Adaptive resampling strength distribution transformer structure response reconstruction method and adaptive resampling strength distribution transformer structure response reconstruction system

Abstract

The invention discloses a distribution transformer structure response reconstruction method and system of self-adaptive resampling strength, and relates to the technical field of power system monitoring and optimization, and the method comprises the steps: collecting the data of a distribution transformer, and carrying out the preprocessing; performing conjoint analysis on the power data and the non-power data, extracting time sequence features and static features, outputting advanced fusion features, inputting the advanced fusion features into a deep neural network for prediction, and dynamically adjusting model parameters and structures; and outputting a heavy overload state prediction result by adopting a game theory optimization strategy. According to the method, multi-dimensional information can be comprehensively analyzed, and the utilization efficiency and the feature expression ability of data are improved. And feature extraction and adaptive adjustment are performed on the data by using a deep learning model, so that the model can dynamically adapt to different load characteristics and environment changes, and the accuracy and robustness of prediction are improved. Through a game theory optimization strategy, the structure and parameters of the model are dynamically optimized, and the optimization of the model performance under different operation conditions is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring and optimization, and in particular to a method and system for reconstructing distribution transformer structure response with adaptive resampling intensity. Background Art

[0002] With the continuous growth of electricity demand and the increasing scale of power grid, the safe operation of power system faces unprecedented challenges. As an important part of power system, the operation status of distribution network is directly related to the stability and power supply reliability of the whole power grid. In the distribution network, distribution transformer is one of the most critical equipment, which is responsible for converting electricity from high-voltage network to low-voltage electricity suitable for users. Therefore, the health status of distribution transformer is crucial to the normal operation of power grid.

[0003] However, with the change of seasonal load and the growing demand for electricity, distribution transformers often face heavy overload during operation. Heavy overload will not only cause overheating and insulation aging of the transformer, but also increase the failure rate of the equipment and even cause serious accidents such as fire. At present, the solutions to the problem of heavy overload of distribution transformers mainly include equipment modification, load adjustment and condition monitoring. However, traditional condition monitoring methods usually rely on static data analysis and on-site inspections, which is not only inefficient, but also unable to timely and accurately reflect the actual operating status of the distribution transformer.

[0004] In recent years, with the continuous improvement of data acquisition technology and computing power, intelligent monitoring methods based on big data analysis and machine learning have gradually become a research hotspot. In particular, the rapid development of deep learning technology has provided new ideas and methods for state prediction and fault diagnosis of power systems. By learning and modeling massive historical data, deep learning models can effectively mine the potential laws in the data, thereby achieving accurate prediction of the state of distribution transformers and fault warning. However, existing deep learning methods still have many problems when dealing with heavy overload state monitoring of distribution transformers. For example, how to efficiently integrate power data and non-power data, how to dynamically adjust the model structure to adapt to different load characteristics, and how to optimize model parameters to improve prediction accuracy and robustness.

[0005] In addition, most existing model optimization strategies use fixed parameter settings and lack the ability to adaptively adjust to different data characteristics, making it difficult to provide stable and reliable prediction results in a complex and changeable power environment. Therefore, it is particularly important to develop a technology that can adaptively adjust the model structure and parameters, effectively integrate multiple data types, and accurately predict the heavy overload state of distribution transformers. Summary of the invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by the present invention is: how to build a heavy overload state monitoring system for power distribution transformers that can dynamically adjust the model structure and parameters based on multimodal data fusion technology and deep learning models, so as to improve the accuracy and real-time performance of the prediction and optimize the operation and maintenance of the power system.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity, which comprises the following steps:

[0009] Collect distribution transformer data through sensors and perform preprocessing;

[0010] Based on the preprocessing results, multimodal data fusion technology is used to jointly analyze power data and non-power data, extract time series features and static features, evaluate and output advanced fusion features;

[0011] Use high-level fusion features as input to deep neural networks for prediction and dynamically adjust model parameters and structures;

[0012] A game theory optimization strategy is adopted within the DNN model to dynamically adjust the weights of different layers of the model and output the optimal heavy overload state prediction results.

