Method, apparatus and storage medium for fault diagnosis of energy microgrid

Abstract

The present disclosure discloses a method, an apparatus, and a storage medium for fault diagnosis of an energy microgrid. The method includes: obtaining data of the energy microgrid; based on the obtained data of the energy microgrid, performing fault diagnosis using a learning network, and the learning network is trained through the following training steps. After obtaining a first training sample data set: randomly selecting some sample data from the first training sample data set according to a predetermined probability distribution; based on the partial sample data, determining a first local gradient of a loss function with respect to parameters of the learning network; based on the first local gradient, restoring a first full gradient of the loss function with respect to the parameters of the learning network for the first training sample data set; and updating the parameters of the learning network based on the first full gradient.

Description

Technical Field

[0001] The present disclosure relates to the application of artificial intelligence technology in the field of energy control management, and in particular to a method, device and storage medium for fault diagnosis of an energy microgrid. Background Art

[0002] Recently, the energy internet has developed rapidly. Due to the large-scale incorporation of renewable energy, electric vehicles, energy storage, etc. into energy microgrids, energy microgrid fault diagnosis is a great challenge encountered in the development of energy internet. On the one hand, the combination of microgrids with computers, communication technologies, etc. has led to a large increase in the amount of microgrid data, which may contain thousands of nodes, thousands or even tens of thousands of data, and a significant increase in data dimensions, which may be as high as thousands of dimensions or more. On the other hand, the fault information contained in high-dimensional complex data is more hidden and difficult to mine. This leads to low accuracy and slow speed in energy microgrid fault diagnosis, and even the inability to diagnose the faulty subgrid or line, which results in the partial or even overall inability of the microgrid to operate, resulting in huge economic losses.

[0003] Artificial intelligence technology has attracted more and more attention due to its advanced, intelligent and robust nature, as well as its ability to process large amounts of data and learn from data. Deep neural networks (DNNs), which are neural networks with multiple hidden layers, are an important branch of artificial intelligence research, with powerful data analysis, clustering, prediction and other capabilities. However, deep neural networks have low operating efficiency when processing large amounts of data in the energy Internet, especially in training.

[0004] Although the stochastic gradient descent method is proposed for training, this training method seriously loses computational accuracy and reliability. Summary of the invention

[0005] The present disclosure is proposed to solve the above-mentioned technical problems existing in the prior art. The present disclosure aims to provide a method and device for fault diagnosis of energy microgrids based on an improved DNN, which can significantly improve the operating efficiency of automatic diagnosis, improve the timeliness of diagnosis, and improve the accuracy of fault diagnosis in view of the specific situation that the data volume of energy microgrid fault diagnosis is large, the fault information is more complex and hidden, and the computational load is large.

[0006] The present disclosure adopts a method and device for energy microgrid fault diagnosis based on an improved DNN to alleviate the above problems or completely overcome the above technical problems or other technical problems. The technical problems that the proposed technology can solve are not limited to the technical problems specifically mentioned above, and may also be able to solve other technical problems that are not mentioned above but do exist in the prior art.

[0007] The first aspect of the present disclosure provides a method for fault diagnosis of an energy microgrid. The method may include: obtaining data of the energy microgrid; based on the obtained data of the energy microgrid, using a learning network for fault diagnosis, and the learning network is trained through the following training steps: randomly selecting some sample data from a first training sample data set according to a predetermined probability distribution, where each sample data includes data of the energy microgrid and corresponding fault diagnosis results; based on the partial sample data, determining a first local gradient of a loss function with respect to parameters of the learning network; based on the first local gradient, restoring a first full gradient of the loss function of the first training sample data set with respect to parameters of the learning network; and based on the first full gradient, updating the parameters of the learning network.

[0008] By using this method for fault diagnosis of an energy microgrid, since some sample data are randomly selected from the first training sample data set according to a predetermined probability distribution, a first local gradient of the loss function with respect to parameters of the learning network is determined based on the partial sample data, then a first full gradient of the loss function of the first training sample data set with respect to parameters of the learning network is restored based on the first local gradient, and the learning network is updated based on the first full gradient, it not only reduces the computational load, improves the operation efficiency of fault diagnosis, but also ensures a relatively high computational accuracy, so the accuracy and reliability of fault diagnosis are improved.

[0009] The second aspect of the present disclosure provides a device for fault diagnosis of an energy microgrid. The device may include: an interface for obtaining data of the energy microgrid through the interface; a processor configured to: based on the obtained data of the energy microgrid, use a learning network for fault diagnosis, and the learning network is trained through the following training steps: randomly selecting some sample data from a first training sample data set according to a predetermined probability distribution, where each sample data includes data of the energy microgrid and corresponding fault diagnosis results; based on the partial sample data, determining a first local gradient of a loss function with respect to parameters of the learning network; based on the first local gradient, restoring a first full gradient of the loss function of the first training sample data set with respect to parameters of the learning network; and based on the first full gradient, updating the parameters of the learning network.

