Method and device for realizing edge side reconstruction of electric power integrated energy data

Abstract

The invention relates to the technical field of electric power integrated energy data, solves the problems of great load pressure and high time delay brought by network bandwidth and distribution network data to a middle station, and discloses a method and a device for realizing edge side reconstruction of electric power integrated energy data. The method comprises the following steps: extracting feature information data among original data sources; calling a residual error connection convolutional neural network model without a BN layer to carry out nonlinear mapping on the characteristic value; synthesizing and screening the feature information data; and sending the screening result and the feature information data output by the last residual block to a sub-pixel convolutional neural network for data reconstruction. According to the application, a global feature multiplexing module is established on the basis of parameters of an activation function ReLU of a traditional residual convolutional neural network, and participation of a BN layer in each residual connection module is reduced; therefore, the pressure of data processing and storage is effectively reduced.

Description

Technical Field

[0001] This application relates to the field of integrated power energy data processing technology, and in particular to a method and apparatus for edge-side reconstruction of integrated power energy data. Background Technology

[0002] With the large-scale integration of renewable energy sources, primarily wind, solar, and energy storage, into the distribution network, the network architecture has undergone profound changes. It has shifted from a closed-loop design and open-loop operation with a single power source sector structure to a system supporting multiple power sources and loads. Therefore, a more flexible and powerful source-grid-load-storage collaborative processing architecture is needed on the distribution network side. On the user side, regional integrated energy systems, with integrated energy stations as the main body, are gradually becoming a major form of local renewable energy consumption and regulation, serving as a microgrid architecture for optimizing the coordinated operation of multiple energy sources within the periphery of the large power grid.

[0003] However, a prominent problem for regional integrated energy stations is that as these stations operate on the distribution network side, they generate massive amounts of multi-source, heterogeneous energy data, causing the scale of time-series data to increase exponentially. The current centralized data processing method based on the distribution network platform involves directly reporting massive amounts of data to the platform via the network for unified classification and processing. This places a huge load on network bandwidth and the distribution network data platform, and also causes significant time delays, which is unacceptable for time-sensitive energy dispatching. This is gradually becoming one of the core issues hindering the development of new power systems.

[0004] To address this issue, there is an urgent need to propose a deep data reconstruction technology for massive energy data at the distribution network edge, along with supporting edge computing terminals, to reduce the processing and storage pressure on the middleware platform and minimize data latency. Since the large amount of time-series data collected by energy stations can describe the real-time operating status of the entire distribution network's power generation, grid, load, and storage system, effectively reconstructing this heterogeneous data at the energy station edge can significantly improve the collaborative optimization application of the distribution network's power generation, grid, load, and storage capabilities and its responsiveness on the user side. This includes predicting user-side demand, forecasting energy storage processing and future energy trading trends, monitoring residential power quality, and monitoring energy prices, thereby effectively enhancing the collaborative optimization capabilities of the distribution network's power generation, grid, load, and storage systems. Summary of the Invention

[0005] The purpose of this application is to overcome the problems of the centralized data processing method based on the distribution network platform in the prior art, which directly reports massive amounts of data to the platform via the network, resulting in a huge load pressure on the platform due to network bandwidth and distribution network data, as well as high time delay. The application provides a method and device for realizing edge-side reconstruction of integrated power energy data.

[0006] Firstly, a method for edge-side reconstruction of integrated power energy data is provided, including:

[0007] Obtain the raw data sources of energy and use convolutional neural networks to extract feature information data between the various raw data sources;

[0008] The residual connection convolutional neural network model with BN layers removed is invoked to perform nonlinear mapping on the feature values. Here, BN stands for Batch Normalization. Generally, BN layers are directly connected after convolution operations. BN layers can standardize the original data to accelerate the convergence speed of the network.

[0009] A global feature reuse module is added to the residual connection convolutional neural network model with the BN layer removed in order to synthesize and filter feature information data;

[0010] The filtering results and the feature information data output by the last residual block are fed into a subpixel convolutional network for data reconstruction.

[0011] Furthermore, the convolutional neural network includes convolutional layers and activation layers.

