Differential protection fault identification method and device based on machine learning model

Abstract

The embodiment of the invention discloses a differential protection fault identification method and device based on a machine learning model, and relates to the technical field of power system relay protection, and the method comprises the steps: obtaining current signals at two sides of a protected device, and extracting a time sequence feature sequence; a feedforward neural network comprising a memory module is adopted for processing, and the memory module generates historical information representation by applying learnable non-cyclic transformation to the historical hidden state of the hidden layer; the historical information representation is spliced with the current hidden state, and a result is transmitted to the next layer; and finally, training by adopting a standard error back propagation algorithm. The invention further provides a corresponding device. By introducing a full feed-forward structure of the memory module, time sequence characteristics of current signals can be effectively captured, meanwhile, the gradient problem of a cyclic network is avoided, stable and efficient training is achieved, high accuracy and strong generalization ability are achieved in a small sample scene, and the reliability of differential protection is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system relay protection, and in particular to a differential protection fault identification method and device based on a machine learning model. BACKGROUND

[0002] Differential protection is the main protection for key power equipment such as generators, transformers and busbars, and its reliability is crucial to the safe and stable operation of the power system. When a short-circuit fault occurs in the power system, the short-circuit current may contain a large non-periodic component, i.e. a direct current component. This direct current component will cause the core of the current transformer to enter a saturated state, resulting in severe distortion of the current waveform output at the secondary side of the current transformer, such as waveform clipping and zero value platform. This distortion will produce a differential current with a huge amplitude, which may cause the differential protection device to misjudge as an internal fault and act incorrectly, seriously threatening the safety of the power grid.

[0003] To solve this problem, traditional differential protection methods usually rely on analysis and discrimination of the steady-state or fundamental frequency phasor characteristics of electrical quantities. However, in the complex transient process caused by current transformer saturation, these traditional methods are difficult to effectively extract key features from the transient signal, and thus cannot accurately distinguish between real internal faults and pseudo-differential currents caused by saturation, and their reliability needs to be improved.

[0004] In recent years, some research has attempted to apply machine learning techniques to the field of fault identification. For example, some methods use recurrent neural networks or long short-term memory networks to process the time series characteristics of current signals. However, such models based on recurrent structures require backpropagation algorithms along the time direction during training, and when processing long sequence data, they are prone to gradient vanishing or gradient explosion problems, which not only lead to unstable model training processes and convergence difficulties, but also consume a lot of time. In addition, in the field of relay protection, the amount of real fault sample data available for model training is usually limited, and traditional deep learning models are prone to overfitting on small data sets, which affects the generalization ability of the model and the final performance of fault identification. SUMMARY

[0005] The main purpose of the present application is to provide a differential protection fault identification method based on a machine learning model, which can solve the problem of insufficient reliability of current differential protection methods in the case of current transformer saturation, and the problems of unstable training, low efficiency and easy overfitting on small data sets when applying existing machine learning methods based on recurrent structures to this scenario.

[0006] To achieve the above purpose, the first aspect of the present application provides a differential protection fault identification method based on a machine learning model, which comprises: obtaining current signals on both sides of a protected device, and extracting a time series sequence containing a plurality of time points from the current signals as input of a model; processing the time series sequence by using a feedforward neural network, the feedforward neural network comprising at least one hidden layer containing a memory module, the memory module being configured to generate a historical information representation by applying a learnable acyclic transformation to a preset number of historical hidden states of the hidden layer; concatenating the historical information representation with a current hidden state of the hidden layer, and passing the concatenated result to a next layer for processing; training the feedforward neural network end to end by using a standard error back propagation algorithm to obtain a target model for differential protection fault identification.

[0007] The second aspect of the present application provides a differential protection fault identification device based on a machine learning model, comprising: an obtaining module configured to obtain current signals on both sides of a protected device, and extract a time series sequence containing a plurality of time points from the current signals as input of a model; a processing module connected with the obtaining module, the processing module being internally disposed with a trained feedforward neural network, the feedforward neural network being configured to process the time series sequence to output a fault identification result; the feedforward neural network comprising at least one hidden layer containing a memory module; the memory module being configured to generate a historical information representation by applying a learnable acyclic transformation to a preset number of historical hidden states of the hidden layer; the feedforward neural network being further configured to concatenate the historical information representation with a current hidden state of the hidden layer, and pass the concatenated result to a next layer for processing.

[0008] The third aspect of the present application provides an electronic device comprising a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to cause the processor to perform the steps of the first aspect and any possible implementation manner thereof.

[0009] The fourth aspect of the present application provides a computer readable storage medium, the computer program being executed by the processor to cause the processor to perform the steps of the first aspect and any possible implementation manner thereof.

