Generation method of power grid service identification model and power grid service identification method

Abstract

The invention discloses a power grid service identification model generation method and a power grid service identification method. Comprising the steps of obtaining a training sample set formed by power grid service data; inputting the training sample and the type of the model task into a teacher network model for identification to obtain a first identification result, and inputting the training sample and the type of the model task into a to-be-trained student network model for identification to obtain a second identification result; determining a first loss function and a second loss function according to the first identification result, the second identification result and the label data; calculating a fitting loss function according to the first loss function and the second loss function, and performing back propagation on the to-be-trained student network model based on the fitting loss function to obtain a to-be-trained student network model for next iteration until an iteration convergence condition is met, thereby obtaining a power grid service identification model; the problems that model recognition is low in accuracy and high in delay, and multi-service data cannot be processed are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution Internet of Things, and in particular to a method for generating a power grid service recognition model and a power grid service recognition method. Background Art

[0002] In recent years, with the large-scale development of distributed new energy in the distribution network and the continuous increase in the types of user-side loads, the operation and control of the distribution network have shown high complexity and uncertainty, and power data also exhibits characteristics such as scale, randomness, and diversity. Deep learning models are less affected by noise. For massive data, they can accurately extract the features of the input data, the output results are relatively stable, and they can fully explore the correlation between data, and have been widely used in the construction of distribution networks. However, in the face of massive distributed data, the traditional method of deploying deep learning models in the cloud to process massive data will bring huge pressure on network bandwidth, resulting in a significant reduction in efficiency and interaction delay; with the continuous deepening of the digital transformation of the power grid, emerging services are constantly emerging in the distribution network, and the computing and storage resources of edge intelligent fusion terminals are limited, and the edge side cannot meet the increasing demand for business computing resources.

[0003] At present, there are mainly two methods for deploying deep neural network models on the edge side of the distribution network: one is to deploy small models that meet the terminal computing power on the edge side, but the small models cannot fully explore the data correlation, resulting in a decrease in accuracy; the other is to process and compress large-scale data on the edge side and then transmit the data to the cloud for model training. Although this type of method significantly reduces the amount of communication data, transmitting data from the far edge side to the cloud for calculation will still bring non-negligible delays. Moreover, the current two deployment methods are only applicable to the deployment of single-service models in the distribution network and cannot process multi-service data. Summary of the Invention

[0004] The present invention provides a method for generating a power grid service recognition model and a power grid service recognition method to solve the problems of low model recognition accuracy, high latency, and inability to process multi-service data.

[0005] According to one aspect of the present invention, there is provided a method for generating a power grid service recognition model, including:

[0006] Obtaining a training sample set formed by power grid service data, where the training sample set includes at least two training samples and label data corresponding to the training samples, and each training sample corresponds to at least two types of model tasks, and the types of the model tasks at least include classification tasks and regression tasks;

[0007] Input the training samples in the current iteration and the types of the corresponding model tasks into a pre-trained teacher network model for recognition to obtain a first recognition result, and input the training samples in the current iteration and the types of the corresponding model tasks into a student network model to be trained for recognition to obtain a second recognition result;

[0008] Determine a first loss function and a second loss function according to the first recognition result, the second recognition result and label data, where the first loss function is a loss function based on response-based knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation;

[0009] Calculate a fitting loss function according to the first loss function and the second loss function, and perform backpropagation on the student network model to be trained based on the fitting loss function to obtain a student network model to be trained for the next iteration, until the iteration convergence condition is satisfied to obtain a power grid service recognition model.

[0010] According to another aspect of the present invention, there is provided a power grid service recognition method, which is applied to edge devices in a distribution network and includes:

[0011] Obtain power grid service data to be recognized and the types of the corresponding model tasks of the power grid service data to be recognized;

[0012] Input the power grid service data to be recognized and the types of the model tasks into a power grid service recognition model, where the power grid service recognition model is obtained by training a model using the generation method of the power grid service recognition model according to any embodiment of the present invention;

[0013] Determine a service recognition result according to the output of the power grid service recognition model;

[0014] Perform an operation and maintenance response according to the service recognition result.

[0015] According to another aspect of the present invention, there is provided a device for generating a power grid service recognition model, including:

[0016] A sample set acquisition module, configured to acquire a training sample set formed by power grid service data, where the training sample set includes at least two training samples and the label data corresponding to the training samples, and each training sample corresponds to at least two types of model tasks, and the types of the model tasks at least include a classification task and a regression task;

[0017] A sample recognition module, configured to input the training samples in the current iteration and the types of the corresponding model tasks into a pre-trained teacher network model for recognition to obtain a first recognition result, and input the training samples in the current iteration and the types of the corresponding model tasks into a student network model to be trained for recognition to obtain a second recognition result;

[0018] A loss function calculation module, configured to determine a first loss function and a second loss function according to the first recognition result, the second recognition result, and label data, where the first loss function is a loss function based on response knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation;

[0019] A loss function fitting module, configured to calculate a fitting loss function according to the first loss function and the second loss function, perform backpropagation on the to-be-trained student network model based on the fitting loss function, and obtain the to-be-trained student network model for the next iteration until the iteration convergence condition is satisfied, so as to obtain a power grid service recognition model.

[0020] According to another aspect of the present invention, there is provided a power grid service recognition device, which is applied to an edge device in a distribution network and includes:

[0021] A to-be-recognized data acquisition module, configured to acquire to-be-recognized power grid service data and the type of the model task corresponding to the to-be-recognized power grid service data;

[0022] A model input module, configured to input the to-be-recognized power grid service data and the type of the model task into a power grid service recognition model, where the power grid service recognition model is obtained by training a model using the generation method of the power grid service recognition model according to any embodiment of the present invention;

[0023] A service recognition result determination module, configured to determine a service recognition result according to the output of the power grid service recognition model;

[0024] An operation and maintenance response module, configured to perform an operation and maintenance response according to the service recognition result.

[0025] According to another aspect of the present invention, there is provided an electronic device, where the electronic device includes:

[0026] At least one processor, and a memory communicatively connected to the at least one processor;

[0027] Wherein, the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the generation method of the power grid service recognition model or the power grid service recognition method according to any embodiment of the present invention.

[0028] According to another aspect of the present invention, there is provided a computer-readable storage medium, where the computer-readable storage medium stores computer instructions, and the computer instructions are used to implement the generation method of the power grid service recognition model or the power grid service recognition method according to any embodiment of the present invention when executed by a processor.

[0029] According to another aspect of the present invention, there is provided a computer program product, which includes a computer program that, when executed by a processor, implements the method for generating a power grid service recognition model or the power grid service recognition method according to any embodiment of the present invention.

[0030] The technical solution of the embodiment of the present invention is as follows: By obtaining a training sample set formed by power grid service data, the training sample set includes at least two training samples and the label data corresponding to the training samples. Each training sample corresponds to at least two types of model tasks, and the types of the model tasks at least include classification tasks and regression tasks; inputting the training samples in the current iteration and the types of the corresponding model tasks into a pre-trained teacher network model for recognition to obtain a first recognition result, and inputting the training samples in the current iteration and the types of the corresponding model tasks into a student network model to be trained for recognition to obtain a second recognition result; determining a first loss function and a second loss function according to the first recognition result, the second recognition result and the label data, the first loss function is a loss function based on response-based knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation; calculating a fitting loss function according to the first loss function and the second loss function, and performing backpropagation on the student network model to be trained based on the fitting loss function to obtain the student network model to be trained for the next iteration until the iteration convergence condition is satisfied, and obtaining a power grid service recognition model, which solves the problems of low model recognition accuracy, high latency and inability to process multi-service data. By generating training samples through at least two types of model tasks, the types of the model tasks at least include classification tasks and regression tasks, multi-task recognition can be realized, and thus the processing of multi-service data can be realized; inputting the training samples and the types of the corresponding model tasks into the trained teacher network model and the student network model to be trained respectively for recognition to obtain a first recognition result and a second recognition result, guiding the training of the student network model by the teacher network model, calculating the first loss function and the second loss function, the first loss function is a loss function based on response-based knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation, calculating a fitting loss function according to the first loss function and the second loss function and performing backpropagation on the student network model to be trained. The embodiment of the present application can take into account different types of model tasks during the model training process, ensure that the student network model to be trained can accurately recognize different types of service data, and perform model training through the method of knowledge distillation. The obtained power grid service recognition model has a high recognition accuracy, and the model is small and can be directly deployed on edge devices, so that the edge devices can directly recognize the power grid service data, reduce the latency time of data recognition, and improve the data processing efficiency.

