Deep learning data transmission method and device based on edge terminal and terminal equipment

Abstract

The application provides a kind of based on edge terminal's deep learning data transmission method, device and terminal equipment, it is related to power distribution network data transmission technical field, method includes: at least one power distribution network operating parameter time series data is collected;With at least one power distribution network operating parameter time series data as input, at least one power distribution network operating parameter time series data is processed by the dimension reduction of pre-constructed encoder model, and the corresponding dimension reduction feature vector is obtained;Dimension reduction feature vector is sent to main station, so that main station obtains at least one power distribution network operating parameter time series data by decoding dimension reduction feature vector through pre-constructed decoder model;Encoder model and decoder model are obtained by training deep learning neural network from the history time series data of different power distribution network operating parameters.The application effectively improves the data transmission efficiency of power distribution network edge side equipment, and reduces the load of main station.

Description

Technical Field

[0001] This application relates to the field of power distribution network data transmission technology, specifically to a deep learning data transmission method based on a side-side terminal, a deep learning data transmission method based on a master station, a deep learning data transmission device based on a side-side terminal, a deep learning data transmission device based on a master station, a machine-readable storage medium, and a terminal device. Background Technology

[0002] With the continuous construction of new power systems and the Internet of Things (IoT) for distribution networks, millions of Distribution Terminal Units (DTUs) are deployed in various substations, switching stations, and ring main units to perform critical tasks such as real-time data collection of power grid operating status, execution of remote control commands, and reporting of telemetry data. These DTUs need to periodically transmit a large number of collected electrical parameters (such as voltage, current, frequency, switch status, and fault information) to the master station system for centralized processing and decision support. In traditional communication architectures, data transmission between DTUs and the master station mostly relies on dedicated lines or 4G / 5G public network communication. Data is reported in plaintext or with simple encoding. While this meets basic functional requirements, it also presents the following problems:

[0003] 1. Poor communication security. Most existing DTU devices do not effectively encrypt data, making them vulnerable to network security threats such as man-in-the-middle attacks, data tampering or forgery, especially in public network transmission scenarios, which seriously affects system stability.

[0004] 2. Increasingly heavy communication load. As the functions of smart grids become more diverse, the dimensions and frequency of data collected by DTUs are constantly increasing. The transmitted content includes various types such as telemetry, remote signaling, remote control, remote adjustment, images, and video, leading to a surge in the receiving pressure on the master station. Especially in scenarios where multiple DTUs report concurrently, communication bandwidth becomes the system bottleneck.

[0005] 3. Lack of effective compression mechanisms. Currently, most DTU devices still use fixed structures or simple compression algorithms (such as Zlib) to encode data, failing to perform deep compression processing based on the temporal correlation and spatial redundancy characteristics of power distribution data. This results in high coding redundancy and low compression ratio, which is not conducive to efficient communication in resource-constrained scenarios.

[0006] To alleviate the aforementioned problems, current methods typically involve reducing the data sampling frequency, using distributed database relay to ease the load on the main station, or setting up local caching for batch reporting to reduce the number of communications. However, these methods can affect data real-time performance or increase system complexity, and they do not fundamentally solve the problems of large data volumes and high communication costs. Therefore, there is an urgent need for a new data compression and transmission method that can leverage the characteristics of deep learning models to efficiently compress data collected by the DTU, and combine this with a secure encryption transmission mechanism to significantly improve transmission efficiency, reduce the load on the main station, and optimize the overall performance of the power distribution automation system while ensuring communication security. Summary of the Invention

[0007] The purpose of this application is to provide a deep learning data transmission method based on a side terminal, a deep learning data transmission method based on a master station, a deep learning data transmission device based on a side terminal, a deep learning data transmission device based on a master station, a machine-readable storage medium, and a terminal device to solve the above-mentioned problems.

[0008] To achieve the above objectives, the first aspect of this application provides a deep learning data transmission method based on a side-side terminal, comprising:

[0009] Collect time-series data of at least one power distribution network operating parameter;

[0010] Using the time-series data of at least one distribution network operating parameter as input, the time-series data of the at least one distribution network operating parameter is dimensionality-reduced through a pre-built encoder model to obtain the corresponding dimensionality-reduced representation vector.

[0011] The dimensionality reduction representation vector is sent to the master station so that the master station can decode the dimensionality reduction representation vector through a pre-built decoder model to obtain the time series data of the at least one distribution network operation parameter;

[0012] The encoder model and the decoder model are obtained by training a deep learning neural network with historical time-series data of different power distribution network operating parameters.

[0013] Optionally, the training steps of the encoder model and the decoder model include:

[0014] Construct an initial encoder model and an initial decoder model based on a deep learning neural network;

[0015] Acquire historical time-series data of different power distribution network operating parameters, and preprocess the historical time-series data of different power distribution network operating parameters;

[0016] The initial encoder model is input with historical time-series data of different power distribution network operating parameters, and the corresponding dimensionality-reduced representation vector is output by the initial encoder model.

[0017] The initial decoder model receives the dimensionality reduction representation vector output by the initial encoder model, decodes the received dimensionality reduction representation vector, and obtains the reconstructed time series data.