[0010] The technical scheme provided by the application has the following beneficial effects: 1. The accuracy and reliability of fault identification are improved. The application can effectively capture and process the complex time sequence characteristics of the current signal under the saturation condition of the current transformer by introducing a memory module in the feedforward network, thereby significantly improving the accuracy of distinguishing real faults from saturation interference and enhancing the reliability of differential protection. 2. Stable and efficient model training is achieved. The application adopts a fully feedforward network structure, which can be directly trained using the standard error back propagation algorithm, fundamentally avoiding the common gradient disappearance or gradient explosion problem in recurrent neural networks, making the model training process more stable, faster in convergence speed and shorter in training time. 3. The generalization ability in the small sample scene is enhanced. The model structure proposed by the application is relatively simple, and in the small data set scene of limited fault samples in the field of relay protection, it shows superior anti-overfitting ability and generalization performance, solving the problem that traditional deep learning models rely on massive data and having more engineering practical value. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical schemes in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0012] Among them: Figure 1 A differential protection fault identification system architecture provided by the embodiment of the application; Figure 2 A differential protection fault identification method based on a machine learning model provided by the embodiment of the application; A flowchart of the differential protection fault identification method based on a machine learning model provided by the embodiment of the application; Figure 3 A network framework diagram of the FSMN provided by the embodiment of the application; Figure 4 A diagram showing the change of F1 score of the sFSMN and vFSMN network with the order provided by the embodiment of the application; Figure 5 A diagram showing the change of AUC of the FSMN and vFSMN network with the order provided by the embodiment of the application; Figure 6 A structure diagram of a differential protection fault identification device based on a machine learning model provided by the embodiment of the application; Figure 7 A structure diagram of an electronic device provided by the embodiment of the application. DETAILED DESCRIPTION

[0013] In order to make the person skilled in the art better understand the scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor fall within the scope of protection of the present application.

[0014] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0015] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily independent or alternative embodiments to each other. The person skilled in the art explicitly and implicitly understands that the embodiments described herein can be combined with other embodiments .

[0016] The principle of differential protection involved in the embodiments of the present application is introduced as follows: The theoretical basis of differential protection is the balance relationship of electrical quantities. Taking line differential protection as an example, suppose that the transformation ratios of the current transformers at both ends of the line are K 1 and K 2, the primary side currents are I 1 and I 2, and the secondary side currents transformed by the current transformers are I 1 ' and I 2 ' When normally operating or a fault occurs outside the protected area, it should satisfy I 1 ' + I 2 ' = 0. When a fault occurs in the protected area, the vector sum of the currents at both ends is no longer zero, forming a differential current Id =| I 1 ' + I 2 ' |

[0017] In differential protection, the core criterion for whether the protection action is triggered is: (1) wherein, I set is the action threshold value; I res is the braking current, the value of which is usually half of the sum of the absolute values of the two-terminal currents; K is the braking coefficient, generally taken as 0.5-0.7.

[0018] In actual application, sampling value differential protection is usually adopted for operation. The sampling value differential protection is a protection implementation mode based on instantaneous sampling values, and has higher real-time performance compared with the traditional effective value differential protection.

[0019] The sampling value differential protection implementation mode is to compare the current values of each sampling point, and the criterion form is: (2) wherein, id ( k ) is the differential current of the i th sampling point, k res( i ) represents the braking current, k set( i ) is the instantaneous action threshold value. In order to avoid malfunctions caused by interference, the sampling value differential protection usually adopts a "majority voting" mechanism. If the number of points satisfying the criterion is not less than n in continuous m sampling points, the protection action is triggered. k N M

[0020] The embodiments of the present application will be described below in combination with the drawings in the embodiments of the present application.

[0021] First, the technical environment to which the embodiments of the present application are applied is described.

[0022] Referring to FIG. 1, a differential protection fault identification system architecture according to an embodiment of the present application is shown. The system can be applied in a power system to protect key power equipment 1 such as a transformer, a generator, a bus or a transmission line. On both sides of the power equipment 1, current transformers 10a and 10b are connected, respectively, the functions of which are to measure the line currents flowing through both sides of the power equipment 1 in real time, and to convert them into secondary current signals suitable for processing by the protection device in proportion. Figure 1 The secondary side output ends of the current transformers 10a and 10b are both connected to a data acquisition and processing unit 20.

[0023] The secondary side output ends of the current transformers 10a and 10b are both connected to a data acquisition and processing unit 20.

