Distributed resource portrait construction method based on RMT and deep learning fusion driving

Abstract

The invention relates to a distributed resource portrait construction method based on RMT and deep learning fusion driving, and the method comprises the following steps: obtaining distributed resources from different sources in a power system, carrying out the preprocessing, and constructing an initial distributed resource portrait; based on the initial distributed resource portrait, performing distributed resource classification by adopting fuzzy clustering; based on a distributed resource classification result, key features are extracted by adopting a random matrix theory; and based on the key features, adopting a deep neural network to reconstruct the initial distributed resource portrait to obtain a final distributed resource portrait. Compared with the prior art, the method has the advantages of dynamically describing the state and characteristics of the distributed resources in multiple dimensions and the like.

Description

Technical Field

[0001] The present invention relates to the field of distributed resource management technology, and in particular to a distributed resource portrait construction method based on RMT and deep learning fusion drive. Background Art

[0002] With the gradual increase in the penetration rate of renewable energy (such as wind power, solar energy, etc.) in the power system, the traditional power system faces huge challenges in load forecasting, dispatch optimization, etc. Especially in scenarios such as virtual power plants, extreme weather and atypical working conditions, how to accurately model and analyze user behavior characteristics has become the key to solving power system optimization problems.

[0003] At present, the user modeling method in the power system is still mainly based on traditional statistical models, lacking in-depth mining of dynamic data, especially for distributed resources and user behavior in complex environments, there is still a lack of effective modeling methods. Existing user portrait construction methods are often limited to static data and cannot fully adapt to the complex needs of new power systems for user behavior.

[0004] On the other hand, in the new power system, the diversity, dynamics and interconnectivity of distributed resources require us to not only rely on a single quantitative indicator, but also to conduct accurate portrait analysis through a comprehensive and multi-dimensional feature characterization system. Therefore, it is necessary to propose a multi-dimensional portrait feature characterization system for distributed resources to dynamically describe user behavior and provide solid data support and technical guarantee for the analysis and modeling of distributed resources. Summary of the invention

[0005] The purpose of this invention is to provide a distributed resource portrait construction method based on the fusion of RMT and deep learning, which can dynamically and accurately characterize the status and characteristics of distributed resources.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A distributed resource portrait construction method based on RMT and deep learning fusion drive, including the following steps:

[0008] Obtain distributed resources from different sources in the power system, perform preprocessing, and construct an initial distributed resource profile;

[0009] Based on the initial distributed resource portrait, fuzzy clustering is used to classify the distributed resources;

[0010] Based on the results of distributed resource classification, random matrix theory is used to extract key features;

[0011] Based on the key features, a deep neural network is used to reconstruct the initial distributed resource portrait to obtain the final distributed resource portrait.

[0012] Furthermore, the initial distributed resource portrait includes multiple dimensions, namely: basic load characteristics, regional network structure, flexible resource characteristics, demand response capabilities and non-electrical quantity characteristics.

[0013] Furthermore, the basic load characteristics include average load, peak-to-valley difference / peak-to-valley ratio, maximum / low load, and important load; the regional network structure includes load rate, limit margin, transfer capacity, and stability margin; the flexible resource characteristics include air-conditioning load proportion, electric vehicle load proportion, storage capacity, controllable load proportion, and flexible resource feasible domain; the demand response capability includes demand response potential, response time, load regulation capability, and load peak shaving and valley filling ratio; the non-electrical quantity characteristics include seasonal fluctuation factors, meteorological factors, ambient temperature factors, and traffic factors.

