Deep learning based low resolution load data super-resolution reconstruction method and system

Abstract

The application discloses a kind of based on deep learning's low-resolution load data super-resolution reconstruction method and system, belong to power system distribution network operation monitoring technical field;The application constructs the end-to-end constraint generative adversarial network model, embeds differentiable convex optimization CVXPY layer in generator, and embeds load amplitude boundary constraint, adjacent sampling point change rate constraint as hard constraint into model training process.The application only uses 15 minute interval low-resolution aggregated load data collected by smart meter, can train to generate 1 minute interval high-resolution load injection time series data conforming to power system operation characteristics, with strict time correlation, without deploying a large number of high-speed sensors at load node.The generated data can be directly used for distribution network high-speed state estimation, real-time voltage control and other core scenarios, greatly improving the engineering application value of AMI data, with strong generalization ability and practicality.

Description

Technical Field

[0001] This invention belongs to the field of power system distribution network operation monitoring technology, and relates to a method and system for super-resolution reconstruction of low-resolution load data based on deep learning. Background Technology

[0002] With the large-scale integration of distributed energy resources (DERs) into distribution networks, traditional passive distribution networks are gradually transforming into active, market-participating grid entities. The inherent power output volatility of DERs places extremely high demands on the reliable operation and real-time monitoring capabilities of distribution networks. Smart meters in Advanced Metering Systems (AMIs) provide the basic data source for distribution network observation, but their core applications remain concentrated in energy-related businesses such as monthly electricity billing. The data they collect is mostly average power consumption values ​​at 15-minute intervals, failing to capture the rapid fluctuations in instantaneous load power. Furthermore, smart meter data suffers from transmission delays, making it unsuitable for core scenarios requiring millisecond- or minute-level decision-making, such as high-speed distribution state estimation (DSSE) and real-time voltage control. This severely limits the application of AMI data in the real-time operation and control of distribution networks.

[0003] Existing research has confirmed that offline training and online operation can be decoupled through learning-based methods. This involves using historical AMI data for offline model training, while the online phase requires only a small amount of high-speed sensor data, such as data acquisition and monitoring control systems (SCADA) and micro phasor measurement units (MicroPMUs) at the feeder head, to achieve rapid state estimation of unobservable distribution networks. This method outperforms traditional methods based on load forecasting pseudo-measurements. However, training with these learning-based methods relies on a high-resolution sample library capable of covering minute-level load fluctuation characteristics, which is difficult for existing AMI systems to directly provide.

[0004] To address this issue, the industry has introduced super-resolution theory from the field of image processing, attempting to generate high-resolution data from slow-timescale low-resolution data using data-driven generative models. Generative Adversarial Networks (GANs) are a mainstream technique for load profile generation; for example, existing methods like ProfileSR-GAN can upsample load profiles from low to high resolution, recovering high-frequency components lost during downsampling. However, existing methods have shortcomings. They do not embed the physical characteristics and operational constraints of the power system into the model and ignore the inherent temporal correlation of load data. This leads to significant and unreasonable fluctuations in the generated high-resolution data within a fixed time window. Although the aggregated 15-minute data has statistical characteristics similar to the real data, the minute-level data completely fails to reflect the actual operating patterns of the load and cannot be used for core scenarios such as real-time state estimation and control of distribution networks.

[0005] In the field of deep learning constrained optimization, existing methods mostly implement soft boundary constraints by adding penalty terms to the loss function. Such methods only penalize infeasible solutions and cannot guarantee the feasibility of the solutions. The few hard constraint implementation schemes mostly involve post-processing the output of unconstrained models, which cannot achieve end-to-end training, make it difficult to embed the physical knowledge of the system into the model generation process, and cannot fundamentally guarantee the physical authenticity of the generated data. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for super-resolution reconstruction of low-resolution load data based on deep learning, so as to solve the technical problems of scarcity of high-resolution load data in existing distribution networks, lack of physical authenticity and temporal correlation in data generated by traditional super-resolution methods, and difficulty in meeting the requirements of high-speed state estimation and real-time control of distribution networks.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for super-resolution reconstruction of low-resolution load data based on deep learning, comprising the following steps: Historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes are collected simultaneously, and the data is filtered and preprocessed according to the same scenario. Construct an end-to-end constrained generative adversarial network model, which is a two-player minimax game architecture, including a generator and a discriminator; Hard physical constraints on load operation are embedded in the generator to construct a constrained convex optimization objective function, which includes load amplitude boundary constraints and adjacent sampling point change rate constraints. Based on the preprocessed low-resolution aggregated load data and high-resolution measured load data, the network parameters of the discriminator and generator are updated in an alternating iterative manner, and the constrained generative adversarial network model is trained end-to-end until the model converges. Using a trained generator and a uniformly distributed random noise vector as input, a high-resolution load injection time series profile is obtained, completing the super-resolution reconstruction from low resolution to high resolution.