[0012] Furthermore, extracting feature values ​​from various data sources using convolutional neural networks includes: performing a convolution operation between the input original data source Y and n1 convolutional kernels to extract n1 feature values, where the feature extraction formula is:

[0013] B0 = σ(W1*Y + b1)

[0014] Where * represents the convolution operation; W1 is the convolution kernel of the first convolutional layer; b1 is the bias of the first convolutional layer, and the dimension of the bias term should be consistent with the dimension of the convolution kernel; B0 is the input feature of the first residual block extracted from the input Y; σ is the activation function, which uses the modified linear unit ReLU, and the formula for calculating the activation function σ is:

[0015]

[0016] Where x is the input of the ReLU activation function and σ(x) is the output of the ReLU activation function.

[0017] Furthermore, the computational formula for the residual connection convolutional neural network model with the BN layer removed is as follows:

[0018] H m =G m (H m-1 )=F(H m-1 W m )+H m-1

[0019] Among them, H m-1 and H m are the input and output of the m-th residual unit, respectively; F is the learned residual mapping, i.e.:

[0020]

[0021] in, Let σ be the weight of the i-th learned convolutional layer, σ be the activation function, and * be the convolution operation.

[0022] Furthermore, the synthesis and filtering of feature information data includes:

[0023] The feature information data obtained through feature extraction and the feature information data output by the first three residual modules are both extracted;

[0024] The extracted feature information data is then combined into a larger amount of feature information data;

[0025] Filter a large amount of feature information data using convolution kernels;

[0026] The filtering results, along with the feature information data output from the last residual block, are fed into a subpixel convolutional network for data reconstruction.

[0027] Furthermore, the last residual block is the last residual block in the residual connection convolutional neural network model with the BN layer removed.

[0028] Furthermore, subpixel convolutional networks perform data reconstruction including:

[0029] Based on the selection results and the feature information data output by the last residual block, a convolutional network of size H×W is obtained and convolution is performed.

[0030] After convolution, we get r. 2 The feature information matrix of ×H×W;

[0031] The channels corresponding to each element of the feature information matrix are rearranged into an r×r matrix, corresponding to an r×r sub-block in a nearby data source, so as to optimize the size of r... 2 The feature information matrix ×H×W is rearranged into a feature information matrix 1×rH×rW.

[0032] Secondly, a device for edge-side reconstruction of integrated power energy data is provided, comprising:

[0033] Storage unit for storing data processing programs that can run on the data processing unit;

[0034] The data access unit is a circuit module with multiple data protocols and interface forms;

[0035] A data processing unit is configured to execute the data processing program, wherein the data processing program, when executed by the data processing unit, implements the method as described in any of the implementations of the first aspect.

[0036] The communication unit is used to report the reconstructed data processed by the data processing unit to the background computer in real time.

[0037] The data access unit's output is connected to the data processing unit's input, the data processing unit's output is connected to the communication unit, and the storage unit is connected to the data processing unit.

[0038] Furthermore, it also includes a power supply unit, and the data access unit, the data processing unit, and the communication unit are all electrically connected to the power supply unit.

[0039] Furthermore, the data access unit includes a wireless data access module, a serial data access module, an Ethernet data access module, an HPLC data access module, a data aggregation module, and a data output module. The output terminals of the wireless data access module, the serial data access module, the Ethernet data access module, and the HPLC data access module are all connected to the input terminal of the data aggregation module, and the output terminal of the data aggregation module is connected to the data output module.

[0040] This application has the following beneficial effects: It establishes a global feature reuse module based on the parameters of the ReLU activation function of the traditional residual convolutional neural network. At the same time, in order to effectively deploy the neural network under the condition of limited resources on the edge side, it reduces the participation of the BN layer in each residual connection module, so that the local residual can increase the amount of network information and depth while promoting the information flow in the network. This allows the network to obtain more data reconstruction information during training and learning, thereby effectively reducing the pressure of data processing and storage. It solves the problem of reconstructing a large amount of multi-source heterogeneous energy data in regional integrated energy stations, reduces the load pressure on the distribution network platform, and provides effective technical support for the implementation and construction of digital traction new power systems on the distribution network side. Attached Figure Description

[0041] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0042] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the overall process of implementing the edge-side reconstruction method for integrated power energy data according to Embodiment 1 of this application.