[0014] Furthermore, the fuzzy clustering is fuzzy C-means clustering, and the step of classifying distributed resources includes:

[0015] 1) Randomly initialize the membership matrix U = [u ij ],satisfy where u ij For distributed resources x i Belongs to cluster c j The membership degree of , C is the number of clusters;

[0016] 2) Calculate the cluster center according to the membership matrix:

[0017]

[0018] In the formula, c j is the cluster center, m>1 is the fuzzy coefficient, which is used to control the fuzziness of clustering, and N is the number of distributed resources;

[0019] 3) Update the membership matrix:

[0020]

[0021] In the formula, c k is the cluster center vector, that is, the center point of the kth cluster;

[0022] 4) Calculate the objective function and determine whether the convergence condition is met. If so, stop the iteration and output the clustering result, that is, the distributed resource classification result. If not, repeat steps 2)-4) until the convergence condition is met, wherein the convergence condition is:

[0023] ||U (t+1) -U(t) ||<∈

[0024] In the formula, ∈ is the set threshold.

[0025] Furthermore, the objective function of the fuzzy C-means clustering is:

[0026]

[0027] In the formula, J m is the target value.

[0028] Furthermore, the step of extracting key features includes:

[0029] Based on the distributed resource classification results, a coupling analysis matrix and a reference matrix are constructed using random matrix theory;

[0030] Based on a preset rolling time window, at each sampling moment, respective window data are obtained from the coupling analysis matrix and the reference matrix;

[0031] Based on the respective window data, the characteristic value and the LES statistic are calculated, and the LES difference is calculated, and the key feature is determined according to the LES difference. The calculation expression of the LES difference is:

[0032]

[0033] in:

[0034]

[0035] In the formula, Δ LES (t) is the LES difference, are the LES statistics of the coupling analysis matrix and the reference matrix, is the detection function, λ i is the eigenvalue of the matrix, N is the number of eigenvalues, C is the coupling analysis matrix, C r is the reference matrix.

[0036] Furthermore, the step of constructing the coupling analysis matrix and the reference matrix includes:

[0037] Based on the distributed resource classification result, a main measurement and an augmented measurement are selected, and further constructed into a basic measurement matrix and an augmented measurement matrix;

[0038] The augmented measurement matrix is ​​dimensionally expanded, random noise is introduced, an extended augmented matrix is ​​formed, and a signal-to-noise ratio of the extended augmented matrix is ​​set, wherein the expression of the extended augmented matrix is:

[0039]

[0040] Where M e To expand the augmented matrix, is the augmented measurement matrix after dimensionality increase, with dimension (k×N a )×T, where [·] indicates the floor operator, N a is the number of measurements in the augmented part, M b is the basic measurement matrix, M a is the augmented measurement matrix, α is the noise amplitude, and N is the dimension (k×N a )×T white noise matrix, T is the number of sampling times;

[0041] The expression of the signal-to-noise ratio of the extended augmented matrix is:

[0042]

[0043] In the formula, ρ(M e ) is the expanded augmented matrix M e The signal-to-noise ratio of , Tr(·) is the trace of the matrix;

[0044] The basic measurement matrix and the extended augmented matrix are combined to construct a coupling analysis matrix, and the expression of the coupling analysis matrix is:

[0045]

[0046] Where, C is the coupling analysis matrix;

[0047] The basic measurement matrix and the white noise matrix are combined to construct a reference matrix, and the expression of the reference matrix is:

[0048]

[0049] In the formula, C r is the reference matrix.

[0050] Furthermore, the step of extracting key features includes:

[0051] Before calculating the eigenvalues ​​and LES statistics, the respective window data are normalized.

[0052] Furthermore, in the process of obtaining the final distributed resource portrait, a deep neural network is used to perform nonlinear mapping between the key features and the initial distributed resource portrait to complete the portrait reconstruction process and obtain the final distributed resource portrait.

[0053] Furthermore, the deep neural network includes an input layer, a hidden layer and an output layer, and the training steps of the deep neural network include:

[0054] 1) Randomly initialize parameters;

[0055] 2) Through the forward propagation method, the key features are input into the input layer, and combined with the hidden layer, layer by layer calculation is performed until the output layer, and the prediction result of the output layer is obtained. The forward propagation process is:

[0056] h l =f(W l h l-1 +b l ),l=1,2,…,L,

[0057] In the formula, h l is the output activity of the lth layer, f is the activation function, W l and b l are the weight matrix and bias vector of the lth layer respectively;

[0058] The prediction result of the output layer is:

[0059]

[0060] In the formula, The prediction result is the output distributed resource portrait.