[0031] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become readily understood from the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0033] Figure 1 is a flowchart of a method for generating a power grid service identification model according to Embodiment 1 of the present invention;

[0034] Figure 2 is a flowchart of a method for generating a power grid service identification model according to Embodiment 2 of the present invention;

[0035] Figure 3 is a flowchart of a method for identifying a power grid service according to Embodiment 3 of the present invention;

[0036] Figure 4 is an implementation example diagram of model deployment according to Embodiment 3 of the present invention;

[0037] Figure 5 is a schematic structural diagram of a device for generating a power grid service identification model according to Embodiment 4 of the present invention;

[0038] Figure 6 is a schematic structural diagram of a device for identifying a power grid service according to Embodiment 5 of the present invention;

[0039] Figure 7 is a schematic structural diagram of an electronic device according to Embodiment 6 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] In order to enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0041] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0042] Embodiment 1

[0043] Figure 1 The figure is a flowchart of a method for generating a power grid service recognition model provided in Embodiment 1 of the present invention. This embodiment is applicable to the situation of accurately recognizing power grid services. This method can be executed by a device for generating a power grid service recognition model. The device for generating a power grid service recognition model can be implemented in the form of hardware and / or software, and the device for generating a power grid service recognition model can be configured in an electronic device. As Figure 1 shown, the method includes:

[0044] S101. Obtain a training sample set formed by power grid service data. The training sample set includes at least two training samples and label data corresponding to the training samples. Each training sample corresponds to at least two types of model tasks, and the types of model tasks at least include classification tasks and regression tasks.

[0045] In this embodiment, the power grid service data may be data generated in a distribution network. For example, it may be one or more of power grid operation data, power enterprise marketing data, power grid model data, power grid real-time data, non-electrical data, power big data, etc.; among them, the power grid operation data includes one or more of the operation status of the power grid, equipment monitoring and detection data. The power enterprise marketing data includes one or more of data related to transaction electricity prices, electricity sales volumes, and electricity customers; the power grid model data mainly includes one or more of equipment topologies, equipment static parameter information, etc., and these data are usually stored in a relational database. The power grid real-time data includes one or more of equipment telecontrol information, telemetry information, etc., and these data reflect the current operation status of the power grid; the non-electrical data includes one or more of meteorology, waterlogging, substation monitoring, etc.; the power big data includes not only structured data, such as one or more of transaction electricity prices, electricity sales volumes, etc., but also unstructured data, such as video surveillance images.

[0046] The training sample set can be understood as a data set composed of training samples; the label data is the true label of the training samples; the model task can be understood as a model training task, for example, topology recognition, power quality monitoring, fault warning, load forecasting, etc. The types of model tasks at least include classification tasks and regression tasks. For example, topology recognition and load forecasting are regression tasks, and power quality monitoring and fault warning are classification tasks.

[0047] Pre-collect the grid service data generated during the operation of the distribution network, process the grid service data to form training samples, and label the training samples to determine the label data corresponding to the training samples. A training sample set is formed according to the training samples and the label data corresponding to the training samples. In the embodiments of the present application, data corresponding to different types of model tasks can be collected to form training samples.

[0048] S102. Input the training samples in the current iteration and the types of their corresponding model tasks into a pre-trained teacher network model for recognition to obtain a first recognition result. Input the training samples in the current iteration and the types of their corresponding model tasks into the student network model to be trained for recognition to obtain a second recognition result.

[0049] In this embodiment, the student network model to be trained can be understood as a student network model that needs to be trained. The student network model can learn from the teacher network model through knowledge distillation. The first recognition result can be understood as the data obtained by the teacher network model for recognizing the training samples; the second recognition result can be understood as the data obtained by the student network model to be trained for recognizing the training samples.

[0050] Pre-train the teacher network model. Input the training samples in the current iteration and the types of their corresponding model tasks into the trained teacher network model. The teacher network model recognizes the training samples according to the learned knowledge to obtain a first recognition result; input the training samples in the current iteration and the types of their corresponding model tasks into the student network model to be trained for recognition to obtain a second recognition result output by the student network model to be trained.

[0051] S103. Determine a first loss function and a second loss function according to the first recognition result, the second recognition result, and the label data. The first loss function is a loss function based on response knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation.

[0052] In this embodiment, the first loss function and the second loss function are loss functions calculated through two different methods. Among them, the first loss function is calculated by using the method of knowledge distillation based on response. For example, the first loss function is calculated according to the neural response of the last layer of the teacher network model. The second loss function is calculated by using the method of channel knowledge distillation. For example, the second loss function is calculated according to the data of each channel identified by the teacher network model.

[0053] Taking the first recognition result as the soft label, using the calculation method of the loss function of knowledge distillation based on response, calculate the loss between the first recognition result and the second recognition result, and taking the label data as the true label, using the calculation method of the loss function of knowledge distillation based on response, calculate the loss between the second recognition result and the label data, and comprehensively process the two losses to obtain the first loss function. Similarly, taking the first recognition result as the soft label, using the calculation method of the loss function of channel knowledge distillation based on response, calculate the loss between the first recognition result and the second recognition result, and taking the label data as the true label, using the calculation method of the loss function of channel knowledge distillation based on response, calculate the loss between the second recognition result and the label data, and comprehensively process the two losses to obtain the second loss function.

[0054] S104. Calculate the fitting loss function according to the first loss function and the second loss function, and perform backpropagation on the student network model to be trained based on the fitting loss function to obtain the student network model to be trained for the next iteration until the iteration convergence condition is met, and obtain the power grid service recognition model.

[0055] In this embodiment, the fitting loss function can be understood as a loss function obtained by fitting multiple loss functions. The power grid service recognition model can be understood as a neural network model for recognizing service data in the distribution network. For example, it can recognize equipment failures, topological structures, etc. according to the power grid service data. The iteration convergence condition can be one or more conditions such as the accuracy of the model exceeding the accuracy threshold, the training time exceeding the time threshold, and the number of training times exceeding the number threshold.

[0056] Perform a comprehensive operation on the first loss function and the second loss function. For example, perform a weighted sum on the first loss function and the second loss function, or substitute them into a preset fitting formula for calculation, etc. By fitting the two loss functions, obtain the fitting loss function. Perform backpropagation on the student network model to be trained based on the fitting loss function, adjust the network parameters of the student network model to be trained, and obtain the student network model to be trained for the next iteration until the iteration convergence condition is met, and obtain the power grid service recognition model.

[0057] When the method provided by the embodiment of the present application is implemented, a teacher model with good performance but complexity can be pre-trained in the cloud in advance; then, the teacher model guides the training of a student model with a simple structure and small computational load, and uses the channel knowledge distillation algorithm based on multi-task to ensure that the effect of the student model is infinitely close to that of the teacher model and is applicable to most of the business of the distribution network; finally, the distilled student model is sent down and deployed on the edge side with limited computing resources in the distribution Internet of Things, and data processing and recognition are directly performed at the edge end, improving the processing efficiency of the massive data in the distribution network and reducing the interaction delay. It solves the problems of low processing efficiency of massive data, large interaction delay, and limited computing resources on the edge side in the current distribution network, and provides an optimized deployment scheme for the deep neural network model on the edge side of the distribution network, which can be applied to various services in the distribution network.

[0058] The embodiment of the present invention provides a method for generating a power grid service recognition model, which solves the problems of low recognition accuracy, high latency, and inability to process multi-service data of the model. Training samples are generated through at least two types of model tasks, and the types of model tasks at least include classification tasks and regression tasks, which can realize multi-task recognition and thus realize the processing of multi-service data; the training samples and the types of their corresponding model tasks are simultaneously input into the trained teacher network model and the student network model to be trained for recognition respectively to obtain the first recognition result and the second recognition result. The training of the student network model is guided by the teacher network model, and the first loss function and the second loss function are calculated. The first loss function is the loss function based on response knowledge distillation, and the second loss function is the loss function based on channel knowledge distillation. The fitting loss function is calculated according to the first loss function and the second loss function, and backpropagation is performed on the student network model to be trained. In the model training process of the embodiment of the present application, different types of model tasks can be taken into account simultaneously to ensure that the student network model to be trained can accurately recognize different types of service data. The model training is carried out by means of knowledge distillation, and the recognition accuracy of the obtained power grid service recognition model is relatively high, and the model is small and can be directly deployed on the edge device so that the edge device can directly recognize the power grid service data, reducing the delay time of data recognition and improving the data processing efficiency.