[0008] Furthermore, the steps of simultaneously collecting historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes, and performing scenario-based filtering and preprocessing, specifically include: Historical low-resolution aggregated load data of smart meters at the target distribution network load nodes are collected, wherein the low-resolution aggregated load data is the average power consumption data at 15-minute intervals; simultaneously, high-resolution measured load data of the corresponding load nodes are collected, wherein the high-resolution measured load data is the instantaneous power injection data at 1-minute intervals; For low-resolution aggregated load data and high-resolution measured load data, the same scenario is selected, and load samples from the same season and the same time period in historical data are selected to remove outliers and missing values. The load data after filtering for the same scenario is preprocessed by normalization. Based on the maximum and minimum values ​​of historical loads, the load data is normalized by 0-1.

[0009] Furthermore, the generator sequentially sets a fully connected hidden layer, a differentiable convex optimization CVXPY layer, and an aggregation layer along the data flow direction. The generator is used to generate high-resolution load injection time-series data with a 1-minute interval that conforms to physical constraints from random noise input, and aggregates and outputs simulated smart meter data with a 15-minute interval. The discriminator is a fully connected neural network with Sigmoid activation output, used to distinguish whether the input 15-minute interval data is real smart meter data or simulated data output by the generator.

[0010] Furthermore, the generator has two fully connected hidden layers, each containing 128 neurons. Each layer includes batch normalization and a dropout layer with a dropout rate of 0.3. The CVXPY layer is the penultimate layer of the generator, with 15 neurons corresponding to 15 high-resolution sampling points at 1-minute intervals. The aggregation layer is a non-training layer that calculates the average value of the 15 1-minute sampling points output by the CVXPY layer, outputting a 15-minute interval of simulated smart meter aggregated data to match the input dimension of the discriminator.

[0011] Furthermore, the discriminator has two fully connected hidden layers, each containing 128 neurons. The output layer of the discriminator uses the Sigmoid activation function, and the output range is a probability value of 0-1. Here, output 1 represents that the input is real smart meter data, and output 0 represents that the input is simulated data generated by the generator.

[0012] Furthermore, the expression for the constrained convex optimization objective function is:

[0013] In the formula, Optimize the high-resolution load vector output for the CVXPY layer. This corresponds to 15 one-minute sampling points; For the first i The generated values ​​of 1-minute sampling points These are the input values ​​of the CVXPY layer, i.e., the output vectors of the fully connected hidden layer; The expression for the load amplitude boundary constraint is:

[0014] The expression for the constraint on the rate of change of adjacent sampling points is:

[0015] In the formula, This is the minimum value of the historical 15-minute aggregated load data. This represents the maximum value of the historical 15-minute aggregated load data; and This is the amplitude scaling factor. This is the maximum permissible rate of change coefficient; The overall objective function of the constrained generative adversarial network model is a two-player minimax game function, expressed as:

[0016] in, For generator, For discriminator, This is based on real 15-minute smart meter data. This is the random noise vector input to the generator. For the mapping function of the convex optimization layer, The function for calculating the average value of the aggregation layer. The operator is computed to the expectation.

[0017] Furthermore, the step of updating the network parameters of the discriminator and generator using an alternating iterative method to train the constrained generative adversarial network model end-to-end until the model converges specifically includes: In each training round, the generator parameters are first fixed, and real smart meter data and simulated data output by the generator are input to update the discriminator parameters and minimize the discriminator's cross-entropy loss. Then, the discriminator parameters are fixed, a random noise vector is input, and the generator parameters are updated to minimize the generator's adversarial loss. The above iterative process is repeated until the model loss converges and the distribution difference between the simulated data output by the generator and the real data reaches a preset threshold, thus completing the training of the constrained generative adversarial network model.

