Multi-source heterogeneous data fusion method and device for power distribution network, terminal and medium

Abstract

The invention discloses a power distribution network multi-source heterogeneous data fusion method and device, a terminal and a medium, and relates to the technical field of power system data processing. According to the scheme provided by the invention, a self-attention neural network model integrating a self-attention unit and a neural network unit adopts a PSO algorithm to quickly optimize and initialize model network parameters; the problem of local optimization caused by random initialization in a traditional deep learning method is avoided, the global search capability of the model is improved, then feature extraction is performed through a neural network by using multi-source data of the power distribution network, and feature weights are dynamically distributed through a self-attention mechanism, so that data fusion is more accurate, heterogeneous data errors are effectively reduced, and the method is suitable for large-scale popularization and application. And the fusion precision and stability are improved.

Description

Technical Field

[0001] The present application relates to the technical field of power system data processing, and in particular to a method, device, terminal and medium for fusion of multi-source heterogeneous data of a distribution network. Background Art

[0002] With the rapid development of emerging technologies such as artificial intelligence and the Internet of Things, technical support has been provided for the intelligent upgrade of distribution networks. Based on smart meters, smart switches, smart sensors and other equipment, comprehensive perception, real-time monitoring and intelligent control of distribution networks can be achieved. In addition, real-time collection and analysis of user electricity consumption information can be carried out to provide higher quality power services. At the same time, during the operation and maintenance of intelligent distribution networks, a large amount of monitoring data needs to be obtained to support decision-making analysis. In actual application, since these data usually come from different data sources, their formats and structures may be inconsistent, and there may even be data noise, missing or errors, which leads to technical problems such as low data fusion accuracy and stability in the existing intelligent distribution networks. Summary of the invention

[0003] The present application provides a method, device, terminal and medium for fusion of multi-source heterogeneous data of a distribution network, which are used to solve the technical problems of low data fusion accuracy and stability in the existing intelligent distribution network.

[0004] In order to solve the above technical problems, the first aspect of the present application provides a method for fusion of multi-source heterogeneous data of a distribution network, comprising:

[0005] Initializing a self-attention neural network model, and generating a plurality of particle swarm objects based on model parameters of the self-attention neural network model;

[0006] Based on the particle swarm object, parameter optimization is performed according to the PSO optimization logic to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization;

[0007] Obtaining preset first distribution network multi-source data, and then performing model training on the self-attention neural network model based on the first distribution network multi-source data, and when a preset training termination condition is met, obtaining a self-attention neural network model with optimized model parameters as a distribution network multi-source data fusion model;

[0008] Actual second distribution network multi-source data is acquired to input the second distribution network multi-source data into the distribution network multi-source data fusion model to obtain feature fusion data through the operation of the distribution network multi-source data fusion model.

[0009] Preferably, the performing model training on the self-attention neural network model after the model parameters are optimized based on the first distribution network multi-source data comprises:

[0010] According to the data type of the first power distribution network multi-source data, extracting data features of the power distribution network multi-source data through the neural network unit in the self-attention neural network model;

[0011] Through the self-attention unit in the self-attention neural network model, the data features of each of the multi-source data of the power distribution network are spliced ​​along the channel dimension to obtain an input feature sequence;

[0012] According to the transposed vector of the input feature sequence, combined with the preset self-attention weight calculation formula, the self-attention weight is obtained;

[0013] The input feature sequence is added to the self-attention weight to obtain the output fusion feature, wherein the fusion feature is used to input into the trained power equipment state classification model to obtain the power equipment health state prediction label, so as to adjust the optimization of the particle swarm object through back propagation through the error between the equipment health state prediction label and the true label.

[0014] Preferably, according to the data type of the multi-source data of the power distribution network, extracting data features of the multi-source data of the power distribution network by using a neural network unit in the self-attention neural network model includes:

[0015] If the multi-source data of the distribution network is image data, the image data is processed by three sets of nonlinear mapping according to the two-dimensional convolution through the convolution model logic, and then the tensor obtained after the nonlinear mapping is stretched into a one-dimensional tensor, and then input into the Linear fully connected layer, so as to convert the input one-dimensional tensor into another one-dimensional tensor F containing 256 features through the fully connected layer. x After output;

[0016] If the multi-source data of the distribution network is point cloud data, the point cloud data is processed by three groups of nonlinear mapping according to one-dimensional convolution through convolution logic, and then the tensors obtained after the nonlinear mapping processing are batch standardized and output;

[0017] If the multi-source data of the distribution network is time series monitoring data, the data features of the time series monitoring data in the time domain are extracted through the LSTM model logic.