[0013] As a preferred solution of the adaptive resampling intensity distribution transformer structural response reconstruction method described in the present invention, wherein: the distribution transformer data includes operation data and environmental data;

[0014] The preprocessing is to use a filtering algorithm to denoise the collected distribution transformer operation data and environmental data, normalize the denoised data, convert them to the same scale, and detect and eliminate abnormal values ​​through statistical analysis methods.

[0015] As a preferred solution of the adaptive resampling intensity distribution transformer structure response reconstruction method described in the present invention, wherein: the joint analysis is to use an improved dynamic time warping algorithm combined with fast Fourier transform to perform time alignment of data, input power data and non-power data into an adaptive convolutional neural network respectively, design an adaptive convolution kernel mechanism in the convolution layer, dynamically adjust the size and shape of the convolution kernel according to the local time-frequency characteristics of the input data, and extract the time series characteristics of power data and the spatial characteristics of non-power data respectively through a multi-level feature extraction strategy;

[0016] A dual attention mechanism is introduced in the fusion layer, and the fusion ratio of power and environmental features is dynamically adjusted through the feature weighted combination method to generate high-level fusion features.

[0017] As a preferred solution of the adaptive resampling intensity distribution transformer structure response reconstruction method described in the present invention, the improved dynamic time warping algorithm introduces weights into the traditional dynamic time warping algorithm, and combines the characteristics of time domain and frequency domain to align data; the power data and non-power data are respectively subjected to fast Fourier transform to obtain their respective frequency domain representations, calculate the difference between the time domain and frequency domain characteristics, and construct a weighted cumulative distance matrix, which is expressed as follows:

[0018] C(i,j)=w t ·|x i -y j |+w f |FFT(x i )-FFT(y j )|

[0019] Among them, x i and j are the sampling points of power data and non-power data in the time domain, FFT(x i ) and FFT(y j ) are their representations in the frequency domain, w t and w f are the weight coefficients in time domain and frequency domain;

[0020] The path that minimizes the weighted cumulative distance is calculated through a dynamic programming algorithm, and a cumulative distance matrix is ​​constructed. The optimal alignment path is obtained by reverse tracing the matrix, so that power data and non-power data are synchronously aligned in the time dimension and frequency domain characteristics.

[0021] The dual attention mechanism is to use the dual attention mechanism in the fusion layer after the convolution layer extracts the features, calculate the weights of the features in the same modality, use the cross-modal attention mechanism to calculate the correlation between power data and non-power data, and output high-level fusion features.

[0022] As a preferred solution of the adaptive resampling intensity distribution transformer structure response reconstruction method described in the present invention, wherein: the advanced fusion features are used as the input of the deep neural network for prediction, which is to construct the advanced fusion features into a feature set F fused =[f 1 ,f 2 ,…,f n ], regularized LASSO regression with dynamic weighting strategy is used to screen the output weights of advanced fusion features, and the expression is:

[0023]

[0024] Among them, δ(f i ) represents the feature f i The change of Δfi Represents the feature f i The rate of change of , λ is the adjustment parameter;

[0025] The high-level fusion feature set is adjusted according to the output feature weights. The expression is:

[0026] F weighted =[ω(f 1 )·f 1 ,ω(f 2 )·f 2 ,…,ω(f n )·f n ]

[0027] Among them, F weighted is the optimized feature vector;

[0028] The optimized high-level fusion feature F weighted Discretize into several intervals, calculate the probability distribution of eigenvalues ​​in each interval, and output the corresponding entropy value. The expression is:

[0029]

[0030] Among them, X i represents the i-th feature interval, P(x ik ) is the interval X i Medium feature x ik The probability distribution of

[0031] Based on the entropy value of each interval, a suitable activation function is dynamically selected for each neuron in the neural network, and the adjusted high-level fusion feature set F weighted As the input of the deep neural network model, it is trained using the selected activation function, dynamically adjusting the structure and parameters according to the distribution of features in different intervals, monitoring the training performance of the model in real time, and dynamically adjusting the model parameters and structure.