[0059] Embodiment 2

[0060] Figure 2 It is a flowchart of a method for generating a power grid service recognition model provided by Embodiment 2 of the present invention, and this embodiment is refined on the basis of the above embodiment. As Figure 2 shown, the method includes:

[0061] S201. Obtain a training sample set formed by power grid service data. The training sample set includes at least two training samples and the label data corresponding to the training samples. Each training sample corresponds to at least two types of model tasks, and the types of model tasks at least include classification tasks and regression tasks.

[0062] S202. Input the training samples in the current iteration and the types of their corresponding model tasks into a pre-trained teacher network model for recognition to obtain a first recognition result, and input the training samples in the current iteration and the types of their corresponding model tasks into a student network model to be trained for recognition to obtain a second recognition result.

[0063] After determining the first recognition result and the second recognition result, calculate the first loss function in the manner of S203 - S207, and calculate the second loss function in the manner of S208 - S213.

[0064] S203. Determine the first neural response result output by the last layer of the teacher network model according to the first recognition result.

[0065] In this embodiment, the first neural response result can be understood as the neural response output by the last layer of the teacher network model. The first recognition result output by the teacher network model can include the neural responses output by each layer of the teacher network model, and the neural responses of each layer are respectively identified for easy distinction. Analyze the first recognition result to determine the first neural response result output by the last layer of the teacher network model.

[0066] S204. Determine the second neural response result output by the last layer of the student network model to be trained according to the second recognition result.

[0067] In this embodiment, the second neural response result can be understood as the neural response output by the last layer of the student network model to be trained. The second recognition result output by the student network model to be trained can include the neural responses output by each layer of the student network model to be trained, and the neural responses of each layer are respectively identified for easy distinction. Analyze the second recognition result to determine the second neural response result output by the last layer of the student network model to be trained.

[0068] S205. Calculate a first target loss function according to the first neural response result and the second neural response result.

[0069] In this embodiment, the first target loss function can be understood as a kind of loss function, and the first target loss function can be a cross-entropy loss function. Substitute the first neural response result and the second neural response result into the calculation formula of the loss function for calculation to obtain the first target loss function.

[0070] S206. Calculate the second target loss function according to the second neural response result and the label data.

[0071] In this embodiment, the second target loss function can be understood as a kind of loss function, and the second target loss function can be a cross-entropy loss function. Substitute the second neural response result and the label data into the calculation formula of the loss function for calculation to obtain the second target loss function.

[0072] S207. Weight the first target loss function and the second target loss function to obtain the first loss function.

[0073] Determine the weight, and weight the first target loss function and the second target loss function according to the weight to obtain the first loss function. The weight can be preset.

[0074] Knowledge based on responses usually refers to the neural responses output by the last layer of the teacher model. The main idea of this type of knowledge distillation algorithm is to use the soft knowledge of the final prediction result of the teacher model to guide the training of the student model. This method is usually used for classification tasks. For the data with real labels in the training set, when training the student network, using both the real labels of the student model and the soft labels generated by the teacher model can produce better results. Therefore, the loss function for the teacher to train the student is composed of the loss function of the student itself and the distillation loss function.

[0075] Exemplarily, an embodiment of the present application provides a calculation formula for the first loss function:

[0076] L KD (x, W) = λ × L CE (y, σ(z S ; T = 1)) + (1 - λ) × L CE (σ(z T ; T = τ), σ(z S ; T = τ)) (1)

[0077] Wherein, x is the input data; W represents the parameters of the student network model to be trained; λ is the weight coefficient; L CE is the cross-entropy loss; y is the real label; σ is the softmax function after adding the temperature parameter T; τ is the custom distillation temperature; z S and z T are the output Logits of the student network model to be trained and the teacher model respectively; where λ and τ are adjusted according to different tasks.

[0078] S208. Determine the first feature activations corresponding to each first feature layer in the teacher network model according to the first recognition result. The first feature layer includes at least one first channel, each first channel contains at least one first feature activation, and each first feature activation corresponds to a spatial position.

[0079] In this embodiment, the first feature layer can be understood as the feature layer of the teacher network model; the first feature activation can be understood as a kind of feature data obtained after being processed by the first feature layer; the first channel can be understood as the channel of each layer of the feature layer of the teacher network model; in a neural network model, multiple channels refer to multiple channels output by a convolutional layer, allowing the model to process data from multiple dimensions. Each feature layer includes at least one first channel, each first channel contains at least one first feature activation, and each first feature activation corresponds to a spatial position.

[0080] When the teacher network model processes power grid service data, the data is sequentially processed through different first feature layers, and each first feature layer outputs the first feature activations corresponding to this feature layer. Analyze the first recognition result to determine the first feature activations corresponding to each first feature layer in the teacher network model. The number of channels corresponding to the data output by each feature layer is at least one, that is, the data output by each feature layer includes at least one first channel, each first channel contains at least one first feature activation, and each first feature activation corresponds to a spatial position.

[0081] For example, the teacher network model includes 5 first feature layers, the data output by each first feature layer includes the first feature activations of 3 first channels, each first channel includes W*H first feature activations, W and H are the sizes of the data output by the first feature layer, and each first feature activation corresponds to a spatial position.

[0082] S209. Determine the second feature activations corresponding to each second feature layer in the student network model to be trained according to the second recognition result. The second feature layer includes at least one second channel, each second channel contains at least one second feature activation, and each second feature activation corresponds to a spatial position.

[0083] In this embodiment, the second feature layer can be understood as the feature layer of the teacher network model; the second feature activation can be understood as a kind of feature data obtained after being processed by the second feature layer; the second channel can be understood as the channel of each layer of the feature layer of the student network model to be trained; each feature layer includes at least one second channel, each second channel contains at least one second feature activation, and each second feature activation corresponds to a spatial position.

[0084] When the learning network model to be trained processes power grid service data, the data is sequentially processed through different second feature layers, and each second feature layer outputs the second feature activation corresponding to this feature layer. Analyze the second recognition result to determine the second feature activation corresponding to each second feature layer in the teacher network model. The number of channels corresponding to the data output by each feature layer is at least one, that is, the data output by each feature layer includes at least one second channel, each second channel includes at least one second feature activation, and each second feature activation corresponds to a spatial position.

[0085] S210. Perform feature layer and channel mapping on the first feature layer and the first channel to determine at least one first feature layer corresponding to each second feature layer, and at least one first channel corresponding to each second channel of the second feature layer.

[0086] Since the structures of the teacher network model and the student network model to be trained are different, generally, the number of their feature layers is also different. Generally, the structure of the teacher network model is more complex, so the number of its feature layers is more than that of the student network model to be trained, that is, the number of the first feature layers is more than that of the second feature layers. The number of channels of the data output by the feature layers of the teacher network model and the student network model to be trained may also be different, and channel mapping is also required.

[0087] Perform mapping on the feature layer and the channel respectively to determine at least one first feature layer corresponding to each second feature layer, and at least one first channel corresponding to each second channel of the second feature layer; for example, the first feature layer includes T1, T2, T3, and T4, and the second feature layer includes S1 and S2. Through feature layer mapping, it is determined that the second feature layer S1 corresponds to the first feature layers T1 and T2, and the second feature layer includes S2 corresponding to the first feature layers T3 and T4; or, the first feature layer includes T1, T2, and T3, and the second feature layer includes S1 and S2. Through feature layer mapping, it is determined that the second feature layer S1 corresponds to the first feature layers T1 and T2, and the second feature layer includes S2 corresponding to the first feature layers T2 and T3. For example, the first channel includes R1, R2, and R3, and the second channel includes A1 and A2. Through channel mapping, it is determined that the second channel A1 corresponds to the first channels R1 and R2, and the second channel A2 corresponds to the first channel R3.

[0088] S211. Calculate the loss function based on the probability distribution according to at least one first feature layer corresponding to the second feature layer and at least one first channel corresponding to each second channel of the second feature layer to obtain the third target loss function.

[0089] In this embodiment, the third target loss function can be understood as a loss function, and the third target loss function is a loss function based on probability distribution. Analyze each second feature layer and its corresponding first feature layer to unify the feature layer data. At the same time, analyze at least one first channel corresponding to each second channel of the second feature layer to unify the channel data, determine the first feature activation corresponding to the second feature activation, and then calculate the loss function according to the second feature activation and its corresponding first feature activation. The loss function is calculated by using the loss function based on probability distribution to obtain the third target loss function.

[0090] S212. Calculate the mean error loss function according to the model recognition result and the label data in the second recognition result to obtain the fourth target loss function.

[0091] In this embodiment, the fourth target loss function can be understood as a loss function, and the fourth target loss function is the mean error loss function. The model recognition result is the recognition result of the final output of the student network model to be trained, that is, the data output by the last layer of the student network model to be trained. Substitute the model recognition result and the label data in the second recognition result into the loss function, calculate the mean error loss function, and obtain the fourth target loss function.