[0010] With the device for fault diagnosis of the energy microgrid, since the processor uses an improved learning network for fault diagnosis, and since a part of the sample data is randomly selected from the first training sample dataset according to a predetermined probability distribution, the first local gradient of the loss function with respect to the parameters of the learning network is determined based on the part of the sample data. Then, the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network is restored based on the first local gradient, and the learning network is updated based on the first full gradient. This not only reduces the computational load of the processor and improves the operating efficiency of fault diagnosis, but also ensures a high computational accuracy, thus improving the accuracy and reliability of fault diagnosis.

[0011] The third aspect of the present disclosure provides a non-transitory computer-readable storage medium storing instructions that, when executed by a processor, implement a method for fault diagnosis of an energy microgrid. The method includes: obtaining data of the energy microgrid; based on the obtained data of the energy microgrid, using a learning network for fault diagnosis, and the learning network is trained through the following training steps. After obtaining the first training sample dataset: randomly selecting a part of the sample data from the first training sample dataset according to a predetermined probability distribution; determining the first local gradient of the loss function with respect to the parameters of the learning network based on the part of the sample data; restoring the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network based on the first local gradient; and updating the parameters of the learning network based on the first full gradient.

[0012] The present disclosure proposes a fault diagnosis classifier based on an improved learning network, which can automatically and efficiently perform fault diagnosis of an energy microgrid. Using this classifier, while ensuring a significant improvement in operating efficiency, sufficient computational accuracy can be guaranteed, and the processing complexity and operational performance are effectively improved. Therefore, the automatic fault diagnosis technology of the energy microgrid based on the improved DNN proposed in the present disclosure can improve the accuracy, speed and efficiency of fault diagnosis, and further improve the functional stability and reliability of the energy microgrid. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The features, advantages, and technical and industrial significance of the exemplary embodiments of the present disclosure will be described below with reference to the drawings, in which the same reference numerals represent the same elements, and the drawings are not necessarily drawn to scale, and in which:

[0014] Figure 1 is a schematic diagram showing the overall architecture of the energy Internet according to an embodiment of the present disclosure, where the energy Internet may include several energy microgrids;

[0015] Figure 2is a block diagram showing the structure of an energy router in an energy microgrid according to an embodiment of the present disclosure, where the energy router includes various functional components;

[0016] Figure 3 is an overall flowchart showing a method for fault diagnosis of an energy microgrid according to an embodiment of the present disclosure;

[0017] Figure 4 is a flowchart showing the training process of a learning network for fault diagnosis of an energy microgrid according to an embodiment of the present disclosure;

[0018] Figure 5 is a flowchart showing the process of restoring the full gradient from the local gradient according to an embodiment of the present disclosure; and

[0019] Figure 6 is a block diagram showing the structure of a device for fault diagnosis of an energy microgrid according to an embodiment of the present disclosure. Detailed Embodiments

[0020] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. The descriptions of these embodiments are exemplary and illustrative, rather than restrictive.

[0021] Figure 1 shows a schematic diagram of the overall architecture of an energy Internet according to an embodiment of the present disclosure. The energy Internet may include several energy microgrids 102-i (collectively referred to as energy microgrids 102 when not needing to be distinguished from each other), where i may take an integer from 1 to n, and n is the number of energy microgrids. Each energy microgrid 102 may include an energy router 103-i (collectively referred to as energy router 103 when not needing to be distinguished from each other). For example, energy microgrid 102-1 includes energy router 103-1, and so on. Each energy microgrid 102-i may further include several energy local area networks 104. Each energy local area network 104 may include an edge server 105 (Edge Server, abbreviated as ES) and several terminal devices 106 (End Device, abbreviated as ED). Each energy microgrid 102-i may be communicatively coupled to the cloud 101 through the energy router 103-i, and the cloud 101 may include devices such as cloud servers. For example, the energy router 103 may be communicatively coupled to the edge server 105 in the energy local area network 104 to exchange energy microgrid data and the like.

[0022] When a fault occurs in the energy microgrid, it is necessary to perform fault diagnosis on the energy microgrid to locate and identify the cause of the fault, so as to eliminate the fault and restore normal operation. Faults may occur at multiple locations in any energy microgrid for various complex reasons. For example, faults may occur in the energy local area network, in the servers or terminal devices of a certain energy local area network, or on a certain communication line, etc. In some embodiments, various services such as renewable energy, electric vehicles, and energy storage can be incorporated into the energy microgrid. As the incorporated services increase, the energy microgrid is more prone to faults, and the causes of the faults are more complex and hidden. Further, various computer and communication technologies can be integrated into the energy microgrid, which will significantly increase the amount of data interacted and processed in the energy microgrid, further increasing the difficulty of fault diagnosis.