[0032] As a preferred solution of the adaptive resampling intensity distribution transformer structure response reconstruction method described in the present invention, wherein: the dynamic adjustment of model parameters and structure is to dynamically adjust the number of layers of the neural network by calculating the entropy value of the input feature and the rate of change of the training error in real time;

[0033] when And H(X)>O 2 When , the number of hidden layers is increased;

[0034] when And H(X) <O 4 When , the number of hidden layers is reduced;

[0035] Among them, O 1 , O2 , O 3 , O 4 is the training threshold.

[0036] As a preferred solution of the adaptive resampling intensity distribution transformer structure response reconstruction method described in the present invention, the game theory optimization strategy adopted in the DNN model is to dynamically adjust the number of layers of the neural network to achieve stability, enter the later stage of model training, and use the utility function to further optimize each layer of the DNN model. The expression is:

[0037] U l (w l )=-Loss(M l )+λ·Complexity(M l )

[0038] Among them, U l (w l ) is the utility function of the lth layer, w l is the weight of the output of the lth layer, Loss(M l ) represents the contribution of the l-th layer output to the final prediction error, Complexity (M l ) indicates that the complexity λ of the lth layer is a tuning parameter;

[0039] According to the utility function of each layer, the parameters and structure of each layer of the DNN model are optimized. When optimizing the utility function of each layer, different layers are regarded as players in the game. Each layer maximizes its positive contribution to the final prediction result by adjusting its own output weights and parameters. The strategy of each layer is regarded as the setting of its output weights and parameters. The utility function U l (w l ) is the profit under the current strategy. The Nash equilibrium point in the game is solved through an iterative optimization algorithm so that each layer can reach the optimal state of the utility function under the current configuration;

[0040] After solving the Nash equilibrium point, the outputs of each layer of the DNN model are weighted and combined according to the optimal output weight of each layer to construct the final output feature, which is expressed as:

[0041]

[0042] Among them, h final is the final output feature of the DNN model, h l is the output feature of the lth layer, is the optimal weight of the lth layer;

[0043] After implementing the BoYun optimization strategy, the hierarchical structure and parameters of the DNN model are optimized. The optimized model is used to predict the heavy overload state. The final prediction result is based on the weighted combination of outputs at different levels to form a comprehensive prediction result, which is expressed as:

[0044]

[0045] in, is the final prediction result of the heavy overload state, W final and b final are the final optimized weights and biases.

[0046] Another object of the present invention is to provide a distribution transformer structural response reconstruction system with adaptive resampling intensity, which can collect and analyze the operating data and environmental data of the distribution transformer in real time, use a deep learning model to extract and fuse the data, and combine the game theory optimization strategy to dynamically adjust the structure and parameters of the model, and finally output the optimal heavy overload state prediction result, solving the problems of low data fusion efficiency, insufficient model optimization capability, and poor prediction result accuracy in the prior art.

[0047] In order to solve the above technical problems, the present invention provides the following technical solutions: a distribution transformer structure response reconstruction system with adaptive resampling intensity, comprising: a data acquisition and preprocessing module, a multimodal data fusion and feature extraction module, an adaptive deep neural network prediction module, a game theory optimization and strategy integration module, and a prediction result fusion and output module;

[0048] The data acquisition and preprocessing module collects the operating data and environmental data of the distribution transformer through sensors, and performs denoising, normalization and outlier removal on the collected raw data;

[0049] The multimodal data fusion and feature extraction module performs joint analysis on different types of data and uses multimodal data fusion technology to integrate data from different sources into advanced fusion features;

[0050] The adaptive deep neural network prediction module uses a deep neural network model to train and predict advanced fusion features, and dynamically adjusts the number of layers of the model according to the distribution characteristics of entropy;

[0051] The game theory optimization and strategy integration module regards each layer of the DNN model as a player in the game, uses game theory methods to solve the optimal weight and configuration of each layer, and optimizes the parameters and structure of each layer of the DNN model;

[0052] The prediction result fusion and output module generates the final heavy overload state prediction result using a weighted combination strategy based on the output results of each layer after game theory optimization.