[0092] S213. Weight the third target loss function and the fourth target loss function to obtain the second loss function.

[0093] Determine the weight coefficient, and perform weighted summation on the third target loss function and the fourth target loss function according to the weight coefficient to obtain the second loss function.

[0094] The knowledge distillation method based on response is used for classification problems. The mean square error function is usually used for model training and prediction in regression problems, and it is impossible to directly use the neural response output by the last layer of the teacher model to guide the training of the student model. Therefore, channel knowledge distillation is introduced to use the information hint of the intermediate feature layer of the teacher network to help the training of the student network.

[0095] The channel knowledge distillation method normalizes the feature activation values of each channel in a certain feature layer of the teacher and the student to obtain the corresponding probability distribution, and then minimizes the KL divergence between the teacher and student channel probability distributions, guiding the student network to pay more attention to the regions with significant activation values in each channel during the distillation process. Exemplarily, represent the teacher model as t and the student model as s, and the corresponding feature activations are y t and y s , and the distillation loss L CD (that is, the third target loss function) is calculated as follows:

[0096]

[0097] Among them, c is the channel index, and C is the total number of channel indexes; i corresponds to the spatial position of channel c, T is the distillation temperature, and increasing T makes the probability distribution softer, indicating that a wider spatial area in each channel is concerned. W and H are the width and height respectively.

[0098] The logits scores output by each channel are converted into a probability distribution through formula (3) to eliminate the influence of the amplitude scale between the teacher network and the student network. If the number of channels between the teacher and the student does not match, a 1×1 convolutional layer is used to upsample the number of channels of the student network. In the channel knowledge distillation method, the loss value L for the teacher model to guide the training of the student model CWD is shown in formula (4).

[0099] L CWD = αL MSE +(1 - α)L CD (4)

[0100] Among them, L CWD is the second loss function, L MSE is the mean error loss function for the normal training of the student model, that is, the fourth target loss function, and α is the weight coefficient.

[0101] S214. Determine the first weight and the second weight according to the first loss function and the second loss function.

[0102] Analyze the first loss function and the second loss function, calculate the first weight corresponding to the first loss function, and the second weight corresponding to the second loss function. For example, pre - generate the weights corresponding to loss functions of different sizes, and select the matching weights as the first weight and / or the second weight according to the size of the first loss function and / or the second loss function; or, preset the sizes of the two weights, compare the sizes of the first loss function and the second loss function, and select the matching weights according to the size, etc.

[0103] As an optional embodiment, this optional embodiment further optimizes determining the first weight and the second weight according to the first loss function and the second loss function to: sum the first loss function and the second loss function to obtain the sum of the loss functions; calculate the ratio of the first loss function to the sum of the loss functions to obtain the first weight; calculate the ratio of the second loss function to the sum of the loss functions to obtain the second weight.

[0104] S215. Weight the first loss function according to the first weight to obtain the first weighted loss function.

[0105] In this embodiment, the first weighted loss function can be understood as a loss function obtained by weighting the first loss function. The first loss function is weighted according to the first weight, that is, the first weight and the first loss function are multiplied to obtain the first weighted loss function.

[0106] S216. Weight the second loss function according to the second weight to obtain a second weighted loss function.

[0107] In this embodiment, the second weighted loss function can be understood as a loss function obtained by weighting the second loss function. The second loss function is weighted according to the second weight, that is, the second weight and the second loss function are multiplied to obtain the second weighted loss function.

[0108] S217. Sum the first weighted loss function and the second weighted loss function to obtain a fitting loss function.

[0109] Exemplarily, an embodiment of the present application provides a calculation formula for a fitting loss function:

[0110] L MT-DL = ωL CWD +(1 - ω)L KD (5)

[0111]

[0112] where, L MT-DL is the fitting loss function, ω is the second weight, and L CWD is the second loss function; 1 - ω is the first weight, and L KD is the first loss function.

[0113] In an embodiment of the present application, to further enhance the effect of the student model learning from the teacher model, a student model that meets the multi-service requirements of the distribution network is trained, which can simultaneously meet the classification and regression problems of the distribution network services. On the basis of response knowledge distillation and channel knowledge distillation, a multi-task-based hybrid knowledge distillation algorithm is added. The student model simultaneously learns the response knowledge and channel-level knowledge of the teacher model, improving the ability of the student model to learn multiple types of tasks.

[0114] S218. Perform backpropagation on the student network model to be trained based on the fitting loss function to obtain the student network model to be trained for the next iteration until the iteration convergence condition is met, and obtain a power grid service recognition model.

[0115] As an alternative embodiment, this alternative embodiment is further optimized to include: obtaining a service recognition result, which is obtained by a power grid service recognition model according to the power grid service data to be recognized and the type of the model task corresponding to the power grid service data to be recognized; detecting that the service recognition result meets the model update condition, generating new training samples according to the service recognition result; and training and updating the power grid service recognition model according to the new training samples.

[0116] In this embodiment, the service recognition result can be understood as the result obtained by the trained power grid service recognition model for recognizing the power grid service data. The power grid service data to be recognized can be understood as the power grid service data with recognition requirements; the model update condition can be understood as the condition for determining whether to update the power grid service recognition model. For example, whether the service recognition result is accurate, whether the service recognition result is for new service data, and so on.

[0117] The trained power grid service recognition model has high accuracy, small volume, and less resource occupation, so it can be deployed to edge devices. The power grid service data to be recognized and the type of the model task corresponding to the power grid service data to be recognized are input into the power grid service recognition model for recognition, and the power grid service recognition model performs recognition according to the learned knowledge to obtain the service recognition result. This execution device obtains the service recognition result obtained by the power grid service recognition model for recognizing the power grid service data to be recognized. The model update condition is preset, the service recognition result is analyzed to determine whether it meets the model update condition. If it meets, new training samples are generated according to the service recognition result; the power grid service recognition model is continuously trained according to the new training samples to update the power grid service recognition model. When continuing to train the power grid service recognition model, the teacher network model can be trained first according to the new training samples. After the teacher network model is trained, the power grid service recognition model is guided to train through the teacher network model.

[0118] Optionally, the model update condition includes at least one of the following:

[0119] The confidence level of the service recognition result is not greater than the preset confidence level threshold and the service recognition result is for new service data;

[0120] The comparison result between the service recognition result and its corresponding actual service data does not meet the service accuracy condition and the service recognition result is for new service data.

[0121] In this embodiment, the confidence threshold can be set in advance; the business accuracy condition can be that the accuracy rate of the business recognition result is not lower than the set threshold. After receiving the business recognition result, if the business recognition is a classification task, analyze the confidence of the business recognition result. If the confidence of the business recognition result is not greater than the preset confidence threshold, determine whether the business recognition result is new business data. If the confidence of the business recognition result is not greater than the preset confidence threshold and the business recognition result is new business data, it can be considered that the model update condition is satisfied. After receiving the business recognition result, if the business recognition result is a regression task, monitor the actual business data corresponding to the business recognition result, and compare the business recognition result with its corresponding actual business data when idle to determine whether the comparison result meets the business accuracy condition. If not, determine whether the business recognition result is new business data; if the comparison result between the business recognition result and its corresponding actual business data does not meet the business accuracy condition and the business recognition result is new business data, it can be considered that the model update condition is satisfied. Alternatively, when both of the above two conditions are met, it is confirmed that the model update condition is satisfied.

[0122] As an alternative embodiment, in this alternative embodiment, further optimize generating new training samples according to the business recognition result to: construct a business mathematical model according to the business recognition result; perform simulation according to the business mathematical model to generate simulation data corresponding to the business recognition result; generate new training samples according to the simulation data.

[0123] In this embodiment, the business mathematical model can be understood as a mathematical model for describing business data, and the relationship between business data can be described through the business mathematical model. Analyze the business recognition result, and construct a business mathematical model according to the data type corresponding to the business recognition result, the relationship between the business recognition result and other data, etc.; perform simulation on the business recognition result through the business mathematical model to generate simulation data corresponding to the business recognition result; generate new training samples according to the simulation data.