[0044] Figure 2 This is a flowchart illustrating the specific process of implementing the edge-side reconstruction method for integrated power energy data according to Embodiment 1 of this application.

[0045] Figure 3 This is a schematic diagram of the reconstruction model in the edge-side reconstruction method for integrated power energy data according to Embodiment 1 of this application;

[0046] Figure 4 This is a schematic diagram of the ReLU excitation function in the edge-side reconstruction method for integrated power energy data according to Embodiment 1 of this application;

[0047] Figure 5 This is a schematic diagram of the p-ReLU excitation function with parameters in the method for reconstructing integrated power energy data at the edge, as described in Embodiment 1 of this application.

[0048] Figure 6 This is a schematic diagram of the nonlinear mapping structure of the ResNets network in the implementation of the edge-side reconstruction method for integrated power energy data in Embodiment 1 of this application;

[0049] Figure 7 This is a schematic diagram of the residual connection convolutional neural network model with BN layers removed in the implementation of the edge-side reconstruction method for integrated power energy data in Embodiment 1 of this application;

[0050] Figure 8 This is a schematic diagram of the residual connection convolutional network structure of the edge-side reconstruction method for integrated power energy data in Embodiment 1 of this application, which embeds a global feature reuse module.

[0051] Figure 9 This is a schematic diagram of mid-subpixel convolutional reconstruction of the edge-side reconstruction method for integrated power energy data according to Embodiment 1 of this application;

[0052] Figure 10 This is a structural block diagram of the device for reconstructing integrated power energy data at the edge, according to Embodiment 2 of this application.

[0053] Figure 11 This is a structural block diagram of the data access unit in the device for reconstructing integrated power energy data at the edge, according to Embodiment 2 of this application.

[0054] Figure label:

[0055] 100. Storage unit; 200. Data access unit; 201. Wireless data access module; 202. Serial data access module; 203. Ethernet data access module; 204. HPLC data access module; 205. Data aggregation module; 206. Data output module; 300. Data processing unit; 400. Communication unit; 500. Power supply unit. Detailed Implementation

[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] Example 1

[0058] This application's embodiment 1 discloses a method for edge-side reconstruction of integrated power energy data, comprising: acquiring the original energy data source and extracting feature information data between each original data source using a convolutional neural network; calling a residual connected convolutional neural network model with BN layers removed to perform nonlinear mapping on the feature values; adding a global feature reuse module to the residual connected convolutional neural network model with BN layers removed to synthesize and filter the feature information data; and sending the filtering results and the feature information data output by the last residual block into a sub-pixel convolutional network for data reconstruction. This method establishes a global feature reuse module based on the parameters of the ReLU activation function of the traditional residual convolutional neural network. Simultaneously, to effectively deploy the neural network under resource-constrained conditions at the edge, the participation of the BN layer in each residual connected module is reduced, allowing local residuals to promote information flow in the network while increasing the amount and depth of network information. This enables the network to obtain more data reconstruction information during training and learning, thereby effectively reducing the pressure on data processing and storage.