[0018] Secondly, the present invention provides a deep learning-based system for super-resolution reconstruction of low-resolution load data, comprising: The data acquisition and processing module is used to simultaneously acquire historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes, and to perform same-scenario filtering and preprocessing. The model building module is used to build an end-to-end constrained generative adversarial network model, which is a two-player minimax game architecture and includes a generator and a discriminator. The constraint construction module is used to embed hard physical constraints of load operation in the generator to construct a constrained convex optimization objective function, wherein the constraints include load amplitude boundary constraints and adjacent sampling point change rate constraints. The model training module is used to update the network parameters of the discriminator and generator respectively using an alternating iterative method based on the preprocessed low-resolution aggregated load data and high-resolution measured load data, and train the constrained generative adversarial network model end-to-end until the model converges. The super-resolution reconstruction module is used to obtain a high-resolution load injection time series profile by taking a uniformly distributed random noise vector as input from a trained generator, thus completing the super-resolution reconstruction from low resolution to high resolution.

[0019] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the deep learning-based low-resolution overhead data super-resolution reconstruction method described above.

[0020] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the deep learning-based low-resolution overhead data super-resolution reconstruction method described above.

[0021] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a method and system for super-resolution reconstruction of low-resolution load data based on deep learning. By constructing an end-to-end constrained generative adversarial network (GAN) model and embedding a differentiable convex optimization (CVXPY) layer into the generator architecture of the GAN, it achieves end-to-end embedding of physical hard constraints of the power system. Compared with traditional soft constraint methods, it can strictly ensure that the generated high-resolution load data conforms to the actual operating characteristics, completely avoiding the unreasonable large jumps in data generated by traditional GANs. The rate of change and amplitude boundaries of adjacent sampling points of the generated data are completely consistent with the real load data. The high-resolution load data generated by this invention has strict temporal correlation and statistical authenticity. The aggregated 15-minute data distribution highly matches the distribution of real smart meter data, and the minute-level data completely conforms to the fluctuation law of actual load. It can be directly used as training data and input data for algorithms such as high-speed state estimation, real-time voltage control, and demand response analysis of distribution networks, significantly improving the application value of AMI data in the real-time operation control of distribution networks.

[0022] Furthermore, this invention adapts to the resolution improvement requirements of 15 minutes to 5 minutes and 15 minutes to 10 minutes by adjusting the number of neurons and constraint parameters in the CVXPY layer. Correspondingly, the sampling rate and calibration values ​​of the constraint parameters are adjusted to achieve high-resolution load profile generation in different scenarios. The model of this invention is an end-to-end trainable architecture with strong generalization performance. It eliminates the need to build separate load models for different types of load nodes; only fine-tuning based on historical smart meter data for the corresponding nodes is required to adapt to application scenarios with different load characteristics, demonstrating strong engineering practicality.

[0023] Furthermore, this invention only requires the use of low-resolution aggregated data collected by smart meters at 15-minute intervals to complete model training, eliminating the need to deploy a large number of high-speed sensors at load nodes. This significantly reduces the hardware cost of high-resolution observation of the distribution network and solves the industry pain point of scarce high-resolution load training data for the distribution network. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system of the present invention; Figure 3 This is a schematic diagram of the generator structure of the present invention; Figure 4 This is a schematic diagram of the discriminator structure of the present invention; Figure 5 This is a schematic diagram of the alternating iterative training process in an embodiment of the present invention; Figure 6 This is a schematic diagram of statistical characteristic verification in an embodiment of the present invention; Figure 7 This is a schematic diagram of time correlation verification in an embodiment of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0027] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0029] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0030] This invention provides a deep learning-based method for super-resolution reconstruction of low-resolution workload data, which can be executed by an electronic device, such as a terminal or server. The terminal can be a smartphone, tablet, laptop, or other similar device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. It is understood that this invention does not limit the specific entity executing this deep learning-based method for super-resolution reconstruction of low-resolution workload data.

[0031] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments described below are used to explain the technical solution of this application and are not intended to limit actual use.