[0018] Preferably, the self-attention weight calculation formula is specifically:

[0019]

[0020]

[0021]

[0022] In the formula, is the self-attention weight, and represents the initialization weight and initialization bias term, scores is the attention score of the input feature sequence, i and j are the sequence numbers of the input feature sequence, The input feature sequence The transposed sequence of is the attention weight obtained by normalizing the attention score, and exp() is an exponential function.

[0023] Preferably, based on the particle swarm object, performing parameter optimization according to the PSO optimization logic includes:

[0024] Initializing the position parameters and speed parameters of the particle swarm object;

[0025] According to a preset fitness function, the fitness value of each particle swarm object is calculated to optimize the particle swarm object according to the fitness value;

[0026] When the preset update termination condition is not met, the position parameters and speed parameters of the particle swarm object are updated, and then the optimization is performed according to the fitness function. When the preset update termination condition is met, the optimal particle swarm object is determined, so as to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization.

[0027] Preferably, the fitness function is calculated as follows:

[0028]

[0029] In the formula, y i is the true label value of the equipment health status of the i-th sample in the multi-source data of the first distribution network, is a device health status prediction label value based on the i-th sample output, where n represents a training set sample provided for neural network training.

[0030] Preferably, it also includes:

[0031] According to the feature fusion data, the preset power equipment status classification model is input for prediction and classification to obtain the power equipment health status assessment level, and according to the power equipment health status assessment level, combined with the preset distribution power equipment reliability function, the optimal maintenance and overhaul time is determined, wherein the power equipment status classification model is specifically a multi-layer perceptron model obtained by training based on preset power equipment health status samples.

[0032] At the same time, the second aspect of the present application provides a distribution network multi-source heterogeneous data fusion device, including:

[0033] A particle swarm initialization unit, used to initialize the self-attention neural network model, and generate a plurality of particle swarm objects based on the model parameters of the self-attention neural network model;

[0034] A model parameter optimization unit, used to perform parameter optimization based on the particle swarm object according to the PSO optimization logic, so as to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization;

[0035] A fusion model training unit is used to obtain preset first distribution network multi-source data, and then perform model training on the self-attention neural network model based on the first distribution network multi-source data, and when a preset training termination condition is met, a self-attention neural network model with optimized model parameters is obtained as a distribution network multi-source data fusion model;

[0036] The multi-source data fusion unit is used to obtain actual second distribution network multi-source data, so as to input the second distribution network multi-source data into the distribution network multi-source data fusion model, so as to obtain feature fusion data through the operation of the distribution network multi-source data fusion model.

[0037] A third aspect of the present application provides a distribution network multi-source heterogeneous data fusion terminal, including: a memory and a processor;

[0038] The memory is used to store program codes, and the program codes are used to implement a method for fusion of multi-source heterogeneous data of a distribution network as provided in the first aspect of the present application;

[0039] The processor is used for reading and executing the program code.

[0040] The fourth aspect of the present application provides a computer-readable storage medium, in which a program code is stored. The program code is used to be read and executed by a processor to implement a distribution network multi-source heterogeneous data fusion method as provided in the first aspect of the present application.

[0041] It can be seen from the above technical solutions that this application has the following advantages:

[0042] The solution provided in this application uses a PSO algorithm to quickly optimize and initialize the model network parameters through a self-attention neural network model that integrates self-attention units and neural network units, avoiding the local optimal problem caused by random initialization in traditional deep learning methods, and improving the global search capability of the model. Then, multi-source data of the distribution network is used to extract features through a neural network, and feature weights are dynamically allocated through a self-attention mechanism to make data fusion more accurate, effectively reduce heterogeneous data errors, and improve fusion accuracy and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0044] Figure 1 A flow chart of an embodiment of a method for fusion of multi-source heterogeneous data in a distribution network provided in the present application.

[0045] Figure 2 A logical schematic diagram of PSO-based optimization of network parameters in a distribution network multi-source heterogeneous data fusion method provided in this application.

[0046] Figure 3 A schematic diagram of the overall architecture of a self-attention neural network model and a power equipment status classification model in a distribution network multi-source heterogeneous data fusion method provided in this application.

[0047] Figure 4 Schematic diagram of the CPS model framework for intelligent maintenance of electrical equipment.

[0048] Figure 5 A schematic diagram of the structure of an embodiment of a multi-source heterogeneous data fusion device for a distribution network provided in the present application.