[0053] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned method for reconstructing the response of a matching variable structure with adaptive resampling intensity are implemented.

[0054] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for reconstructing the response of a variable structure with adaptive resampling intensity.

[0055] Beneficial effects of the present invention: By integrating power data and non-power data, the present invention can comprehensively analyze multi-dimensional information, thereby improving the utilization efficiency and feature expression ability of data. The deep learning model is used to extract features and adaptively adjust the data, so that the model can dynamically adapt to different load characteristics and environmental changes, thereby improving the accuracy and robustness of the prediction. Through the game theory optimization strategy, the structure and parameters of the model are dynamically optimized to ensure the optimization of the model performance under different operating conditions. The present invention can predict the heavy overload state of the distribution transformer in real time and efficiently, improving the problems of untimely data analysis, low model prediction accuracy and poor adaptability in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0057] Figure 1 An overall flow chart of a method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity provided by the first embodiment of the present invention;

[0058] Figure 2 A structural diagram of a distribution transformer structure response reconstruction system with adaptive resampling intensity provided in a second embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0060] Example 1, reference Figure 1According to an embodiment of the present invention, a method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity is provided, comprising:

[0061] Collect distribution transformer data through sensors and perform preprocessing;

[0062] Based on the preprocessing results, multimodal data fusion technology is used to jointly analyze power data and non-power data, extract time series features and static features, evaluate and output advanced fusion features;

[0063] Use high-level fusion features as input to deep neural networks for prediction and dynamically adjust model parameters and structures;

[0064] A game theory optimization strategy is adopted within the DNN model to dynamically adjust the weights of different layers of the model and output the optimal heavy overload state prediction results.

[0065] It should be further explained that the distribution transformer data includes operation data and environmental data.

[0066] Preprocessing is to use filtering algorithms to denoise the collected distribution transformer operation data and environmental data, normalize the denoised data, convert them to the same scale, and detect and eliminate outliers through statistical analysis methods.

[0067] By using an improved dynamic time warping algorithm combined with fast Fourier transform to time align data, power data and non-power data are input into the adaptive convolutional neural network respectively. An adaptive convolution kernel mechanism is designed in the convolution layer to dynamically adjust the size and shape of the convolution kernel according to the local time-frequency characteristics of the input data. The time series characteristics of power data and the spatial characteristics of non-power data are extracted respectively through a multi-level feature extraction strategy.

[0068] A dual attention mechanism is introduced in the fusion layer, and the fusion ratio of power and environmental features is dynamically adjusted through the feature weighted combination method to generate high-level fusion features.

[0069] It is further explained that:

[0070] The improved dynamic time warping algorithm introduces weights into the traditional dynamic time warping algorithm and combines the features of time domain and frequency domain to align data. The power data and non-power data are respectively subjected to fast Fourier transform to obtain their respective frequency domain representations, and the difference between the time domain and frequency domain features is calculated to construct a weighted cumulative distance matrix, which is expressed as follows:

[0071] C(i,j)=w t ·|x i -y j |+w f |FFT(x i)-FFT(y j )|

[0072] Among them, x i and j are the sampling points of power data and non-power data in the time domain, FFT(x i ) and FFT(y j ) are their representations in the frequency domain, w t and w f are the weight coefficients in time domain and frequency domain;

[0073] The path that minimizes the weighted cumulative distance is calculated through a dynamic programming algorithm, and a cumulative distance matrix is ​​constructed. The optimal alignment path is obtained by reverse tracing based on the matrix, so that power data and non-power data can be synchronously aligned in the time dimension and frequency domain characteristics.