[0024] ​​​​The data acquisition and processing unit 20 is the core execution entity of the method described in the embodiments of the present application, and its hardware implementation can include but is not limited to a digital signal processor, a field programmable gate array, an application specific integrated circuit, or an industrial computer composed of a central processing unit, a graphics processing unit, a memory and the like. The data acquisition and processing unit 20 is used to receive the secondary current sampling values from the current transformers 10a and 10b, and perform a series of processing on the collected data. Specifically, the data acquisition and processing unit 20 internally deploys the feedforward sequence memory network model 30 proposed in the embodiments of the present application. The model is pre-trained and can analyze the input current data features, and then identify whether the current system state is normal operation, external fault or internal fault requiring protection action. When the data acquisition and processing unit 20 determines that an internal fault occurs through the feedforward sequence memory network model 30, it will immediately issue a control instruction through the protection action output 40, such as sending a trip signal to the circuit breaker, thereby quickly removing the faulty element to ensure the safety and stability of the power equipment 1 and the entire power system.

[0025] Please refer to Figure 2 which is the overall flowchart of a differential protection fault identification method provided by the embodiments of the present application. As shown in Figure 2 the method mainly includes the following steps: Step S101, obtaining and preprocessing current signals. In this step, the data acquisition and processing unit 20 continuously collects the instantaneous sampling values of the secondary current from the current transformers 10a and 10b. The preprocessing process can include filtering the signals to remove high-frequency noise, and synchronizing the data to ensure that the current data on both sides are aligned in time.

[0026] Step S102, extracting a time sequence of features. From the preprocessed current signals, a series of time sequence features that can reflect the system operating state and the saturation characteristics of the current transformer are calculated and extracted. These features constitute a feature sequence or a feature vector sequence at multiple consecutive time points.

[0027] Step S103, inputting the feature sequence to the feedforward sequence memory network model. The time sequence of features containing multiple time points extracted in step S102 is fed as input to the trained feedforward sequence memory network model 30.

[0028] Step S104, performing model processing. The feedforward sequence memory network model 30 performs forward propagation calculation on the input time sequence of features. In this process, the memory module inside the model aggregates historical information and combines it with the information at the current time to capture the dynamic time sequence dependence of the signal.

[0029] Step S105, output the fault identification result. The model outputs a discriminant value, such as a value representing the probability of an internal fault, at the output layer according to the calculation result of the forward propagation. The data acquisition and processing unit 20 finally determines whether it is an internal fault according to the comparison result of the value and the preset threshold, and decides whether to issue an action instruction through the protection action outlet 40.

[0030] The embodiment details the specific implementation of a differential protection fault identification method based on a vector feedforward sequence memory network. The method corresponds to the complete process shown in FIG. 1. Figure 2

[0031] Specifically, in step S101, the current signal is acquired and preprocessed. The data acquisition and processing unit 20 acquires the instantaneous values of the three-phase currents from the current transformers 10a and 10b at a high sampling frequency (e.g., 128 points per power frequency cycle), and the acquired data forms a time sequence. In order to facilitate model processing, a fixed-length time window (e.g., containing data of the past 2 power frequency cycles) is usually selected as the input for one discrimination. The original data collected is normalized, such as using the minimum-maximum normalization method, to scale all data values to the interval [0, 1] or [-1, 1]. This is intended to avoid the adverse effects of features with different dimensions on model training, and to help accelerate model convergence.

[0032] Subsequently, in step S102, the time sequence of features is extracted. For each sampling time point, a group of rich features is extracted from the normalized current signal to form a feature vector at that time point. Combining the feature vectors of consecutive time points forms the time sequence of features input to the model. In this embodiment, in order to comprehensively represent the system state, especially the signs of current transformer saturation, the extracted features can include but are not limited to: Primary and secondary current amplitudes: the amplitudes of the currents at the same side are obtained by appropriate transformation (e.g., for transformer protection, the compensation of transformation ratio and wiring mode needs to be considered).

[0033] Differential current: the vector difference between the primary and secondary currents after scaling. Its value is small in normal or external fault, but significantly increases in internal fault or current transformer saturation.

[0034] Ratio and phase difference: reflects the difference in amplitude and phase between the primary and secondary currents, and is an important criterion for traditional differential protection.

[0035] DC component: the non-periodic component extracted from the current signal by a low-pass filter or moving average method, which is the main cause of current transformer saturation.

[0036] ​Harmonic components: The third and fifth harmonic components in the current signal are calculated using methods such as Fast Fourier Transform. Current transformer saturation generates abundant harmonics, especially odd harmonics.