[0049] Figure 6 A schematic diagram of the structure of an embodiment of a distribution network multi-source heterogeneous data fusion terminal provided in this application. DETAILED DESCRIPTION

[0050] The embodiments of the present application provide a method, device, terminal and medium for fusion of multi-source heterogeneous data of a distribution network, which are used to solve the technical problems of low data fusion accuracy and stability in the existing intelligent distribution network.

[0051] In order to make the purpose, features, and advantages of the invention of this application more obvious and easy to understand, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described below are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0052] See also Figure 1 , a method for fusion of multi-source heterogeneous data of a distribution network provided in an embodiment of the present application includes:

[0053] Step 101, initializing a self-attention neural network model, and generating a plurality of particle swarm objects based on the model parameters of the self-attention neural network model;

[0054] It should be noted that, based on the initialized self-attention neural network model and the model parameters of the self-attention neural network model, several particle swarm objects are generated. It can be understood that each particle swarm object corresponds to a set of model parameters.

[0055] Step 102: Based on the particle swarm object, parameter optimization is performed according to the PSO optimization logic to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization;

[0056] It should be noted that the PSO algorithm is used to optimize the initial parameters of the neural network to improve the convergence speed and global search capability and prevent falling into the local optimum.

[0057] More specifically, the basic principle of the Particle Swarm Optimization (PSO) algorithm is to initialize a group of random particles and search for the optimal solution to the problem through information sharing and mutual influence between individuals. In each iteration, the particles optimize themselves by comparing two "extreme values", the first one is the optimal solution searched by the particle itself (individual extreme value), and the other is the optimal solution selected after comparison of the entire population (global extreme value). Figure 2 , the specific process is as follows:

[0058] (1) Initialize the weights and thresholds of the CNN-SA neural network, assign a random initial position and velocity to each particle, and initialize the initial position and velocity of each particle within the definition domain:

[0059]

[0060]

[0061] where x i (k) represents the i-th particle in the k-th iteration, Represents x i The nth dimension in (k), v i (k) represents its search direction, specifically the velocity vector of the i-th particle in the k-th iteration, with dimension n.

[0062] (2) Calculate the objective function f of each particle at the kth iteration i(k), calculate the fitness of the particle according to the fitness function, and compare the fitness value of each particle to determine whether it is an individual extreme point and a global optimal extreme point. In PSO-CNN-SA, the fitness function is selected as the mean square error. The smaller the fitness function, the higher the prediction accuracy of the model. The formula is:

[0063]

[0064] Among them, y i is the true label value, representing the true equipment health status label (normal, caution, abnormal, severe) of the i-th sample in the training dataset - the first distribution network multi-source data; is the predicted label value of the equipment health status output based on the i-th sample, representing the predicted value of the equipment health status obtained by the model operating on the i-th sample, and y i The corresponding physical quantity dimensions are consistent, n represents the number of training set samples provided for neural network training, and represents the size of the current training batch or the complete training data set.

[0065] If the current objective function is better than the historical value, the current position is updated to the individual extreme value of the particle, and the particle after the updated position is recorded as , where p i (k) represents the optimal position of the i-th particle in the k-th iteration. The superscript represents the particle number; the subscript represents the component of the particle in the n-dimensional parameter space;

[0066] Compare the objective functions of the optimal positions of all particles and select the optimal position as the global extreme value, that is, , where p g (k) represents the global optimal position of the current population at the kth iteration, the superscript g represents the optimal particle number, and the subscript represents the component of the particle in the n-dimensional parameter space.

[0067] (3) Update the position and velocity information of each particle through the formula. Specifically, the velocity and position of each particle can be updated according to the following formula:

[0068]

[0069]

[0070] Where i=1,2,…,m, j=1,2,…,n, m represents the number of particles in the particle swarm, n represents the parameter dimension contained in each particle, and ω represents the inertia weight, which is used to balance the global search and local search capabilities. and is a non-negative acceleration constant used to control the particle's dependence on the individual optimal solution and the global optimal solution. rand(0,a1) and rand(0,a2) are random numbers with uniform distribution in the range of [0,a1] and [0,a2].

[0071] (4) If the maximum number of iterations or the preset maximum error threshold is not reached, return to step (2). If the maximum number of iterations is met, stop iterating, and take the value of the last iteration as the global optimal solution.

[0072] (5) After the particle swarm optimization algorithm is iterated, the global optimal particle position obtained is the optimal weight and threshold of the CNN-SA neural network. Substitute it into the CNN-SA neural network and continue training until the preset accuracy requirement is reached.

