Power data super-resolution sensing method based on improved diffusion model
By improving the diffusion model combined with U-Net and LSTM networks, the limitations of high-precision power data in the acquisition, transmission and storage process are solved, high-precision reconstruction of low-resolution data is realized, and the analysis capability and transmission of power systems are improved. Storage performance.
Patent Information
- Application Number
- CN202310753882.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-25
- Publication Date
- 2024-12-31
AI Technical Summary
There are limitations in the acquisition, transmission and storage of high-precision power data, which cannot be used in practice. In particular, the transmission and storage of high-frequency data have high requirements for system performance and throughput, and the traditional super-resolution perception method has limited detailed recovery capabilities. .
Using an improved diffusion model combined with U-Net and LSTM networks, the noise is gradually added and the features are learned through the diffusion process. The inverse diffusion process recovers high-resolution data from pure Gaussian noise, realizing super-resolution reconstruction of power data, and using low resolution Data is reconstructed with high precision.
It improves the system transmission storage performance, realizes higher precision power system modeling and analysis, overcomes the limitations of explicit modeling of a large amount of prior knowledge and features in traditional methods, and has good generalization ability and flexibility.
Smart Images
Figure BDA0004302281960000031 
Figure FDA0004302281940000021 
Figure HDA0004302281980000011
Abstract
Description
Technical Field
[0001] The invention relates to a power data super-resolution perception method based on an improved diffusion model, and belongs to the technical field of power big data processing. Background Art
[0002] In recent years, my country's renewable energy projects have made significant progress, and the competitiveness of renewable energy has been continuously enhanced, maintaining its leading position in the world. The high proportion of renewable energy connected to the grid will become a basic feature of the future power system. However, compared with traditional power sources, renewable energy has the characteristics of output uncertainty and volatility, which poses a great challenge to the stable operation of the power system. How to deal with the uncertainty brought about by the high proportion of renewable energy connected to the grid and meet the needs of the future power system for safety, economy, flexibility and stability is of great significance.
[0003] With the interactive development of the digital revolution and the energy revolution, big data has been widely used in power systems, mainly in prediction, evaluation, pattern recognition, model building and decision analysis. The diversity and volatility of the system due to the grid connection of renewable energy often require higher-precision power data for mining and analysis. Compared with low-precision power data, high-frequency data contains richer feature information, which can deeply explore the deep characteristics and change laws of renewable energy, and is of great significance for system state perception, management decision-making, performance evaluation, and fault diagnosis. For example: in the state monitoring system of power transmission and transformation equipment, in order to accurately diagnose the insulation discharge and other conditions, the sampling frequency of the signal needs to be above 200kHz, and the ultra-high frequency detection needs to reach GHz; in the prediction of wind power ramping, more attention is paid to the short-term and large changes in wind power caused by the sudden change of average wind speed in extreme meteorological events, and the use of low-frequency data will result in low prediction accuracy, so it is necessary to obtain the implicit features in high-frequency data to improve the prediction accuracy.
[0004] The data granularity collected by the Supervisory Control and Data Acquisition (SCADA) system is generally at the minute level. With the grid connection of renewable energy, the uncertainty and randomness of the system increase, and data of this time scale will not be able to meet many challenges. At present, phasor measurement units (PMUs) are installed in power plants and new energy grid connection stations, which can continuously transmit PMU data at a high frequency of 50Hz. However, PMU equipment is expensive and cannot replace most traditional meters. In addition, the long-term continuous transmission of high-frequency data by PMU will have a huge impact on the transmission network. The growing demand for data storage has made the storage pressure faced by the data storage center increasingly greater, and the transmission and writing requests of high-speed data have also put higher and higher requirements on the performance and throughput of the storage system. Due to various limitations of data acquisition equipment, transmission lines and storage capacity, high-frequency power data is difficult to play a practical role in the power grid. Therefore, this paper considers the use of super-resolution perception technology, using low-frequency data in the process of data acquisition, transmission and storage. When high-frequency data is needed, high-frequency details are reconstructed from low-frequency data to generate high-precision power data, which greatly improves the system transmission and storage performance while making full use of high-frequency data to achieve higher-precision power system modeling and analysis.