[0037] Discontinuity angle: refers to the electrical angle corresponding to the length of time during which the secondary current waveform is zero within one power frequency cycle during the saturation period of a current transformer. The magnitude of the discontinuity angle directly reflects the severity of saturation.

[0038] Excitation current: According to the basic principle of current transformer, the excitation current is equal to the primary current minus the secondary current (both referred to the same side). Under saturation conditions, the excitation current will increase sharply and become distorted.

[0039] Hysteresis loop parameters: The magnetic flux is obtained by integrating the secondary side voltage (which can be estimated by the secondary current through the resistance and leakage reactance of the transformer winding), and the magnetic flux-excitation current relationship curve (i.e., the hysteresis loop) is plotted. Then, parameters such as area, slope, and inflection point are extracted from the curve as features.

[0040] 10% error curve parameters: Based on the 10% error curve of the current transformer, parameters such as the transient saturation factor are calculated in conjunction with the current.

[0041] Transient time constant: The transient time constant of the primary system, used to reflect the decay rate of the DC component in the short-circuit current. The values ​​calculated from the above features at each sampling point are combined into a feature vector, and the feature vectors of multiple consecutive sampling points constitute the time-series feature sequence of the input model.

[0042] Next, in steps S103 and S104, the feedforward sequence memory network model is used for processing. The feedforward sequence memory network model constructed in this embodiment can be a fully feedforward neural network, which includes: an input layer for receiving the temporal feature sequence generated in step S102, multiple hidden layers, and an output layer for outputting the final classification result.

[0043] In one embodiment of this application, at least one hidden layer includes a memory module 210. As an optional implementation, the hidden layer containing the memory module 210 is the first hidden layer of the network. This design enables the capture and encoding of temporal information directly from the original input feature sequence at an early stage of the network.

[0044] The feedforward sequential memory network model in the embodiments of this application can be FSMN (Feedforward Sequential Memory Network).

[0045] Figure 3 This is a schematic diagram of a network framework for an FSMN provided in an embodiment of this application.

[0046] Specifically, the FSMN network mentioned in this application embodiment achieves the function of memorizing time information by adding several memory modules in the hidden layer. It is essentially a standard fully connected feedforward neural network. Figure 3 (a) gives an example of the first An FSMN network with a memory module added to each hidden layer. The memory module uses, for example... Figure 3 (b) shows a tap delay structure, the function of which is to delay the tap of the first tap. Hidden layer N Each historical state is represented as a fixed-dimensional code (called...). N The current hidden layer state (FSMN) is input into the next hidden layer along with the current hidden layer state.

[0047] Depending on the encoding method, the FSMN network proposed in this paper can be divided into the following two forms: Scalar FSMN (sFSMN): Uses scalars as coefficients for historical state encoding; Vector FSMN (vFSMN): Uses vectors as coefficients to encode historical states.

[0048] For the input sequence X = { x 1,..., xt ,..., xT} ,in Indicates time t The input, the first The output sequence of each hidden layer is represented as follows: ,in, .exist N In the sFSMN of order 1, when time is t Time memory module output It can be represented as: i (3) in, It is a set of scalar coefficients that can be learned through training.

[0049] exist N In the vFSMN of order, then use N +1 vector coefficient to encode the historical state: h i (4) Among them, symbols This represents the Hadamard Product, or dot product.

[0050] This is a set of vector coefficients. In sFSMN, the outputs of all hidden nodes at the same time share the same encoding. coefficient Its advantage lies in the ability to extend the model to higher orders by introducing fewer parameters. And in vFSMN... In this model, different coefficients can be assigned to the outputs of different hidden nodes, which significantly improves the expressive power of the model.

[0051] Furthermore, the FSMN network can be expanded into a bidirectional structure, simultaneously incorporating historical and future information:

[0052] in, N 1 represents the order of the lookback sequence. N 2 represents the look-ahead order. Theoretically, bidirectional FSMNs have stronger modeling capabilities. However, because bidirectional structures require future information, they introduce greater delays, making them unsuitable for scenarios with stringent latency requirements, such as differential protection.

[0053] In the FSMN network, the output of the memory module Compared with the output of the current hidden layer If the input is passed to the next layer, the output of the next hidden layer will be: (6) in, and For the weight matrix and bias vector, For memory module to the next hidden layer The weight matrix, f (.) denotes the activation function. The paper uses the Sigmoid function as the activation function, i.e.: f (7) The FSMN in this embodiment can adopt a pure feedforward structure and can be trained by the standard error backpropagation algorithm (BP). The learning process is more stable and efficient, and there is no need to use the backpropagation algorithm along time, thus avoiding the problems of gradient vanishing or gradient exploding.