[0032] See Figure 1 This invention discloses a method for super-resolution reconstruction of low-resolution load data based on deep learning, comprising the following steps: S1, synchronously collect historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes, and perform same-scenario filtering and preprocessing; Specifically, historical low-resolution aggregated load data from smart meters of target distribution network load nodes is collected. This low-resolution aggregated load data consists of average power consumption data at 15-minute intervals. Simultaneously, high-resolution measured load data from corresponding load nodes is collected for constraint parameter calibration and model validation. This high-resolution measured load data consists of instantaneous power injection data at 1-minute intervals. The collected data undergoes scenario-based filtering and normalization preprocessing to construct model training and validation sets, respectively. Scenario-based filtering involves selecting load samples from the same season and time period in historical data, removing outliers and missing values ​​to ensure scenario consistency in the training samples. Preprocessing involves normalizing the load data to 0-1 based on the maximum and minimum historical load values.

[0033] S2, Construct an end-to-end constrained generative adversarial network model, wherein the constrained generative adversarial network model is a two-player minimax game architecture, including a generator and a discriminator; Specifically, the generator is used to generate high-resolution load injection time-series data with a 1-minute interval that conforms to physical constraints from random noise input, and aggregates and outputs analog smart meter data with a 15-minute interval; the generator sequentially sets a fully connected hidden layer, a differentiable convex-optimized CVXPY layer, and a non-trainable aggregation layer along the data flow direction; the generator is as follows: Figure 3 As shown; the generator has two fully connected hidden layers, each containing 128 neurons, and each layer includes batch normalization and dropout, with a dropout rate of 0.3; the CVXPY layer is the penultimate layer of the generator, with 15 neurons corresponding to 15 high-resolution sampling points at 1-minute intervals; the aggregation layer is a non-training layer that calculates the average value of the 15 1-minute sampling points output by the CVXPY layer, outputting a 15-minute interval of simulated smart meter aggregated data to match the input dimension of the discriminator.

[0034] The discriminator is a fully connected neural network with Sigmoid activation output, used to distinguish whether the input 15-minute interval data is real smart meter data or simulated data output by the generator, and finally outputs a continuous probability score of the input data as real data; the discriminator is as follows: Figure 4 As shown, the discriminator has two fully connected hidden layers, each containing 128 neurons. The output layer uses the Sigmoid activation function, and the output range is a probability value of 0-1. Output 1 represents the input as real smart meter data, and output 0 represents the input as simulated data generated by the generator.

[0035] S3, embed hard physical constraints of load operation into the generator to construct a constrained convex optimization objective function, the constraints including load amplitude boundary constraints and adjacent sampling point change rate constraints; Specifically, the expressions for the constrained convex optimization objective function and constraints are as follows:

[0036] Constraint 1: Load amplitude boundary constraint; limiting the power value of each generated 1-minute sampling point to not exceed the upper and lower limits of load fluctuation calibrated based on historical high-resolution data; the specific expression is:

[0037] Constraint 2: Adjacent sampling point change rate constraint, limiting the power change amplitude between two adjacent 1-minute sampling points to not exceed the maximum allowable change rate calibrated based on historical data; the specific expression is:

[0038] in, Optimize the high-resolution load vector output for the CVXPY layer. This corresponds to 15 one-minute sampling points; For the first i The generated values ​​of 1-minute sampling points These are the input values ​​of the CVXPY layer, i.e., the output vectors of the fully connected hidden layer; This is the minimum value of the historical 15-minute aggregated load data. This represents the maximum value of the historical 15-minute aggregated load data; , This is the amplitude scaling factor. The three parameters are user-defined parameters calibrated based on high-resolution measured data, and are the maximum permissible rate of change coefficients.

[0039] S4. Based on the preprocessed low-resolution aggregated load data and high-resolution measured load data, the network parameters of the discriminator and generator are updated in an alternating iterative manner, and the constrained generative adversarial network model is trained end-to-end until the model converges. Specifically, the overall objective function of the constrained generative adversarial network model is a two-player minimax game function, expressed as:

[0040] in, For generator, For discriminator, This is based on real 15-minute smart meter data. This is the random noise vector input to the generator. For the mapping function of the convex optimization layer, The function for calculating the average value of the aggregation layer. The operator is computed to the expectation.

[0041] S5 utilizes the trained generator and inputs a random noise vector that follows a uniform distribution to obtain a high-resolution load injection time series profile, thus completing the super-resolution reconstruction from low resolution to high resolution.