[0005] Super-resolution perception technology originally refers to the technology that can improve the resolution of optical imaging systems. Now it is usually applied to the field of images to refer to methods that can convert single or multiple low-resolution images into high-resolution images after processing. Super-resolution perception technology can be divided into interpolation-based and machine learning-based methods. The traditional interpolation-based method inserts new pixels according to a certain interpolation algorithm based on known pixels. It is simple and easy to implement, but its detail recovery ability is limited and the recovery effect is poor. With the continuous development of machine learning and deep learning technology, various neural networks are gradually applied to the field of super-resolution perception. Dong et al. first proposed a convolutional neural network SRCNN containing three layers of convolution. Compared with traditional methods, it can accurately reconstruct high-precision images. Subsequently, neural networks such as generative adversarial networks (GANs), attention mechanisms, and residual networks have gradually made significant progress in the field of super-resolution perception. The deep learning-based method can directly learn the potential mapping relationship between low-frequency images and high-frequency images, reconstruct the lost feature information in low-frequency images, and reconstruct high-frequency images.
[0006] This paper proposes a super-resolution perception technology for power data based on an improved diffusion model. The addition of a long short-term memory network module (LSTM) enables the diffusion model to fully mine the temporal features in the data, continuously iterate and refine to remove noise, and achieve high-precision data reconstruction of different multiples. It is also applicable to different electrical parameter data and different building power data, and has good generalization ability and high flexibility. Summary of the invention
[0007] The purpose of the invention in this article is to solve the problem that high-precision power data cannot be put into practical use due to various limitations in the collection, transmission and storage processes, and to provide a power data super-resolution perception technology based on an improved diffusion model. When used, high-precision data can be constructed from low-precision data without the need for excessive collection, transmission and storage of high-frequency data.
[0008] The present invention adopts a diffusion model with powerful generation ability, selects the U-Net model as the network for parameter training in the diffusion model, and integrates the LSTM network to extract time series features, fully learns the time series characteristics in the power data, and uses the operation of restoring the length of low-resolution data through traditional interpolation methods to make the super-resolution perception ability diverse. The method of the present invention also has good generalization ability.
[0009] A method for super-resolution perception of power data based on an improved diffusion model is proposed. The method can be divided into two major processes. The diffusion process uses low-resolution power data as a control condition and gradually adds noise to the high-resolution power data until the power data is oriented to pure Gaussian noise; the reverse diffusion process uses the U-Net model parameter sharing to learn the random noise added in each step of the diffusion process, and uses the low-resolution data and the learned noise according to the derived specific formula to recover the high-resolution power data from the pure Gaussian noise, thereby realizing super-resolution reconstruction of power data. The specific steps are as follows:
[0010] Step 1: Preprocess the power data, remove missing data, unify the data length, and perform maximum and minimum normalization;
[0011] Step 2: Construct the power data into three subsets: high-resolution dataset X, low-resolution dataset X lr , interpolation super-resolution dataset X ulr ;
[0012] Step 3: Add very small random noise to the high-resolution data X in T steps, and the data obtained in each step are X1, X2, …X t ,..,X T , each step of X t The interpolation super-resolution data as the control condition is used as input data and input into the U-Net model to learn the features of the random noise added in the tth step;
[0013] Step 4: Generate super-resolution data from random noise according to the parameters learned by the U-Net model in step 3. Randomly generate a pure Gaussian noise, use the interpolation super-resolution data as the control condition, iterate T steps, perform inverse diffusion, and gradually restore it to super-resolution data close to high-resolution data.
[0014] The steps of constructing the power data set are:
[0015] Step 1: Take the power data as high-resolution data;
[0016] Step 2: Downsample the high-resolution data (the downsampling rate is K) as low-resolution data;
[0017] Step 3: Use traditional interpolation methods (such as bicubic interpolation) to restore the data length of the low-resolution data so that its data size is the same as that of the high-resolution data, which is used as the interpolation super-resolution data set.
