Transient overvoltage stability evaluation method based on CK-means and deep learning
Through the CK-means and deep learning 1D-CNN model combined with the focus loss function, the data imbalance and timing characteristics problems in the transient stability evaluation of the power system are solved, and efficient transient overvoltage stability evaluation is achieved.
Patent Information
- Application Number
- CN202510690197.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-27
- Publication Date
- 2025-07-08
AI Technical Summary
The existing deep learning models have problems such as data imbalance, insufficient consideration of timing characteristics and excessive model complexity in the power system's transient stability evaluation, resulting in inaccurate evaluation results.
The CK-means algorithm is used to cluster and decompose stable and instable samples, combined with the deep learning 1D-CNN model, the focus loss function optimization model training is introduced, multi-grained timing features are extracted, data imbalance problem is alleviated, and timing characteristics are preserved.
On the premise of ensuring a stable sample accuracy rate, the accuracy and speed of the transient stability evaluation of the transmission end system of the new energy high-voltage DC transmission system is improved, and the misjudgment rate is reduced.
Smart Images
Figure CN120277446A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electric power, and specifically relates to a transient overvoltage stability evaluation method based on CK-means and deep learning. Background Art
[0002] Traditional transient stability assessment (TSA) methods are usually divided into time-domain simulation methods and direct methods, also known as transient energy function methods. Among them, the time-domain simulation method can achieve accurate evaluation but has a large amount of calculation and is time-consuming; the direct method can quickly make a stability judgment and give the stability degree, but it is necessary to construct an "energy function" to reflect the transient characteristics of the system, which is difficult to be used in complex power systems. With the rapid development of the smart grid supported by big data, how to achieve fast and accurate online assessment of power system transient stability has always been a hot and difficult issue in the research of power grid safety and stability. Wehenkel et al. pointed out that the core of TSA lies in establishing the mapping relationship between the system characteristic quantity X and the transient stability Y. In recent years, artificial intelligence, including artificial neural networks, deep learning, and ensemble learning, has developed rapidly. Compared with the above machine learning algorithms, deep learning can independently combine and extract input features from data, avoiding the subjectivity brought by manual intervention. Existing research has proposed transient stability assessment methods for AC-DC receiving-end power grids based on artificial intelligence such as CNN. Although deep learning methods have good performance on self-built data sets, there are still problems that most of the network structures of deep learning models are oriented to image data or text speech sequence data, and there are few models for power system measurement data, and the problem of sample imbalance leads to a decrease in the precision rate of stable samples, and the overall evaluation performance of the model has not been effectively improved.
[0003] Traditional deep learning models are all designed to analyze image data and speech sequence data, and there is little involvement in the model structure design for power system measurement data; in the power system, the data distribution is uneven. To solve the problem of sample imbalance, most models add penalty weights to the loss function. Although the recall rate of unstable samples is improved to a certain extent, the precision rate of stable samples is reduced, and the overall evaluation performance of the model has not been effectively improved. In the construction of the data set in transient stability assessment (TSA), there will be a large number of problems of inaccurate and uneven data, and at the same time, most studies have not fully considered the time series characteristics, thereby improving the accuracy of the quantitative evaluation results of new energy through the external transmission system. Summary of the Invention
[0004] The object of the present invention is to provide a transient overvoltage stability evaluation method based on CK-means and deep learning, which can improve the problem of unbalanced data samples while ensuring the precision rate of stable samples, and can retain the original time series characteristics, enabling rapid quantitative evaluation, so as to solve the problem of inaccurate, unbalanced, time series asynchronous sensitivity, and too high model complexity in the transient stability evaluation of a new power system, resulting in inaccurate evaluation results.
[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:
[0006] A transient overvoltage stability evaluation method based on CK-means and deep learning includes the following steps:
[0007] S100. Construct the sending-end system of a new energy power system via a high-voltage direct current (HVDC) transmission system. Combine different fault information of the sending-end system to conduct transient simulation of the fault information, and generate simulation data samples of the converter station bus voltage amplitude, voltage phase angle, line active power, and line reactive power of the sending-end system of the new energy power system via the HVDC transmission system during steady-state operation, as well as the converter station bus voltage amplitude, voltage phase angle, line active power, and line reactive power during transient operation.
[0008] S200. Extract the simulation data samples generated in S100 using a phasor measurement unit (PMU) to obtain time series data. Process the time series data using the CK-means algorithm composed of the K-means clustering algorithm assisted by the Canopy algorithm, divide the stable samples into two subclasses, and classify the unstable samples into one subclass. At the same time, optimize the squared error objective functions of the stable sample subclass and the unstable sample subclass.
[0009] S300. Select the time series data processed in S200 as input features and establish a 1D-CNN deep learning model.
[0010] S310. The one-dimensional convolution operation formula of the 1D-CNN model is:
[0011]
[0012]
[0013] where: X l is the input feature, and
[0014] The convolution kernel is: consisting of c2 with a size of is obtained by stacking matrices of convolution kernels. Here, c1 is the number of channels of the input features, and c2 is the number of convolution kernels; t1 and t2 are the sequence lengths of the input features and output features respectively; the subscript l represents the convolution layer number; R is the real number set for the data set; h is the size of the convolution kernel.
[0015] X l+1 is the output feature, and f is the non-linear activation function. is the matrix concatenation operator. is the bias term.
[0016] The parameters to be learned in a single-layer one-dimensional convolutional network of the deep learning model are K l+1 and b l+1 , with a total of c2×h×c1 + c2. At the same time, the time series length t2 of the output feature X l+1 is:
[0017]
[0018] In the above formula: t1 and t2 are the sequence lengths of the input features and output features respectively; h is the size of the convolution kernel; p is the number of convolution padding layers; s is the convolution stride. When p and s are adjusted to a reasonable range, the rapid decrease of t2 and the loss of edge information can be avoided.
[0019] S320, the output of the 1D-CNN model
[0020] Determine the two output nodes of the model, namely transient stability and transient instability. The formula is:
[0021]
[0022] In the formula is the output of the i-th classification node; z i is the input of the i-th classification node.