[0061] 3) Calculate the loss function:

[0062]

[0063] In the formula, L is the loss value, M is the number of samples, and y i is the actual output, is the predicted output;

[0064] 4) Use the back propagation algorithm combined with gradient descent to update the weight matrix and bias vector;

[0065] 5) Repeat steps 2)-4) until the end condition is reached and the training process is completed.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] (1) Aiming at the high complexity of multi-dimensional data of distributed resources, the present invention adopts a fusion method of random matrix theory and deep learning to efficiently extract key features from multi-source features of different distributed resources and construct accurate portraits. Specifically, the present invention goes through three stages: resource classification, key feature extraction and portrait reconstruction. The fuzzy clustering method is used to identify resource categories, the random matrix theory is used to extract key features, and the deep neural network is used to reconstruct resource portraits. A closed-loop analysis framework from data to features and from features to models is formed, which can dynamically and accurately characterize the status and characteristics of distributed resources.

[0068] (2) The present invention systematically characterizes distributed resources from multiple dimensions, such as load characteristics, regional networks, flexible resources, and demand response, providing a comprehensive perspective on the operating status, response capabilities, and environmental impact of resources.

[0069] (3) By combining deep learning with random matrix theory, the present invention can not only dynamically characterize user characteristics and adapt to complex operating conditions in new power systems, such as virtual power plants, extreme weather, and user aggregators, but also improve prediction accuracy and adaptability, which is of great significance to the intelligent scheduling and optimization of new power systems.

[0070] (4) Distributed resources in the power system are represented as user characteristics, that is, the distributed resource portrait is the user portrait, that is, a comprehensive user portrait is constructed based on deep learning, which can more accurately reflect user behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0072] Figure 2 It is a multi-dimensional image feature characterization system of the present invention;

[0073] Figure 3 It is the resource classification result based on fuzzy clustering of the present invention;

[0074] Figure 4 This is the coupling analysis method flow of the present invention. DETAILED DESCRIPTION

[0075] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0076] This embodiment provides a distributed resource portrait construction method based on the fusion of RMT and deep learning to model user behavior in future power systems with renewable energy penetration. This method constructs a multi-dimensional portrait feature characterization system for distributed resources and combines modern technical means to propose an efficient feature selection and portrait reconstruction method. First, by systematically characterizing the characteristics of distributed resources from multiple dimensions such as load characteristics, regional networks, flexible resources, and demand response, a comprehensive perspective is provided for the operating status, response capabilities, and environmental impact of resources. Secondly, in view of the high complexity of multi-dimensional data of distributed resources, a feature selection framework that integrates random matrix theory and deep learning is introduced, and resource classification is performed using the fuzzy C-means clustering algorithm. The core features are extracted through coupling analysis of random matrices, and resource portraits are reconstructed with the help of deep neural networks. Specifically, as Figure 1As shown, the method comprises the following steps:

[0077] S1. Build an initial multi-dimensional distributed resource portrait.

[0078] In order to meet the needs of holographic characterization of distributed resource portraits, it is necessary to build a multi-dimensional, multi-level, and multi-feature portrait feature characterization system. This system is developed around five core dimensions: basic load characteristics, regional network structure, flexible resource characteristics, demand response capabilities, and non-electrical quantity characteristics. It comprehensively characterizes the operating status, response capabilities, and environmental characteristics of distributed resources, laying a solid data foundation for subsequent data-driven modeling and optimization. The specific framework diagram is as follows: Figure 2 The framework diagram includes the following dimensions:

[0079] (1) Basic characteristics of load

[0080] Load characteristics are the basic dimension of distributed resource portraits, which mainly reflect the power demand characteristics and changing patterns of distributed resources. Characteristic characterization includes indicators such as average load, peak-to-valley difference / peak-to-valley ratio, and minimum / low load. Among them, the average load describes the average power demand of resources within a certain period of time, reflecting its normalized operation level; the peak-to-valley difference / peak-to-valley ratio describes the amplitude and fluctuation characteristics of load fluctuations, which is an important indicator for measuring load stability; the minimum / low load characterizes the operating characteristics of resources under low demand conditions, which is suitable for identifying the minimum operating capacity of resources. By characterizing the load characteristics, the regularity and volatility of distributed resource loads can be revealed, providing an important reference for system operation optimization.