[0018] The steps for building the U-Net model are:
[0019] Step 1: Take Xt and the interpolation super-resolution dataset as input data;
[0020] Step 2: In the feature extraction stage, the input data passes through the downsampled convolution residual module (DRB) to obtain high-level features in the data. The data dimension changes in this stage: the data length is reduced by 1 / 2 after the downsampling operation, and the number of data channels is increased by 1 time after the convolution operation;
[0021] Step 3: In the dimension recovery stage, the data is gradually restored through the upsampling convolution residual module (URB): the data length is doubled after the upsampling operation, and the number of data channels is reduced by 1 / 2 after the convolution operation, so that the output data is consistent with the input data dimension;
[0022] Step 4: Feature fusion stage, which runs through the feature extraction in step 2 and the dimension restoration in step 3. The convolution residual module of up and down sampling uses residual network jump connection to preserve the details of the downsampling process.
[0023] The reverse diffusion process is as follows:
[0024] Step 1: Randomly generate pure Gaussian noise ε~N(0,I) as X T ;Xulr is used as the control condition and is denoted as y;
[0025] Step 2: Place the X T , X ulr As input data, according to X T-1 =g(X T ) to get X T-1 , g(X T ) formula is as follows, where α t =1-β t , β t is the preset variance when adding noise, σ 2 Because of β t Represents the variance, z is a standard Gaussian distribution:
[0026]
[0027] Step 3: Iteratively repeat step 2 until X0, i.e., super-resolution power data, is obtained. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 This is a system processing flow chart of the power data super-resolution perception method based on the improved diffusion model.
[0029] Figure 2 It is the U-Net model network architecture that learns the parameters required for the diffusion model. DETAILED DESCRIPTION
[0030] The following is combined with Figure 1 , Attachment Figure 2 The power data super-resolution perception method based on the improved diffusion model is introduced in detail in the embodiments, but the implementation of the present invention is not limited to this.
[0031] Example:
[0032] like Figure 1 As shown, the power data super-resolution perception processing flow based on the improved diffusion model of this embodiment uses the diffusion model to construct the main model, and recovers information details from low-resolution data to obtain super-resolution data.
[0033] like Figure 2 As shown, the U-Net network structure used for the diffusion model noise prediction in this embodiment builds a U-Net model containing feature extraction, dimensionality recovery and feature fusion.
[0034] The steps to build the entire model are as follows:
[0035] (1) Prepare a super-resolution perception dataset for power data: use the foreign public dataset I-BLEND with a sampling frequency of 1 minute
[0036] The active power of the male dormitory for 52 months is constructed in units of days, that is, the length of each data is 24*60*1. (2) Preprocess the data: After removing the empty data and invalid data, a total of 1187 high-resolution power data with a length of 1440 are obtained.
[0037] The force data was normalized using the maximum and minimum normalization method, and low-resolution data with a length of 96 was obtained after downsampling (downsampling rate K = 15), and then interpolation super-resolution data with a length of 1440 was obtained after upsampling using the bicubic interpolation method. The high-resolution data and interpolation super-resolution data constitute the data set used in the experiment. 100 data were randomly selected as the test set, and the remaining data were used as the training set.
[0038] (3) The diffusion process gradually adds random noise and learns the characteristic information of the added noise: the total number of steps T is 1000, and the noise variance added in each step is β1β2,…,β T Set to 10 -6 to 10 -2 Linear growth: the model adds smaller noise in the early stage of the diffusion process and larger noise in the later stage, which makes it easier for the model to learn the detailed features of the high-precision data in the early stage. The number of groups in the group normalization layer is 32, that is, the channels are divided into 32 small groups, and the normalization is performed independently within the groups to reduce the computational cost of the model. The Dropout layer randomly sets the output of some neurons to zero with a loss rate of 0.2 to reduce model overfitting.
[0039] (4) The inverse diffusion process iteratively generates super-resolution data from pure Gaussian noise.