[0042] Specifically, in each training round, the generator parameters are first fixed, and real smart meter data and simulated data output by the generator are input to update the discriminator parameters and minimize the discriminator's cross-entropy loss. Then, the discriminator parameters are fixed again, a random noise vector is input, and the generator parameters are updated to minimize the generator's adversarial loss. This iterative process is repeated until the model loss converges and the distribution difference between the simulated data output by the generator and the real data reaches a preset threshold. In each iteration, the discriminator parameters are updated 5 times, and then the generator parameters are updated once to ensure that the discriminator's discrimination ability always leads the generator.

[0043] Preferably, by adjusting the number of neurons and constraint parameters of the CVXPY layer, the resolution improvement requirements of 15 minutes to 5 minutes and 15 minutes to 10 minutes are adapted, and the sampling rate and the calibration values ​​of the constraint parameters are adjusted accordingly to achieve high-resolution load contour generation in different scenarios.

[0044] See Figure 2This invention discloses a deep learning-based system for super-resolution reconstruction of low-resolution load data, comprising a data acquisition and processing module, a model building module, a constraint building module, a model training module, and a super-resolution reconstruction module. The data acquisition and processing module is used to simultaneously acquire historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes, and perform scene-based filtering and preprocessing. The model building module is used to construct an end-to-end constrained generative adversarial network model, which is a two-player minimax game architecture, including a generator and a discriminator. The constraint building module is used to generate... The system embeds hard physical constraints on load operation to construct a constrained convex optimization objective function, which includes load amplitude boundary constraints and adjacent sampling point change rate constraints. A model training module is used to update the network parameters of the discriminator and generator respectively using an alternating iterative method based on preprocessed low-resolution aggregated load data and high-resolution measured load data, training the constrained generative adversarial network model end-to-end until the model converges. A super-resolution reconstruction module is used to obtain a high-resolution load injection time series profile by inputting a uniformly distributed random noise vector into the trained generator, completing the low-resolution to high-resolution super-resolution reconstruction.In this embodiment, the system first uses a data acquisition and processing module to simultaneously acquire historical 15-minute low-resolution aggregated load data and 1-minute high-resolution measured load data of the target distribution network load nodes. Abnormal and missing samples are removed through scenario-based filtering, and normalization is performed based on historical extreme values ​​to eliminate dimensional differences, thus providing a high-quality, scenario-consistent data foundation for model training. Subsequently, the model building module constructs an end-to-end constraint generative adversarial network. Its generator sequentially deploys two 128-dimensional fully connected hidden layers, a 15-neuron CVXPY differentiable convex optimization layer, and a non-training average aggregation layer along the data flow. The discriminator uses two 128-dimensional fully connected layers with a Sigmoid output head, forming a standard two-player zero-sum game architecture. Based on this, the constraint building module embeds two hard physical constraints—load amplitude boundary and rate of change of adjacent sampling points—into the CVXPY layer. This transforms the physical law that the load in actual power system operation cannot experience instantaneous jumps or exceed limits into the solution boundary of the convex optimization objective function, mathematically strengthening the constraint. The generated data conforms to physical common sense. Next, the model training module employs an alternating iterative strategy: in each round, the generator is fixed and the discriminator is updated 5 times to enhance its discrimination ability, then the discriminator is fixed and the generator is updated 1 time to optimize the generation quality, until the discriminator output probability stabilizes around 0.5 and the cumulative distribution function of the generated data highly overlaps with that of the real data, achieving stable model convergence. Finally, the super-resolution reconstruction module inputs a uniformly distributed random noise vector into the trained generator. Through fully connected layer feature mapping, CVXPY layer physical constraint solving, and aggregation layer reconstruction, it directly outputs a 1-minute high-resolution load injection time series profile that conforms to the physical characteristics of the power grid. This invention not only overcomes the bottleneck of traditional interpolation methods failing to restore high-frequency fluctuation characteristics and effectively solves the computational divergence problem caused by information loss in high-speed state estimation and real-time voltage control of low-resolution data, but also significantly improves the engineering credibility of the generated data by embedding hard physical constraints, providing high-value data support for refined operation and maintenance of the distribution network without relying on expensive hardware upgrades.