[0054] As a specific implementation, the feedforward sequence memory network model in this embodiment can be configured to contain 3 hidden layers, each containing 256 neurons. For the vector feedforward sequence memory network, based on experimental optimization, the order N of its memory module can be set to 9, that is, to aggregate the historical hidden states of the past 9 time points.

[0055] During the model training phase, the standard error backpropagation algorithm is used to train the entire network end-to-end. For example, 500 simulated or measured data samples containing different fault types, fault locations, and saturation levels can be prepared and divided into training and test sets according to a required ratio (e.g., 8:2). During training, stochastic gradient descent can be used as the optimizer, with a learning rate set to 0.005. The loss function can be the cross-entropy loss function, used to measure the difference between the fault probability predicted by the model and the true label (e.g., 1 for in-zone faults, 0 for out-of-zone faults). The training process is iterative, for example, setting the maximum... The maximum number of training rounds is 50, until the model's performance on the validation set no longer improves or reaches the preset convergence criterion.

[0056] Finally, in step S105, the fault identification result is output. The trained model is deployed in the data acquisition and processing unit 20. When a new current signal is input, the model performs a complete forward propagation calculation, and the output layer (usually using the sigmoid activation function) outputs a probability value between 0 and 1. If a threshold is set (e.g., 0.5), when the output probability is greater than 0.5, it is determined to be an in-zone fault, and the data acquisition and processing unit 20 immediately drives the protection action output 40 to operate; if the probability is less than or equal to 0.5, it is determined to be a normal or out-of-zone fault, and the protection remains inactive.

[0057] The solution provided in this embodiment, through the structure of a vector feedforward sequence memory network, can effectively capture the complex temporal dynamics in current signals, especially the transient process of waveform distortion when the current transformer is saturated, thereby achieving high-precision fault identification.

[0058] In an alternative implementation, as a variation of the foregoing embodiment, the difference lies in the specific method by which the memory module 210 generates historical information representations.

[0059] The main difference lies in the model processing stage of step S104. This embodiment uses a scalar feedforward sequence memory network, where the memory module 210 generates historical information representation h- tl At that time, a set of learnable scalar coefficients is used to perform a weighted summation of the historical hidden states, the mathematical expression of which is: (8) in, It is still the hidden state vector of the l-th layer at past time ti, while the aggregation coefficients Then it is a scalar value. That is, all hidden nodes at time ti (i.e., All elements in a vector share the same weight. .

[0060] The subsequent calculation steps are consistent with those in the aforementioned embodiments: the generated historical information is represented. Compared to the current original hidden state The layers are concatenated, and then the final hidden layer output is obtained through a non-linear activation function. And pass it on to the next layer.

[0061] As a specific implementation, the network structure in this embodiment can also be set to three hidden layers, each with 256 neurons. For scalar feedforward sequence memory networks, due to their relatively small model capacity, longer historical information may be required to achieve similar performance. Based on experimental optimization, the order of its memory module... N can be set to 13. The training process is similar to that described above, using the standard error backpropagation algorithm and the stochastic gradient descent optimizer. Due to the smaller number of model parameters, its training time is usually shorter.

[0062] This embodiment provides a technical solution that achieves a balance between performance and efficiency. Although its refined modeling capability may theoretically be weaker than that of vector models, in many practical application scenarios, especially in embedded protection devices with limited computing resources, this solution can achieve a fault identification accuracy far superior to traditional methods with faster speed and less resource consumption, thus having high practical value.

[0063] In one optional implementation, the timing feature sequence further includes a DC-zero plateau coupling coefficient, which is used to characterize the coupling relationship between the DC component of the secondary current, the average length of the zero plateau of the secondary current within one power frequency cycle, and the transient time constant of the primary system.

[0064] In the feature extraction stage of step S102, in addition to extracting the basic features described above, an additional composite feature, namely the "DC-zero plateau coupling coefficient", is calculated and introduced, denoted as K. dc-zp .

[0065] Specifically, the calculation process for this composite feature may include: The input is the secondary current waveform from the current transformer. This waveform signal is fed in parallel to the DC component extraction module and the zero-value plateau detection module. The DC component extraction module is used to calculate the average DC component magnitude I of the secondary current over one or more power frequency cycles. dc Then, it is normalized to eliminate the amplitude effect, resulting in I. dc,norm The zero-plateau detection module is used to detect the "zero-plateau" phenomenon in the secondary current waveform caused by deep saturation. Specifically, it identifies the time period during which the current value remains within a small threshold range near zero, and calculates the total or average length of all these zero-plateaus within one power frequency cycle, denoted as T. zp .