[0059] For integrated power energy stations, the data types they access include various formats. For example, power sources can be packaged in multiple ways, such as photovoltaics, energy storage, and distribution networks. The load side has various energy-consuming devices, including electrical equipment within the park, as well as heat, water, cooling, and gas systems. The data structures differ significantly between different energy users and suppliers. Using a single standard for measurement leads to redundancy and errors, placing immense pressure on data storage and processing. Therefore, for processing and reconstructing such multi-source, heterogeneous energy data with diverse structures, directly using traditional SRCNN convolutional neural networks for analysis and reconstruction presents challenges. Since SRCNN is a multi-layered perceptual mechanism, more layers mean more features. However, the SRCNN algorithm only uses three convolutional layers, resulting in structural ambiguity between adjacent data and excessively high correlation function eigenvalues ​​in the reconstructed heterogeneous data. Although increasing the network depth can improve performance, the limitations of traditional methods cannot solve these problems. To obtain better reconstructed data feature values, but considering the current limited resources at the edge, the number of network layers cannot be increased indefinitely, otherwise gradient problems will occur. Moreover, when the number of network layers exceeds a certain value, the network becomes more difficult to train, and the accuracy of the network drops rapidly, i.e., network degradation occurs. To address these issues, this application proposes an edge-side multi-source heterogeneous data reconstruction method based on ResNets residual connected convolutional neural networks. Considering that directly using ResNets is highly dependent on hardware resources, because adding a BN layer to each residual module of ResNets increases information loss and data variability when sampling and reconstructing multi-source heterogeneous data, and the ReLU activation function inside the residual module has the problem of boundary convergence causing local neurons to fail to activate, this application proposes an improved reconstruction model for multi-source heterogeneous energy data using residual connected convolutional neural networks. The reconstruction model is as follows: Figure 3 As shown.

[0060] Specifically, Figure 1-2 The flowchart of the method for reconstructing integrated power energy data at the edge, as shown in Embodiment 1 of the application, includes:

[0061] S101. Obtain the original data source of energy and use a convolutional neural network to extract the feature information data between the various original data sources;

[0062] For multi-source heterogeneous raw energy data collected from multiple sensors, to obtain feature values ​​from each data source, compared to manually selecting features in traditional algorithms, using convolutional neural networks (CNNs) is more effective in obtaining reconstructed features during data deduplication and redundancy removal. The feature extraction using a CNN consists of one convolutional layer and one activation layer. By convolving the input raw data source Y with n1 convolutional kernels, n1 feature values ​​can be obtained. The feature extraction formula is:

[0063] B0=σ(W1*Y+b1) (1)

[0064] In formula (1): * represents the convolution operation; W1 is the convolution kernel of the first convolutional layer; b1 is the bias of the first convolutional layer, and the dimension of the bias term should be consistent with the dimension of the convolution kernel; B0 is the input feature of the first residual block extracted from the input Y; σ is the activation function, which uses the Rectified Linear Unit (ReLU), and the calculation formula is as follows:

[0065] σ(x)=max(0,x) (2)

[0066] In formula (2), x is the input of the ReLU activation function; σ(x) is the output of the ReLU activation function. However, a serious problem arises when using ReLU: directly setting inputs less than zero to zero, such as... Figure 4 As shown, some neurons cannot be activated and cannot participate in subsequent network training.

[0067] Therefore, in formula (2), this application has modified the linear excitation function ReLU, as follows: Figure 5 As shown, this function is an improvement on the ReLU function, setting the first half of the σ input that is less than 0 as ax. i Therefore, formula (2) can be modified as follows:

[0068]

[0069] Unlike the residual unit in ResNets, the improved residual unit uses the parameterized activation function PReLU only once to effectively activate all neurons involved in the computation, thus significantly improving the accuracy of data computation and prediction.

[0070] S102. Call the residual connection convolutional neural network model with BN layers removed to perform nonlinear mapping on the feature values;

[0071] In SRCNN and ResNets neural network reconstruction tasks, when training deep convolutional neural networks, the detailed information of the data is lost after multiple convolutional processing layers due to excessive network depth, causing network performance degradation and preventing optimal training results. In ResNets, to preserve data source features at different levels, the nonlinear mapping process uses local residual connections, repeatedly using Batch Normalization (BN) layers within each residual block, such as... Figure 6 As shown.