[0081] (2) Regional network structure

[0082] The operation of distributed resources is inseparable from the support of the network structure, and the characteristics of the regional network structure are an important part of the portrait system. The characteristic characterization includes indicators such as the limit load rate and the stability margin. The limit load rate reflects the load pressure of the network operation. The higher the value, the greater the risk of network operation; the stability margin describes the system's carrying capacity and recovery ability in the case of load fluctuations or failures, reflecting the resilience and reliability of the network. By analyzing the characteristics of the regional network structure, the carrying status and dynamic characteristics of distributed resources in the network can be revealed, providing data support for network planning and optimization.

[0083] (3) Flexible resource characteristics

[0084] Flexible resource characteristics are an important dimension of distributed resource portraits, mainly describing the adjustability and dynamic response capabilities of resources. The characteristic characterization includes indicators such as the proportion of electric vehicle load and the proportion of air-conditioning load. The proportion of electric vehicle load reflects the proportion of electric vehicles as flexible loads in the overall load, reflecting its potential for system regulation; the proportion of air-conditioning load describes the proportion of air-conditioning load and its contribution to load fluctuations. The characterization of these characteristics can provide important data basis for the demand-side management of flexible resources.

[0085] (4) Demand responsiveness

[0086] Demand response capability is an important capability for distributed resources to participate in power system regulation, which mainly includes indicators such as response rate, response depth, response time, and response smoothness. The response rate represents the degree of resource execution of demand response instructions; the response depth describes the amplitude of power adjustment during the response process; the response time reflects the time that the resource continues to provide services during the response process; and the response smoothness evaluates the continuity and stability of power adjustment during the response process. The characterization of these characteristics can quantify the demand response capability of resources and provide a reliable basis for power market transactions and ancillary services.

[0087] (5) Non-electrical quantity characteristics

[0088] Non-electrical quantity characteristics are a supplementary dimension of distributed resource portraits, reflecting the environmental factors affecting resource operation and external constraints. The characteristic characterization includes indicators such as traffic factors and environmental factors. Traffic factors reflect the impact of traffic flow, road conditions, etc. on the load of resources such as electric vehicles; environmental factors describe the impact of temperature, humidity, weather conditions, etc. on the operating status of resources. Through the integrated analysis of non-electrical quantity characteristics, we can deeply understand the impact of the external environment on distributed resources and provide more comprehensive information support for resource coordination and optimization.

[0089] S2. Construct a feature selection system driven by the fusion of random matrix theory and deep learning.

[0090] This embodiment is based on the initial multi-dimensional distributed resource portrait constructed in the S1 stage, and collects multi-dimensional data such as the user's basic load characteristics, regional network structure, flexible resource characteristics, demand response capabilities, and non-electrical quantity characteristics in real time through smart meters, sensors and other equipment. The multi-dimensional portrait features of distributed resources face great processing difficulties due to their high data dimensions, complex characteristics, and strong correlation. In order to efficiently extract the key features of different distributed resource multi-source features and construct accurate portraits, this embodiment proposes a feature selection system driven by the fusion of random matrix theory and deep learning. The system covers three stages: resource classification, key feature extraction, and portrait reconstruction. It identifies resource categories through fuzzy clustering methods, extracts core features using random matrix theory, and reconstructs resource portraits in combination with deep neural networks, forming a closed-loop analysis framework from data to features and from features to models. The specific methods, formulas, and implementation details of each stage are described in detail below.

[0091] S2.1. Resource classification based on fuzzy clustering.