[0018] Determine the reconstruction error between the historical time-series data input to the initial encoder model and the reconstructed time-series data output by the initial decoder model. Update the model parameters of the initial encoder model and the initial decoder model based on the reconstruction error until the reconstruction error meets the convergence condition, thereby obtaining the encoder model and the decoder model.

[0019] Optionally, the initial encoder model includes:

[0020] The embedding layer is used to encode the time-series data of the distribution network operating parameters, so as to convert the time-series data of the distribution network operating parameters into the corresponding feature vector representation;

[0021] The attention layer is used to perform QKV transformation on the corresponding feature vector to obtain a query matrix, a key matrix, and a value matrix. Each value in the value matrix is ​​set to 1. Attention is calculated on the corresponding feature vector through the query matrix and the key matrix to obtain the attention matrix of the corresponding feature vector. The attention output matrix of the corresponding feature vector is obtained by weighted summation of the value matrix through the attention matrix.

[0022] At least one fully connected layer is used to perform dimensionality reduction processing on the attention output matrix to obtain the dimensionality reduction representation vector of the corresponding feature vector.

[0023] Optionally, historical time-series data of different distribution network operating parameters are preprocessed, including:

[0024] Identify missing values ​​in the current historical time series data;

[0025] Determine the first trend of change of multiple consecutive data values ​​before the current missing value, and determine the second trend of change of multiple consecutive data values ​​after the current missing value;

[0026] Based on the first trend and the second trend, the predicted value of the current missing value is determined, and the predicted value is used to assign a value to the current missing value.

[0027] Optionally, determining the predicted value of the current missing value based on the first trend and the second trend includes:

[0028] If the first trend of change is the same as the second trend of change, determine the average first change of multiple consecutive data values ​​before the current missing value and the average second change of multiple consecutive data values ​​after the current missing value;

[0029] A positive prediction value for the current missing value is constructed by summing the first valid data value and the mean of the first change, and a negative prediction value for the current missing value is constructed by the difference between the second valid data value and the mean of the second change, wherein the first valid data value is the previous valid data value of the current missing value, and the second valid data value is the next valid data value of the current missing value.

[0030] The first weight of the positive prediction value and the second weight of the negative prediction value are constructed based on the first mean of the change and the second mean of the change, wherein the first weight is positively correlated with the first mean of the change and the second weight is positively correlated with the second mean of the change.

[0031] The positive and negative predicted values ​​are weighted and summed using the first and second weights to obtain the predicted value of the current missing value.

[0032] Optionally, determining the predicted value of the current missing value based on the first trend and the second trend includes:

[0033] If the first trend of change is different from the second trend of change, determine the ranking value of the current missing value in the current missing segment and the number of missing values ​​in the current missing segment;

[0034] The prediction weight of the current missing value is determined based on the ranking value of the current missing value in the current missing segment and the number of missing values ​​in the current missing segment. The prediction weight is positively correlated with the ranking value of the current missing value in the current missing segment.

[0035] Determine the change between the next valid data value and the previous valid data value of the current missing value. Weight the change value with the prediction weight to obtain the predicted adjustment value of the current missing value. Sum the predicted adjustment value and the previous valid data value of the previous valid data value to obtain the predicted value of the current missing value.

[0036] A second aspect of this application provides a deep learning data transmission method based on a master station, including the aforementioned deep learning data transmission method based on a side terminal. The deep learning data transmission method based on a master station includes:

[0037] The encoder model receives a dimension-reduced representation vector from at least one side-side terminal. The dimension-reduced representation vector is obtained by the encoder model after performing dimension reduction processing on the time-series data of at least one distribution network operation parameter.

[0038] Using the reduced-dimensional representation vector as input, the reduced-dimensional representation vector is decoded by a pre-built decoder model to obtain the time-series data of the at least one distribution network operation parameter;

[0039] The current operating status of the distribution network is determined based on the time-series data of at least one of the distribution network operating parameters.

[0040] A third aspect of this application provides a deep learning data transmission device based on an edge terminal, which applies the above-described deep learning data transmission method based on an edge terminal. The device includes:

[0041] The data acquisition module is configured to acquire time-series data of at least one power distribution network operating parameter;

[0042] The data encoding module is configured to take the time-series data of the at least one distribution network operating parameter as input, and perform dimensionality reduction processing on the time-series data of the at least one distribution network operating parameter through a pre-built encoder model to obtain the corresponding dimensionality reduction representation vector.

[0043] The data transmission module is configured to send the dimensionality reduction representation vector to the master station, so that the master station can decode the dimensionality reduction representation vector through a pre-built decoder model to obtain the time-series data of the at least one distribution network operation parameter;

[0044] The encoder model and the decoder model are obtained by training a deep learning neural network with historical time-series data of different power distribution network operating parameters.

[0045] In a fourth aspect, this application provides a deep learning data transmission device based on a master station, applying the aforementioned deep learning data transmission method based on a master station, the device comprising:

[0046] The data receiving module is configured to receive a dimensionality reduction representation vector sent by the encoder model of at least one side terminal. The dimensionality reduction representation vector is obtained by the encoder model after performing dimensionality reduction processing on the time-series data of at least one distribution network operation parameter.