[0072] Figure 6In ResNets, the nonlinear mapping process consists of multiple residual modules, each composed of multiple residual units, and each residual unit has a batch normalization (BN) layer. Since the BN layer effectively restricts the distribution range of the input values, placing them within the region of large derivatives of the activation function, it not only effectively alleviates gradient vanishing but also accelerates network convergence. However, for the deduplication and reconstruction of heterogeneous energy data, adding a BN layer to each residual unit is counterproductive. This is mainly because data reconstruction requires more detailed information and data variability, and the use of the BN layer not only destroys this information but also ignores the absolute differences in the data. Furthermore, adding the BN layer makes the network more complex, requiring more training time and higher hardware resources, making it unsuitable for resource-constrained edge environments. Therefore, this application addresses... Figure 6 Significant modifications have been made. Due to the temporal similarity of multi-source heterogeneous data, there are not many differences between structural feature values. Removing the BN layer does not reduce the gradient vanishing problem of the data source, while improving training efficiency and reducing dependence on hardware resources. The improved residual connection convolutional neural network model (i.e., the residual connection convolutional neural network model with BN layers removed) is as follows: Figure 7 As shown.

[0073] Figure 7 In this context, the residual connection convolutional neural network model with the Batch Normalization (BN) layer removed can be represented as:

[0074] H m =G m (H m-1 )=F(H m-1 W m )+H m-1 (4)

[0075] In formula (4): H m-1 and H m are the input and output of the m-th residual unit, respectively; F is the learned residual mapping, i.e.:

[0076]

[0077] In formula (5): σ represents the weights of the i-th convolutional layer learned. To simplify the formula, the bias term is omitted. σ is the PReLU of formula (3), i.e., the adjusted activation function. * represents the convolution operation.

[0078] S103. Add the global feature reuse module to the residual connection convolutional neural network model with the BN layer removed in order to synthesize and filter feature information data;

[0079] To fully utilize the feature information of heterogeneous energy data in both time and spatial dimensions, and to obtain reconstructed data with better deduplication performance at the edge under resource-constrained conditions, in addition to removing the Batch Normalization (BN) layer, this invention also proposes adding a global feature reuse module inside the residual network after removing the BN layer, such as... Figure 8 As shown, the data feature information obtained through feature extraction and the data information output by the first three residual modules are extracted and synthesized into a large amount of feature data information. This information is then filtered by a 1×1 convolution kernel. The filtered results and the feature information output by the last residual block are then fed into the sub-pixel convolutional network for data reconstruction.

[0080] S104. The filtering results and the feature information data output by the last residual block are fed into the sub-pixel convolutional network for data reconstruction.

[0081] Specifically, subpixel convolutional reconstruction is essentially a magnification process, generally used in the field of image reconstruction. This invention applies linolenic acid convolutional reconstruction to multi-source heterogeneous energy data, simulating a two-dimensional image in both time and space dimensions. The steps of linolenic acid convolutional reconstruction are:

[0082] (1) Input a two-dimensional spatial data source and a convolutional network of size ;

[0083] (2) The feature information matrix obtained after convolution is then rearranged into a matrix corresponding to a sub-block of a similar data source, so that the feature information matrix of size is rearranged into a feature information matrix.

[0084] It should be noted that subpixel convolution only modifies the 2D data in the last layer; the numerous preceding convolution operations are performed on the original data. Furthermore, this operation can automatically learn the interpolation functions implicit in the preceding convolutional layers. Therefore, subpixel convolution reconstruction reduces network complexity, shortens training time, and offers high efficiency. The subpixel convolution reconstruction structure is as follows: Figure 9 As shown.

[0085] Example 2

[0086] like Figure 10 As shown, the device for edge-side reconstruction of integrated power energy data according to Embodiment 2 of this application includes:

[0087] Storage unit 100 is used to store data processing programs that can run on the data processing unit 300;

[0088] For example, storage unit 100 includes 8GB DDR4 SDRAM and 32GB eMMC; storage unit 100 mainly stores the RT Linux operating system, data acquisition program, data processing program and intermediate data.

[0089] The data access unit 200 is a circuit module with multiple data protocols and interface forms;

[0090] like Figure 11 As shown, the data access unit 200 includes a wireless data access module 201, a serial data access module 202, an Ethernet data access module 203, an HPLC data access module 204, a data aggregation module 205, and a data output module 206. The output terminals of the wireless data access module 201, the serial data access module 202, the Ethernet data access module 203, and the HPLC data access module 204 are all connected to the input terminal of the data aggregation module 205, and the output terminal of the data aggregation module 205 is connected to the data output module 206.