[0023] After normalization
[0024] Take as the output, and judge whether is 1 at this time. If it is 1, it means that the sending end system of the new energy through the HVDC transmission system is in stable operation; otherwise, the sending end system of the new energy through the HVDC transmission system enters the unstable operation state.
[0025] S330. Perform stability judgment on the input data of the 1D-CNN model
[0026] Introduce the judgment index A svsi , an index for judging voltage instability when the sending end system of the new energy through the HVDC transmission system suffers a large disturbance.
[0027]
[0028] In the formula, A svsi is the transient voltage stability evaluation coefficient; U min is the electrical performance of the system under fault conditions; U s is the minimum value at which voltage collapse occurs; U d is the degree of voltage fluctuation; T u is the factor determining whether the system can recover stability, T s is the time in actual engineering when the voltage is allowed to be lower than the minimum value U s at which voltage collapse occurs, generally 10 s;
[0029] A svsi being 1 indicates that the voltage of the sending - end system of the new - energy through HVDC transmission system is stable. If A svsi is less than 1, it indicates that the transient voltage of the sending - end system of the new - energy through HVDC transmission system is unstable.
[0030] Preferably, in step S100, when the sending - end system of the new - energy through HVDC transmission system operates stably, the reactive power of the sending - end system of the new - energy through HVDC transmission system is balanced with the rectifier - side reactive power of the sending - end system of the new - energy through HVDC transmission system. The reactive - power consumption of the converter station is as shown in the following formula:
[0031]
[0032] In the formula, Q dr is the reactive power of the sending - end system of the new - energy through HVDC transmission system, P dr is the active power of the sending - end system of the new - energy through HVDC transmission system, α is the firing angle of the rectifier side of the sending - end system of the new - energy through HVDC transmission system, and μ is the commutation overlap angle of the rectifier side of the sending - end system of the new - energy through HVDC transmission system. When DC commutation failure occurs on the inverter - side bus, the firing angle α of the sending - end system of the new - energy through HVDC transmission system increases, resulting in a decrease in the active power P dr of the sending - end system of the new - energy through HVDC transmission system, and the reactive power Q dr of the sending - end system of the new - energy through HVDC transmission system also decreases accordingly;
[0033] In the case of commutation failure, the transient voltage rise on the rectifier - side bus is negatively correlated with the short - circuit ratio. The specific formula is as follows:
[0034]
[0035] In the formula, ΔU P is the change in the rectifier - side bus voltage of the sending - end system of the new - energy through HVDC transmission system before and after the fault occurs; U h is the voltage at the grid - connection point of the thermal power unit; P dFor the DC transmission capacity of the HVDC transmission system, the bus voltage U is selected. P , the active power p on the line acp , and the reactive power Q acp are used as evaluation criteria; K CR is the short-circuit ratio of the DC system; ΔQ is the surplus reactive power of the filtered AC system; ΔP is the change in active power of the filtered AC system.
[0036] Preferably, in the step S100, the Newton-Raphson method is used for system power flow analysis. When the sending-end system of the HVDC transmission system for new energy operates normally, according to power balance, the calculation formula for the active power of the sending-end system of the HVDC transmission system for new energy is:
[0037] P acp = P PV1 + P wind1 + P h1 + P st1
[0038] Where: P acp is the active power generated by the new energy system, P PV1 is the active power transmitted by the photovoltaic power station, P wind1 is the active power transmitted by the doubly-fed wind power, P h1 is the active power transmitted by the thermal power unit, P st1 is the active power transmitted by the energy storage;
[0039] The reactive power Q acp of the sending-end system of the HVDC transmission system for new energy has the following calculation formula:
[0040] Q acp = Q PV1 + Q wind1 + Q h1 + Q st1
[0041] Where: Q PV1 is the reactive power transmitted by the photovoltaic power station, Q wind1 is the reactive power transmitted by the doubly-fed wind power, Q h1 is the reactive power transmitted by the thermal power unit, Q st1 is the reactive power transmitted by the energy storage, Q dr is the reactive power of the DC transmission system;
[0042] The reactive power Q dr of the HVDC transmission system has the following calculation formula:
[0043] Q dr = Q acp + Q cp
[0044] Where Q cp is the reactive power generated by the filter AC system;
[0045] Meanwhile, under the normal operating condition, the bus voltage U at the rectifier side of the sending-end system of the new energy through the HVDC transmission system is p as follows:
[0046]
[0047] P h1 is the active power transmitted by the thermal power unit; Q h1 is the reactive power transmitted by the thermal power unit; U h is the voltage at the grid connection point of the thermal power unit; X sn is the reactance of the thermal power transmission line.
[0048] Preferably, in the step S100, when a commutation failure fault occurs in the sending-end system of the new energy through the HVDC transmission system, it is equivalent to a short circuit on the inverter side, the DC current increases rapidly, and the rectifier side consumes more reactive power. The constant current controller reduces the DC current I by increasing the trigger angle α dc ; during the transient overvoltage period, the power of the sending-end system of the new energy through the HVDC transmission system still satisfies the balance. The calculation formula for the active power P′ of the sending-end system of the new energy through the HVDC transmission system is: acp as follows:
[0049] P′ acp = P′ PV1 + P′ wind1 + P h1 ′ + P st1 ′
[0050] Where: P′ PV1 is the active power transmitted by the PV power station after the fault occurs, P′ wind1 is the active power transmitted by the doubly-fed wind power after the fault occurs, P h1 ′ is the active power transmitted by the thermal power unit after the fault occurs, P st1 ′ is the active power transmitted by the energy storage after the fault occurs, Q st1 ′ is the reactive power transmitted by the energy storage after the fault occurs;
[0051] The calculation formula for the reactive power of the sending-end system of the new energy through the HVDC transmission system is:
[0052] Q′ acp = Q′ PV1 + Q′ wind1 + Q h1 ′ + Q st1 ′
[0053] Where: Q′ acpQ′ is the reactive power generated by the new energy system after a fault occurs. PV1 Q′ is the reactive power transmitted by the photovoltaic power station after a fault occurs. wind1 Q is the reactive power transmitted by the doubly-fed wind power after a fault occurs. h1 Q′ is the reactive power transmitted by the thermal power unit after a fault occurs.