[0047] The data decoding module is configured to take the dimensionality reduction representation vector as input and decode the dimensionality reduction representation vector through a pre-built decoder model to obtain the time series data of the at least one distribution network operation parameter;

[0048] The status identification module is configured to determine the current operating status of the distribution network based on the time-series data of the at least one distribution network operating parameter.

[0049] In a fifth aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform either the aforementioned deep learning data transmission method based on a side-end terminal or the aforementioned deep learning data transmission method based on a master station.

[0050] In a sixth aspect, this application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned deep learning data transmission method based on a side-side terminal, or implements the steps of the aforementioned deep learning data transmission method based on a main station.

[0051] This application constructs an encoder and decoder based on deep learning. By deploying encoder models on each side device and a decoder model on the main station, the side devices encode the collected power distribution network operating parameters through the encoder model and then transmit them, thereby achieving dimensionality reduction and encryption of the collected data. The main station decodes the received low-dimensional data to reconstruct the original data and predicts the operating status of the power distribution network based on the reconstructed data. This application can uniformly compress multiple types of operating data, effectively improving the data transmission efficiency of the power distribution network side devices and reducing the load on the main station.

[0052] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings:

[0054] Figure 1 A flowchart illustrating a preferred embodiment of this application for a deep learning data transmission method based on a side-side terminal;

[0055] Figure 2 A schematic diagram of data compression provided for a preferred embodiment of this application;

[0056] Figure 3 A schematic diagram of data transmission provided for a preferred embodiment of this application;

[0057] Figure 4 This is a schematic diagram of the encoder and decoder structure provided in a preferred embodiment of this application;

[0058] Figure 5 A schematic diagram of the attention mechanism provided in a preferred embodiment of this application;

[0059] Figure 6 This is a schematic diagram of encoder and decoder training provided in a preferred embodiment of this application;

[0060] Figure 7 This is a schematic diagram of the fault prediction model structure provided in a preferred embodiment of this application;

[0061] Figure 8 A flowchart illustrating a preferred embodiment of this application for a deep learning data transmission method based on a master station;

[0062] Figure 9 A schematic diagram of the system structure of the power distribution network data processing system provided in the preferred embodiment of this application;

[0063] Figure 10 A schematic diagram of a deep learning data transmission device based on a side-side terminal provided in a preferred embodiment of this application;

[0064] Figure 11 A schematic diagram of a deep learning data transmission device based on a master station provided in a preferred embodiment of this application;

[0065] Figure 12 A schematic diagram of a terminal device provided for a preferred embodiment of this application.

[0066] Explanation of reference numerals in the attached figures

[0067] 10 - Terminal device, 100 - Processor, 101 - Memory, 102 - Computer program. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0069] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0070] With the continuous development of smart distribution networks, the types of data collected by distribution automation terminals (DTUs) are becoming increasingly diverse, expanding from basic electrical data to event logs, images, and even videos. This has led to a rapid increase in communication data volume, posing a severe challenge to the communication links and data processing capabilities of the master station. Currently, communication between DTUs and the master station generally uses fixed formats or simple compression methods, lacking intelligent compression techniques optimized based on data characteristics. This results in low data upload efficiency, high bandwidth consumption, and a heavy decoding burden on the master station. Especially when processing large-scale data for intelligent analysis tasks such as fault detection and load forecasting, master station resources are often under strain. Existing solutions mainly rely on adding master station servers or expanding cloud resources, but these methods are costly, inefficient, and unable to meet real-time transmission and processing requirements. Furthermore, current data transmission between DTUs and the master station lacks effective data encryption and compression strategies, making it vulnerable to attacks or data leaks. Existing compression algorithms are also ill-suited for the complex structure of distribution operation data.

[0071] To solve the above problems, such as Figure 1 As shown, the first aspect of this application provides a deep learning data transmission method based on a side-end terminal, comprising: collecting time-series data of at least one distribution network operating parameter; using the time-series data of at least one distribution network operating parameter as input, performing dimensionality reduction processing on the time-series data of at least one distribution network operating parameter through a pre-constructed encoder model to obtain a corresponding dimensionality-reduced representation vector; sending the dimensionality-reduced representation vector to a master station, so that the master station decodes the dimensionality-reduced representation vector through a pre-constructed decoder model to obtain the time-series data of at least one distribution network operating parameter; wherein, the encoder model and the decoder model are obtained by training a deep learning neural network with historical time-series data of different distribution network operating parameters.

[0072] Thus, this application constructs an encoder and decoder based on deep learning. By deploying encoder models on each side device and a decoder model on the main station, the side devices encode the collected power distribution network operating parameters through the encoder model and then transmit them, thereby achieving dimensionality reduction and encryption of the collected data. The main station decodes the received low-dimensional data to reconstruct the original data and predicts the operating status of the power distribution network based on the reconstructed data. This application can uniformly compress multiple types of operating data, effectively improving the data transmission efficiency of the power distribution network side devices and reducing the load on the main station.

