an output layer; the input end of the long short-term memory recursive neural network is used to input feature quantity data, and the output end is connected to the feature vector weighted overlay layer; the input end of the convolutional neural network is used to input action data, and the output end is connected to the feature vector weighted overlay layer; the feature vector weighted overlay layer is connected to the output layer through the fully connected layer and the regression layer in sequence.

5. The method according to claim 4, characterized in that In the channel of the long short-term memory recurrent neural network, the feature data P with strong temporal correlation is i 、V i 、F i For feature learning, the long short-term memory recurrent neural network is composed of multiple identical sub-units connected in series. Each sub-unit contains an input gate i, an output gate o, a forget gate f and a storage unit c structure. The forget gate f and the storage unit c jointly determine whether the input is important enough to be remembered and whether it can be output by establishing a connection relationship between the previous moment and the current moment; the long short-term memory recurrent neural network finally outputs the node feature data (P i ,V i ,F i )’s feature vector T1, where randomly discarding neural units dropout is used to enhance the generalization performance of the model; In the channel of the convolutional neural network, the action data A i The local feature data is obtained through the convolution layer, the pooling layer performs dimensionality reduction on the local feature data, and finally the feature vector T2 is output through the fully connected layer; T1 and T2 are weighted summed and fully connected layer, and the frequency prediction value of the receiving system is regressed and output.

6. The method according to claim 3, characterized in that The step 4 comprises: Before a fault occurs, all available frequency stability control measures in the receiving system are integrated and put into operation in sequence according to the priority relationship in the time dimension. M1 and M2 should be put into operation as much as possible. According to the different load shedding amounts, a load shedding measure M3 is designed. i The construction method obtains a total of l multi-measure coordinated frequency stability control actions, which constitute the frequency stability control action set A={A1,A2,…,A i ,…,A l }stand-by.

7. The method according to claim 6, characterized in that The step 5 comprises: The first round screens out reliable actions in the frequency stability control action set; when a fault causes a shortage of active power in the receiving system, the active power time series data, voltage time series data, and frequency time series data of the nodes in the receiving system are collected within the observation time window ΔT. The sampling time interval is Δt, and the lost information in the information transmission is supplemented with 0 to obtain the characteristic matrix {P, V, F}; {P, V, F} is combined with the frequency stability control data matrix A i Integrate sequentially to obtain l input data combinations {P, V, F, A i } is sent to the trained model to predict the frequency of the receiving system. If the frequency prediction value y i The frequency of the receiving system reaches the safe range, and the frequency stability control action A in the current input data combination i Recorded as reliable action; P i =[P 11 ...,P ij ,...,P ah ] is the active power matrix, V i =[V 11 ...,V ij ,...,V ah ] is the voltage vector matrix, F i =[F 11 ...,F ij ,...,F ah ] is the frequency vector matrix.

Images