[0023] In some embodiments, the automatic fault diagnosis function can be implemented in the energy router 103.

[0024] Figure 2 Shown is a schematic structural diagram of an energy router 103 according to another embodiment of the present disclosure. As Figure 2 shown, the energy router 103 may include a processor 210, a memory 220, a solid-state transformer 230, an energy storage battery 240, and an interface 250.

[0025] The processor 210 may be a processing unit including more than one general-purpose processing device, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, the processor 210 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor running other instruction sets, or a processor running a combination of instruction sets, but is not limited to these. The processor 210 may also be more than one dedicated processing device, such as an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), a system on a chip (SoC), etc. The processor 210 may be communicatively coupled to the memory 220 and configured to execute computer-executable instructions stored thereon to perform functions such as energy microgrid fault diagnosis calculation processing, fault diagnosis classification, and prediction.

[0026] The memory 220 may be a non-transitory computer-readable medium, such as read-only memory (ROM), random access memory (RAM), phase change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), electrically erasable programmable read-only memory (EEPROM), other types of random access memory (RAM), flash drives or other forms of flash memory, caches, registers, static memories, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical memories, cassette tapes or other magnetic storage devices, or any other possible non-transitory medium used to store information or instructions accessible by a computer device, etc. For example, in some embodiments, a fault diagnosis classifier may be stored in the memory 220, and the fault diagnosis classifier can implement functions such as energy microgrid fault diagnosis classification and prediction.

[0027] The core component of the energy router 103 is the solid-state transformer 230. The solid-state transformer 230 is the main physical device for controlling electric energy. It is a power device that combines power electronics conversion technology and electric energy exchange technology based on the principle of electromagnetic induction to realize the conversion of electric energy with one power characteristic into electric energy with another power characteristic. The so-called power characteristics include, for example, the amplitude, phase, frequency, number of phases, and waveform of voltage (or current), etc. The solid-state transformer 230 may include an AC / DC rectifier, a DC / AC inverter, a high-frequency transformer, an AC / DC converter, a low-voltage DC bus parallel module, and a DC / AC inverter. The AC / DC rectifier is connected to the DC / AC inverter through a high-voltage DC bus. The DC / AC inverter, the high-frequency transformer, and the AC / DC converter are connected in sequence. The AC / DC converter is connected to the low-voltage DC bus parallel module through a low-voltage DC bus, and the low-voltage DC bus parallel module is connected to the DC / AC inverter.

[0028] The energy storage battery 240 can be used as the electric energy storage module of the energy router 103, which can provide power quality control compensation or provide active power when there is a power supply fault in the microgrid system, and can also play a role in power balance of the power system. The energy storage battery 240 can be electrically connected to the solid-state transformer 230 through a DC / AC inverter.

[0029] The interface 250 may include multiple interfaces such as an electric energy interface, a communication interface, a water and electricity interface, etc., for docking and communicating with corresponding external devices respectively.

[0030] According to one embodiment, a classifier for fault diagnosis of an energy microgrid can be constructed based on a learning network such as a DNN. As Figure 3As shown, the process of constructing the energy microgrid fault classifier may include an energy microgrid data collection stage 301, a DNN training stage 302 (which can be executed offline), and a stage 303 of using the classifier for fault diagnosis (executed online). In stage 301, data of the energy microgrid can be obtained. The amount of data of the energy microgrid is usually large and the data dimension is high. Specifically, the energy microgrid may include a number of nodes at the thousand-level or above. Correspondingly, the data of the energy microgrid has a number of dimensions at the thousand-level or above. In stage 302, based on at least part of the sample data from the training sample database 305, the classifier can be trained, where each sample data includes the data of the energy microgrid and the corresponding fault diagnosis result. The trained classifier 304 can be called in stage 303 to perform fault diagnosis on the energy microgrid, so as to take corresponding measures according to the fault diagnosis result to perform fault elimination and recovery processing. The present disclosure improves the training process of stage 302. Specifically, part of the sample data is randomly selected from the training sample database 305 according to a predetermined probability distribution. The following steps are sequentially executed in each training iteration step. Based on the part of the sample data, determine the first local gradient of the loss function with respect to the parameters of the classifier; based on the first local gradient, the first full gradient of the loss function of the training sample data set to which the part of the sample data belongs with respect to the parameters of the learning network can be restored. Subsequently, based on the first full gradient, the parameters of the classifier can be updated. Through the selection of part of the sample data, for example, in some embodiments, even for a training sample data set, only a single sample data can be selected, and the original full gradient can be accurately restored and approximated via, for example, but not limited to, the compressive sensing algorithm. The classifier obtained by such iterative training significantly improves the calculation accuracy compared with the stochastic gradient descent method, and significantly reduces the operation load and calculation delay compared with the ordinary gradient descent method of directly calculating the full gradient based on the training sample data set and updating the parameters of the classifier accordingly. In some cases, the operation load and calculation delay can even be comparable to the stochastic gradient descent method. The classifier obtained by such training can automatically perform correct, timely and effective fault diagnosis in the energy microgrid where the data dimension is usually very large, at the thousand-level or above, the number of nodes is usually at the thousand-level or above, and faults occur frequently and the causes are hidden and difficult to predict, so as to maintain the excellent performance of the system and avoid incalculable losses caused by the failure to detect faults in time and overcome them.