[0073] Understandably, there are usually multiple DTUs (Distribution Units) on the side of a distribution network, such as... Figure 2 and Figure 3As shown, each DTU is equipped with a pre-built encoder model, and the decoder model is deployed at the master station. The DTU collects distribution network operating parameters and encodes and reduces the dimensionality of these parameters using the encoder model. The encoder model then transmits the dimensionality-reduced low-dimensional data to the decoder model at the master station. The decoder model decodes the received low-dimensional data to reconstruct and restore the original distribution network operating parameters. In this application, the encoder and decoder models can be built based on deep learning. Specifically, the encoder model supports quantization (such as INT8) and pruning techniques, enabling it to run efficiently on resource-constrained DTU devices. The decoder model is deployed at the master station, which can utilize its strong computing power to complete large-scale reconstruction tasks.

[0074] Understandably, the encoder and decoder need to be trained before deployment. In this application, the training steps for the encoder and decoder models include:

[0075] S100. Construct an initial encoder model and an initial decoder model based on a deep learning neural network. Specifically, the initial encoder model and the initial decoder model are constructed according to the large model network framework. The encoder structure includes an embedding layer, a multi-layer self-attention structure, and a dimensionality reduction projection layer, which are used to compress high-dimensional data into a low-dimensional representation. The decoder has a symmetric structure and uses a multi-head attention layer and linear transformation to restore the compressed representation to approximately the original data.

[0076] S200. Acquire historical time-series data of different distribution network operating parameters and preprocess the historical time-series data of different distribution network operating parameters. Training data can be collected from the historical operating parameters of the distribution network in actual operation, directly acquiring multi-dimensional sensor data from different DTUs, including time-series data such as current, voltage, frequency, and active / reactive power. For example, multiple sets of historical time-series data of different dimensions can be collected from different DTUs within a specified historical period, such as multiple sets of historical current time-series data, multiple sets of historical voltage time-series data, etc. Understandably, to ensure training effectiveness, all training data must be standardized and missing value processed before training the encoder and decoder, thereby improving the data reconstruction accuracy. This application adopts a self-supervised learning approach, using the construction of a compression reconstruction task as the training objective. The loss function can be selected as mean squared error (MSE) or weighted MSE. Simultaneously, strategies such as Dropout and residual connections can be introduced during training to improve the model's generalization ability.

[0077] S300 uses historical time-series data of different power distribution network operating parameters as input to the initial encoder model, and outputs the corresponding dimensionality-reduced representation vector. It can be understood that each encoder model can include multiple encoding units to process operating parameters of different dimensions. The encoding units perform feature segmentation and projection processing on the input data based on a self-attention mechanism to map the input data to a low-dimensional space, and extract features based on the attention mechanism to generate a dimensionality-reduced representation vector of the input data, thereby achieving compression of the input data.

[0078] S400: The initial decoder model receives the dimensionality-reduced representation vector output by the initial encoder model, decodes the received dimensionality-reduced representation vector, and obtains the reconstructed time-series data. Understandably, the decoder model deployed at the main station can also include multiple decoding units to process the corresponding time-series data.

[0079] S500: Determine the reconstruction error between the historical time-series data input to the initial encoder model and the reconstructed time-series data output by the initial decoder model. Update the model parameters of the initial encoder and decoder models based on the reconstruction error until the reconstruction error meets the convergence condition, thus obtaining the encoder and decoder models. For example, the model parameters can be self-attention layer parameters, such as weight matrices and bias terms. The process of adjusting the model parameters based on the prediction error is existing technology and is not limited here.

[0080] In this application, the initial encoder model includes: an embedding layer for encoding time-series data of distribution network operating parameters to convert the time-series data of distribution network operating parameters into corresponding feature vector representations; an attention layer for performing QKV transformation on the corresponding feature vectors to obtain a query matrix, a key matrix, and a value matrix, setting each value in the value matrix to 1, performing attention calculation on the corresponding feature vectors using the query matrix and key matrix to obtain the attention matrix of the corresponding feature vectors, and performing a weighted summation of the value matrix using the attention matrix to obtain the attention output matrix of the corresponding feature vectors; and at least one fully connected layer for dimensionality reduction processing of the attention output matrix to obtain the dimensionality-reduced representation vector of the corresponding feature vectors. Specifically, as shown... Figure 4As shown, in this application, the encoder model, namely the Encorder encoder, includes an embedding layer, an attention layer, a first fully connected layer Fc1, and a second fully connected layer Fc2. In this application, the DTU data acquisition cycle can be set to 1 time / second, and the data acquired every minute is encoded and decoded once. The DTU Encorder first adds two bits of data and data type checksum to the acquired 60 data points to complete the encoding to 64 data points, and then performs embedding encoding to convert the input data into the corresponding feature vector representation to obtain the initial input embedding matrix. Then, the processed data is processed through an N*M*K multi-head attention matrix, such as a 64*128*8 multi-head attention matrix, to learn data features, and the learned features are merged together to obtain the weighted data features of the input sequence, i.e., the attention output matrix. Here, 64 represents the length of the input sequence, 128 represents the output feature dimension of the attention head, and 8 represents the number of attention heads. It is understood that the process of attention calculation through Q, V, K matrices is existing technology, and this application does not limit it. Then, the obtained weighted data features are compressed for dimensionality reduction through an Fc1 fully connected matrix, and then compressed for dimensionality reduction again through an Fc2 fully connected matrix to obtain the transmitted data code, i.e., the dimensionality-reduced data.