[0054] Preferably, in step S100, due to the large inertia time of thermal power output and the relatively long excitation regulation response time compared with the commutation failure transient process, the terminal voltage of the thermal power unit remains unchanged. At this time, the bus voltage on the rectifier side of the HVDC transmission system is:
[0055]
[0056] In the formula, U′ p is the bus voltage on the rectifier side after a fault occurs in the system, P h1 ′ is the active power transmitted by the thermal power unit after a fault occurs, Q h1 ′ is the reactive power transmitted by the thermal power unit after a fault occurs, U h is the voltage at the grid connection point of the thermal power unit during normal operation, X sn is the reactance of the thermal power transmission line.
[0057] Preferably, in step S100, when the voltage of the converter station filter increases, its reactive power compensation capacity increases. When the commutation failure recovers, the reactive power consumed by the converter station changes. The actual reactive power transmitted by the converter station to the sending-end system during the transient process is;
[0058]
[0059] In the formula: U′ p is the bus voltage on the rectifier side after a fault occurs in the system, U p is the bus voltage on the rectifier side of the HVDC transmission system during normal operation of the system, Q cp is the reactive power generated by the AC filter system, Q’ cp is the actual reactive power transmitted from the new energy through the sending-end system of the HVDC transmission system to the AC filter system during the transient process;
[0060] Compared with the normal operation state, the calculation formula for the surplus reactive power at the sending-end AC bus during the transient overvoltage period is:
[0061] ΔQ = Q h1 -Q′ h1
[0062] In the formula, ΔQ is the surplus reactive power of the filtering AC system, Q h1 is the reactive power transmitted by the thermal power unit during normal operation, Q h1Q' is the reactive power transmitted by the thermal power unit after a fault occurs.
[0063] Preferably, in the step S100, the change amount ΔU of the bus voltage on the rectifier side of the sending end system of the new energy through the HVDC transmission system before and after a fault occurs P The calculation formula is:
[0064] The transient voltage change amount at the AC bus on the rectifier side of the HVDC transmission system is:
[0065]
[0066] In the formula: ΔQ is the surplus reactive power of the filtering AC system, X sn is the reactance of the thermal power transmission line, ΔP is the change amount of the active power of the filtering AC system, U h is the voltage at the grid connection point of the thermal power plant during normal operation.
[0067] Preferably, the construction of the input of the 1D-CNN model in the step S300 includes the following steps:
[0068] Collect the numerical values of the voltage and power on the rectifier side bus node, use the time-series measurement data of the synchronized vector measurement unit PMU as the input, and use the bus voltage amplitude, voltage phase angle, line active power and line reactive power of the sending end converter station of the sending end system of the new energy through the HVDC transmission system as the initial features to establish the initial features of the time-series measurement data, as shown in the following formula,
[0069] X = [U p1 ,..., U pm , θ p1 ,..., θ pm ,
[0070] P acp1 ,…, P acpn , Q acp1 ,…, Q acpn ,
[0071] (U pm , θ pm , P acpn , Q acpn ∈R d )
[0072] In the formula: X is the time-series measurement data, U p is the effective value of the phase voltage of the DC rectifier side converter bus, θ p is the phase angle of the rectifier side bus voltage, and the serial numbers 1... m in the above symbols are the bus node numbers; P acp is the active power transmitted by a single phase, Q acpis the reactive power for single-phase transmission. In the above symbols, the serial numbers 1...n are the transmission line numbers; R d is the real number set under d sampling points, and d is the number of sampling points;
[0073] The process of decomposing data using the CK-means algorithm includes: using the Canopy algorithm to roughly cluster the data set, determining the number of clusters of the CK-means algorithm, obtaining the initial cluster centers, and then calculating according to the distance from the sample to the center, updating the cluster centers through calculation. The objective function of the CK-means algorithm uses the squared error criterion function, and after multiple iterations, the objective function E converges; the formula of the objective function is:
[0074]
[0075] In the formula: E is the sum of the squared errors of the data set composed of bus voltage, active power, and reactive power, x is the position of the sample data, is the average value of cluster C i ;
[0076] Classify the data, and the stable samples are divided into two subclasses, and the unstable samples are divided into one subclass;
[0077] Preferably, when performing cluster analysis on the collected data in S200, the focal loss function FL is introduced to replace the cross-entropy CE to enhance the prediction accuracy of difficult samples, and the 1D-CNN model is trained using the training sample set to obtain a transient overvoltage judgment model for the sending-end system of the new energy through the HVDC transmission system; input the parameters of the bus voltage amplitude, voltage phase angle, line active power, and line reactive power of the converter station bus of the sending-end system of the new energy through the HVDC transmission system constructed by the time-series data into the 1D-CNN model to obtain the transient voltage stability evaluation coefficient of the sending-end system of the new energy through the HVDC transmission system;
[0078] The loss function increases the accuracy. The formula of the focal loss function is:
[0079]
[0080] In the formula: α∈(0, 1) is the balance factor, γ is the modulation factor, and the weight coefficient formed with the model output can reduce the loss function value of simple samples, enabling it to have better prediction for difficult samples; y i is the stable sample label.
[0081] The fault information described in the present invention includes AC line short-circuit faults, DC commutation failure faults, and DC blocking faults.
[0082] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0083] The present invention constructs a model of the sending-end system of a new energy device through a high-voltage direct current transmission system LCC-HVDC to obtain the bus voltage amplitude, voltage phase angle, and active and reactive power of the line. The PUM is used to directly extract the data, and the CK-means algorithm is used to cluster and decompose the stable and unstable samples. The stable samples are divided into two subclasses, and the unstable samples are divided into one subclass, alleviating the problem of data imbalance.
[0084] The CK-means algorithm is a clustering method composed of the Canopy algorithm and the K-means algorithm, mainly used to solve the problem of data imbalance. The Canopy algorithm is used for rough clustering. By setting two distance thresholds, the data is divided into multiple clusters to quickly determine the initial clustering centers and the number of K-means. On the basis of the Canopy algorithm, the K-means algorithm is used to further optimize the clustering. By iteratively calculating to minimize the squared error function, the final clustering result is obtained.