[0045] Example: (a) Implementation environment and dataset preparation This embodiment uses measured load injection data as its data source, including high-resolution measured data of residential load nodes at 1-minute intervals and corresponding aggregated smart meter data at 15-minute intervals, perfectly matching the application scenario of this invention. To ensure the consistency of the training samples, load data from 12:00-13:00 daily during the summer period was selected for the target load nodes, as the characteristics of residential loads are similar during this period, which can improve the training effect of the model. Missing values ​​and outliers in the data were removed, and a dataset containing 1840 valid samples was finally constructed, of which 80% was used as the training set and 20% as the validation set. Finally, the constraint parameters were calibrated using the 1-minute high-resolution measured data, and the specific calibration results are as follows: The measured power injection data for the target load node fluctuates between 0 and 10.26 kW over 1 minute, while the corresponding aggregated data fluctuates between 0.55 and 7.38 kW over 15 minutes. Therefore, the historical minimum value is set. Historical maximum value ; Amplitude scaling factor: Setting , This ensures that the maximum amplitude of the generated 1-minute data does not exceed 10.26kW, and is completely consistent with the boundary of the measured data; Maximum permissible rate of change coefficient: Statistically, the maximum rate of power change between adjacent sampling points in a 1-minute measurement does not exceed 50%, therefore, it is set as follows: Strictly limit the magnitude of adjacent jumps in the generated data.

[0046] (II) Specific Construction of Constrained Generative Adversarial Network Model The generator is constructed. The input layer takes a random noise vector uniformly distributed in the range [-1, 1]. Two fully connected hidden layers are used, each containing 128 neurons. Each layer is configured with batch normalization and a dropout layer with a dropout rate of 0.3. The ReLU activation function is used. The CVXPY layer is a differentiable convex optimization layer with 15 neurons, corresponding to 15 high-resolution sampling points at 1-minute intervals. The layer embeds the aforementioned amplitude boundary constraints and adjacent sampling point change rate constraints to solve a constrained convex optimization problem, outputting a 1-minute load data vector that meets the constraints. The aggregation layer is a non-training layer that calculates the arithmetic mean of the 15 1-minute sampling points output by the CVXPY layer, outputting a 15-minute interval simulated smart meter aggregated data, with dimensions perfectly matching the real smart meter data.

[0047] Construct the discriminator: The input layer takes real smart meter data at 15-minute intervals as input, or simulated aggregated data output by the generator; the fully connected hidden layer has two fully connected layers, each containing 128 neurons, and the activation function is LeakyReLU; the output layer has one neuron, uses the Sigmoid activation function, and outputs a continuous probability value between 0 and 1, representing the probability that the input data is real smart meter data.

[0048] (III) End-to-end training of the model The training hardware environment was a computing platform equipped with an Intel Xeon 6246R CPU (3.40GHz), 256GB of memory, and an Nvidia Quadro RTX 5000 16GB GPU. The Adam optimizer was used for training parameter settings, with a learning rate of 0.0002, a batch size of 64, and 1000 training iterations.

[0049] Alternating iterative training process and convergence criteria: 1. In each iteration, the discriminator is trained first: the parameters of the generator are fixed, and batches of real smart meter data and simulated data generated by the generator are input respectively. The cross-entropy loss of the discriminator is calculated, and the network parameters of the discriminator are updated through backpropagation to maximize the discriminator's ability to distinguish between real data and generated data. 2. Then train the generator: fix the parameters of the discriminator, input a random noise vector, output simulated aggregated data through the generator, input it into the discriminator, calculate the adversarial loss of the generator, update the fully connected hidden layer parameters of the generator through backpropagation, and minimize the discriminator's recognition accuracy on the generated data; 3. In each iteration, the discriminator parameters are updated 5 times, and then the generator parameters are updated once to ensure that the discriminator's discrimination ability always leads the generator's, which conforms to the training logic of GAN. The alternating iterative training process is as follows: Figure 5 As shown; 4. Model convergence determination: When the cumulative distribution function (CDF) of the simulated aggregated data output by the generator is highly consistent with that of the real smart meter data, and the probability value output by the discriminator is stable at around 0.5, the model is determined to have converged and training is stopped. In this embodiment, the model reaches the convergence state after about 6 hours of training.