[0081] like Figure 5 As shown, this application further improves the attention mechanism. Traditional attention learns data features using three matrices: Q, K, and V. Q (query) retrieves the corresponding feature value K (key) and provides the value V (value) of K. In distribution network systems, every feature value is crucial for DTU-transmitted data. Therefore, this application sets the V matrix to 1, treating all learned feature values ​​as equally important and compressing them for transmission. This reduces computation by one-third, improving data compression efficiency.

[0082] Understandably, the decoder model in this application, namely the DTU Decorder decoder, adopts a similar structural connection to the encoder, amplifying the feature data layer by layer to obtain the true data. Specifically, the decoder also includes a fully connected layer Fc2, a fully connected layer Fc1, an attention layer, and an embedding layer. The fully connected layer Fc2 of the decoder receives the dimensionality-reduced data Code transmitted by the encoder, performs dimensionality-upgrading on the Code in the corresponding dimension, and then performs a second dimensionality-upgrading on the Code in the corresponding dimension through the fully connected layer Fc1, thereby restoring the dimension of the feature data. Finally, the attention layer, with a multi-head attention matrix of the same size as the encoder, restores the dimensionality-upgraded feature data to obtain the feature vector representation of the input sequence. Finally, the embedding layer restores the feature vector to obtain the input sequence. Figure 6 As shown, this application achieves optimal performance for both the Encorder and Decorator models through joint training. This is achieved by forward training the data from encoding to decoding, followed by backward training using the data decoder to encode the original data. The encoder and decoder in this application can also be used to construct fault prediction models, such as... Figure 7 As shown, the decoder can be trained for fault prediction. For example, using the time series data of distribution network parameters corresponding to different faults as training samples, the encoder performs dimensionality reduction on the time series data of distribution network parameters and transmits it to the decoder of the master station. The decoder performs dimensionality increase on the dimensionality-reduced data through a fully connected layer and performs fault prediction on the distribution network through the dimensionality-increased feature data. In this way, the data transmitted by the encoder can be decoded into a decoding model indicating whether there is a fault after passing through a decoder layer.

[0083] In step S200, the historical time-series data of different distribution network operating parameters are preprocessed, including:

[0084] S210. Determine the missing values ​​in the current historical time series data. For example, in the received running parameter time series data, if the data is empty or abnormal at the corresponding time point, it can be regarded as a missing value.

[0085] S220. Determine the first trend of change for multiple consecutive data values ​​preceding the current missing value, and determine the second trend of change for multiple consecutive data values ​​following the current missing value. The trend can be determined based on the rate or magnitude of change of multiple consecutive data values. For example, for the N consecutive data values ​​preceding the current missing value, calculate the difference between each pair of adjacent data values. If the number of positive differences is greater than a threshold (e.g., greater than 55%), the trend of change for the N consecutive data values ​​is determined to be an upward trend. If the number of negative differences is greater than a threshold (e.g., greater than 55%), the trend of change for the N consecutive data values ​​is determined to be a downward trend. If the ratio of positive to negative differences is between 45% and 55%, the trend of change for the N consecutive data values ​​is considered to be a stationary trend. Alternatively, the trend can be determined by calculating the slope of the N consecutive data values ​​using linear regression.

[0086] S230. Determine the predicted value of the current missing value based on the first trend and the second trend, and assign the predicted value to the current missing value.

[0087] In this embodiment, if the first trend and the second trend are the same, the average of the first changes of multiple consecutive data values ​​before the current missing value and the average of the second changes of multiple consecutive data values ​​after the current missing value are determined. For example, if the trend of the N consecutive data values ​​before the current missing value is upward, and the trend of the N consecutive data values ​​after the current missing value is also upward, then the two trends are determined to be the same; if the trend of the N consecutive data values ​​before the current missing value is upward, and the trend of the N consecutive data values ​​after the current missing value is downward, then the two trends are determined to be different. In this application, N can be set to 2 or more. A positive prediction value for the current missing value is constructed by the sum of the first valid data value and the average of the first changes, and a negative prediction value for the current missing value is constructed by the difference between the second valid data value and the average of the second changes. The first valid data value is the previous valid data value of the current missing value, and the second valid data value is the next valid data value of the current missing value. The positive prediction value can be represented as (y 前 +△y 前 The reverse prediction value can be expressed as (y 后 -△y 后 ), where y 前 Δy represents the previous valid data value before the current missing value. 前 y represents the mean of the first change. 后 Δy represents the previous valid data value before the current missing value. 后This represents the mean of the second change. Based on the mean of the first and second changes, a first weight for the positive prediction and a second weight for the negative prediction are constructed. The first weight is positively correlated with the mean of the first change, and the second weight is positively correlated with the mean of the second change. Specifically, the first and second weights can be constructed using the softmax function, where the first weight can be represented as w. 前 =e △y前 / (e △y前 +e △y后 The second weight can be represented as w. 后 =e △y后 / (e △y前 +e △y后 ), where w 前 w represents the first weight. 后 Let represent the second weight, and e be the natural constant. The predicted value for the current missing value is obtained by weighting the positive and negative predicted values ​​with the first and second weights and summing the results. Specifically, the predicted value for the current missing value can be represented as y. 缺失 =w 前 (y 前 +△y 前 )+w 后 (y 后- △y 后 ).