[0031] Next, refer to Figure 4 and Figure 5 to describe in detail stage 302 for training the classifier. The classifier can adopt various learning network architectures. In some embodiments, the classifier may be a deep neural network (DNN) including multiple hidden layers, and its main operation process includes two parts: forward propagation and backward propagation.

[0032] First, the coefficient matrix ω and the bias vector b of the DNN can be initialized by using a random input method. The training sample data can be input from the input layer into the neural network, and the forward propagation process can be used for calculation. The calculation result can be output at the output layer, and the error between the training sample output and the true sample label can be compared, so as to calculate the loss function. If the error does not meet the training stop condition, the coefficient matrix ω and b can be modified by using the backpropagation algorithm.

[0033] Specifically, the process of backpropagation is essentially a process of finding the minimum value of the loss function. Usually, the gradient descent method is used to find the extreme value. Since the training samples are usually huge in scale, the energy microgrid may include nodes in the thousands or more, and the data of the energy microgrid has dimensions in the thousands or more. Therefore, when calculating the gradient information in each loop, if the traditional gradient descent method is used, it will require a huge computational cost and a time delay that affects the timeliness of fault diagnosis. The stochastic gradient descent method considers using partial gradients to replace the overall gradient. In each calculation of the gradient, the gradients of some sample points are randomly selected for solution to replace the global gradient, and then the loop is repeated until the loss function reaches the minimum value. At this time, the training data result approaches the true sample label.

[0034] As described above, although the stochastic gradient descent method improves the computational efficiency, it seriously loses the computational accuracy of the classifier it trains and is not applicable in application scenarios such as energy microgrids with a huge number of nodes, frequent faults, and difficult to predict. The low computational accuracy of the classifier obtained by the stochastic gradient descent method leads to false positives in fault detection, misdetection, and missed detection of fault categories in the energy microgrid, thus seriously affecting the overall performance of the energy system and even bringing immeasurable losses.

[0035] In some embodiments, some sample data can be randomly selected according to a predetermined probability distribution, and after obtaining the local gradient information using the stochastic gradient based on the selected partial sample data in each loop of backpropagation, the full gradient information can be restored based on the local gradient. The inventors considered the sparse characteristics of the gradient vectors (such as the full gradient vector, local gradient vector, etc.) of the above learning network. By randomly selecting some sample data according to a predetermined probability distribution (including but not limited to Gaussian distribution, binomial distribution, lognormal distribution, etc.), sampling of some sample data can be achieved at a sampling rate far lower than the traditional Nyquist sampling theorem, and the full gradient vector can still be accurately restored via the optimization algorithm, so as to improve the computational efficiency without losing the computational accuracy. In this way, the computational accuracy is significantly improved compared with the stochastic gradient descent method, and the computational cost is greatly reduced compared with the traditional gradient descent method.

[0036] The following specifically describes an example of the implementation process of using an improved DNN for energy microgrid fault diagnosis.

[0037] This process generally can include the following several stages.

[0038] First, given the training sample data of the energy microgrid as input where M is the total number of samples, x i is the energy microgrid training sample data, and y i is the label data. Then, the number of layers of the DNN, the number of neurons, and the activation function σ(z) of each layer can be determined, where the variable is the linear transformation part with respect to the data {x i}, and ω and b are the weight coefficients and biases. In some embodiments, the activation function can adopt RELU, sigmoid, etc. Sigmoid can be used as the activation function to achieve the improvement of the calculation efficiency when adopting the cross-entropy loss function.

[0039] Next, the weight coefficients ω and biases b of each layer of the neural network of the DNN can be initialized. In some embodiments, the initial values of the weight coefficients ω and biases b can be randomly given.