[0050] (iv) Generation and verification of high-resolution load data The aggregation layer in the trained generator is removed, and the input is a random noise vector that follows a uniform distribution. After being mapped by a fully connected hidden layer, 15 high-resolution load injection time series data with a 1-minute interval are directly output through the CVXPY layer to complete the super-resolution reconstruction from 15 minutes of low-resolution data to 1 minute of high-resolution data. Statistical property verification, such as Figure 6 As shown, the CDF curves of the generated 15-minute simulated aggregated data and the real smart meter data almost completely overlap, proving that the aggregated statistical characteristics of the generated data are consistent with the real data. Physical authenticity verification. Statistical analysis was conducted on 500 sets of generated 1-minute high-resolution load profiles. As shown in Table 1, the maximum change rate of adjacent sampling points in the constrained GAN-generated data of this invention consistently did not exceed 50%, and the maximum generated power value remained stable at 10.26kW, completely consistent with the indicators of the measured real data. In contrast, the maximum change rate of adjacent sampling points in the traditional unconstrained GAN-generated data could reach 5000%, and the maximum generated power value reached 13.21kW, exhibiting serious violations of physical characteristics. Table 1. Maximum percentage change between two consecutive samples and maximum instantaneous values ​​of 500 high-resolution load injection time-series profiles.

[0051] Time correlation verification, such as Figure 7 As shown, the 1-minute load profile generated by this invention has continuous and smooth time variation characteristics, without unreasonable abrupt changes, and fully conforms to the actual electricity consumption patterns of residential loads. It can be directly used in core scenarios such as high-speed state estimation and real-time voltage control of distribution networks.

[0052] (v) Extended applications of the model In this embodiment, the core scenario is a resolution improvement from 15 minutes to 1 minute. Those skilled in the art can adapt to different resolution improvement requirements, such as 15 minutes to 5 minutes or 15 minutes to 10 minutes, by adjusting the number of neurons and constraint parameters of the CVXPY layer based on the core architecture of this invention. Only the value of the magnification m and the calibration value of the constraint parameters need to be adjusted accordingly to achieve high-resolution load contour generation in different scenarios, all of which fall within the protection scope of this invention.

[0053] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may 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. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a deep learning-based low-resolution overload data super-resolution reconstruction method.

[0054] This invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the deep learning-based low-resolution overhead data super-resolution reconstruction method described in the above embodiments.

[0055] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention 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.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for super-resolution reconstruction of low-resolution load data based on deep learning, characterized in that, Includes the following steps: Historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes are collected simultaneously, and the data is filtered and preprocessed according to the same scenario. Construct an end-to-end constrained generative adversarial network model, which is a two-player minimax game architecture, including a generator and a discriminator; Hard physical constraints on load operation are embedded in the generator to construct a constrained convex optimization objective function, which includes load amplitude boundary constraints and adjacent sampling point change rate constraints. Based on the preprocessed low-resolution aggregated load data and high-resolution measured load data, the network parameters of the discriminator and generator are updated in an alternating iterative manner, and the constrained generative adversarial network model is trained end-to-end until the model converges. Using a trained generator and a uniformly distributed random noise vector as input, a high-resolution load injection time series profile is obtained, completing the super-resolution reconstruction from low resolution to high resolution.

2. The method for super-resolution reconstruction of low-resolution load data based on deep learning according to claim 1, characterized in that, The steps of synchronously collecting historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes, and performing scene-based filtering and preprocessing, specifically include: Historical low-resolution aggregated load data of smart meters at the target distribution network load nodes are collected, wherein the low-resolution aggregated load data is the average power consumption data at 15-minute intervals; simultaneously, high-resolution measured load data of the corresponding load nodes are collected, wherein the high-resolution measured load data is the instantaneous power injection data at 1-minute intervals; For low-resolution aggregated load data and high-resolution measured load data, the same scenario is selected, and load samples from the same season and the same time period in historical data are selected to remove outliers and missing values. The load data after filtering for the same scenario is preprocessed by normalization. Based on the maximum and minimum values ​​of historical loads, the load data is normalized by 0-1.