[0085] The present invention uses the subclasses of the clustered dataset as the input features of deep learning, takes transient stability and instability as two output nodes, and at the same time introduces the focal loss function to train the dataset. A deep learning network with multi-size convolutional kernels is used to extract multi-granularity time series features, improving the model's ability to extract time series information and reducing redundant parameters, ensuring the time series features and data features of the evaluation model.
[0086] The present invention introduces the focal loss function (FL) into the deep learning model to optimize model training, enhancing the identification ability of unstable samples. Description of the Drawings
[0087] Figure 1 is the transient evaluation flow chart;
[0088] Figure 2 is the schematic structural diagram of the sending-end system of a new energy device through a high-voltage direct current transmission system LCC-HVDC;
[0089] Figure 3 is the schematic diagram of the deep learning model;
[0090] Figure 4 is the schematic diagram of the CK-means algorithm dataset classification process;
[0091] Figure 5 is the schematic diagram of the training process. Detailed Embodiment
[0092] The following further describes the present invention in detail with reference to the drawings.
[0093] As Figure 1A transient overvoltage stability evaluation method based on CK-means and deep learning is shown as follows, including the following steps:
[0094] S100. Build a direct current (DC) transmission system model of a new energy source via a line-commutated converter based high voltage direct current (LCC-HVDC) transmission system. The new energy source includes photovoltaic power generation, wind power, energy storage power stations, thermal power, etc.
[0095] When the LCC-HVDC transmission system operates stably, the reactive power of the new energy source at the sending end of the new energy source via the HVDC transmission system is balanced with the rectifier side. The reactive power consumption of the converter station is shown as follows:
[0096]
[0097] In the formula, Q dr is the reactive power of the DC transmission system, P dr is the active power of the DC transmission system, α is the firing angle of the rectifier side of the DC system, and μ is the commutation overlap angle of the rectifier side of the DC system. When commutation failure occurs at the inverter side bus, the extinction angle will gradually decrease and the firing angle will increase, resulting in a decrease in active power and a decrease in the reactive power consumed by the DC.
[0098] Adopt the Newton-Raphson method for power flow analysis of the transmission system. When the sending end system of the new energy source via the HVDC transmission system operates under normal conditions, according to the power balance, the calculation formulas for the active power and reactive power of the sending end system of the new energy source via the HVDC transmission system are as follows:
[0099] P acp = P PV1 + P wind1 + P h1 + P st1 (2)
[0100] Q acp = Q PV1 + Q wind1 + Q h1 + Q st1 (3)
[0101] Q dr = Q acp + Q cp (4)
[0102] In the formula: P acp is the active power generated by the new energy system, P PV1 is the active power transmitted by the photovoltaic power station, P wind1 is the active power transmitted by the doubly-fed wind power, P h1 is the active power transmitted by the thermal power unit, P st1 is the active power transmitted by the energy storage.
[0103] Q acp is the reactive power generated by the new energy system, Q PV1 is the reactive power transmitted by the photovoltaic power station, Q wind1 is the reactive power transmitted by the doubly-fed wind power, Q h1 is the reactive power transmitted by the thermal power unit, Q st1 is the reactive power transmitted by the energy storage, Q dr is the reactive power of the DC transmission system; where Q cp is the reactive power generated by the filter AC system;
[0104] At the same time, the bus voltage at the rectifier side of the HVDC transmission system under normal operating conditions is obtained as:
[0105]
[0106] When a fault such as commutation failure occurs in the sending-end system of the LCC-HVDC of the HVDC transmission system, it is equivalent to a short circuit on the inverter side, the DC current increases rapidly, the rectifier station consumes more reactive power, and the constant current controller reduces the DC current I by increasing the trigger angle α dc ; During the transient overvoltage period, the power of the sending-end system still satisfies the balance, and the specific formula is as follows:
[0107] P′ acp =P′ PV1 +P′ wind1 +P h1 ′+P st1 ′ (6)
[0108] Q′ acp =Q′ PV1 +Q′ wind1 +Q h1 ′+Q st1 ′ (7)
[0109] Where: P′ acp is the active power generated by the new energy system after the fault occurs, P′ PV1 is the active power transmitted by the photovoltaic power station after the fault occurs, P′ wind1 is the active power transmitted by the doubly-fed wind power after the fault occurs, P h1 ′ is the active power transmitted by the thermal power unit after the fault occurs, P st1 ′ is the active power transmitted by the energy storage after the fault occurs, Q st1 ′ is the reactive power transmitted by the energy storage after the fault occurs; Where: Q′ acp is the reactive power generated by the new energy system after the fault occurs, Q′ PV1 is the reactive power transmitted by the photovoltaic power station after the fault occurs, Q′ wind1 is the reactive power transmitted by the doubly-fed wind power after the fault occurs, Q′h1 The reactive power transmitted by the thermal power unit after a fault occurs;
[0110] Due to the large inertia time of the thermal power and the relatively long response time of the excitation regulation compared to the transient process of commutation failure, the terminal voltage of the thermal power unit remains unchanged. At this time, the voltage at the rectifier-side bus of the HVDC transmission system is:
[0111]
[0112] Where U′ p Is the rectifier-side bus voltage after a fault occurs in the system;
[0113] When the voltage of the converter station filter increases, its reactive power compensation capacity increases. When commutation failure is restored, the reactive power consumed by the converter station changes. The actual reactive power transmitted by the converter station to the sending-end system during the transient process is;
[0114]
[0115] In the formula: U′ p Is the rectifier-side bus voltage after a fault occurs in the system, U p Is the rectifier-side bus voltage of the HVDC transmission system during normal operation of the system, Q cp Is the reactive power generated by the AC filter system, Q’ cp Is the actual reactive power transmitted from the new energy to the AC filter system through the sending-end system of the HVDC transmission system during the transient process;
[0116] Compared with the normal operation state, there is a reactive power surplus at the sending-end AC bus during the transient overvoltage period, and its value is:
[0117] ΔQ = Q h1 -Q′ h1 (10)
[0118] In the formula, ΔQ is the surplus reactive power of the filtering AC system, Q h1 Is the reactive power transmitted by the thermal power unit during normal operation, Q h1 ′ Is the reactive power transmitted by the thermal power unit after a fault occurs.