[0092] In the feature selection of distributed resources, resource classification is a basic link, and its purpose is to divide resources into multiple categories according to their characteristics. In order to better deal with the problem of fuzzy boundaries and multi-attribute intersection of distributed resource characteristics, this embodiment adopts the fuzzy C-means clustering (FCM) method. FCM is a soft clustering algorithm that allows samples to belong to multiple categories at the same time and quantifies this affiliation through a membership matrix.

[0093] (1) Principle of Fuzzy C-means Clustering Algorithm

[0094] The goal of the FCM algorithm is to minimize the following objective function:

[0095]

[0096] Where: x i is the characteristic vector activity of the i-th resource; c j is the jth cluster center; u ij For resource x i Belongs to cluster c j The membership degree satisfies m>1 is the fuzzy coefficient, which is used to control the fuzziness of clustering; ||x i -c j || 2 is the Euclidean distance from the resource to the cluster center.

[0097] (2) Fuzzy C-means clustering algorithm steps

[0098] 1) Initialize the membership matrix U = [u ij ], usually randomly generated, satisfying

[0099] 2) Calculate the cluster center based on the membership matrix:

[0100]

[0101] 3) Update the membership matrix:

[0102]

[0103] 4) Repeat steps 2) and 3) until the convergence condition ||U is met (t+1) -U (t) ||<∈.

[0104] Through the above process, fuzzy C-means clustering can identify the categories of resources and their fuzzy affiliations, providing a grouping basis for subsequent feature extraction.

[0105] This step finally obtains the resource classification result based on fuzzy clustering as follows Figure 3 shown.

[0106] S2.2. Key feature extraction of distributed resources based on random matrix.

[0107] After resource classification is completed, the goal of feature extraction is to remove redundant information and noise from high-dimensional feature data and retain the core features that are decisive for resource status. This embodiment extracts key features based on random matrix theory (RMT). The linear eigenvalue statistic (LES) is constructed by setting a fixed detection function To extract and evaluate features of high-dimensional spatial data, the general expression of LES is shown as follows:

[0108]

[0109] In the formula, λ i is the eigenvalue of a random matrix.

[0110] According to the law of large numbers, It converges to the following formula according to probability:

[0111]

[0112] Where ρ(λ) is the probability density function of the eigenvalue of the random matrix.

[0113] From the above formula, we can find that the key to designing the LES index is to select a suitable detection function Compared with traditional low-dimensional indicators that characterize the coupling between data, such as the Pearson correlation coefficient and the Spearman correlation coefficient, the LES indicator has better analytical robustness for discontinuous data with outliers. Although this article has designed a standardized cleaning method suitable for power grid data in the previous section, it is impossible to directly remove the abnormal data values ​​in the sampling, so the coupling calculation indicator needs to have better robustness. In addition, the LES indicator can directly process high-dimensional matrix inputs and can better grasp the nonlinear correlation in the data than low-dimensional indicators.

[0114] In summary, this step will establish a correlation analysis method for the multi-source heterogeneous characteristics of distributed resources based on random matrix theory. This method uses the distributed resource operation data as the basic part of the coupling analysis matrix, the coupling measurement as the augmented part, and compares the LES index of the analysis matrix with and without coupling measurement to reflect the impact of coupling measurement on basic measurement.

[0115] First, some distributed resources in the power grid are used as the main measurement for this coupling analysis, and their operating data are used as the basic part of the coupling analysis matrix. Then, a certain characteristic measurement is used as the augmented part, and together they constitute the coupling analysis matrix as the data source for subsequent analysis. However, in order to ensure the accuracy of subsequent coupling analysis in general, the augmented part can only contain one-dimensional measurements or a type of measurement with similar physical meaning, such as using only ambient temperature or using both the maximum temperature and the minimum temperature. Taking the above differences into account, when cleaning data, it is possible to consider using the augmented part measurement as the main cleaning measurement to simplify the data normalization operation during coupling analysis.