3. The method for super-resolution reconstruction of low-resolution load data based on deep learning according to claim 1, characterized in that, The generator is configured with a fully connected hidden layer, a differentiable convex optimization CVXPY layer, and an aggregation layer sequentially along the data flow direction. The generator is used to generate high-resolution load injection time-series data with a 1-minute interval that conforms to physical constraints from random noise input, and aggregates and outputs simulated smart meter data with a 15-minute interval. The discriminator is a fully connected neural network with Sigmoid activation output, used to distinguish whether the input 15-minute interval data is real smart meter data or simulated data output by the generator.

4. The method for super-resolution reconstruction of low-resolution load data based on deep learning according to claim 3, characterized in that, The generator has two fully connected hidden layers, each containing 128 neurons. Each layer includes batch normalization and a dropout layer with a dropout rate of 0.

3. The CVXPY layer is the penultimate layer of the generator, with 15 neurons corresponding to 15 high-resolution sampling points at 1-minute intervals. The aggregation layer is a non-training layer that calculates the average value of the 15 1-minute sampling points output by the CVXPY layer, outputting a 15-minute interval of simulated smart meter aggregated data to match the input dimension of the discriminator.

5. The method for super-resolution reconstruction of low-resolution load data based on deep learning according to claim 3, characterized in that, The discriminator has two fully connected hidden layers, each containing 128 neurons. The output layer of the discriminator uses the Sigmoid activation function, and the output is a probability value ranging from 0 to 1. Here, an output of 1 represents that the input is real smart meter data, and an output of 0 represents that the input is simulated data generated by the generator.

6. The method for super-resolution reconstruction of low-resolution load data based on deep learning according to claim 1, characterized in that, The expression for the constrained convex optimization objective function is: In the formula, Optimize the high-resolution load vector output for the CVXPY layer. This corresponds to 15 one-minute sampling points; For the first i The generated values ​​of 1-minute sampling points These are the input values ​​of the CVXPY layer, i.e., the output vectors of the fully connected hidden layer; The expression for the load amplitude boundary constraint is: The expression for the constraint on the rate of change of adjacent sampling points is: In the formula, This is the minimum value of the historical 15-minute aggregated load data. This represents the maximum value of the historical 15-minute aggregated load data; and This is the amplitude scaling factor. This is the maximum permissible rate of change coefficient; The overall objective function of the constrained generative adversarial network model is a two-player minimax game function, expressed as: in, For generator, For discriminator, This is based on real 15-minute smart meter data. This is the random noise vector input to the generator. For the mapping function of the convex optimization layer, The function for calculating the average value of the aggregation layer. The operator is computed to the expectation.

7. The method for super-resolution reconstruction of low-resolution load data based on deep learning according to claim 1, characterized in that, The step of updating the network parameters of the discriminator and generator using an alternating iterative method to train the constrained generative adversarial network model end-to-end until the model converges specifically includes: In each training round, the generator parameters are first fixed, and real smart meter data and simulated data output by the generator are input to update the discriminator parameters and minimize the discriminator's cross-entropy loss. Then, the discriminator parameters are fixed, a random noise vector is input, and the generator parameters are updated to minimize the generator's adversarial loss. The above iterative process is repeated until the model loss converges and the distribution difference between the simulated data output by the generator and the real data reaches a preset threshold, thus completing the training of the constrained generative adversarial network model.

8. A deep learning-based super-resolution reconstruction system for low-resolution load data, characterized in that, include: The data acquisition and processing module is used to simultaneously acquire historical low-resolution aggregated load data and high-resolution measured load data of the target distribution network load nodes, and to perform same-scenario filtering and preprocessing. The model building module is used to build an end-to-end constrained generative adversarial network model, which is a two-player minimax game architecture and includes a generator and a discriminator. The constraint construction module is used to embed hard physical constraints of load operation in the generator to construct a constrained convex optimization objective function, wherein the constraints include load amplitude boundary constraints and adjacent sampling point change rate constraints. The model training module is used to update the network parameters of the discriminator and generator respectively using an alternating iterative method based on the preprocessed low-resolution aggregated load data and high-resolution measured load data, and train the constrained generative adversarial network model end-to-end until the model converges. The super-resolution reconstruction module is used to obtain a high-resolution load injection time series profile by taking a uniformly distributed random noise vector as input from a trained generator, thus completing the super-resolution reconstruction from low resolution to high resolution.

9. A computer 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-based low-resolution load data super-resolution reconstruction method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based low-resolution load data super-resolution reconstruction method as described in any one of claims 1-7.

Images