[0040] The loss function of the DNN with the current weight coefficients ω and biases b compared to the label data of the training sample data can be calculated through the forward propagation process. Specifically, the training sample data {x i} can be used as the input of the DNN with the current weight coefficients ω and biases b to calculate the corresponding result Then, based on the corresponding result and the label data y i the loss function J(x, y; w, b) can be calculated. In some embodiments, the cross-entropy loss function J(x, y; w, b) can be used, where the cross-entropy loss function can be defined as the following formula (1):

[0041]

[0042] Then, the backpropagation process can be executed to determine the iteration step size α, the maximum number of iterations, and the iteration stop threshold ε. Next, in the backpropagation loop process, m sample data can be randomly selected from M samples according to the Gaussian distribution to find the local gradient vectors of the loss function J with respect to ω and b Where m is much smaller than M. Note that the sample data can also be randomly selected according to other probability distribution functions, which will not be elaborated here. By selecting according to the Gaussian distribution, the selected sample data can be controlled within about one-thousandth of M or below, and still be accurately restored to the full gradient vector through the optimization algorithm in the following text, so as to significantly improve the operation efficiency without loss of operation accuracy.

[0043] Next, based on the local gradient vector the full gradient vector g of the loss function with respect to the parameters of the DNN network can be restored ω ,g b .

[0044] This process can be specifically realized by solving the following optimization problem:

[0045] such that

[0046] such that

[0047]

[0048] where ω and b represent the parameters of the learning network, respectively represent the first full gradient coefficients of ω and b, respectively represent the sparse transformation matrices of ω and b, respectively represent the first local gradients of ω and b, Φ ω , Φ b respectively represent the information matrices of ω and b, g ω ,g b respectively represent the first full gradients of ω and b, l1 is the first norm, l2 is the second norm, and τ is the threshold.

[0049] Subsequently, gradient descent is performed in the direction of the full gradient vector, that is:

[0050] ω = ω - αg ω ; b = b - αg b Formula (5)

[0051] where α represents the step size of descent, so as to gradually approximate and solve the minimum value of the loss function J.

[0052] When the loop process reaches the stop threshold or the maximum number of iteration steps, the iteration can be stopped, and the network parameters ω and b of the trained fault classifier of the energy microgrid are obtained, that is, the trained fault classifier of the energy microgrid is obtained. Then, the trained energy microgrid fault classifier can be used to perform online fault diagnosis on the energy microgrid. It should be noted that the cross-entropy loss function is only an example form of the loss function, and other forms of loss functions can be used, such as the mean square error loss function, etc.

[0053] Figure 4 is a flowchart of a process for training a learning network such as a DNN according to an embodiment of the present disclosure.

[0054] As Figure 4 shown, in step S40, after obtaining the first training sample data set, a part of the sample data can be randomly selected from the first training sample data set according to a predetermined probability distribution. In some embodiments, by appropriately selecting the predetermined probability distribution, such as selecting a Gaussian distribution, the part of the sample data can be reduced to the case of only a single sample data.

[0055] In step S42, based on the part of the sample data, a loss function can be determined, and its first local gradient with respect to the parameters of the learning network can be calculated

[0056] In step S44, based on the first local gradient the first full gradient g of the loss function J of the first training sample data set with respect to the parameters ω and b of the learning network can be restored ω ,g b .

[0057] In step S46, based on the first full gradient g ω ,g b , the parameters ω and b of the learning network can be updated. In some embodiments, the part of the sample data can account for about 0.1% to less than 1% of the total amount of sample data in the first training sample data set.

[0058] In some embodiments, the predetermined probability distribution can include a Gaussian distribution. In addition, the predetermined probability distribution can include various common probability distribution types, such as binomial distribution, exponential distribution, Chebyshev's inequality, etc.

[0059] In Figure 5The flowchart of the process of restoring the full gradient from the local gradient according to an embodiment of the present disclosure is shown. First, in step S50, a second training sample data set is obtained. In step S52, based on the second training sample data set, a second full gradient of the loss function with respect to the parameters of the learning network is determined. In step S54, the second full gradient is subjected to a sparse transformation to decompose it into a sparse transformation matrix and second full gradient coefficients

[0060] Then, the process of restoring the first full gradient from the first local gradient in step S44 above may specifically include the following steps. In step S56, when the second norm l2 of the difference between the vector obtained by processing the first full gradient coefficient ω , Φ b and the sparse transformation matrix and the first local gradient is less than the threshold τ, the first norm l1 of the first full gradient coefficient is minimized to determine the first full gradient coefficient For example, specifically, reference can be made to formulas (2) and (3).

[0061] Among them, the predetermined information matrix Φ ω , Φ b is generated according to the predetermined probability distribution. In step S58, based on the first full gradient coefficient the first full gradient g is determined by using the sparse transformation matrix ω , g b For example, specifically, reference can be made to formula (4).

[0062] In one embodiment, the above-mentioned sparse transformation matrix can be determined based on the principal component analysis method. The principal component analysis (PCA) can adopt various methods, including linear decomposition or non-linear decomposition, etc. For example, through the principal component analysis (PCA), the eigenvector corresponding to the maximum eigenvalue of the high-dimensional full gradient vector can be obtained and used as the sparse transformation matrix.