[0124] An embodiment of the present invention provides a method for generating a power grid service recognition model, which solves the problems of low model recognition accuracy, high latency, and inability to process multi-service data. Training samples are generated through at least two types of model tasks, and the types of model tasks at least include classification tasks and regression tasks, which can realize the recognition of multi-tasks and then realize the processing of multi-service data. In the process of model training in this application embodiment, different types of model tasks can be taken into account at the same time to ensure that the student network model to be trained can accurately recognize different types of service data. Based on multi-task hybrid knowledge distillation, model distillation for problems such as classification and regression can be carried out, improving the generalization ability of the student model and meeting the multi-service requirements of the distribution network. Through the method of knowledge distillation for model training, the obtained power grid service recognition model has a high recognition accuracy, and the model is small and can be directly deployed on edge devices, so that the edge devices can directly recognize the power grid service data, improving the generalization ability of the edge-side model and meeting the multi-service requirements of the distribution network. On the premise of limited edge computing resources, the accuracy of the distribution network edge-side model can be guaranteed; the data transmission volume of the distribution network is reduced, the latency caused by the transmission distance is reduced, and the processing efficiency of the massive data of the distribution network is improved.

[0125] Embodiment III

[0126] Figure 3 The figure is a flowchart of a power grid service recognition method provided by Embodiment III of the present invention. This embodiment is applicable to the situation of accurately recognizing power grid services. This method is applied to edge devices in the distribution network and can be executed by a power grid service recognition device. The power grid service recognition device can be implemented in the form of hardware and / or software, and the power grid service recognition device can be configured in an electronic device. As Figure 3 shown, the method includes:

[0127] S301. Obtain the power grid service data to be recognized and the type of the model task corresponding to the power grid service data to be recognized.

[0128] The power grid service data to be recognized can be power grid service data collected in real time by a hardware device, or power grid service data processed or calculated by software, etc. Obtain the power grid service data to be recognized and the type of its corresponding model task. The type of the model task can be determined by classifying and recognizing the power grid service data to be recognized, or can be manually input by a staff member or determined by other means.

[0129] S302. Input the power grid service data to be recognized and the type of the model task into the power grid service recognition model. The power grid service recognition model is obtained by training the model using the method for generating a power grid service recognition model according to any embodiment of the present invention.

[0130] Previously, a model is trained through knowledge distillation by the method provided in any embodiment of this application to obtain a power grid service recognition model. The trained power grid service recognition model is deployed on edge devices in the distribution network. The power grid service data to be recognized and the type of the corresponding model task are input into the power grid service recognition model for processing. The power grid service recognition model recognizes the power grid service data to be recognized based on the type of the pattern task and the knowledge learned during the training process, and obtains and outputs the recognition result.

[0131] S303. Determine the service recognition result according to the output of the power grid service recognition model.

[0132] After completing the data recognition, the power grid service recognition model can directly output the service recognition result. For example, if the model task is topology recognition, the corresponding service recognition result can be the topological structure; if the model task is power quality monitoring, the corresponding service recognition result can be power quantity disturbance; if the model task is fault warning, the corresponding service recognition result can be whether there is a fault, the type of the fault, the location where the fault occurs, etc.; if the model task is load forecasting, the corresponding service recognition result can be the trend of voltage and / or current.

[0133] S304. Perform operation and maintenance response according to the service recognition result.

[0134] Analyze the service recognition result to determine whether relevant operation and maintenance operations need to be performed to achieve operation and maintenance response. For example, if the service recognition result is that there is a fault, control the faulty device to restart. Or, upload the service recognition result to the cloud and receive the control instructions sent by the cloud to perform operation and maintenance response.

[0135] The embodiment of the present invention provides a power grid service recognition method, which solves the problems of low model recognition accuracy, high latency, and inability to process multi-service data. By pre-training the model through knowledge distillation, a power grid service recognition model with a simple structure and small computational complexity is obtained. The power grid service recognition model is directly deployed on edge devices to recognize power grid service data, with high recognition accuracy and the ability to achieve multi-task recognition, thereby realizing the processing of multi-service data; based on multi-task hybrid knowledge distillation for model training, model distillation for classification and regression problems can be performed, improving the generalization ability of the student model and meeting the multi-service requirements of the distribution network; on the premise of limited edge computing resources, the accuracy of the distribution network edge-side model can be guaranteed; the data transmission volume of the distribution network is reduced, the latency caused by the transmission distance is reduced, and the processing efficiency of the massive data of the distribution network is improved.

[0136] Optionally, the type of the model task is a classification task. Performing operation and maintenance response according to the service recognition result includes:

[0137] A1. Determine the confidence level of the service recognition result, and upload the service recognition result and the confidence level to the cloud.

[0138] When the type of the model task is a classification task, the power grid service recognition model outputs the confidence level of the service result while outputting the service recognition result, and the edge device uploads the service recognition result and the confidence level to the cloud.

[0139] A2. Receive the first control instruction sent by the cloud. The first control instruction is generated by the cloud according to the service recognition result when the confidence level is greater than the preset confidence level threshold, or the first control instruction is generated by the cloud according to the user operation when the confidence level is not greater than the preset confidence level threshold.

[0140] In this embodiment, the first control instruction can be understood as an instruction for controlling the hardware devices in the power grid accordingly. The confidence level threshold can be set in advance, for example, 0.9; the confidence level threshold can be set according to the accuracy requirements of the actual service type. After receiving the service recognition result and its corresponding confidence level, the cloud compares the confidence level with the confidence level threshold. If the confidence level is greater than the confidence level threshold, the service recognition result is reliable, and the cloud can directly generate the first control instruction according to the service recognition result; if the confidence level is not greater than the confidence level threshold, it is considered that the service recognition result is unreliable, and the service recognition result is displayed to the user for viewing, or the user is prompted to view it. After the user views the service recognition result and triggers a relevant operation, the cloud receives the user operation and generates the corresponding first control instruction according to the user operation. The cloud sends the first control instruction to the edge device, and the edge device receives the first control instruction sent by the cloud.

[0141] A3. Control the device accordingly according to the first control instruction.

[0142] Determine the device that needs to be controlled, and control the device accordingly according to the first control instruction. For example, control the device to restart, update and correct the parameters of the device, etc.

[0143] Optionally, when the type of the model task is a regression task, perform operation and maintenance response according to the service recognition result, including:

[0144] B1. Upload the service recognition result to the cloud.

[0145] When the type of the model task is a regression task, directly upload the service recognition result to the cloud.

[0146] B2. Receive the second control instruction sent by the cloud. The second control instruction is generated by the cloud according to the service recognition result.

[0147] In this embodiment, the second control instruction can be understood as an instruction for controlling the devices in the distribution network. The cloud receives the service recognition result, analyzes and processes the service recognition result to determine the second control instruction, and sends the second control instruction to the edge devices.

[0148] B3. Perform corresponding control on the devices according to the second control instruction.

[0149] Determine the devices that need to be controlled, and perform corresponding control on the devices according to the first control instruction. For example, control the device to restart, update and correct the parameters of the device, etc.

[0150] The power grid service recognition model adopted in the embodiments of the present application can recognize different types of power grid service data, output different prediction results for different types of model tasks, provide effective prediction data for different types of tasks, and assist in the operation and maintenance work. The cloud sends the first control instruction or the second control instruction according to the output result, instructs the edge device to perform operation and maintenance response, and re-trains the model when the model accuracy is low and the data is new data to improve the model recognition / prediction accuracy.

[0151] Exemplarily, Figure 4 An implementation example diagram of model deployment is provided, including the following steps:

[0152] S1. The cloud collects and generates a historical data set.

[0153] The cloud generates a data set with labels according to the mathematical model and adds it to the historical collection data set to provide training samples for the original teacher model and student model in the cloud. The cloud can also generate a simulation data set and generate training samples according to the simulation data set and the historical data set.

[0154] S2. The teacher network model guides the student network model to train according to the knowledge distillation algorithm.

[0155] The cloud constructs a convolutional neural network (the teacher network model and the student network model can be selected according to different service requirements of the distribution network), inputs the generated training samples into the teacher model, first trains the teacher model, and then uses the multi-task based hybrid knowledge distillation algorithm to let the teacher network model guide the student network model to train. After the training is completed, the model is saved to the cloud server.

[0156] S3. Deployment of the student network model.

[0157] The trained student network model in the cloud is sent to each intelligent fusion terminal on the edge side by means of wireless communication to complete the model deployment on the edge side. The student network model is the trained power grid service recognition model.

[0158] S4. After the distribution network data is transmitted to the edge device, data preprocessing is performed.

[0159] S5. The student network model performs identification / prediction.

[0160] S6. Determine whether the identification / prediction is accurate. If so, execute S7; otherwise, execute S8.