[0091] For example, the wireless data access module 201 includes a 433M wireless sensor access module and a LoRa490M wireless sensor access module, supporting simultaneous access from over 200 sensors via 433M and 490M wireless channels; the access protocol supports various industrial protocols such as Modbus RTU, NetLink, ProfiNet, IEC101, and IEC104; the serial data access module 202 includes standard industrial serial interfaces such as RS485, RS422, RS232, USB, and CAN, with RS485 / 422 / 232 adaptive matching, automatically matching according to the access level and physical interface of the external device; the access protocol supports various industrial protocols such as Modbus RTU, NetLink, ProfiNet, IEC101, IEC104, and CANBUS; the Ethernet data access module 203 includes a 10 / 100 / 1000M adaptive Ethernet port, supporting Modbus Ethernet industrial protocols such as TCP, SNMP, IP / TCP, Profibus, Interbus, FTP, and HTTPS are supported. The HPLC data access module 204 mainly includes an HPLC acquisition module, supporting HPLC meter reading and adaptive access for various data formats (electricity meters, water meters, and gas meters, etc.). The data aggregation module 205 is an STM32F105 processor, which mainly collects reported data from various front-end data access modules, performs protocol parsing on the data, repackages it, and then reports it to the data processing unit 300 through the data output module 206. The data output module 206 is an H-SPI interface that connects to the data processing unit 300 at high speed with a bandwidth of 50M. The synchronization clock comes from the HSPI controller inside the data processing unit 300.

[0092] Data processing unit 300 is used to execute the data processing program, which, when executed by data processing unit 300, implements the method described in any of the implementations in Embodiment 1;

[0093] For example, the data processing unit 300 is a core processing module integrating a multi-core processor and a large-capacity memory. The multi-core processor unit consists of an integrated ARM quad-core Cortex-A72 and a NEOP neural network convolutional operation core. Data acquisition, task scheduling, data caching, and communication are completed by the Cortex-A72 quad-core, while the NEOP neural network core is mainly responsible for data processing, namely the data reconstruction model based on the improved residual connection convolutional neural network proposed in this invention, including computation, training, and feedback. The core also integrates a 2MB high-speed cache for direct instruction-level reading and storage of convolutional operation data. The cache communicates with external DDR4 memory through a dual-channel DMA burst mode. The NEOP core can process up to 2 TOPS of computation speed, which basically meets the requirements of the improved multi-source heterogeneous data reconstruction algorithm of the neural network model under resource-constrained conditions proposed in this invention.

[0094] The communication unit 400 is used to report the reconstructed data processed by the data processing unit 300 to the background computer in real time.

[0095] For example, the communication unit 400 is a communication module of a 4G / 5G wireless + wired fiber optic module. The communication unit 400 mainly reports the reconstructed data processed by the multi-core processor to the background computer in real time.

[0096] The output terminal of the data access unit 200 is connected to the input terminal of the data processing unit 300, the output terminal of the data processing unit 300 is connected to the communication unit 400, and the storage unit 100 is connected to the data processing unit 300.

[0097] In a further embodiment, a power supply unit 500 is also included, and the data access unit 200, the data processing unit 300 and the communication unit 400 are all electrically connected to the power supply unit 500.

[0098] For example, the power supply unit 500 is a series of AC-CDC and DC-CDC conversion circuits. The AC-CDC conversion circuit can directly convert 110V-380V AC power into 12V DC power, and the DC-CDC conversion circuit can convert 12V DC power into five sets of power supplies, namely 5V, 3.3V, 1.8V, 1.2V and 1V, for use by various modules inside the device.

[0099] The above are merely preferred embodiments of this application; however, the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and its improved concept, should be covered within the scope of protection of this application.

Claims

1. A method for edge-side reconstruction of integrated power energy data, characterized in that, include: Obtain the raw data sources of energy and use convolutional neural networks to extract feature information data between the various raw data sources; The residual connection convolutional neural network model with BN layers removed is invoked to perform a nonlinear mapping on the feature values; A global feature reuse module is added to the residual connection convolutional neural network model with the BN layer removed in order to synthesize and filter feature information data; The filtering results and the feature information data output by the last residual block are fed into a subpixel convolutional network for data reconstruction.