[0119] The transient voltage change at the rectifier-side AC bus is:
[0120]
[0121] In the formula ΔU P Is the change in the rectifier-side bus voltage;
[0122] The short-circuit capacity in the AC system is
[0123]
[0124] Where: S sc is the short-circuit capacity of the AC system; K CR is the short-circuit ratio of the DC system; Z is the equivalent impedance of the AC system;
[0125] Substitute equations (12) and (13) into equation (11) to obtain the transient voltage calculation method considering the short-circuit ratio, which is used to quantitatively analyze the impact of the short-circuit ratio on the system;
[0126] In the case of commutation failure, the transient voltage rise at the rectifier-side bus of the HVDC transmission system is negatively correlated with the short-circuit ratio, and the specific formula is as follows:
[0127]
[0128] The quantities generated on the rectifier-side bus can directly reflect the system overvoltage and stability. Select the bus voltage U P , the active power P acp on the line, and the reactive power Q acp as the evaluation criteria;
[0129] S200. Establish a deep learning model, and its one-dimensional convolution operation formula is:
[0130]
[0131] Where: X l is the input feature, and
[0132] The convolution kernel is: Stacked by c2 convolution kernels with a size of . c1 is the number of channels of the input feature, and c2 is the number of convolution kernels; t1 is the sequence length of the input feature; the subscript l is the convolution layer number;
[0133] X l+1 is the output feature, f is the non-linear activation function, is the matrix splicing operator, is the bias, and the subscript i represents the i-th component of the corresponding identifier;
[0134] The parameters to be learned in a single-layer one-dimensional convolution network are K l+1 and b l+1 , with a total of c2×h×c1 + c2. At the same time, the time series length t2 of the system output feature X l+1 is:
[0135]
[0136] Where: p is the number of convolution padding layers; s is the convolution step. When p and s are adjusted to a reasonable range, the rapid decline of t2 and the loss of edge information can be avoided;
[0137] In this embodiment, the size of the first convolutional kernel in the deep learning convolutional neural network structure is 1×3, the number of output channels is 16, and 16 kinds of time series features are extracted. The window size of the first pooling layer is 1×2, and the output size is reduced to 1 / 2 of the original size in the width direction. After pooling, the time dimension of the sequence is halved. The size of the second convolutional kernel is 1×3, and the number of output channels is 32; the second pooling layer is similar to the first pooling layer, and the output size is further reduced by 1 / 2 in the width direction. The fully connected layer 1 reduces the dimension of the 64-dimensional global time series features, and finally the transient stability evaluation classification result is output from a single output unit.
[0138] The present invention uses a deep learning network with multi-size convolutional kernels to extract multi-granularity time series features, improves the ability of the model to extract time series information and reduces redundant parameters, ensuring the time series features and data features of the evaluation model; uses the subclass of the clustered dataset as the input features of deep learning, takes transient stability and instability as two output nodes, trains the dataset using a deep learning model, introduces a focal loss function to enhance the sensitivity to unstable samples, and quickly quantitatively evaluates the transient stability of the new energy power system;
[0139] S300. Transient overvoltage evaluation based on CK-means and deep learning
[0140] S310. Model input construction
[0141] S311. Taking the time series measurement data of the synchronous vector measurement unit PMU as the input, and taking the bus voltage amplitude, voltage phase angle, line active power and line reactive power of the sending-end converter station as the initial features, establish an initial feature formula for the time series measurement data, as shown in Equation 18 specifically:
[0142]
[0143] In the formula: U p is the effective value of the phase voltage of the converter bus on the DC rectifier side; θ p is the phase angle of the bus voltage on the rectifier side; P acp is the active power transmitted by a single phase, Q acp is the reactive power transmitted by a single phase;
[0144] m is the bus node number; n is the transmission line number; d is the number of sampling points;
[0145] S312. Decompose the data using the CK - means algorithm. Coarsely cluster the dataset using the Canopy algorithm to determine the number of clusters in CK - means, obtain the initial cluster centers, and then calculate based on the distance from the samples to the centers. Update the cluster centers through calculation. The objective function of the CK - means algorithm uses the squared error criterion function, and after multiple iterations, the objective function converges. The formula for the squared error criterion function is:
[0146]
[0147] where: E is the sum of the squared errors of the dataset composed of bus voltage, active and reactive power, and short - circuit ratio, etc.; x is the position of the sample data, and x i is the average value of cluster C i .
[0148] Decompose the data using the CK - means algorithm. The specific steps are as follows:
[0149] a. Input the dataset X, and set distance thresholds T1 and T2, where T1 > T2;
[0150] b. Establish a new cluster. Select any data point in the outgoing system dataset X and add it to the cluster as the cluster center, and remove x from X;
[0151] c. Calculate the distance d(x, y) from all data points y in X to the data point x;
[0152] d. If d(x, y) < T1, then assign the corresponding data point to the current Canopy. If d(x, y) < T2, then remove the corresponding data;
[0153] e. Repeat steps a - d until the coarse clustering of the samples in X is completed, forming k Canopies and corresponding cluster centers m1, m2, m3... m k ;
[0154] f. Use the cluster centers obtained from Canopy as the initial cluster centers;
[0155] g. Calculate the distance d(x, m1) between each sample data x and the k cluster centers, and classify the results into the categories of the neighboring cluster centers;
[0156] h. Update the cluster centers and calculate the squared error criterion function E of the function;
[0157] i. Repeat steps g and h until E converges.