[0066] At the same time, the system also receives an external input, namely the transient time constant of the primary system. This parameter can be obtained by pre-calculating or online estimating the line parameters and is used as the input time constant.

[0067] Finally, based on the normalized DC component Idc,norm, the average length of the zero plateau Tzp, and the primary side time constant... The coupling coefficient for these three quantities is calculated using the following formula:

[0068] The final output K is obtained. dc-zp .

[0069] The design of this composite feature incorporates relevant physical mechanisms. DC component (I dc,norm The zero-value plateau length (T) is a contributing factor to current transformer saturation. zp This is a direct reflection of the severity of saturation, and one instance... Side time constant ( The decay rate of the DC component is determined by the DC component, which in turn affects the duration of saturation. Coupled with these three factors into a single feature, the model can be provided with highly condensed and discriminative information that links the cause and effect of saturation and its dynamic process.

[0070] After feature extraction step S102 is completed, the calculated K will be... dc-zp This enhanced temporal feature sequence is concatenated with all other features at each time point to form a higher-dimensional feature vector. This enhanced temporal feature sequence is then fed into the aforementioned vector feedforward sequence memory network model for training and recognition. The model structure and training parameters can remain unchanged.

[0071] Because the input information includes prior knowledge based on physical mechanisms, the model can more easily learn the intrinsic laws governing saturation phenomena, rather than simply fitting data from a large amount of data. Using the method in this embodiment helps improve the model's fault identification accuracy, especially in distinguishing between deep saturation caused by faults outside the severe saturation zone and faults within the true saturation zone, potentially achieving higher performance metrics.

[0072] In an alternative implementation, the training method of the model is improved based on the foregoing embodiments by introducing a fault identification method based on physical constraint regularization. This method incorporates known physical laws as prior knowledge into the model training process, aiming to improve the model's generalization ability and robustness on small sample data.

[0073] This embodiment is completely identical to the previous one in terms of data acquisition (step S101), feature extraction (step S102), and model architecture (step S103). The core difference lies in the model training stage in step S104, specifically in the design of the loss function.

[0074] In standard model training, the loss function typically contains only one classification loss term L for the supervised main task (i.e., fault classification). cls For example, cross-entropy loss. In contrast, this embodiment employs a composite loss function L. total Its form is as follows: (10) The composite loss function consists of two parts: 1. Classification loss term L cls : Used to measure the difference between the fault category predicted by the model and the true label, driving the model to learn how to classify correctly.

[0075] 2. Physical constraint loss term L phy This is a newly added regularization term that uses a fundamental physical law of current transformers as a constraint to guide the learning direction of the model.

[0076] Specifically, this physical constraint is based on the ampere-turn balance principle of current transformers, that is, the ampere-turns of the primary current are approximately equal to the sum of the ampere-turns of the secondary current and the ampere-turns of the magnetizing current. After neglecting the turns ratio, this relationship can be simplified to: Magnetizing current (iex) = Primary current (i1) - Secondary current (i2).

[0077] To embed this law into the model, an auxiliary task can be designed. For example, after one intermediate hidden layer of the backbone network (such as the vector feedforward sequence memory network in Example 1), an auxiliary output head consisting of one or more fully connected layers can be added. The task of this auxiliary output head is to predict the current excitation current value i at the current moment based on the state of the hidden layers it is connected to.ex,pred .

[0078] Accordingly, the physical constraint loss term L phy Defined as the model-predicted excitation current i ex,pred With the actual excitation current i ex,true The deviation between them. Among them, the actual excitation current i ex,true It can be directly calculated from the input primary and secondary current data. The deviation can be measured using methods such as mean square error (MSE).

[0079] in T It is the length of the time window; λ is a hyperparameter used to balance the importance of the main task loss and the physical constraint loss. During training, the gradient of the total loss Ltotal with respect to the network parameters is calculated using the backpropagation algorithm. Thus, when updating the weights, the model must not only minimize the classification error but also ensure that its internal feature representation can accurately predict the excitation current that conforms to the physical laws.

[0080] This approach provides a physical guide to the model's learning process, ensuring that the learned feature representations are not only effective for classification tasks but also maintain physical consistency and interpretability. When training data is limited, this physical prior knowledge helps prevent the model from learning spurious associations in the data, thereby improving the model's generalization ability and reliability in unseen conditions.

[0081] By implementing this embodiment, the trained feedforward sequence memory network model is constrained to some extent by physical laws in its internal decision-making process, making the fault identification results more reliable.