[0088] In this embodiment, if the first trend and the second trend are different, the ranking value of the current missing value in the current missing segment and the number of missing values ​​in the current missing segment are determined. For example, if there are M missing values ​​in the current missing segment, the number of missing values ​​is M. The prediction weight of the current missing value is determined based on its ranking value and the number of missing values ​​in the current missing segment. The prediction weight is positively correlated with the ranking value of the current missing value in the current missing segment. Specifically, the prediction weight can be expressed as n / (M+α), where n represents the ranking value of the current missing value in the current missing segment. For example, if the current missing segment includes 3 missing values, and the current missing value is ranked 1 according to the temporal order of the time series data, then n is 1, and so on; M represents the number of missing values, and α is an adjustment factor, which can be set empirically, for example, it can be set to 1. Determine the change between the next valid data value and the previous valid data value for the current missing value. Weight this change using prediction weights to obtain the predicted adjustment value for the current missing value. Specifically, the predicted adjustment value can be expressed as (n / (M+α))(y 后 -y 前 The predicted value for the current missing value is obtained by summing the predicted adjusted value and the previous valid data value, which can be represented as y. 缺失 =(n / (M+α))(y 后 -y前 )+y 前 Understandably, once the current missing value is determined, it can be considered a valid data value.

[0089] In this way, this application dynamically selects the interpolation method by judging the changing trend of valid data before and after the missing value. When the changing trend of valid data before and after is the same, the missing value is predicted through forward prediction and backward prediction. At the same time, the weights of forward prediction and backward prediction are dynamically constructed according to the changing trend of valid data before and after. Based on the dynamic weights, the forward and backward predicted values ​​are weighted, so that the predicted value incorporates the influence of the changing trend of data before and after, and is closer to the true value. When the changing trend of valid data before and after is different, the missing value is linearly interpolated based on the valid data before and after, which can more effectively calculate the missing value in this case. Through the dynamic interpolation method of this application, the accuracy of training data can be effectively guaranteed, thereby improving the training effect of encoder model and decoder model.

[0090] like Figure 8 As shown, in a second aspect of this application, a deep learning data transmission method based on a master station is provided, including the aforementioned deep learning data transmission method based on a side terminal. The deep learning data transmission method based on a master station includes: receiving a dimensionality reduction representation vector sent by an encoder model of at least one side terminal, wherein the dimensionality reduction representation vector is obtained by the encoder model after performing dimensionality reduction processing on time-series data of at least one distribution network operating parameter; using the dimensionality reduction representation vector as input, decoding the dimensionality reduction representation vector through a pre-constructed decoder model to obtain time-series data of at least one distribution network operating parameter; and determining the current operating state of the distribution network based on the time-series data of at least one distribution network operating parameter.

[0091] like Figure 9As shown, the distribution network data processing system of this application includes multiple distributed DTUs and a master station. This application embeds a trained and compressed optimized encoding model (DTU Encoder) into the DTU terminal. After model deployment, the DTU can directly complete feature construction and standardization processing locally after collecting operational data, and input the data into the encoder model to generate a low-dimensional compressed vector. Only the corresponding decoder model can recognize the compressed vector, and the data is encrypted during transmission, thus ensuring privacy protection for data transmission. The master station deploys a decoding model (DTU Decoder) to restore the compressed low-dimensional vector to the original data structure. The master station can improve decoding efficiency by integrating heterogeneous acceleration resources such as GPUs / TPUs. Simultaneously, the decoded data is accessed through a unified data analysis platform for subsequent business processing. After data decoding and aggregation are completed at the master station, the decoded structured data can enter the data fusion and intelligent analysis stage. The master station can identify the operating status of the distribution network based on the received distribution network operating parameters. For example, it can support the following business scenarios:

[0092] Distribution network anomaly identification: By analyzing compressed and restored real-time operating data, it identifies electrical anomalies such as short circuits, grounding, and voltage fluctuations, and supports alarm and linkage processing;

[0093] Equipment fault diagnosis: Combining historical equipment operation data and current compressed data, machine learning algorithms are used to determine the health status of equipment such as distribution transformers, switches, and DTUs;

[0094] Load forecasting analysis: Short-term / medium-to-long-term load forecasting based on time series reconstruction data to assist in load balancing scheduling and operation optimization;

[0095] Energy consumption analysis and trend assessment: Supports modeling of user electricity consumption behavior in power distribution areas, enabling energy-saving potential exploration and regional energy efficiency assessment;

[0096] Edge quality monitoring: The restored DTU operation data is integrated with GIS, power quality monitoring and other systems to achieve intelligent profiling of the operation quality of edge devices.

[0097] In summary, this application applies the encoding-decoding mechanism from deep learning to the compression and restoration process of distribution data, implementing edge intelligent encoding at the DTU end and completing decoding and analysis at the master station. This constructs a collaborative processing architecture for on-site compression and centralized intelligent decoding of distribution network data. By introducing a deep learning compression mechanism, it can automatically adapt to different data characteristics, improving encoding efficiency. Simultaneously, this application adopts a "edge compression + central decoding" data transmission and processing structure. Focusing on the efficient communication needs between the distribution automation station terminal (DTU) and the master station, it significantly reduces communication load and master station computational pressure while ensuring data accuracy. Through model training, lightweight deployment, and business integration, it enhances the efficiency, security, and intelligence level of the distribution network data transmission system.