[0063] In some embodiments, the sparse transformation matrix can be determined based on other common sparse transformation methods. For example, it can be based on orthogonal sparse transformation, sparse Fourier transform, sparse fast Fourier transform, wavelet transform, sparse manifold transform, etc. In the embodiment, sparse transformation can also be performed based on the dictionary learning method. For example, by performing singular value decomposition, the sparse expression basis function of the first full gradient is determined, and through a large number of second sample data, the sparse expression dictionary of the data is learned.

[0064] The predetermined information matrix Φ ω , Φ b can be generated according to the predetermined probability distribution. For example, according to some embodiments, a part of sample data can be randomly selected from the first training sample dataset according to the predetermined probability distribution. If the Gaussian probability distribution is selected to select m sample numbers, then the information matrix is a Gaussian random matrix with an m x M dimension.

[0065] In an embodiment of the present disclosure, as Figure 1 and Figure 2 shown, the energy router 103 can perform the process of training a learning network such as DNN as shown in Figure 4 and the process of restoring the full gradient from the local gradient as shown in Figure 5 through the processor 210. The energy router 103 can obtain the data of the energy microgrid through the interface 250 and then store it in the memory 220. The obtained energy microgrid fault diagnosis classifier can also be stored in the memory 220.

[0066] Figure 6 is a block diagram showing the structure of the energy microgrid fault diagnosis device 60 according to an embodiment of the present disclosure. As Figure 6 shown, the energy microgrid fault diagnosis device 60 may include a processor 600, a memory 601, a non-volatile memory (NVM) 602, a user interface 603, a network interface 604, and a storage interface 605. It should be understood that Figure 6 the structural composition of the energy microgrid fault diagnosis device 60 shown is exemplary in nature and can therefore be simplified for the purpose of explanation. Depending on the specific implementation and application, the energy microgrid fault diagnosis device 60 can be implemented as a laptop computer, a desktop computer, a tablet computer, a server, an embedded computing device, or other electronic devices.

[0067] As Figure 6As shown, the memory 601 can store various programs, such as routines for training a learning network such as a DNN, and various data such as the collected energy microgrid datasets. The processor 600 can be configured to access these programs and data to perform the learning network training process as described herein. In addition to the memory 601 coupled to the processor 600, the energy microgrid fault diagnosis device 60 may further include an NVM 602. The NVM 602 can be a permanent cache or cache memory located outside the memory 601, and can be implemented to store data regardless of the power status of the energy microgrid fault diagnosis device 60. The user interface 603 may include various input devices and output devices for the energy microgrid fault diagnosis device 60 to interface with the user, facilitating input and / or output interactions between the user and the energy microgrid fault diagnosis device 60. The network interface 604 included in the energy microgrid fault diagnosis device 60 can perform network communication according to one or more communication protocols such as Ethernet. Additionally, the energy microgrid fault diagnosis device 60 can communicate with an external storage device (not shown) through the storage interface 605.

[0068] In some embodiments, the energy microgrid fault diagnosis device 60 may store a trained fault diagnosis classifier. The training of the classifier can be performed offline, and according to some embodiments, the training of the classifier can also be performed online. After obtaining the fault diagnosis classifier, online fault diagnosis of the energy microgrid can be performed by means of the trained fault diagnosis classifier based on the input energy microgrid data received from the user interface 603, so as to be able to timely know the cause of the fault and take corresponding measures.

[0069] Furthermore, the processor 600 can be configured to perform the following processes, such as including a fault diagnosis classifier training process and a fault diagnosis process. For example, based on the energy microgrid data pre-stored in the memory 601 or received through the user interface 603, online fault diagnosis can be performed using the trained fault diagnosis classifier based on a learning network, where the learning network can be trained through the following steps. After obtaining the first training sample dataset: randomly select a part of the sample data from the first training sample dataset according to a predetermined probability distribution; based on the part of the sample data, determine the first local gradient of the loss function with respect to the parameters of the learning network; based on the first local gradient, restore the first full gradient of the loss function with respect to the parameters of the learning network for the first training sample dataset; based on the first full gradient, update the parameters of the learning network.

[0070] The training steps performed by the processor 600 may further include: obtaining a second training sample data set; determining, based on the second training sample data set, a second full gradient of the loss function with respect to the parameters of the learning network; performing a sparse transformation on the second full gradient to decompose it into a sparse transformation matrix and a first full gradient coefficient; wherein, based on the first local gradient, restoring the first full gradient of the loss function with respect to the parameters of the learning network for the first training sample data set further includes: when the second norm of the difference between the vector obtained by processing the first full gradient coefficient with a predetermined information matrix and the sparse transformation matrix and the first local gradient is less than a threshold, minimizing the first norm of the first full gradient coefficient to determine the first full gradient coefficient, wherein the predetermined information matrix is generated according to the predetermined probability distribution; and determining the first full gradient by using the sparse transformation matrix based on the first full gradient coefficient.