[0082] To evaluate the performance of the proposed algorithm, the following six algorithms were selected for comparative experiments: the CART decision tree-based method (CART method for short), the support vector regression-based method (SVR method for short), and the support vector regression-based method (SVR method for short). The methods are referred to as SVR method, FCN method, LSTM method, sFSMN method, and vFSMN method.

[0083] In the experiments, the parameter settings for each comparison algorithm were as follows: In the CART method, the number of decision trees was set to 100, the maximum tree depth was set to 5 layers, the number of samples required for internal node splitting and the minimum number of samples required for leaf node splitting were both 5, the attribute splitting criterion was information entropy gain, and the dropout rate was set to 10% to prevent overfitting. In the SVR method, a binomial kernel function was used, with a regularization parameter of 0.8, a kernel bandwidth of 1.0, a maximum number of iterations of 100, and a maximum error of 0.001. In the FCN method, 3 hidden layers were used, each with 256 nodes, a sigmoid activation function was used, and the learning rate was set to 0.005. In the LSTM method, 3 hidden layers were also used, each containing 256 units, with a learning rate of 0.005, the Adam algorithm was used for training, the regularization coefficient was set to 0.2, the dropout rate was set to 10%, and the sigmoid function was also used as the activation function. In both sFSMN and vFSMN methods, the hidden layers are set to 3 layers with 256 nodes per layer, the learning rate is set to 0.005, and the memory module is placed in the first hidden layer, using the sigmoid activation function. In FCN, sFSMN, and vFSMN, the backpropagation algorithm is used for training, with a maximum of 50 training epochs.

[0084] Experimental results: In this experiment, the following three evaluation metrics were used to evaluate the above model: F1 score (classification threshold set to 0.5), AreaUndertheCurve (AUC), and training time (training time is the average of 20 training iterations).

[0085] In sFSMN and vFSMN networks, the order N The choice of order is quite important. Therefore, the experiment first investigates the order. N Impact on the performance of sFSMN and vFSMN methods.

[0086] Figure 4 , Figure 5 The F1 scores and AUC of the sFSMN and vFSMN methods on the test set as a function of order are presented respectively. N The changes.

[0087] from Figure 4 , Figure 5 As can be seen, the performance of both sFSMN and vFSMN methods increases with the order. N The value increases first and then decreases. Among them, the sFSMN method... N Performance is optimal at 13, while vFSMN The method is inN The performance is optimal when the order is 9. Therefore, in subsequent experiments, the order of the sFSMN method is... N The order is set to 13, which is the order of the vFSMN method. N Set it to 9.

[0088]

[0089] Table 1 Table 1 shows the performance of different algorithms. As can be seen from Table 1, due to their lower modeling ability, the CART and SVR methods perform significantly worse on the training set than other methods. Among the remaining four methods, the LSTM method outperforms the other three on the training set due to its strong Long Short-Term Memory (LSTM) capability. Among FCN, sFSMN, and vFSMN methods, vFSMN performs best on the training set, while FCN performs worst. This is because the order of strength in modeling time series sequences is vFSMN, sFSMN, and FCN. On the test set, the LSTM method performs significantly worse than its performance on the training set, but is comparable to the FCN method, indicating a severe overfitting problem. Furthermore, the LSTM network is prone to gradient vanishing, resulting in a significantly longer training time than other methods. vFSMN performs best on the test set, followed by sFSMN. However, the training time of vFSMN is longer than that of sFSMN. This is because both sFSMN and vFSMN are fully feedforward networks with simple network structures and memory modules, which have a significant advantage in processing small time series datasets.

[0090] In summary, since there is currently no publicly available dataset on differential protection of current transformers, this application proposes scalar feedforward sequence memory networks (sFSMN) and vector feedforward sequence memory networks (vFSMN) to address the overfitting and gradient vanishing problems that easily occur in machine learning algorithms on small datasets. Both the proposed sFSMN and vFSMN methods belong to fully feedforward neural networks, have simple structures, can be trained using the standard backpropagation algorithm, and can better avoid the gradient vanishing problem. Furthermore, the introduction of memory modules makes them more suitable for processing time-series sequences. Simulation results show that the proposed sFSMN and vFSMN methods outperform other comparative methods. Furthermore, it boasts relatively high training efficiency. Among them, the vFSMN method has stronger modeling capabilities, resulting in superior performance. However, due to its more complex structure, its training efficiency is lower than that of the sFSMN method.