[0116] Assume that the basic part of the coupling analysis matrix contains N b The augmented part includes N a If the number of measurements after normalization is T, the basic data dimension is N. b ×T, hereinafter referred to as the basic measurement matrix M b , similarly, the dimension of the augmented data is N a ×T, referred to as the augmented measurement matrix M a .

[0117] Based on the coupling analysis matrix measurement selection principle mentioned above, in general, the number of basic measurements is greater than the number of augmented measurements, that is, N b >N a In order to improve the accuracy of coupling analysis, the dimension of the augmented measurement matrix is ​​expanded. First, the dimension of the augmented measurement matrix is ​​increased, as shown in the following formula.

[0118]

[0119] In the formula, The dimension is (k×N a )×T, where [·] indicates the rounding down operator. In order to prevent repeated data from interfering with subsequent coupling analysis, it is necessary to Introducing random noise.

[0120]

[0121] Where M e is the expanded matrix; N is the dimension (k×N a )×T white noise matrix, the noise distribution is standard normal distribution; α is the noise amplitude, which has a direct impact on the result of coupling analysis. If the noise amplitude is too small, the repeated data used in the augmented measurement dimension will still affect the coupling analysis. On the contrary, excessive noise will distort the measurement in the expanded augmented matrix and turn it into an abnormal measurement.

[0122] In order to quantitatively evaluate and better select the amplitude of white noise, the signal-to-noise ratio of the expanded augmented matrix is ​​defined as:

[0123]

[0124] In the formula, Tr(·) is the trace of the matrix. At the same time, in order to ensure the accuracy and consistency of the coupling analysis, the signal-to-noise ratio should be kept uniform throughout the coupling analysis process.

[0125] The basic measurement matrix and the extended augmented matrix are combined to construct a coupling analysis matrix as the data source for analysis.

[0126]

[0127] As a comparison, the reference matrix for coupling analysis is constructed using the basic measurement matrix and the white noise matrix, which can reduce the interference caused by white noise to a certain extent:

[0128]

[0129] After completing the construction of the coupling analysis matrix and the reference matrix, the LES statistics of the two matrices under each window can be calculated based on the rolling time window. By comparing the difference between the LES of the two, the coupling relationship between the augmented measurement and the main measurement can be obtained. The calculation formula of the LES difference is shown as follows:

[0130]

[0131] In the formula, are the LES statistics of the coupling analysis matrix and the reference matrix, respectively.

[0132] In summary, the coupling analysis method based on random matrix designed in this embodiment has the following steps: Figure 4 As shown in Figure 2, the potential correlation structure between resources is revealed, redundant information is removed, and only the most representative features for reconstructing resource behavior profiles are retained.

[0133] S2.3. Image reconstruction based on deep neural network.

[0134] After the key features are extracted, it is necessary to transform the extracted features into a visual resource portrait through a portrait reconstruction method to depict the state and characteristics of the resource. This embodiment uses a deep neural network (DNN) for portrait reconstruction.

[0135] (1) Structure and formula of deep neural network

[0136] The deep neural network consists of an input layer, a hidden layer, and an output layer. Assume that the input feature is The output image is The forward propagation process of DNN can be expressed as:

[0137] h l =f(W l h l-1 +b l ),l=1,2,…,L,

[0138] Where: h l is the output activity of layer l; W l and b l are the weight matrix and bias vector of the lth layer respectively; f is the activation function (such as ReLU, Sigmoid, etc.).

[0139] The prediction result of the output layer is:

[0140]

[0141] (2) Loss Function and Optimization

[0142] The goal of model training is to minimize the predicted output The difference between the image y and the real image y. The commonly used loss function is the mean square error (MSE):

[0143]

[0144] Where M is the sample size.

[0145] The weight matrix and bias vector are updated using the back-propagation algorithm combined with gradient descent or Adam optimizer.

[0146] (3) Model training and image generation

[0147] 1) Randomly initialize parameters;

[0148] 2) Through the forward propagation method, the key features are input into the input layer, and combined with the hidden layer, layer by layer calculation is performed until the output layer, and the prediction result of the output layer is obtained;

[0149] 3) Calculate the loss function;

[0150] 4) Use the back propagation algorithm combined with gradient descent to update the weight matrix and bias vector;

[0151] 5) Repeat steps 2)-4) until the end condition is reached and the training process is completed.