[0161] After receiving the terminal data, the intelligent fusion terminal at the edge node performs simple data processing on it, adjusts it into a data form that can be input into the model, and temporarily saves it, and then inputs it into the student network model on the edge side. Classification task judgment method: Obtain the output value of each path of softmax, and judge whether the probability of the category with the highest output value exceeds 90%, that is, judge whether the identification is accurate. If it exceeds 90%, continue with S7; if it is less than 90%, go to S8; Regression task judgment method: The prediction fitting result of the student model on the edge side directly performs S7. When the system is idle, the actual data and the predicted service identification result are compared at the edge. If the data point fitting accuracy rate exceeds 90%, the data is uploaded and saved in the cloud historical database; if it is less than 90%, go to S8.

[0162] S7. Perform operation and maintenance response according to the model determination. The response process is as follows: Upload the service identification result to the cloud, then the cloud issues corresponding instructions, and then the edge side conveys the instructions to the terminal to control the device status.

[0163] S8. Upload data to the cloud.

[0164] S9. The cloud determines whether to trigger model update. If so, execute S10; otherwise, execute S7.

[0165] S10. Generate new simulation data.

[0166] S11. Retrain the teacher model based on the established model, and return to execute S2.

[0167] When the edge side cannot identify or predict accurately, it is necessary to upload the signal data to the cloud in time to trigger model update, construct a mathematical model for the newly appeared signal data, and generate simulation data of the signal in the cloud. Based on the established teacher model, use the new simulation data to train and update it, that is, repeat steps S1 - S3 to update the student model on the edge side.

[0168] Embodiment 4

[0169] Figure 5 It is a schematic structural diagram of a device for generating a power grid service identification model provided by Embodiment 4 of the present invention. As Figure 5 shown, the device includes: a sample set acquisition module 41, a sample identification module 42, a loss function calculation module 43, and a loss function fitting module 44.

[0170] A sample set acquisition module 41, configured to acquire a training sample set formed by power grid service data, where the training sample set includes at least two training samples and label data corresponding to the training samples, each of the training samples corresponds to at least two types of model tasks, and the types of the model tasks at least include a classification task and a regression task;

[0171] A sample recognition module 42, configured to input the training samples in the current iteration and the types of the corresponding model tasks into a pre-trained teacher network model for recognition to obtain a first recognition result, and input the training samples in the current iteration and the types of the corresponding model tasks into a student network model to be trained for recognition to obtain a second recognition result;

[0172] A loss function calculation module 43, configured to determine a first loss function and a second loss function according to the first recognition result, the second recognition result, and the label data, where the first loss function is a loss function based on response knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation;

[0173] A loss function fitting module 44, configured to calculate a fitting loss function according to the first loss function and the second loss function, and perform backpropagation on the student network model to be trained based on the fitting loss function to obtain a student network model to be trained for the next iteration until an iteration convergence condition is satisfied, so as to obtain a power grid service recognition model.

[0174] An embodiment of the present invention provides a device for generating a power grid service recognition model, which solves the problems of low model recognition accuracy, high latency, and inability to process multi-service data. Training samples are generated through at least two types of model tasks, and the types of model tasks at least include classification tasks and regression tasks, which can realize the recognition of multi-tasks, and then realize the processing of multi-service data; the training samples and the types of the corresponding model tasks are simultaneously input into a trained teacher network model and a student network model to be trained for recognition respectively, obtaining a first recognition result and a second recognition result. The training of the student network model is guided by the teacher network model, and a first loss function and a second loss function are calculated. The first loss function is a loss function based on response-based knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation. The fitting loss function is calculated according to the first loss function and the second loss function, and backpropagation is performed on the student network model to be trained. In the process of model training in this application embodiment, different types of model tasks can be taken into account at the same time, ensuring that the student network model to be trained can accurately recognize different types of service data. The model is trained by means of knowledge distillation, and the obtained power grid service recognition model has a high recognition accuracy and a small model size, which can be directly deployed on edge devices, so that the edge devices can directly recognize the power grid service data, reduce the latency time of data recognition, and improve the data processing efficiency.

[0175] Optionally, the loss function calculation module 43 includes:

[0176] The first response determination unit is configured to determine a first neural response result output by the last layer of the teacher network model according to the first recognition result;

[0177] The second response determination unit is configured to determine a second neural response result output by the last layer of the student network model to be trained according to the second recognition result;

[0178] The first objective function calculation unit is configured to calculate a first objective loss function according to the first neural response result and the second neural response result;

[0179] The second objective function calculation unit is configured to calculate a second objective loss function according to the second neural response result and the label data;

[0180] The first loss function calculation unit is configured to weight the first objective loss function and the second objective loss function to obtain a first loss function.

[0181] Optionally, the loss function calculation module 43 includes:

[0182] The first feature activation determination unit is configured to determine the first feature activation corresponding to each first feature layer in the teacher network model according to the first recognition result. The first feature layer includes at least one first channel, each first channel includes at least one first feature activation, and each first feature activation corresponds to a spatial position;

[0183] The second feature activation determination unit is configured to determine the second feature activation corresponding to each second feature layer in the student network model to be trained according to the second recognition result. The second feature layer includes at least one second channel, each second channel includes at least one second feature activation, and each second feature activation corresponds to a spatial position;

[0184] The mapping unit is configured to perform feature layer and channel mapping on the first feature layer and the first channel to determine at least one first feature layer corresponding to each second feature layer, and at least one first channel corresponding to each second channel of the second feature layer;

[0185] The third objective function calculation unit is configured to calculate a loss function based on a probability distribution according to at least one first feature layer corresponding to the second feature layer and at least one first channel corresponding to each second channel of the second feature layer, and obtain a third objective loss function;

[0186] The fourth objective function calculation unit is configured to calculate a mean error loss function according to the model recognition result in the second recognition result and the label data, and obtain a fourth objective loss function;

[0187] The second loss function calculation unit is configured to weight the third objective loss function and the fourth objective loss function to obtain a second loss function.

[0188] Optionally, the loss function fitting module 44 includes:

[0189] The weight determination unit is configured to determine a first weight and a second weight according to the first loss function and the second loss function;

[0190] The first weighting unit is configured to weight the first loss function according to the first weight to obtain a first weighted loss function;

[0191] The second weighting unit is configured to weight the second loss function according to the second weight to obtain a second weighted loss function;

[0192] The function fitting unit is configured to sum the first weighted loss function and the second weighted loss function to obtain a fitting loss function.

[0193] Optionally, the weight determination unit is specifically configured to:

[0194] Sum the first loss function and the second loss function to obtain the sum of loss functions;

[0195] Calculate the ratio of the first loss function to the sum of loss functions to obtain the first weight;

[0196] Calculate the ratio of the second loss function to the sum of loss functions to obtain the second weight.

[0197] Optionally, the apparatus further includes:

[0198] A service recognition result acquisition module, configured to acquire a service recognition result, where the service recognition result is a power grid service recognition model;

[0199] A training sample update module, configured to detect that the service recognition result meets the model update condition, and generate new training samples according to the service recognition result;

[0200] A model update module, configured to train and update the power grid service recognition model according to the new training samples.

[0201] Optionally, the model update condition includes at least one of the following:

[0202] The confidence level of the service recognition result is not greater than a preset confidence level threshold and the service recognition result is new service data;

[0203] The comparison result between the service recognition result and its corresponding actual service data does not meet the service accuracy condition and the service recognition result is new service data.

[0204] Optionally, the training sample update module includes:

[0205] A model construction unit, configured to construct a service mathematical model according to the service recognition result;

[0206] A simulation unit, configured to perform simulation according to the service mathematical model to generate simulation data corresponding to the service recognition result;

[0207] A training sample update unit, configured to generate new training samples according to the simulation data.

[0208] The power grid service recognition model generation apparatus provided by the embodiments of the present invention can execute the power grid service recognition model generation method provided by any embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method.

[0209] Embodiment 5

[0210] Figure 6The following is a schematic structural diagram of a power grid service recognition device provided in Embodiment 5 of the present invention, which is applied to edge devices in a distribution network. As Figure 6 shown, the device includes: a data to be recognized acquisition module 51, a model input module 52, a service recognition result determination module 53, and an operation and maintenance response module 54.

[0211] The data to be recognized acquisition module 51 is used to acquire the power grid service data to be recognized and the type of the model task corresponding to the power grid service data to be recognized;

[0212] The model input module 52 is used to input the power grid service data to be recognized and the type of the model task into a power grid service recognition model, and the power grid service recognition model is obtained by training the model using the generation method of the power grid service recognition model according to any one of claims 1-5;

[0213] The service recognition result determination module 53 is used to determine the service recognition result according to the output of the power grid service recognition model;

[0214] The operation and maintenance response module 54 is used to perform operation and maintenance response according to the service recognition result.