[0073] (6) After training, the optimized CNN-SA neural network model is used for multi-source data feature fusion and state classification.

[0074] (7) Parameter setting of particle swarm optimization algorithm. In PSO, the population size affects the optimization range and convergence speed of the algorithm. If the population size is too large, the search range will increase, which will increase the probability of finding the optimal solution, but may reduce the convergence speed; if the size is too small, the convergence speed will increase, but the search range will be limited, and the global optimal solution may not be obtained. In the PSO-CNN-SA model, the CNN-SA network is regarded as a particle in the population. Considering its complex topological structure and device computing power, the population size is set to 20 to ensure a balance between the search space and the optimization speed. The maximum number of iterations is set to 100, the inertia weight w=0.8, and the maximum particle speed V max =1, learning factor c1=c2=1.2. Based on the existing CNN-SA neural network, a CNN-SA neural network prediction model based on particle swarm optimization is constructed.

[0075] Step 103: obtaining preset first distribution network multi-source data, and then training the self-attention neural network model based on the first distribution network multi-source data. When the preset training termination condition is met, the self-attention neural network model with optimized model parameters is obtained as the distribution network multi-source data fusion model.

[0076] It should be noted that, after step 102, the model parameters of the self-attention neural network model are optimized by using the PSO algorithm, and then, the preset multi-source data of the distribution network is obtained, and the features of the multi-source data of the distribution network are extracted. Then, based on the self-attention neural network model, the extracted data features are used as the input features of the model, and all the features are connected in series through the connection layer in the model, and the self-attention (Self-Attention, SA) unit is used to assign weights of different input features to improve the expression ability of important features. Among them, the overall architecture of the self-attention neural network model mentioned in this embodiment is as follows Figure 3 shown.

[0077] Among them, regarding the feature extraction part, the method provided in this embodiment can also use different feature extraction methods for parallel processing for different types of input data. For example, a convolutional neural network (CNN) is used to extract the spatial features of graph data, while a time series neural network (such as LSTM) is used to process time series data.

[0078] More specifically, the feature extraction of graph data can be further divided into feature extraction of image data and feature extraction of point cloud data. The specific extraction implementation method can refer to the following examples:

[0079] (1) Image feature extraction:

[0080] The image data is used as the input of the image feature module and first undergoes three sets of nonlinear mapping operations:

[0081]

[0082]

[0083]

[0084] Among them, the input X of image feature extraction is the original image tensor, f 3×3 It is a standard two-dimensional convolution operation with a convolution kernel size of 3×3 and a stride of 1. It adds a row and a column of zero-valued pixels around the input tensor. MaxPool is the maximum pooling layer with a pooling kernel size of 2×2 and a stride of 2. It downsamples the feature map, reducing the size and depth of the feature map to half of the original while retaining important features. The three sets of nonlinear mapping operations change the size and number of channels of the feature map. Next, perform a linear mapping operation on it:

[0085]

[0086] Flatten is a flattening layer that stretches the tensor obtained by the nonlinear mapping operation into a one-dimensional tensor so that it can be input into the Linear fully connected layer, and the fully connected layer converts it into another one-dimensional tensor F containing 256 features. x is the final output.

[0087] (2) Point cloud feature extraction:

[0088] The 3D point cloud is used as the input of the point cloud feature extraction module and undergoes three sets of nonlinear mapping operations:

[0089]

[0090]

[0091]

[0092] Among them, the input P of point cloud feature extraction is the point cloud coordinates and reflectivity vector, f 1 It is a standard one-dimensional convolution operation with a convolution kernel size of 1 and a step size of 1. The convolution operation does not change the size of the feature map, and the number of channels changes from 3 to 256. BN is a batch normalization operation that can standardize the input of a batch to make the distribution of the data more stable. ReLU is an activation function that can convert the input value into a non-negative value.

[0093] The image feature extraction module uses two-dimensional convolutional layers and maximum pooling layers to compress the input image layer by layer, extract key features, and map them to the vector space through a fully connected layer, which not only retains the original information but also reduces the feature dimension, improving computational efficiency and generalization capabilities. The point cloud feature extraction module combines the characteristics of three-dimensional point clouds with the advantages of convolutional neural networks, uses three layers of one-dimensional convolution for feature extraction and downsampling, and cooperates with batch normalization and nonlinear activation to improve the point cloud expression ability and enhance the feature extraction effect.