2. The method for edge-side reconstruction of integrated power energy data according to claim 1, characterized in that, The convolutional neural network includes convolutional layers and activation layers.

3. The method for edge-side reconstruction of integrated power energy data according to claim 2, characterized in that, Extracting feature values ​​from various data sources using a convolutional neural network involves performing a convolution operation between the input raw data source Y and n1 convolutional kernels, which extracts n1 feature values. The feature extraction formula is as follows: B0 = σ(W1*Y + b1) Where * represents the convolution operation; W1 is the convolution kernel of the first convolutional layer; b1 is the bias of the first convolutional layer, and the dimension of the bias term should be consistent with the dimension of the convolution kernel; B0 is the input feature of the first residual block extracted from the input Y; σ is the activation function, which uses the modified linear unit ReLU, and the formula for calculating the activation function σ is: Where x is the input of the ReLU activation function and σ(x) is the output of the ReLU activation function.

4. The method for edge-side reconstruction of integrated power energy data according to claim 3, characterized in that, The computational formula for the residual connection convolutional neural network model with the BN layer removed is as follows: H m =G m (H m-1 )=F(H m-1 ,W m )+H m-1 Among them, H m-1 and H m are the input and output of the m-th residual unit, respectively; F is the learned residual mapping, i.e.: in, Let σ be the weight of the i-th learned convolutional layer, σ be the activation function, and * be the convolution operation.

5. The method for edge-side reconstruction of integrated power energy data according to claim 1, characterized in that, The synthesis and filtering of feature information data includes: The feature information data obtained through feature extraction and the feature information data output by the first three residual modules are both extracted; The extracted feature information data is then combined into a larger amount of feature information data; Filter a large amount of feature information data using convolution kernels; The filtering results, along with the feature information data output from the last residual block, are fed into a subpixel convolutional network for data reconstruction.

6. The method for edge-side reconstruction of integrated power energy data according to claim 1 or 5, characterized in that, The last residual block is the last residual block in the residual connection convolutional neural network model with the BN layer removed.

7. The method for edge-side reconstruction of integrated power energy data according to claim 1, characterized in that, Subpixel convolutional networks perform data reconstruction including: Based on the selection results and the feature information data output by the last residual block, a convolutional network of size H×W is obtained and convolution is performed. After convolution, we get r. 2 The feature information matrix of ×H×W; The channels corresponding to each element of the feature information matrix are rearranged into an r×r matrix, corresponding to an r×r sub-block in a nearby data source, so as to optimize the size of r... 2 The feature information matrix ×H×W is rearranged into a feature information matrix 1×rH×rW.

8. A device for edge-side reconstruction of integrated power energy data, characterized in that, include: Storage unit for storing data processing programs that can run on the data processing unit; The data access unit is a circuit module with multiple data protocols and interface forms; A data processing unit is configured to execute the data processing program, wherein the data processing program, when executed by the data processing unit, implements the method as described in any one of claims 1-7; The communication unit is used to report the reconstructed data processed by the data processing unit to the background computer in real time. The data access unit's output is connected to the data processing unit's input, the data processing unit's output is connected to the communication unit, and the storage unit is connected to the data processing unit.

9. The device for edge-side reconstruction of integrated power energy data according to claim 8, characterized in that, It also includes a power supply unit, and the data access unit, the data processing unit and the communication unit are all electrically connected to the power supply unit.

10. The device for edge-side reconstruction of integrated power energy data according to claim 8, characterized in that, The data access unit includes a wireless data access module, a serial data access module, an Ethernet data access module, an HPLC data access module, a data aggregation module, and a data output module. The output terminals of the wireless data access module, the serial data access module, the Ethernet data access module, and the HPLC data access module are all connected to the input terminal of the data aggregation module, and the output terminal of the data aggregation module is connected to the data output module.

Images