[0040] The present invention provides a method for super-resolution perception of electric power data based on an improved diffusion model, and belongs to the technical field of electric power big data processing. The method utilizes the improved diffusion model to fully exploit the time series characteristics of electric power data, and can realize the collection of low-frequency data for transmission and storage. When in use, the improved diffusion model is used for up-frequency reconstruction to obtain high-frequency data for analysis or prediction. The present invention is completely data-driven, and overcomes the limitations of mathematical reconstruction models based on business mechanisms in terms of adaptability, algorithm efficiency, and accuracy in complex distribution network environments. The use of this method avoids the need to make distribution assumptions and explicit feature modeling of data based on a large amount of prior knowledge in the traditional missing reconstruction process, and has a high accuracy of data super-resolution reconstruction.
Claims
1. A method for super-resolution perception of power data based on an improved diffusion model, characterized in that: The diffusion model in deep learning is used to learn the detailed information in the power data in the process of continuous noise addition, and the learned noise addition information is used to gradually denoise the pure Gaussian noise. The LSTM layer is integrated into the model, which retains the ability of the convolution layer in the diffusion model to extract high-level data features while adding the LSTM's ability to mine time series features. The time series characteristics in the power data are fully learned to achieve super-resolution perception of the power data. The specific steps of this method are as follows: Step 1: Preprocess the power data, remove missing data, unify the data length, and perform maximum and minimum normalization; Step 2: Construct the power data into three subsets: high-resolution dataset X, low-resolution dataset X lr , interpolation super-resolution dataset X ulr ; Step 3: Add very small random noise to the high-resolution data X in T steps, and the data obtained in each step are X1, X2, …X t ,..,X T , each step of X t The interpolation super-resolution data as the control condition is used as input data and input into the U-Net model to learn the features of the random noise added in the tth step; Step 4: Generate super-resolution data from random noise according to the parameters learned by the U-Net model in step 3. Randomly generate a pure Gaussian noise, use the interpolation super-resolution data as the control condition, iterate T steps, perform inverse diffusion, and gradually restore it to super-resolution data close to high-resolution data.
2. According to claim 1, a method for super-resolution perception of power data based on an improved diffusion model is characterized in that: The steps of constructing the power data into three subsets are as follows: Step 1: Take the power data as high-resolution data; Step 2: Downsample the high-resolution data (the downsampling rate is K) as low-resolution data; Step 3: Use traditional interpolation methods (such as bicubic interpolation) to restore the data length of the low-resolution data so that its data size is the same as that of the high-resolution data, which is used as the interpolation super-resolution data set.
3. According to claim 1, a method for super-resolution perception of power data based on an improved diffusion model is characterized in that: The steps of building the U-Net model are as follows: Step 1: Take Xt and the interpolation super-resolution dataset as input data; Step 2: In the feature extraction stage, the input data passes through the downsampled convolution residual module (DRB) to obtain high-level features in the data. The data dimension changes in this stage: the data length is reduced by 1 / 2 after the downsampling operation, and the number of data channels is increased by 1 time after the convolution operation; Step 3: In the dimension recovery stage, the data is gradually restored through the upsampling convolution residual module (URB): the data length is doubled after the upsampling operation, and the number of data channels is reduced by 1 / 2 after the convolution operation, so that the output data is consistent with the input data dimension; Step 4: Feature fusion stage, which runs through the feature extraction in step 2 and the dimension restoration in step 3. The convolution residual module of up and down sampling uses residual network jump connection to preserve the details of the downsampling process.
4. According to claim 1, a method for super-resolution perception of power data based on an improved diffusion model is characterized in that: The reverse diffusion process steps are: Step 1: Randomly generate pure Gaussian noise ε~N(0,I) as X T ;Xulr is used as the control condition and is denoted as y; Step 2: Place the X T , X ulr As input data, according to X T-1 =g(X T ) to get X T-1 , g(X T ) formula is as follows, where α t =1-β t , β t is the preset variance when adding noise, σ 2 Because of β t Represents the variance, z is a standard Gaussian distribution: Step 3: Iteratively repeat step 2 until X0, i.e., super-resolution power data, is obtained.