[0094] For time series monitoring data (such as current, voltage, and frequency sequences that change over time), this application uses Long Short-Term Memory (LSTM) neural network for feature extraction. LSTM is an improved recurrent neural network (RNN) structure with a memory gating mechanism that can effectively capture long-term dependency information and avoid the gradient vanishing problem that is prone to occur in traditional RNN training in long sequences.

[0095] Specifically, the time series monitoring data is input into the LSTM unit in units of time steps. The network regulates the flow of information through the input gate, forget gate, and output gate to generate a hidden state vector that reflects the historical change trend. This vector, as the semantic representation vector of the time series data, can be further input into the fusion model for multi-source feature splicing.

[0096] In this embodiment, the input of LSTM is a preprocessed time series data vector with a dimension of (T, D), where T is the number of time steps and D is the feature dimension of each step (such as physical quantities such as voltage and current); the output is a hidden state sequence of (T, H), where H is the hidden layer dimension of LSTM, which is used for subsequent feature fusion processing.

[0097] Regarding the feature fusion part, given an input sequence , the self-attention mechanism generates output through the following steps:

[0098] (1) Calculate the attention score: For each position i, calculate the degree of association with other positions j. This is usually done by calculating a score e ij This score represents the degree of association between position i and position j.

[0099] (2) Calculate the attention weight: Normalize the attention score through the softmax function to obtain the attention weight a ij , and keep the sum of weights of all positions equal to 1.

[0100] (3) Weighted summation: Use the attention weights to weight the information at all positions and obtain the final output representation y i The output representation here is the weighted average of all positions in the input sequence.

[0101] It should be noted that the core process of the self-attention mechanism is to calculate the attention weight through Q (query) and K (key), and then act on V (value) to get the entire weight and output. For the input Q, K and V, the calculation formula for its output vector is:

[0102]

[0103] Where Q, K and V are three matrices, d k is the second dimension of K, divided by d k The process is the Scale operation in the self-attention mechanism structure. The reason for this operation is that for larger d k For example, after completing QK r After that, a very large value will be obtained, which will result in a very small gradient after the softmax operation, which is not conducive to network training.

[0104] Input features of the self-attention mechanism , , where concat represents the concatenation operation of the specified channel dimension, and transpose represents the transposition operation of the specified dimension, which can be obtained by the following calculation:

[0105] , Data features representing n different distribution network multi-source data;

[0106]

[0107] During the forward propagation process, and represents the initialized weights and biases, and during the model training process, the weights and biases can be dynamically updated through the back-propagation algorithm to optimize the prediction performance. scores is the attention score used to indicate the importance of each sequence, which can be obtained by the following calculation:

[0108]

[0109] Apply softmax to normalize the attention scores and obtain the attention weights, which put the elements of the feature vector in the range [0,1] in the specified dimension and sum to 1. The calculation formula is:

[0110]

[0111] Weight soft and Multiply and sum over the specified dimension to get the final self-attention vector. The weighted summed self-attention vector is obtained by the following calculation formula:

[0112]

[0113] Finally, the original feature vector is added to the self-attention weight to obtain the output fusion feature. The process can be expressed as:

[0114]

[0115] It should be noted that the fused features are only the forward propagation results in the self-attention neural network model, and will not be used directly as the final output during the model training phase. During the training process, the fused features output from each round of iteration will be input into the multi-layer perceptron MLP, i.e., the power equipment status classification model, to generate the equipment health status prediction results, and compare them with the true label values ​​to calculate the prediction error. Subsequently, the system uses the backpropagation algorithm to perform gradient calculation and update on the parameters (including weights and bias terms) in the model according to the error function, and continues to iterate the training until the model meets the training termination conditions. The self-attention neural network model with the optimal model parameters can be obtained, and this self-attention neural network model can be used as the multi-source data fusion model for the distribution network.

[0116] The distribution network multi-source data fusion model of this embodiment uses a feature fusion module based on the self-attention mechanism to fuse multiple different types of input features, so that the network can more effectively utilize different types of feature information and better complete multi-source heterogeneous data fusion.

[0117] The self-attention mechanism can adaptively weight the information at different positions of the input sequence, improve the efficiency of feature utilization and network accuracy, filter irrelevant data, and enhance noise resistance. The multi-layer perceptron learns complex features through linear and nonlinear transformations and optimizes feature space mapping to improve prediction capabilities.

[0118] Step 104: Acquire actual second distribution network multi-source data to input the second distribution network multi-source data into a distribution network multi-source data fusion model to obtain feature fusion data through operation of the distribution network multi-source data fusion model.