[0152] Through training on large-scale distributed resource data sets, deep neural networks can learn the nonlinear mapping relationship between key features and portraits. After training, new key features can be input to generate corresponding resource portraits to fully characterize resource status.

[0153] The method of the present invention integrates multi-dimensional data such as historical electricity consumption data, environmental variables, and socio-economic factors, and combines deep learning with probabilistic statistical models to dynamically characterize user characteristics and adapt to complex working conditions in new power systems, such as virtual power plants, extreme weather, and user aggregators. Its core content includes:

[0154] 1. A multi-dimensional distributed resource portrait construction method that combines different data sources to build dynamic portraits.

[0155] In the power system, distributed resources refer to the user's usage characteristics of the power system, that is, the distributed resource portrait is the user portrait.

[0156] 2. The user behavior prediction model based on deep learning provides accurate user behavior prediction under complex working conditions. It can be dynamically adjusted according to factors such as environmental changes and user behavior changes to improve the adaptability of the prediction model.

[0157] The present invention integrates random matrix theory with deep learning through three stages: resource classification, key feature extraction and portrait reconstruction. The deep neural network outputs the prediction results, adjusts the user portrait in time, and provides data support for power system optimization.

[0158] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0159] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The schemes in the embodiments of the present invention may be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0160] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0161] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0162] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0163] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0164] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A distributed resource portrait construction method based on RMT and deep learning fusion drive, characterized in that: The following steps are involved: Obtain distributed resources from different sources in the power system, perform preprocessing, and construct an initial distributed resource profile; Based on the initial distributed resource portrait, fuzzy clustering is used to classify the distributed resources; Based on the results of distributed resource classification, random matrix theory is used to extract key features; Based on the key features, a deep neural network is used to reconstruct the initial distributed resource portrait to obtain the final distributed resource portrait.

2. According to claim 1, a distributed resource portrait construction method based on RMT and deep learning fusion drive is characterized in that: The initial distributed resource portrait includes multiple dimensions, namely: basic load characteristics, regional network structure, flexible resource characteristics, demand response capabilities and non-electrical quantity characteristics.

3. According to claim 2, a distributed resource portrait construction method based on RMT and deep learning fusion drive is characterized in that: The basic load characteristics include average load, peak-to-valley difference / peak-to-valley ratio, maximum / low load, and important load. The regional network structure includes load rate, limit margin, transfer capacity, and stability margin. The flexible resource characteristics include the proportion of air-conditioning load, the proportion of electric vehicle load, storage capacity, the proportion of controllable load, and the feasible domain of flexible resources. The demand response capability includes demand response potential, response time, load regulation capability, and load peak shaving and valley filling ratio. The non-electrical quantity characteristics include seasonal fluctuation factors, meteorological factors, ambient temperature factors, and traffic factors.

4. According to claim 1, a distributed resource portrait construction method based on RMT and deep learning fusion drive, characterized in that: The fuzzy clustering is fuzzy C-means clustering, and the steps of classifying distributed resources include: 1) Randomly initialize the membership matrix U = [u ij ],satisfy where u ij For distributed resources x i Belongs to cluster c j The membership degree of , C is the number of clusters; 2) Calculate the cluster center according to the membership matrix: In the formula, c j is the cluster center, m>1 is the fuzzy coefficient, which is used to control the fuzziness of clustering, and N is the number of distributed resources; 3) Update the membership matrix: In the formula, c k is the cluster center vector, that is, the center point of the kth cluster; 4) Calculate the objective function and determine whether the convergence condition is met. If so, stop the iteration and output the clustering result, that is, the distributed resource classification result. If not, repeat steps 2)-4) until the convergence condition is met, wherein the convergence condition is: ∥U (t+1) -IN (t) ∥<∈ In the formula, ∈ is the set threshold.