[0091] Based on the description of the foregoing method embodiments, this application also provides a differential protection fault identification device based on a machine learning model. For example...Figure 6 As shown, the differential protection fault identification device 600 based on a machine learning model includes: The acquisition module 610 is used to acquire the current signals on both sides of the protected equipment and extract the time-series feature sequence containing multiple time points from the current signals as the input of the model. The processing module 620 is connected to the acquisition module 610. The processing module 620 has a trained feedforward neural network deployed inside it. The feedforward neural network is used to process the time series feature sequence to output the fault identification result. The aforementioned feedforward neural network includes at least one hidden layer containing a memory module; The aforementioned memory module is configured to generate a historical information representation by applying a learnable non-cyclic transformation to a preset number of historical hidden states of the aforementioned hidden layer. The aforementioned feedforward neural network is also configured to concatenate the aforementioned historical information representation with the current hidden state of the aforementioned hidden layer, and pass the concatenated result to the next layer for processing.

[0092] Understandably, this involves Figure 6 The relevant content of each module in the above method embodiments has been described in detail, and you can refer to the content of the method embodiments for details; that is... Figure 6 The provided differential protection fault identification device 600, based on a machine learning model, can perform tasks such as... Figure 2 Any steps in the illustrated embodiment or its variations are not described in detail here.

[0093] In one embodiment of this application, an electronic device is also provided. See also... Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 700 includes a processor 701 and a memory 702. The memory 702 stores a computer program, which, when executed by the processor 701, will perform actions such as... Figure 2 Any step in the method embodiment shown can be a control method step in the experimental process, such as controlling and adjusting the transformer ratio of the voltage regulator, monitoring the phase difference, etc. The electronic device 700 may also include input / output devices, etc. In a specific embodiment, the electronic device can be a terminal device, etc.

[0094] In one embodiment, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor 701, causes the processor 701 to perform any of the steps in the above method embodiments.

[0095] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A differential protection fault identification method based on a machine learning model, characterized in that, The method The law includes: Acquire the current signals on both sides of the protected device, and extract the time-series feature sequence containing multiple time points from the current signals as the input of the model; The time-series feature sequence is processed by a feedforward neural network, which includes at least one hidden layer containing a memory module. The memory module generates a historical information representation by applying a learnable non-cyclic transformation to a preset number of historical hidden states of the hidden layer. The historical information representation is concatenated with the current hidden state of the hidden layer, and the concatenated result is passed to the next layer for processing. The feedforward neural network is trained end-to-end using the standard error backpropagation algorithm to obtain a target model for differential protection fault identification.

2. The differential protection fault identification method based on a machine learning model according to claim 1, characterized in that, The learnable non-cyclic transformation includes at least one of the following: The Hadamard product is performed between a set of learnable vector coefficients and the historical hidden state. Alternatively, the historical hidden states can be weighted and summed using a set of learnable scalar coefficients.

3. The differential protection fault identification method based on a machine learning model according to claim 1, characterized in that, The hidden layer containing the memory module is the first hidden layer of the feedforward neural network.

4. The differential protection fault identification method based on a machine learning model according to claim 1, characterized in that, The time-series feature sequence includes at least one of the following features: The parameters include the amplitude of the primary and secondary currents, differential current, ratio difference, phase difference, DC component, harmonic component, discontinuity angle, excitation current, hysteresis loop parameters, 10% error curve parameters, and transient time constant.

5. The differential protection fault identification method based on a machine learning model according to claim 4, characterized in that, The time-series feature sequence also includes a DC-zero plateau coupling coefficient, which is used to characterize the coupling relationship between the DC component of the secondary current, the average length of the zero plateau of the secondary current in one power frequency cycle, and the transient time constant of the primary system.

6. The differential protection fault identification method based on a machine learning model according to claim 1, characterized in that, The method further includes: The training of the feedforward neural network employs a composite loss function, which includes a classification loss term for supervising the fault identification task and a physical constraint loss term based on the physical laws of current transformers.

7. The differential protection fault identification method based on a machine learning model according to claim 6, characterized in that, The physical constraint loss term is used to measure the deviation between the predicted excitation current represented in the model and the actual value of the excitation current calculated based on the primary and secondary currents.

8. A differential protection fault identification device based on a machine learning model, characterized in that, include: The acquisition module is used to acquire the current signals on both sides of the protected device and extract the time-series feature sequence containing multiple time points from the current signals as the input of the model; A processing module is connected to the acquisition module. The processing module has a trained feedforward neural network deployed in it. The feedforward neural network is used to process the time-series feature sequence to output fault identification results. The feedforward neural network includes at least one hidden layer containing a memory module; The memory module is configured to generate a historical information representation by applying a learnable non-cyclic transformation to a preset number of historical hidden states of the hidden layer. The feedforward neural network is further configured to concatenate the historical information representation with the current hidden state of the hidden layer, and pass the concatenated result to the next layer for processing.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1-7.

Images