[0119] By using the distribution network multi-source data fusion model constructed through steps 101 to 103, in the actual application stage, the actual distribution network multi-source data obtained can be input into the distribution network multi-source data fusion model, so as to perform feature extraction, feature fusion and other operations on these data through the distribution network multi-source data fusion model to obtain feature fusion data.

[0120] This embodiment uses the particle swarm optimization (PSO) algorithm to quickly optimize and initialize the network parameters, avoiding the local optimal problem caused by random initialization in traditional deep learning methods and improving the global search capability of the model. Combined with the convolutional neural network (CNN) for feature extraction, and dynamically assigning feature weights through the self-attention mechanism (SA), data fusion is more accurate, effectively reducing the alignment error of heterogeneous data, and improving fusion accuracy and stability. At the same time, by adopting a multi-input parallel feature extraction strategy, it can be specialized for different types of data (such as time series, images, text, etc.), and unified feature expression through the connection layer to eliminate the impact of inconsistent data formats and coordinate systems. The self-attention mechanism further optimizes the data fusion process, enabling the model to adaptively focus on key information, thereby maximizing the utilization efficiency of multi-source data and providing high-quality data support for the operation status perception, fault prediction and maintenance optimization of the distribution network.

[0121] Furthermore, the method provided in this embodiment may also include:

[0122] Step 105: According to the feature fusion data, a preset power equipment status classification model is input for prediction and classification to obtain the power equipment health status assessment level, and according to the power equipment health status assessment level, combined with the preset distribution power equipment reliability function, the optimal maintenance and repair time is determined;

[0123] Among them, the power equipment status classification model is specifically a multi-layer perceptron model trained based on preset power equipment health status samples.

[0124] It should be noted that based on the above-mentioned distribution network multi-source data fusion model, after obtaining the fused features, the fused features can be further sent to the power equipment status classification model based on multi-layer perceptron training for final prediction or classification, thereby realizing the construction of an information control module based on the historical health and fault characteristic information of electrical equipment, performing pattern discrimination on real-time operation data, realizing health status assessment and prediction, and prompting the equipment to perform self-maintenance or issue early warning signals through control instructions to support maintenance decisions.

[0125] More specifically, Figure 4 The figure shows a logical framework for evaluating the health status of electrical equipment based on an information control module. The information control module itself does not directly perform mode discrimination, but acts as a control center to receive the health status evaluation results from the fusion model, and generates control instructions based on the evaluation results and external set rules. During system operation, the operating status data output by the physical dynamic system of the distribution transformer is used as physical quantity input. After being processed by the fusion model, the fusion features are extracted and input into the preset power equipment status classification model. The classification model outputs the equipment health status level, and the information control module determines the equipment maintenance cycle based on this in combination with the preset reliability function, and sends a service request signal when necessary to achieve dynamic regulation and health management of the operating status of the distribution equipment. The solid arrows in the figure represent the flow of physical quantities, and the dotted arrows represent the flow of information quantities.

[0126] More specifically, the reliability function of the power equipment in the distribution network can be solved according to the following formula:

[0127]

[0128] In the formula, is the reliability function of power equipment in distribution network, is the probability density function of power equipment failure, is the failure rate function of power equipment in the distribution network.

[0129] According to the characteristics of Weibull distribution, the expression of the probability density function f(t) of power equipment failure is:

[0130]

[0131]

[0132] Where a is the shape parameter of the Weibull distribution, b is the scale parameter of the Weibull distribution, and t represents the time t.

[0133] In particular, when λ(t) is a constant, we have:

[0134]

[0135] In the formula, is the failure rate of power equipment in the distribution network.

[0136] Thus, the reliability function distribution characteristics corresponding to the failure rate function of the accidental failure period and the wear-out failure period of the distribution power equipment are obtained. According to the reliability function, the average trouble-free working time of the distribution power equipment within a certain period of time can be calculated. The lower the health status (normal, attention, abnormal, severe), the shorter the average trouble-free working time of the equipment, and the lower the corresponding reliability threshold. With this as the threshold constraint condition, the optimal maintenance and repair cycle of the power equipment can be solved. The reliability threshold of the distribution power equipment is calculated by the formula:

[0137]

[0138] Where MTBF0 represents the mean trouble-free working time of the power distribution equipment, that is, the set initial reliability threshold, T0 is the design life of the power equipment, T max is the maximum time interval for preventive maintenance of power equipment, and R0(t) is the initial reliability of power equipment.