5. According to claim 4, a distributed resource portrait construction method based on RMT and deep learning fusion drive is characterized in that: The objective function of the fuzzy C-means clustering is: In the formula, J m is the target value.

6. According to claim 1, a distributed resource portrait construction method based on RMT and deep learning fusion drive is characterized in that: The step of extracting key features comprises: Based on the distributed resource classification results, a coupling analysis matrix and a reference matrix are constructed using random matrix theory; Based on a preset rolling time window, at each sampling moment, respective window data are obtained from the coupling analysis matrix and the reference matrix; Based on the respective window data, the characteristic value and the LES statistic are calculated, and the LES difference is calculated, and the key feature is determined according to the LES difference. The calculation expression of the LES difference is: in: In the formula, Δ LES (t) is the LES difference, are the LES statistics of the coupling analysis matrix and the reference matrix, is the detection function, λ i is the eigenvalue of the matrix, N is the number of eigenvalues, C is the coupling analysis matrix, C r is the reference matrix.

7. According to claim 6, a distributed resource portrait construction method based on RMT and deep learning fusion drive is characterized in that: The steps of constructing the coupling analysis matrix and the reference matrix include: Based on the distributed resource classification result, a main measurement and an augmented measurement are selected, and further constructed into a basic measurement matrix and an augmented measurement matrix; The augmented measurement matrix is ​​dimensionally expanded, random noise is introduced, an extended augmented matrix is ​​formed, and a signal-to-noise ratio of the extended augmented matrix is ​​set, wherein the expression of the extended augmented matrix is: Where M e To expand the augmented matrix, is the augmented measurement matrix after dimensionality increase, with dimension (k×N a )×T, where [·] represents the floor operator, N a is the number of measurements in the augmented part, M b is the basic measurement matrix, M a is the augmented measurement matrix, α is the noise amplitude, and N is the dimension (k×N a )×T white noise matrix, T is the number of sampling times; The expression of the signal-to-noise ratio of the extended augmented matrix is: In the formula, ρ(M e ) is the expanded augmented matrix M e The signal-to-noise ratio of , Tr(·) is the trace of the matrix; The basic measurement matrix and the extended augmented matrix are combined to construct a coupling analysis matrix, and the expression of the coupling analysis matrix is: Where, C is the coupling analysis matrix; The basic measurement matrix and the white noise matrix are combined to construct a reference matrix, and the expression of the reference matrix is: In the formula, C r is the reference matrix.

8. According to claim 6, a distributed resource portrait construction method based on RMT and deep learning fusion drive is characterized in that: The step of extracting key features comprises: Before calculating the eigenvalues ​​and LES statistics, the respective window data are normalized.

9. According to claim 1, a distributed resource portrait construction method based on RMT and deep learning fusion drive, characterized in that: In the process of obtaining the final distributed resource portrait, a deep neural network is used to perform nonlinear mapping between the key features and the initial distributed resource portrait to complete the portrait reconstruction process and obtain the final distributed resource portrait.

10. A distributed resource portrait construction method based on RMT and deep learning fusion drive according to claim 9, characterized in that: The deep neural network includes an input layer, a hidden layer and an output layer, and the training steps of the deep neural network include: 1) Randomly initialize parameters; 2) Through the forward propagation method, the key features are input into the input layer, and combined with the hidden layer, layer by layer calculation is performed until the output layer, and the prediction result of the output layer is obtained. The forward propagation process is: h l =f(W l h l-1 +b l ),l=1,2,…,L, In the formula, h l is the output activity of the lth layer, f is the activation function, W l and b l are the weight matrix and bias vector of the lth layer respectively; The prediction result of the output layer is: In the formula, The prediction result is the output distributed resource portrait. 3) Calculate the loss function: In the formula, L is the loss value, M is the number of samples, and y i is the actual output, is the predicted output; 4) Use the back propagation algorithm combined with gradient descent to update the weight matrix and bias vector; 5) Repeat steps 2)-4) until the end condition is reached and the training process is completed.

Images