[0098] like Figure 10 As shown, in a third aspect of this application, a deep learning data transmission apparatus based on an edge terminal is provided. Applying the aforementioned deep learning data transmission method based on an edge terminal, the apparatus includes:

[0099] The data acquisition module is configured to acquire time-series data of at least one power distribution network operating parameter;

[0100] The data encoding module is configured to take time-series data of at least one distribution network operating parameter as input, and perform dimensionality reduction processing on the time-series data of at least one distribution network operating parameter through a pre-built encoder model to obtain the corresponding dimensionality reduction representation vector.

[0101] The data transmission module is configured to send the dimensionality reduction representation vector to the master station, so that the master station can decode the dimensionality reduction representation vector through a pre-built decoder model to obtain time-series data of at least one distribution network operation parameter;

[0102] The encoder and decoder models were obtained by training a deep learning neural network with historical time-series data of different power distribution network operating parameters.

[0103] like Figure 11 As shown, in a fourth aspect of this application, a deep learning data transmission device based on a master station is provided. Applying the aforementioned deep learning data transmission method based on a master station, the device includes:

[0104] The data receiving module is configured to receive a dimensionality reduction representation vector sent by the encoder model of at least one side terminal. The dimensionality reduction representation vector is obtained by the encoder model after performing dimensionality reduction processing on the time series data of at least one distribution network operation parameter.

[0105] The data decoding module is configured to take the dimensionality reduction representation vector as input and decode it through a pre-built decoder model to obtain time-series data of at least one distribution network operation parameter.

[0106] The status identification module is configured to determine the current operating status of the distribution network based on time-series data of at least one distribution network operating parameter.

[0107] It is understood that those skilled in the art will clearly recognize that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0108] In a fifth aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform either the aforementioned deep learning data transmission method based on a side-end terminal or the aforementioned deep learning data transmission method based on a master station.

[0109] In a sixth aspect, this application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned deep learning data transmission method based on a side terminal, or implements the steps of the aforementioned deep learning data transmission method based on a main station.

[0110] like Figure 12 The diagram shown is a schematic representation of a terminal device provided in an embodiment of this application. Figure 12 As shown, the terminal device 10 of this embodiment includes a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements the steps in the above method embodiments. Alternatively, when the processor 100 executes the computer program 102, it implements the functions of each module / unit in the above device embodiments.

[0111] For example, computer program 102 may be divided into one or more modules / units, one or more of which are stored in memory 101 and executed by processor 100 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 102 in terminal device 10.

[0112] Terminal device 10 may be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. Terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 12 This is merely an example of terminal device 10 and does not constitute a limitation on terminal device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, terminal device may also include input / output devices, network access devices, buses, etc.

[0113] Processor 100 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0114] The memory 101 can be an internal storage unit of the terminal device 10, such as a hard disk or RAM of the terminal device 10. The memory 101 can also be an external storage device of the terminal device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the terminal device 10. Furthermore, the memory 101 can include both internal and external storage units of the terminal device 10. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 can also be used to temporarily store data that has been output or will be output.

[0115] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0117] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A deep learning data transmission method based on a side-end terminal, characterized in that, include: Collect time-series data of at least one power distribution network operating parameter; Using the time-series data of at least one distribution network operating parameter as input, the time-series data of the at least one distribution network operating parameter is dimensionality-reduced through a pre-built encoder model to obtain the corresponding dimensionality-reduced representation vector. The dimensionality reduction representation vector is sent to the master station so that the master station can decode the dimensionality reduction representation vector through a pre-built decoder model to obtain the time series data of the at least one distribution network operation parameter; The encoder model and the decoder model are obtained by preprocessing historical time-series data of different power distribution network operating parameters and then training a deep learning neural network. Preprocessing of historical time-series data of different power distribution network operating parameters includes: Identify missing values ​​in the current historical time series data; Determine the first trend of change of multiple consecutive data values ​​before the current missing value, and determine the second trend of change of multiple consecutive data values ​​after the current missing value; Based on the first trend and the second trend, the predicted value of the current missing value is determined, and the predicted value is used to assign a value to the current missing value; Determining the predicted value of the current missing value based on the first trend and the second trend includes: If the first trend of change is the same as the second trend of change, determine the average first change of multiple consecutive data values ​​before the current missing value and the average second change of multiple consecutive data values ​​after the current missing value; A positive prediction value for the current missing value is constructed by summing the first valid data value and the mean of the first change, and a negative prediction value for the current missing value is constructed by summing the second valid data value and the mean of the second change, wherein the first valid data value is the previous valid data value of the current missing value, and the second valid data value is the next valid data value of the current missing value; A first weight for the positive prediction value and a second weight for the negative prediction value are constructed based on the first mean change and the second mean change, respectively. The first weight is positively correlated with the first mean change, and the second weight is positively correlated with the second mean change. The first weight is w. 前 =e △y前 / (e △y前 +e △y后 The second weight is w. 后 =e △y后 / (e △y前 +e △y后 ), where w 前 w represents the first weight. 后 This represents the second weight, where e is the natural constant and y is... 前 Indicates the first valid data value, △y 前 y represents the mean of the first change. 后 Indicates the second valid data value, △y 后 This represents the mean of the second change. The positive and negative predicted values ​​are weighted and summed using the first and second weights to obtain the predicted value of the current missing value. Determining the predicted value of the current missing value based on the first trend and the second trend includes: If the first trend of change is different from the second trend of change, determine the ranking value of the current missing value in the current missing segment and the number of missing values ​​in the current missing segment; The prediction weight of the current missing value is determined based on the ranking value of the current missing value in the current missing segment and the number of missing values ​​in the current missing segment. The prediction weight is positively correlated with the ranking value of the current missing value in the current missing segment. Determine the change between the next valid data value and the previous valid data value of the current missing value. Weight the change value with the prediction weight to obtain the predicted adjustment value of the current missing value. Sum the predicted adjustment value and the previous valid data value of the previous valid data value to obtain the predicted value of the current missing value.