[0139] From the last maintenance to a certain time T in the future, the average trouble-free working time of the power distribution equipment during this period is:

[0140]

[0141] Where MTBF represents the reliability threshold of power distribution equipment from the last maintenance to a certain time T in the future, t e k Indicates the equivalent service life of the power equipment after the last maintenance and overhaul, t e Indicates the equivalent service life of the power equipment at the current moment.

[0142] When the average trouble-free working time of the power distribution equipment reaches the threshold, the corresponding time is the best time to perform maintenance and repair on it, let:

[0143]

[0144] By solving the above equation, we can get the optimal maintenance and repair time T of the power distribution equipment: m .

[0145] This embodiment ensures the efficient decision-making ability of the final output through a multi-layer perceptron (MLP) structure, so that the model can quickly respond to the operation needs of the distribution network, realize real-time fusion and intelligent analysis of multi-source heterogeneous data, and is conducive to improving the intelligence level of distribution network monitoring and maintenance.

[0146] The above is a detailed description of an embodiment of a distribution network multi-source heterogeneous data fusion method provided by the present application. The following is a detailed description of an embodiment of a distribution network multi-source heterogeneous data fusion device provided by the present application.

[0147] See also Figure 5 , a distribution network multi-source heterogeneous data fusion device provided in an embodiment of the present application includes:

[0148] A particle swarm initialization unit 201 is used to initialize the self-attention neural network model and generate a plurality of particle swarm objects based on the model parameters of the self-attention neural network model;

[0149] A model parameter optimization unit 202 is used to perform parameter optimization based on a particle swarm object according to a PSO optimization logic, so as to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization;

[0150] A fusion model training unit 203 is used to obtain preset first distribution network multi-source data, and then perform model training on the self-attention neural network model based on the first distribution network multi-source data. When a preset training termination condition is met, a self-attention neural network model with optimized model parameters is obtained as a distribution network multi-source data fusion model;

[0151] The multi-source data fusion unit 204 is used to obtain actual second distribution network multi-source data to input the second distribution network multi-source data into the distribution network multi-source data fusion model to obtain feature fusion data through the operation of the distribution network multi-source data fusion model.

[0152] Furthermore, it also includes:

[0153] The equipment health assessment and maintenance auxiliary unit 205 is used to input the preset power equipment status classification model for prediction and classification based on the feature fusion data to obtain the power equipment health status assessment level, and determine the optimal maintenance and inspection time based on the power equipment health status assessment level and the preset distribution power equipment reliability function.

[0154] In addition, if Figure 6As shown, the present application also provides an embodiment of a distribution network multi-source heterogeneous data fusion terminal, the implementation types of the terminal include but are not limited to: a personal computer, a server and an embedded intelligent device, and the main components of the terminal include: a memory 33 and a processor 31, wherein the memory 33 and the processor 31 can be connected via a communication bus 34;

[0155] The memory 33 is used to store program codes, and the program codes are used to implement a method for fusion of multi-source heterogeneous data of a distribution network as provided in the above embodiment;

[0156] The processor 31 is used to read and execute program codes.

[0157] The present application also provides an embodiment of a computer-readable storage medium, in which a program code is stored. The program code is used to be read and executed by a processor to implement a distribution network multi-source heterogeneous data fusion method as provided in the above embodiment.

[0158] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the terminals, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0159] In the several embodiments provided in the present application, it should be understood that the disclosed terminals, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0160] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0161] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0162] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0163] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0164] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or 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. Various media that can store program codes.

[0165] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for fusion of multi-source heterogeneous data in distribution network, characterized in that: include: Initializing a self-attention neural network model, and generating a plurality of particle swarm objects based on model parameters of the self-attention neural network model; Based on the particle swarm object, parameter optimization is performed according to the PSO optimization logic to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization; Acquire preset first distribution network multi-source data, and then perform model training on the self-attention neural network model based on the first distribution network multi-source data, and when a preset training termination condition is met, obtain a self-attention neural network model with optimized model parameters as a distribution network multi-source data fusion model; Actual second distribution network multi-source data is acquired to input the second distribution network multi-source data into the distribution network multi-source data fusion model to obtain feature fusion data through the operation of the distribution network multi-source data fusion model.

2. A method for fusion of multi-source heterogeneous data in a distribution network according to claim 1, characterized in that: The performing model training on the self-attention neural network model based on the first power distribution network multi-source data includes: According to the data type of the first power distribution network multi-source data, extracting data features of the power distribution network multi-source data through the neural network unit in the self-attention neural network model; Through the self-attention unit in the self-attention neural network model, the data features of each of the multi-source data of the power distribution network are spliced ​​along the channel dimension to obtain an input feature sequence; According to the transposed vector of the input feature sequence, combined with the preset self-attention weight calculation formula, the self-attention weight is obtained; The input feature sequence is added to the self-attention weight to obtain the output fusion feature, wherein the fusion feature is used to input into the trained power equipment state classification model to obtain the power equipment health state prediction label, so as to adjust the optimization of the particle swarm object through back propagation through the error between the equipment health state prediction label and the true label.