[0158] Classify the data. Stable samples are divided into two sub - classes, and unstable samples are divided into one sub - class;
[0159] Collect the voltage amplitude, active power, and reactive power through simulation to obtain 392 effective samples. Use the voltage stability criterion formula 21 to divide the samples into stable and unstable parts, with a data ratio of 3:1. Use the Canopy algorithm to perform rough clustering on the dataset to obtain 3 initial clustering centers, and then use the K-means algorithm to perform fine clustering on the dataset. After processing the data with CK-means, the samples are aggregated into 169. The stable samples are divided into two subclasses with 57 each, and the unstable samples are divided into one subclass with 55 samples. Classify the samples into a test set and a training set at a ratio of 0.7.
[0160] S320, Model Output
[0161] Determine the two output nodes of the model, namely transient stability and transient instability, and its formula is:
[0162]
[0163] In the formula is the output of the i-th classification node; z i is the input of the i-th classification node;
[0164] After normalization
[0165] S330, Stability Criterion
[0166] Introduce an index to determine voltage instability of the system under large disturbances;
[0167]
[0168] In the formula, A svsi is the transient voltage stability evaluation coefficient; U min is the electrical performance of the system under fault conditions; U s is the minimum value of voltage collapse; U d is the degree of voltage fluctuation; T u is the factor for whether the system can recover stability, A svsi When this coefficient is 1, it indicates transient voltage stability, and if it is less than 1, it indicates transient voltage instability.
[0169] S340, Focal Loss Function and Classification Evaluation Index
[0170] Perform clustering processing on the collected data, divide the samples into 3 subclasses, introduce the focal loss function (FL) to replace the cross-entropy (CE) loss function to increase accuracy, and the formula of the focal loss function is:
[0171]
[0172] where: α ∈ (0, 1) is the balance factor, β is the modulation factor, and the weight coefficient composed of the model output can reduce the loss function value of simple samples, enabling better prediction for difficult samples; y i is the label of stable samples;
[0173] The single classification accuracy index is difficult to comprehensively evaluate the performance. Therefore, the accuracy A cc , precision P rec , recall R ecall and the comprehensive index F of recall and precision 1-score are used to comprehensively evaluate the identification ability;
[0174] The formula for calculating the accuracy A cc is:
[0175]
[0176] where A cc is the accuracy; T P is the number of voltage stable samples; F P is the number of misjudged voltage stable samples; F N is the number of misjudged voltage unstable samples; T N is the number of voltage unstable samples;
[0177] The formula for calculating the precision P rec is:
[0178]
[0179] where P rec is the precision; T P is the number of voltage stable samples; F P is the number of misjudged voltage stable samples; F N is the number of misjudged voltage unstable samples; T N is the number of voltage unstable samples;
[0180] The formula for calculating the recall R ecall is:
[0181]
[0182] where R ecall is the recall; T P is the number of voltage stable samples; F P is the number of misjudged voltage stable samples; F N is the number of misjudged voltage unstable samples; T N is the number of voltage unstable samples;
[0183] The formula for calculating the comprehensive index F of recall and precision 1-score is:
[0184]
[0185] Wherein, P rec is the precision rate; R ecall is the recall rate.
[0186] The determination confusion matrix is shown in Table 1;
[0187]
[0188] S350, Transient assessment
[0189] The specific steps are as follows:
[0190] S351. Directly obtain a large amount of data from the time-domain simulation model using PMU;
[0191] S352. Preprocess the obtained data, distinguish the data samples using stability criteria, perform clustering decomposition on voltage, active power, and reactive power data using the CK-means algorithm, optimize the decomposed data set again with the focal loss function aiming to minimize the loss function, and handle the data imbalance;
[0192] S353. Use the data processed in the above step S502 as the input of the deep learning model for offline training;
[0193] S354. After introducing the data set and metrics into the model, take as the output, and judge whether is 1 at this time. If it is 1, it means that the power system is operating stably at this time; otherwise, the power system enters an unstable operating state;
[0194] S355. Judge the classification metrics for the above determination results, including the accuracy rate A cc , precision rate P rec , recall rate R ecall and F 1-score to judge whether the evaluation result is accurate this time. If the classification results are FN and Fp, the system evaluation result is incorrect, and at this time, repeat steps S503 - S505; otherwise, the system evaluation result is accurate.
[0195] According to the above implementation process, using the 169 processed data as the input of the neural network, after deep learning, the accuracy rate of the model is obtained. At the same time, the results of evaluating 64 test samples through the model are as follows. It can be seen that the "false stable" and "false unstable" results of the method used in this implementation are significantly lower than those of the traditional CNN method.
[0196] Table 2 shows the output results of the CK-means combined with the deep learning model.
[0197] Stable (TP) False Positive (FP) False Negative (FN) Unstable (TN) 38 1 5 20
[0198] Table 3 shows the output results of the traditional CNN model.
[0199] Stable (TP) False Positive (FP) False Negative (FN) Unstable (TN) 37 2 10 15
[0200] Under the same conditions, the accuracy and recall rate of other methods are calculated, showing a significant improvement compared with the traditional CNN method, indicating that the evaluation model has good accuracy. The specific training process is as Figure 5 shown in Table 4. It is found through the implementation process that this method can effectively improve the accuracy.
[0201] Table 4 is a comparison table of model evaluation
[0202]
[0203] The above are only the preferred examples of the present invention. It should be pointed out that for those of ordinary skill in the art, other equivalent deformations and improvements can also be made under the technical inspiration provided by the present invention, which should also be regarded as the protection scope of the present invention.