2. The deep learning data transmission method based on a side-end terminal according to claim 1, characterized in that, The training steps for the encoder model and the decoder model include: Construct an initial encoder model and an initial decoder model based on a deep learning neural network; Acquire historical time-series data of different power distribution network operating parameters, and preprocess the historical time-series data of different power distribution network operating parameters; The initial encoder model is input with historical time-series data of different power distribution network operating parameters, and the corresponding dimensionality-reduced representation vector is output by the initial encoder model. The initial decoder model receives the dimensionality reduction representation vector output by the initial encoder model, decodes the received dimensionality reduction representation vector, and obtains the reconstructed time series data. Determine the reconstruction error between the historical time-series data input to the initial encoder model and the reconstructed time-series data output by the initial decoder model. Update the model parameters of the initial encoder model and the initial decoder model based on the reconstruction error until the reconstruction error meets the convergence condition, thereby obtaining the encoder model and the decoder model.

3. The deep learning data transmission method based on a side-end terminal according to claim 2, characterized in that, The initial encoder model includes: The embedding layer is used to encode the time-series data of the distribution network operating parameters, so as to convert the time-series data of the distribution network operating parameters into the corresponding feature vector representation; The attention layer is used to perform QKV transformation on the corresponding feature vector to obtain a query matrix, a key matrix, and a value matrix. Each value in the value matrix is ​​set to 1. Attention is calculated on the corresponding feature vector through the query matrix and the key matrix to obtain the attention matrix of the corresponding feature vector. The attention output matrix of the corresponding feature vector is obtained by weighted summation of the value matrix through the attention matrix. At least one fully connected layer is used to perform dimensionality reduction processing on the attention output matrix to obtain the dimensionality reduction representation vector of the corresponding feature vector.

4. A deep learning data transmission method based on a master station, characterized in that, The deep learning data transmission method based on a side-side terminal, as described in any one of claims 1-3, and the deep learning data transmission method based on a main station, include: The encoder model receives a dimension-reduced representation vector from at least one side-side terminal. The dimension-reduced representation vector is obtained by the encoder model after performing dimension reduction processing on the time-series data of at least one distribution network operation parameter. Using the reduced-dimensional representation vector as input, the reduced-dimensional representation vector is decoded by a pre-built decoder model to obtain the time-series data of the at least one distribution network operation parameter; The current operating status of the distribution network is determined based on the time-series data of at least one of the distribution network operating parameters.

5. A deep learning data transmission device based on a side-side terminal, employing the deep learning data transmission method based on a side-side terminal as described in any one of claims 1-3, characterized in that, The device includes: The data acquisition module is configured to acquire time-series data of at least one power distribution network operating parameter; The data encoding module is configured to take the time-series data of the at least one distribution network operating parameter as input, and perform dimensionality reduction processing on the time-series data of the at least one distribution network operating parameter through a pre-built encoder model to obtain the corresponding dimensionality reduction representation vector. The data transmission module is configured to send the dimensionality reduction representation vector to the master station, so that the master station can decode the dimensionality reduction representation vector through a pre-built decoder model to obtain the time-series data of the at least one distribution network operation parameter; The encoder model and the decoder model are obtained by training a deep learning neural network with historical time-series data of different power distribution network operating parameters.

6. A deep learning data transmission device based on a master station, employing the deep learning data transmission method based on a master station as described in claim 4, characterized in that, The device includes: The data receiving module is configured to receive a dimensionality reduction representation vector sent by the encoder model of at least one side terminal. The dimensionality reduction representation vector is obtained by the encoder model after performing dimensionality reduction processing on the time-series data of at least one distribution network operation parameter. The data decoding module is configured to take the dimensionality reduction representation vector as input and decode the dimensionality reduction representation vector through a pre-built decoder model to obtain the time series data of the at least one distribution network operation parameter; The status identification module is configured to determine the current operating status of the distribution network based on the time-series data of the at least one distribution network operating parameter.

7. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by the processor, this instruction causes the processor to be configured to perform the deep learning data transmission method based on a side terminal as described in any one of claims 1-3, or to perform the deep learning data transmission method based on a master station as described in claim 4.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning data transmission method based on a side terminal as described in any one of claims 1-3, or the steps of the deep learning data transmission method based on a master station as described in claim 4.

Images