[0215] Embodiment of the present invention provides a power grid service recognition device, which solves the problems of low model recognition accuracy, high latency, and inability to process multi-service data. The power grid service recognition model with a simple structure and small computational complexity is obtained by pre-training the model through knowledge distillation. The power grid service recognition model is directly deployed on edge devices to recognize power grid service data, with high recognition accuracy and the ability to realize multi-task recognition, thereby realizing the processing of multi-service data; model training is carried out based on multi-task hybrid knowledge distillation, and model distillation for problems such as classification and regression can be performed, improving the generalization ability of the student model and meeting the multi-service requirements of the distribution network; on the premise of limited edge computing resources, the model accuracy on the edge side of the distribution network can be guaranteed; the data transmission volume of the distribution network is reduced, the latency caused by the transmission distance is reduced, and the processing efficiency of the massive data of the distribution network is improved.

[0216] Optionally, the type of the model task is a classification task, and the operation and maintenance response module includes:

[0217] The first data upload unit is used to determine the confidence level of the service recognition result and upload the service recognition result and the confidence level to the cloud;

[0218] A first instruction receiving unit, configured to receive a first control instruction sent by the cloud. The first control instruction is generated by the cloud according to the service recognition result when the confidence level is greater than a preset confidence level threshold, or the first control instruction is generated by the cloud according to a user operation when the confidence level is not greater than the preset confidence level threshold;

[0219] A first control unit, configured to perform corresponding control on the device according to the first control instruction.

[0220] Optionally, the type of the model task is a regression task. The operation and maintenance response module includes:

[0221] A second data uploading unit, configured to upload the service recognition result to the cloud;

[0222] A second instruction receiving unit, configured to receive a second control instruction sent by the cloud. The second control instruction is generated by the cloud according to the service recognition result;

[0223] A second control unit, configured to perform corresponding control on the device according to the second control instruction.

[0224] The power grid service recognition device provided by the embodiments of the present invention can execute the power grid service recognition method provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution of the method.

[0225] Embodiment Six

[0226] Figure 7 FIG. is a schematic structural diagram of an electronic device according to Embodiment Six of the present invention. The electronic device 60 can be used to implement any method provided by the embodiments of the present invention. The electronic device is intended to represent various forms of digital computers, such as, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, a personal digital assistant, a cellular phone, a smart phone, a wearable device (such as a helmet, glasses, a watch, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described herein and / or claimed.

[0227] As Figure 7As shown, the electronic device 60 includes at least one processor 61 and a memory communicatively connected to the at least one processor 61, such as a read-only memory (ROM) 62, a random access memory (RAM) 63, etc. The memory stores a computer program executable by the at least one processor. The processor 61 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 62 or the computer program loaded from the storage unit 68 into the random access memory (RAM) 63. In the RAM 63, various programs and data required for the operation of the electronic device 60 can also be stored. The processor 61, the ROM 62, and the RAM 63 are connected to each other via a bus 64. An input / output (I / O) interface 65 is also connected to the bus 64.

[0228] Multiple components in the electronic device 60 are connected to the I / O interface 65, including: an input unit 66, such as a keyboard, a mouse, etc.; an output unit 67, such as various types of displays, speakers, etc.; a storage unit 68, such as a magnetic disk, an optical disc, etc.; and a communication unit 69, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 69 allows the electronic device 60 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0229] The processor 61 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 61 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 61 executes the various methods and processes described above, such as the method for generating a power grid service identification model or the power grid service identification method.

[0230] In some embodiments, the method for generating a power grid service identification model or the power grid service identification method can be implemented as a computer program tangibly embodied in a computer-readable storage medium, such as the storage unit 68. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 60 via the ROM 62 and / or the communication unit 69. When the computer program is loaded into the RAM 63 and executed by the processor 61, one or more steps of the method for generating a power grid service identification model or the power grid service identification method described above can be executed. Alternatively, in other embodiments, the processor 61 can be configured to execute the method for generating a power grid service identification model or the power grid service identification method by any other appropriate means (e.g., by means of firmware).

[0231] The various embodiments of the systems and techniques described above in this specification can be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which may be a special-purpose or general-purpose programmable processor that receives data and instructions from, and transmits data and instructions to, a storage system, at least one input device, and at least one output device.

[0232] The computer programs for implementing the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to the processors of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the computer programs, when executed by the processors, cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine, or entirely on the remote machine or server.

[0233] Embodiments of the present invention provide a computer program product, which includes a computer program that, when executed by a processor, implements the method for generating a power grid service identification model or the power grid service identification method according to any embodiment of the present invention.

[0234] In the context of the present invention, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. The computer-readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of the machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0235] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or a trackball) through which the user can provide input to the electronic device. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0236] The systems and techniques described herein can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected to each other by digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), blockchain network, and the Internet.

[0237] A computing system can include a client and a server. The client and the server are generally far from each other and typically interact through a communication network. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system and solves the defects of difficult management and weak business scalability existing in traditional physical hosts and VPS services.

[0238] It should be understood that various forms of the processes shown above can be used, with steps reordered, added, or deleted. For example, the steps recited in the present invention can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved, and no limitation is made herein.

[0239] The above specific embodiments do not constitute a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for generating a power grid service identification model, characterized in that: include: Acquire a training sample set formed by power grid service data, wherein the training sample set includes at least two training samples and label data corresponding to the training samples, each of the training samples corresponds to at least two types of model tasks, and the types of the model tasks include at least classification tasks and regression tasks; Input the training samples in the current iteration and the types of the corresponding model tasks into the pre-trained teacher network model for identification, and obtain a first identification result; input the training samples in the current iteration and the types of the corresponding model tasks into the to-be-trained student network model for identification, and obtain a second identification result; Determine a first loss function and a second loss function according to the first recognition result, the second recognition result and the label data, wherein the first loss function is a loss function based on response knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation; A fitting loss function is calculated according to the first loss function and the second loss function, and the student network model to be trained is back-propagated based on the fitting loss function to obtain the student network model to be trained for the next iteration until the iterative convergence condition is met to obtain a power grid service identification model.

2. The method according to claim 1, characterized in that The determining a first loss function according to the first recognition result, the second recognition result and the label data includes: Determine a first neural response result output by the last layer of the teacher network model according to the first recognition result; Determine a second neural response result output by the last layer of the student network model to be trained according to the second recognition result; Calculate a first objective loss function according to the first neural response result and the second neural response result; Calculate a second objective loss function according to the second neural response result and the label data; The first objective loss function and the second objective loss function are weighted to obtain a first loss function.

3. The method according to claim 1, characterized in that: Determining a second loss function according to the first recognition result, the second recognition result and the label data includes: Determine, according to the first recognition result, a first feature activation corresponding to each first feature layer in the teacher network model, wherein the first feature layer includes at least one first channel, each first channel includes at least one first feature activation, and each first feature activation corresponds to a spatial position; Determine, according to the second recognition result, a second feature activation corresponding to each second feature layer in the student network model to be trained, wherein the second feature layer includes at least one second channel, each second channel includes at least one second feature activation, and each second feature activation corresponds to a spatial position; Performing feature layer and channel mapping on the first feature layer and the first channel to determine at least one first feature layer corresponding to each of the second feature layers and at least one first channel corresponding to each of the second channels of the second feature layer; Calculate a loss function based on probability distribution according to at least one first feature layer corresponding to the second feature layer and at least one first channel corresponding to each second channel of the second feature layer to obtain a third target loss function; Calculate a mean error loss function according to the model recognition result in the second recognition result and the label data to obtain a fourth target loss function; The third objective loss function and the fourth objective loss function are weighted to obtain a second loss function.

4. The method according to claim 1, characterized in that: The calculating the fitting loss function according to the first loss function and the second loss function includes: Determine a first weight and a second weight according to the first loss function and the second loss function; Weighting the first loss function according to the first weight to obtain a first weighted loss function; Weighting the second loss function according to the second weight to obtain a second weighted loss function; The first weighted loss function and the second weighted loss function are summed to obtain a fitting loss function.

5. The method according to claim 4, characterized in that The determining the first weight and the second weight according to the first loss function and the second loss function comprises: Summing the first loss function and the second loss function to obtain a loss function sum; Calculate the ratio of the first loss function to the sum of the loss functions to obtain a first weight; The ratio of the second loss function to the sum of the loss functions is calculated to obtain a second weight.

6. The method according to any one of claims 1 to 5, characterized in that: Also includes: Obtaining a business identification result, where the business identification result is obtained by a power grid business identification model according to the power grid business data to be identified and the type of model task corresponding to the power grid business data to be identified; Detecting that the service identification result meets the model update condition, generating a new training sample according to the service identification result; The power grid service identification model is trained and updated according to the new training samples.