Claims
1. A transient overvoltage stability assessment method based on CK-means and deep learning, characterized in that: It includes the following steps: S100. Construct the sending-end system of the new energy through the HVDC transmission system. Combine different fault information of the sending-end system to conduct transient simulation of the fault information, and generate simulation data samples of the converter station bus voltage amplitude, voltage phase angle, line active power, line reactive power of the sending-end system of the new energy through the HVDC transmission system under steady-state operation, and the converter station bus voltage amplitude, voltage phase angle, line active power, and line reactive power of the sending-end system of the new energy through the HVDC transmission system during transient operation. S200. Extract the simulation data samples generated in S100 using the Phasor Measurement Unit (PMU), extract the time-series data, and process the time-series data using the CK-means algorithm composed of the K-means clustering algorithm assisted by the Canopy algorithm. Divide the stable samples into two subclasses and classify the unstable samples into one subclass, and at the same time optimize the squared error objective functions of the stable sample subclasses and unstable sample subclasses. S300. Select the time-series data processed in S200 as input features and establish a deep learning 1D-CNN model. S310. The one-dimensional convolution operation formula of the 1D-CNN model is: where: X l is an input feature, and The convolutional kernel is: obtained by stacking matrices of c2 convolutional kernels with a size of , where c1 is the number of channels of the input feature, c2 is the number of convolutional kernels; t1 and t2 are the sequence lengths of the input feature and the output feature; the subscript l is the convolutional layer number; R is the real number set for the data set; h is the convolutional kernel size; X l+1 is the output feature, f is the non-linear activation function, is the matrix concatenation operator, is the bias term; The learnable parameters of a single-layer one-dimensional convolutional network in a deep learning model are K l+1 and b l+1 , with a total of c2×h×c1 + c2, and at the same time output the feature X l+1 The time series length t2 of In the above formula: t1 and t2 are the sequence lengths of the input features and output features; h is the convolution kernel size; p is the number of convolution padding layers; s is the convolution step. When p and s are adjusted to a reasonable range, it is possible to avoid the rapid decrease of t2 and loss of edge information. S320. The 1D-CNN model outputs Determine the two output nodes of the model, namely transient stability and transient instability, and its formula is: where is the output of the i-th classification node; z i is the input of the i-th classification node; After normalization Take as the output, and judge whether is 1 at this time. If it is 1, it means that the sending-end system of the new energy through the HVDC transmission system is in stable operation; otherwise, the sending-end system of the new energy through the HVDC transmission system enters an unstable operation state; S330. Conduct stability judgment on the input data of the 1D-CNN model. Introduce the judgment index A svsi , which is an index for judging voltage instability when the sending-end system of the HVDC transmission system for new energy suffers from large disturbances; where A svsi is the transient voltage stability evaluation coefficient; U min is the electrical performance of the system under fault conditions; U s is the minimum value at which voltage collapse occurs; U d is the degree of voltage fluctuation; T u is a factor for whether the system can recover stability, T s is the time when the allowed voltage in actual engineering is lower than the minimum value U at which voltage collapse occurs s , generally 10 s; A svsi When it is 1, it indicates that the voltage of the sending-end system of the new energy through the HVDC transmission system is stable. If A svsi is less than 1, it indicates that the transient voltage of the sending-end system of the new energy through the HVDC transmission system is unstable.
2. The evaluation method according to claim 1, characterized in that: In the step S100, when the sending-end system of the new energy through the HVDC transmission system operates stably, the reactive power of the sending-end system of the new energy through the HVDC transmission system is balanced with the reactive power of the rectifier side of the sending-end system of the new energy through the HVDC transmission system. The reactive power consumption of the converter station is shown in the following formula: Where, Q dr is the reactive power of the sending end system of the new energy through the HVDC transmission system, P dr is the active power of the sending end system of the new energy through the HVDC transmission system, α is the triggering angle of the rectifier side of the sending end system of the new energy through the HVDC transmission system, and μ is the commutation overlap angle of the rectifier side of the sending end system of the new energy through the HVDC transmission system; when DC commutation failure occurs at the inverter side bus, the triggering angle α of the sending end system of the new energy through the HVDC transmission system will increase, resulting in a decrease in the active power P dr of the sending end system of the new energy through the HVDC transmission system, and the reactive power Q dr of the sending end system of the new energy through the HVDC transmission system will also decrease; In the case of commutation failure, the transient voltage rise of the rectifier side bus is negatively correlated with the short-circuit ratio, and the specific formula is as follows: Where ΔU P is the change in the rectifier-side bus voltage of the sending-end system of the new energy through the HVDC transmission system before and after the fault occurs; U h is the voltage at the grid connection point of the thermal power unit; P d is the DC transmission capacity of the HVDC transmission system. The bus voltage U P , the active power P acp on the line, and the reactive power Q acp are selected as the evaluation criteria; K CR is the short-circuit ratio of the DC system; ΔQ is the surplus reactive power of the filtering AC system; ΔP is the change in the active power of the filtering AC system.
3. The evaluation method according to claim 1, wherein: In the step S100, the Newton-Raphson method is used for system power flow analysis. When the sending-end system of the new energy through the HVDC transmission system operates under normal conditions, the calculation formula for the active power of the sending-end system of the new energy through the HVDC transmission system obtained according to power balance is: P acp = P PV1 + P wind1 + P h1 + P st1 Where: P acp is the active power generated by the new energy system, P PV1 is the active power transmitted by the photovoltaic power station, P wind1 is the active power transmitted by the doubly-fed wind power, P h1 is the active power transmitted by the thermal power unit, P st1 is the active power transmitted by the energy storage; Reactive power Q of the sending end system of the new energy through the HVDC transmission system acp The calculation formula is as follows: Q acp = Q PV1 + Q wind1 ++ Q h1 + Q st1 Where: Q PV1 is the reactive power transmitted by the PV power station, Q wind1 is the reactive power transmitted by the doubly-fed wind power, Q h1 is the reactive power transmitted by the thermal power unit, Q st1 is the reactive power transmitted by the energy storage, Q dr is the reactive power of the HVDC transmission system; Reactive power Q of the HVDC transmission system dr The calculation formula is as follows: Q dr = Q acp + Q cp where Q cp is the reactive power generated by the filter AC system; Under normal operating conditions, the bus voltage U of the rectifier side of the sending-end system of the HVDC transmission system for new energy is p as follows: P h1 is the active power transmitted by the thermal power unit; Q h1 is the reactive power transmitted by the thermal power unit; U h is the voltage at the grid connection point of the thermal power unit; X sn is the reactance of the thermal power transmission line.