[0071] According to some embodiments, the training of a learning network such as a DNN can be performed by the processor 600 based on partial sample data (e.g., sample data including the operation data of the energy microgrid and the corresponding fault diagnosis results).

[0072] According to another embodiment of the present disclosure, the above-mentioned energy microgrid fault diagnosis function can be implemented in Figure 1 the cloud 101 shown. In this embodiment, for example, the powerful computing power and fast computing speed of the cloud server in the cloud 101 can be utilized to further improve the efficiency of fault diagnosis. The cloud server can perform the learning network training as described above and perform fault diagnosis based on the fault diagnosis classifier. For example, when a fault occurs in the energy local area network, Figure 1 and Figure 2 the energy router 103 shown in can receive the fault information and relevant microgrid operation data sent by the server of the energy local area network in the microgrid, and after appropriate processing, send the received information and data to the cloud server in the cloud. The cloud server can complete the fault diagnosis and send the diagnosis result back to the energy router 103 for taking corresponding fault safety measures.

[0073] As described above, embodiments of the technology of the present disclosure have been described. However, the present disclosure can be regarded as a method for fault diagnosing an energy microgrid, an apparatus for fault diagnosing an energy microgrid, program instructions for performing energy microgrid fault diagnosis, and a non-transitory computer-readable storage medium storing the program instructions.

[0074] In addition, the processing described as being performed by a single device can be performed by multiple devices in a shared manner. Alternatively, the processing described as being performed by different devices can be performed by a single device. In a computer system, the hardware configuration (e.g., server configuration) for implementing each function can be flexibly changed.

[0075] The energy microgrid fault diagnosis method and device of the present disclosure can be used in the energy Internet, particularly in the microgrid of the energy Internet.

[0076] Each of the above embodiments is intended to be illustrative rather than restrictive. The present invention can be implemented with appropriate modifications without departing from its gist. For example, unless there is a technical contradiction, the processes and functional units described in the present disclosure can be freely combined and implemented. For example, the above examples (or one or more of their aspects) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description. Additionally, in the above specific implementation manners, various features can be grouped together to simplify the present disclosure. This should not be construed as an intention that the disclosed features not claimed are necessary for any claim. On the contrary, the subject matter of the present disclosure can be less than all the features of a particular disclosed embodiment. Thus, the following claims are incorporated herein by way of example or embodiment into the specific implementation manners, where each claim independently serves as a separate embodiment, and considering these embodiments, they can be combined with each other in various combinations or permutations. The scope of the present disclosure should be determined with reference to the appended claims and the full scope of the equivalents to which these claims are entitled.

Claims

1. A method for fault diagnosis of an energy microgrid, characterized in that Including: Obtaining data of the energy microgrid; Based on the obtained data of the energy microgrid, using a learning network for fault diagnosis, and the learning network is trained through the following training steps: Randomly selecting some sample data from the first training sample dataset according to a predetermined probability distribution, where each sample data includes the data of the energy microgrid and the corresponding fault diagnosis result; Based on the part of the sample data, determining the first local gradient of the loss function with respect to the parameters of the learning network; Based on the first local gradient, restoring the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network; Based on the first full gradient, updating the parameters of the learning network; The training steps further include: Obtaining a second training sample dataset, where each sample data includes the data of the energy microgrid and the corresponding fault diagnosis result; Based on the second training sample dataset, determining the second full gradient of the loss function with respect to the parameters of the learning network; Performing a sparse transformation on the second full gradient to decompose it into a sparse transformation matrix and second full gradient coefficients; Wherein, based on the first local gradient, restoring the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network further includes: When the second norm of the difference between the vector obtained by processing the first full gradient coefficients through a predetermined information matrix and the sparse transformation matrix and the first local gradient is less than a threshold, minimizing the first norm of the first full gradient coefficients to determine the first full gradient coefficients, where the predetermined information matrix is generated according to the predetermined probability distribution; Based on the first full gradient coefficients, using the sparse transformation matrix to determine the first full gradient.

2. The method according to claim 1, characterized in that The predetermined probability distribution includes a Gaussian distribution.

3. The method according to claim 1, characterized in that The part of the sample data includes a single sample data.

4. The method according to claim 1, characterized in that The part of the sample data accounts for 0.1% to 1% of the total amount of sample data in the first training sample dataset.

5. The method according to claim 1, characterized in that The sparse transformation matrix is determined based on one of the principal component analysis method and the dictionary learning method.