3. A method for fusion of multi-source heterogeneous data in distribution network according to claim 2, characterized in that: According to the data type of the multi-source data of the power distribution network, extracting data features of the multi-source data of the power distribution network by using the neural network unit in the self-attention neural network model includes: If the multi-source data of the distribution network is image data, the image data is processed by three sets of nonlinear mapping according to the two-dimensional convolution through the convolution model logic, and then the tensor obtained after the nonlinear mapping is stretched into a one-dimensional tensor, and then input into the Linear fully connected layer, so as to convert the input one-dimensional tensor into another one-dimensional tensor F containing 256 features through the fully connected layer. x After output; If the multi-source data of the distribution network is point cloud data, the point cloud data is processed by three groups of nonlinear mapping according to one-dimensional convolution through convolution logic, and then the tensors obtained after the nonlinear mapping processing are batch standardized and output; If the multi-source data of the distribution network is time series monitoring data, the data features of the time series monitoring data in the time domain are extracted through the LSTM model logic.

4. A method for fusion of multi-source heterogeneous data in a distribution network according to claim 2, characterized in that: The self-attention weight calculation formula is specifically: In the formula, is the self-attention weight, and represents the initialized weights and bias items, scores is the attention score of the input feature sequence, i and j are the sequence numbers of the input feature sequence, The input feature sequence The transposed sequence of is the attention weight obtained by normalizing the attention score, and exp() is an exponential function.

5. A method for fusion of multi-source heterogeneous data in distribution network according to claim 1, characterized in that: Based on the particle swarm object, parameter optimization is performed according to the PSO optimization logic, including: Initializing the position parameters and speed parameters of the particle swarm object; According to a preset fitness function, the fitness value of each particle swarm object is calculated to optimize the particle swarm object according to the fitness value; When the preset update termination condition is not met, the position parameters and speed parameters of the particle swarm object are updated, and then the optimization is performed according to the fitness function. When the preset update termination condition is met, the optimal particle swarm object is determined, so as to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization.

6. A method for fusion of multi-source heterogeneous data in a distribution network according to claim 5, characterized in that: The calculation formula of the fitness function is: In the formula, y i is the true label value of the equipment health status of the i-th sample in the multi-source data of the first distribution network, is a device health status prediction label value based on the i-th sample output, and n represents the number of training set samples provided for neural network training.

7. A method for fusion of multi-source heterogeneous data in a distribution network according to claim 2, characterized in that: Also includes: According to the feature fusion data, a preset power equipment status classification model is input for prediction and classification to determine the power equipment health status assessment level, and according to the power equipment health status assessment level, combined with a preset distribution power equipment reliability function, the optimal maintenance and overhaul time is determined, wherein the power equipment status classification model is specifically a multi-layer perceptron model obtained by training based on preset power equipment health status samples.

8. A device for fusion of multi-source heterogeneous data in a distribution network, characterized in that: include: A particle swarm initialization unit, used to initialize the self-attention neural network model, and generate a plurality of particle swarm objects based on the model parameters of the self-attention neural network model; A model parameter optimization unit, used to perform parameter optimization based on the particle swarm object according to the PSO optimization logic, so as to determine the optimal model parameters of the self-attention neural network model according to the optimal particle swarm object obtained by optimization; A fusion model training unit is used to obtain preset first distribution network multi-source data, and then perform model training on the self-attention neural network model based on the first distribution network multi-source data, and when a preset training termination condition is met, a distribution network multi-source data fusion model is obtained; The multi-source data fusion unit is used to obtain actual second distribution network multi-source data, so as to input the second distribution network multi-source data into the distribution network multi-source data fusion model, so as to obtain feature fusion data through the operation of the distribution network multi-source data fusion model.

9. A distribution network multi-source heterogeneous data fusion terminal, characterized in that: include: Memory and processor; The memory is used to store program codes, and the program codes are used to implement a method for fusion of multi-source heterogeneous data of a distribution network according to any one of claims 1 to 7; The processor is used for reading and executing the program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program code, and the program code is used to be read and executed by a processor to implement a distribution network multi-source heterogeneous data fusion method as described in any one of claims 1 to 7.

Images