4. The evaluation method according to claim 1, wherein: In the step S100, when a commutation failure fault occurs in the sending end system of the HVDC transmission system for new energy, it is equivalent to a short circuit on the inverter side, the DC current increases rapidly, and the rectifier side consumes more reactive power. The constant current controller reduces the DC current I by increasing the trigger angle α. dc During the transient overvoltage period, the power of the sending end system of the HVDC transmission system for new energy still satisfies the balance, and the active power P' of the sending end system of the HVDC transmission system for new energy acp The calculation formula is: P′ acp = P′ PV1 + P′ wind1 + P h1 ′ + P st1 ′ Where: P' PV1 is the active power transmitted by the PV power station after a fault, P' wind1 is the active power transmitted by the doubly-fed wind power after a fault, P h1 ' is the active power transmitted by the thermal power unit after a fault, P st1 ' is the active power transmitted by the energy storage after a fault, Q st1 ' is the reactive power transmitted by the energy storage after a fault; The calculation formula for the reactive power of the sending-end system of the new energy through the HVDC transmission system is: Q′ acp = Q′ PV1 + Q′ wind1 + Q h1 ′ + Q st1 ′ Where: Q' acp is the reactive power generated by the new energy system after a fault, Q' PV1 is the reactive power transmitted by the photovoltaic power station after a fault, Q' wind1 is the reactive power transmitted by the doubly-fed wind power after a fault, Q h1 ' is the reactive power transmitted by the thermal power unit after a fault.
5. The evaluation method according to claim 1, wherein: In step S100, due to the large inertia time of thermal power output and the relatively long excitation regulation response time compared with the commutation failure transient process, the terminal voltage of the thermal power unit remains unchanged. At this time, the bus voltage of the rectifier side of the HVDC transmission system is: Where U′ p is the rectifier side bus voltage after the system fails, P h1 ′ is the active power transmitted by the thermal power unit after the fault occurs, Q h1 ′ is the reactive power transmitted by the thermal power unit after the fault occurs, U h is the voltage at the grid connection point of the thermal power unit during normal operation, X sn is the reactance of the thermal power transmission line.
6. The evaluation method according to claim 1, characterized in that: In the step S100, when the voltage of the converter station filter increases, its reactive power compensation capacity increases. When the commutation failure is restored, the reactive power consumed by the converter station changes. The actual reactive power transmitted by the converter station to the sending-end system during the transient process is; Where: U' p is the rectifier-side bus voltage after a system fault, and U p is the rectifier-side bus voltage of the HVDC system during normal system operation, Q cp is the reactive power generated by the AC filter system, and Q' cp is the actual reactive power transmitted from the new energy through the sending-end system of the HVDC system to the AC filter system during the transient process; Compared with the normal operation state, the calculation formula for the surplus reactive power at the sending-end AC bus during the transient overvoltage period is: ΔQ = Q h1 -Q' h1 where ΔQ is the surplus reactive power of the filtered AC system, and Q h1 is the reactive power transmitted by the thermal power unit during normal operation, and Q h1 ′ is the reactive power transmitted by the thermal power unit after a fault occurs.
7. The evaluation method according to claim 1, characterized in that: In the step S100, the voltage change amount ΔU of the rectifier side bus of the sending-end system of the new energy through the HVDC transmission system before and after the fault occurs P is calculated by the formula: The transient voltage change at the rectifier side AC bus of the HVDC transmission system is: Where: ΔQ is the surplus reactive power of the filtered AC system, X sn is the reactance of the thermal power transmission line, ΔP is the change in active power of the filtered AC system, U h is the voltage at the grid connection point of the thermal power plant during normal operation.
8. The evaluation method according to claim 1, wherein: The construction of the input of the 1D-CNN model in step S300 includes the following steps: Collect the values of the voltage and power at the rectifier side bus node. Using the timing measurement data of the phasor measurement unit (PMU) as the input, and taking the bus voltage amplitude, voltage phase angle, line active power, and line reactive power of the sending-end converter station of the new energy through the high-voltage direct current (HVDC) transmission system as the initial features, establish the initial features of the timing measurement data, as shown in the following formula: X = [U p1 ,..., U pm , θ p1 ,..., θ pm , P acp1 ,...,P acpn ,Q acp1 ,...,Q acpn , (U pm , θ pm , P acpn , Q acpn ∈ ℝ d ) Where: X is the timing measurement data, U p is the effective value of the phase voltage of the commutation bus on the DC rectifier side, θ p is the phase angle of the bus voltage on the rectifier side. In the above symbols, the serial numbers 1... m are the bus node numbers; P acp is the active power transmitted in single-phase, Q acp is the reactive power transmitted in single-phase. In the above symbols, the serial numbers 1... n are the transmission line numbers; R d is the real number set at d sampling points, and d is the number of sampling points; The process of decomposing data using the CK-means algorithm includes: using the Canopy algorithm to perform rough clustering on the dataset, determining the number of clusters of the CK-means algorithm to obtain the initial cluster centers, and then calculating based on the distance from the samples to the centers, updating the cluster centers through calculation. The objective function of the CK-means algorithm uses the squared error criterion function, and after multiple iterations, the objective function E converges; the formula of the objective function is: where: E is the sum of the squared errors of the data set composed of bus voltage, active power, and reactive power, and x is the position of the sample data. is the cluster C i average value of; Classify the data, and the stable samples are divided into two subclasses, and the unstable samples are divided into one subclass.
9. The evaluation method according to claim 1, wherein: When performing cluster analysis on the collected data in S200, introduce the focal loss function FL to replace the cross-entropy CE to enhance the prediction accuracy of difficult samples, and use the training sample set to train the 1D-CNN model to obtain a transient overvoltage judgment model for the sending-end system of the new energy through the HVDC transmission system; input the parameters of the bus voltage amplitude, voltage phase angle, line active power, and line reactive power of the converter station of the sending-end system of the new energy through the HVDC transmission system constructed by the time-series data into the 1D-CNN model to obtain the transient voltage stability evaluation coefficient of the sending-end system of the new energy through the HVDC transmission system; The loss function increases the accuracy, and the formula of the focal loss function is: where: α ∈ (0, 1) is the balance factor, γ is the modulation factor, and the weight coefficient composed of the model output can reduce the loss function value of simple samples, enabling it to have better prediction for difficult samples; y i is the label of stable samples.