7. The method according to claim 6, characterized in that The model update condition includes at least one of the following: The confidence level of the service identification result is not greater than a preset confidence level threshold and the service identification result is new service data; A comparison result of the business identification result and its corresponding actual business data does not meet the business accuracy condition and the business identification result is new business data.

8. The method according to claim 6, characterized in that The generating a new training sample according to the service identification result includes: Constructing a business mathematical model according to the business identification result; Perform simulation according to the business mathematical model to generate simulation data corresponding to the business identification result; Generate new training samples according to the simulation data.

9. A method for identifying power grid services, characterized in that: Edge devices used in distribution networks include: Acquire the power grid service data to be identified and the type of model task corresponding to the power grid service data to be identified; Inputting the to-be-identified power grid service data and the type of the model task into a power grid service identification model, wherein the power grid service identification model is obtained by model training using the power grid service identification model generation method according to any one of claims 1 to 5; Determine a service identification result according to the output of the power grid service identification model; An operation and maintenance response is performed according to the business identification result.

10. The method according to claim 9, characterized in that The type of the model task is a classification task, and the operation and maintenance response is performed according to the business identification result, including: Determine the confidence level of the service identification result, and upload the service identification result and the confidence level to the cloud; receiving a first control instruction issued by the cloud, where the first control instruction is generated by the cloud according to the service identification result when the confidence is greater than a preset confidence threshold, or the first control instruction is generated by the cloud according to a user operation when the confidence is not greater than a preset confidence threshold; The device is controlled accordingly according to the first control instruction.

11. The method according to claim 9, characterized in that The type of the model task is a regression task, and the operation and maintenance response is performed according to the business identification result, including: Uploading the business identification result to the cloud; receiving a second control instruction issued by the cloud, where the second control instruction is generated by the cloud according to the service identification result; The device is controlled accordingly according to the second control instruction.

12. A device for generating a power grid service identification model, characterized in that: include: A sample set acquisition module is used to acquire a training sample set formed by power grid business data, wherein the training sample set includes at least two training samples and label data corresponding to the training samples, each of the training samples corresponds to at least two types of model tasks, and the types of the model tasks include at least classification tasks and regression tasks; The sample recognition module is used to input the training samples of the current iteration and the types of the corresponding model tasks into the pre-trained teacher network model for recognition, and obtain a first recognition result; and input the training samples of the current iteration and the types of the corresponding model tasks into the student network model to be trained for recognition, and obtain a second recognition result; A loss function calculation module, used to determine a first loss function and a second loss function according to the first recognition result, the second recognition result and the label data, wherein the first loss function is a loss function based on response knowledge distillation, and the second loss function is a loss function based on channel knowledge distillation; A loss function fitting module is used to calculate a fitting loss function according to the first loss function and the second loss function, and to perform backpropagation on the student network model to be trained based on the fitting loss function to obtain a student network model to be trained for the next iteration until an iterative convergence condition is met to obtain a power grid business identification model.

13. The device according to claim 12, characterized in that The loss function calculation module comprises: A first response determination unit, used to determine a first neural response result output by the last layer of the teacher network model according to the first recognition result; A second response determination unit, used to determine a second neural response result output by the last layer of the student network model to be trained according to the second recognition result; A first objective function calculation unit, configured to calculate a first objective loss function according to the first neural response result and the second neural response result; A second objective function calculation unit, configured to calculate a second objective loss function according to the second neural response result and the label data; The first loss function calculation unit is used to weight the first objective loss function and the second objective loss function to obtain a first loss function.

14. The device according to claim 12, characterized in that The loss function calculation module comprises: A first feature activation determination unit is used to determine a first feature activation corresponding to each first feature layer in the teacher network model according to the first recognition result, wherein the first feature layer includes at least one first channel, each first channel includes at least one first feature activation, and each first feature activation corresponds to a spatial position; A second feature activation determination unit is used to determine, according to the second recognition result, a second feature activation corresponding to each second feature layer in the student network model to be trained, wherein the second feature layer includes at least one second channel, each second channel includes at least one second feature activation, and each second feature activation corresponds to a spatial position; a mapping unit, configured to perform feature layer and channel mapping on the first feature layer and the first channel, and determine at least one first feature layer corresponding to each second feature layer, and at least one first channel corresponding to each second channel of the second feature layer; A third objective function calculation unit, configured to calculate a loss function based on probability distribution according to at least one first feature layer corresponding to the second feature layer and at least one first channel corresponding to each second channel of the second feature layer, so as to obtain a third objective loss function; a fourth objective function calculation unit, configured to calculate a mean error loss function according to the model recognition result in the second recognition result and the label data, so as to obtain a fourth objective loss function; The second loss function calculation unit is used to weight the third objective loss function and the fourth objective loss function to obtain a second loss function.

15. The device according to claim 12, characterized in that The loss function fitting module comprises: A weight determination unit, configured to determine a first weight and a second weight according to the first loss function and the second loss function; A first weighting unit, configured to weight the first loss function according to the first weight to obtain a first weighted loss function; A second weighting unit, configured to weight the second loss function according to the second weight to obtain a second weighted loss function; A function fitting unit is used to sum the first weighted loss function and the second weighted loss function to obtain a fitting loss function.

16. The device according to claim 15, characterized in that The weight determination unit is specifically used for: Summing the first loss function and the second loss function to obtain a loss function sum; Calculate the ratio of the first loss function to the sum of the loss functions to obtain a first weight; The ratio of the second loss function to the sum of the loss functions is calculated to obtain a second weight.

17. The device according to any one of claims 12 to 16, characterized in that: Also includes: A service identification result acquisition module is used to obtain a service identification result, wherein the service identification result is a power grid service identification model; A training sample updating module, used to detect that the service identification result meets the model updating condition and generate a new training sample according to the service identification result; The model updating module is used to train and update the power grid service identification model according to the new training samples.

18. The device according to claim 17, characterized in that The model update condition includes at least one of the following: The confidence level of the service identification result is not greater than a preset confidence level threshold and the service identification result is new service data; A comparison result of the business identification result and its corresponding actual business data does not meet the business accuracy condition and the business identification result is new business data.

19. The device according to claim 17, characterized in that The training sample updating module comprises: A model building unit, used to build a business mathematical model according to the business identification result; A simulation module is used to perform simulation according to the business mathematical model to generate simulation data corresponding to the business identification result; The training sample updating unit is used to generate new training samples according to the simulation data.

20. A power grid service identification device, characterized in that: Edge devices used in distribution networks include: A module for acquiring data to be identified, used to acquire the power grid service data to be identified and the type of model task corresponding to the power grid service data to be identified; A model input module, used for inputting the to-be-identified power grid service data and the type of the model task into a power grid service identification model, wherein the power grid service identification model is obtained by model training using the power grid service identification model generation method according to any one of claims 1 to 5; A service identification result determination module, used to determine a service identification result according to the output of the power grid service identification model; The operation and maintenance response module is used to perform operation and maintenance response according to the business identification result.

21. The device according to claim 20, characterized in that The type of the model task is a classification task, and the operation and maintenance response module includes: A first data uploading unit, used to determine the confidence level of the service identification result, and upload the service identification result and the confidence level to the cloud; A first instruction receiving unit, configured to receive a first control instruction issued by the cloud, wherein the first control instruction is generated by the cloud according to the service identification result when the confidence level is greater than a preset confidence threshold, or the first control instruction is generated by the cloud according to a user operation when the confidence level is not greater than a preset confidence threshold; The first control unit is used to control the device accordingly according to the first control instruction.

22. The device according to claim 20, characterized in that The type of the model task is a regression task, and the operation and maintenance response module includes: A second data uploading unit, used to upload the business identification result to the cloud; A second instruction receiving unit, used to receive a second control instruction issued by the cloud, where the second control instruction is generated by the cloud according to the service identification result; The second control unit is used to control the device accordingly according to the second control instruction.

23. An electronic device, characterized in that: The electronic device comprises: at least one processor, and a memory communicatively coupled to the at least one processor; In which, the memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the method for generating a power grid business identification model described in any one of claims 1-8 or the power grid business identification method described in any one of claims 9-11.

24. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the method for generating a power grid service identification model described in any one of claims 1-8 or the method for identifying power grid services described in any one of claims 9-11 when executed.

25. A computer program product, characterized in that The computer program product comprises a computer program, which, when executed by a processor, implements the method for generating a power grid service identification model according to any one of claims 1 to 8 or the method for identifying power grid services according to any one of claims 9 to 11.

Images