6. The method according to claim 1, characterized in that The loss function includes a cross-entropy loss function.

7. The method according to claim 1, characterized in that The energy microgrid includes a number of nodes at the thousand-level and above, and the data of the energy microgrid has a number of dimensions at the thousand-level and above.

8. The method according to claim 1, characterized in that Using the following formula to calculate the first full gradient: Among them, ω and b represent the parameters of the learning network, respectively represent the first full gradient coefficients of ω and b, respectively represent the sparse transformation matrices of ω and b, respectively represent the first local gradients of ω and b, Φ ω , Φ b respectively represent the information matrices of ω and b, g ω , g b respectively represent the first full gradients of ω and b, l1 is the first norm, l2 is the second norm, and τ is the threshold.

9. A device for fault diagnosis of an energy microgrid, characterized in that Including: An interface, through which data of the energy microgrid is obtained; A processor, which is configured to: Based on the obtained data of the energy microgrid, using a learning network for fault diagnosis, and the learning network is trained through the following training steps: Randomly selecting some sample data from the first training sample dataset according to a predetermined probability distribution, where each sample data includes the data of the energy microgrid and the corresponding fault diagnosis result; Based on the part of the sample data, determining the first local gradient of the loss function with respect to the parameters of the learning network; Based on the first local gradient, restoring the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network; Based on the first full gradient, updating the parameters of the learning network; The training steps further include: Obtain a second training sample dataset, where each sample data includes data of the energy microgrid and the corresponding fault diagnosis result; Based on the second training sample dataset, determine the second full gradient of the loss function with respect to the parameters of the learning network; Perform a sparse transformation on the second full gradient to decompose it into a sparse transformation matrix and second full gradient coefficients; Among them, further including: based on the first local gradient, restoring the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network; When the second norm of the difference between the vector obtained by processing the first full gradient coefficients through a predetermined information matrix and the sparse transformation matrix and the first local gradient is less than a threshold, minimize the first norm of the first full gradient coefficients to determine the first full gradient coefficients, where the predetermined information matrix is generated according to the predetermined probability distribution; Based on the first full gradient coefficients, use the sparse transformation matrix to determine the first full gradient.

10. The device according to claim 9, wherein, The predetermined probability distribution includes a Gaussian distribution.

11. The device according to claim 9, wherein, The partial sample data includes a single sample data.

12. The device according to claim 9, wherein, The partial sample data accounts for 0.1% to 1% of the total amount of sample data in the first training sample dataset.

13. The device according to claim 9, wherein, The sparse transformation matrix is determined based on one of the principal component analysis method and the dictionary learning method.

14. The device according to claim 9, wherein, The loss function includes a cross-entropy loss function.

15. The device according to claim 9, wherein, The energy microgrid includes a number of nodes at the thousand level or above, and the data of the energy microgrid has a number of dimensions at the thousand level or above.

16. The device according to claim 9, wherein, Use the following formula to calculate the first full gradient: ω and b represent the parameters of the learning network, respectively represent the first full gradient coefficients of ω and b, respectively represent the sparse transformation matrices of ω and b, respectively represent the first local gradients of ω and b, Φ ω , Φ b respectively represent the information matrices of ω and b, g ω , g b respectively represent the first full gradients of ω and b, l1 is the first norm, l2 is the second norm, and τ is the threshold.

17. A non - transitory computer - readable storage medium storing instructions that, when executed by at least one processor, implement a method for fault diagnosis of an energy micro - grid, wherein, The method includes: Obtain the data of the energy microgrid; Based on the obtained data of the energy microgrid, perform fault diagnosis using a learning network, and the learning network is trained through the following training steps: Randomly select partial sample data from the first training sample dataset according to a predetermined probability distribution, where each sample data includes data of the energy microgrid and the corresponding fault diagnosis result; Based on the partial sample data, determine the first local gradient of the loss function with respect to the parameters of the learning network; Based on the first local gradient, restore the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network; Based on the first full gradient, update the parameters of the learning network; The training steps further include: Obtain a second training sample dataset, where each sample data includes data of the energy microgrid and the corresponding fault diagnosis result; Based on the second training sample dataset, determine the second full gradient of the loss function with respect to the parameters of the learning network; Perform a sparse transformation on the second full gradient to decompose it into a sparse transformation matrix and second full gradient coefficients; Among them, further including: based on the first local gradient, restoring the first full gradient of the loss function of the first training sample dataset with respect to the parameters of the learning network; When the second norm of the difference between the vector obtained after processing the first full gradient coefficient by the predetermined information matrix and the sparse transformation matrix and the first local gradient is less than the threshold, minimize the first norm of the first full gradient coefficient to determine the first full gradient coefficient, where the predetermined information matrix is generated according to the predetermined probability distribution; Based on the first full gradient coefficient, use the sparse transformation matrix to determine the first full gradient.

Images