[0074] After data alignment, power data and non-power data are input into adaptive convolutional neural network (CNN) respectively. An adaptive convolution kernel mechanism is designed so that the size and shape of the convolution kernel can be dynamically adjusted according to the local statistical characteristics of the input data.

[0075] The convolution kernel is adjusted based on the time-frequency characteristics of the data. By controlling the function φ(X), the convolution kernel size K is dynamically adjusted according to the local characteristics of the data such as the mean, variance, and spectral density. adaptive .

[0076] K adaptive =K base ×φ(X)

[0077] Among them, K base is the basic convolution kernel size, and φ(X) is an adjustment function based on the input data features. This can extract temporal features for power data and spatial features for non-power data.

[0078] After the convolutional layer extracts features, it enters the fusion layer and uses a dual attention mechanism, including self-attention mechanism and cross-modal attention mechanism. The self-attention mechanism is used to calculate the weights of features within the same modality and enhance the correlation within the modality:

[0079]

[0080] Among them, Q self and K self are query and key matrices respectively, α self is the self-attention weight.

[0081] The cross-modal attention mechanism is used to calculate the cross-correlation between power data and non-power data, capturing their mutual influence:

[0082]

[0083] Among them, Q cross and K cross are the cross-modal query and key matrices, and β is the cross-modal attention weight. Through the weight calculation of these two attention mechanisms, the fusion of power data and non-power data features is optimized.

[0084] After the fusion layer completes the feature weighting, the feature weighted combination method is used to optimize the weights through the genetic algorithm, dynamically adjust the fusion ratio of power and environmental features, and generate the final fusion feature representation:

[0085] F fused =γ·(α self F self +βF cross )

[0086] Among them, γ is the global weight optimized by the genetic algorithm, which is used to adjust the effect of the final feature combination.

[0087] The final generated fusion feature F fused Provides comprehensive information for further status analysis and prediction.

[0088] For the high-level fusion feature F fused =[f 1 ,f 2 ,…,f n ] is further selected and optimized, and regularized LASSO regression with a dynamic weighting strategy is used to determine the features that have the greatest impact on the prediction of heavy overload states. The weights of the features are dynamically adjusted according to the changes in real-time data features to enhance the adaptability of the model.

[0089] Feature weighting formula:

[0090]

[0091] Among them, δ(f i ) represents the feature f i The change of Δf i Represents the feature f i The rate of change of , λ is the adjustment parameter.

[0092] Through this weighting mechanism, the importance of features can be adaptively adjusted, making the model's sensitivity to features change dynamically.

[0093] Calculate the entropy value H(X) of each feature interval i ), used to dynamically select the activation function:

[0094] H(X i )=-∑ k P(x ik )logP(xik )f

[0095] Among them, X i represents the i-th feature interval, P(x ik ) is the interval X i Medium feature x ik The probability distribution of .

[0096] Based on the entropy value of each interval, a suitable activation function is dynamically selected for each neuron in the neural network, and the adjusted high-level fusion feature set F weighted As the input of the deep neural network model, it is trained using the selected activation function, dynamically adjusting the structure and parameters according to the distribution of features in different intervals, monitoring the training performance of the model in real time, and dynamically adjusting the number of model layers and structure.

[0097] It should be further explained that according to the entropy value H(X i ), dynamically select the appropriate activation function:

[0098] For low entropy values ​​(feature concentration), linear activation functions (such as ReLU) are selected, which is suitable for situations where features are relatively concentrated.

[0099] The medium entropy value selects a moderate nonlinear activation function (such as Tanh), which is suitable for situations where the features are somewhat dispersed but not complex.

[0100] High entropy value (feature dispersion): Select a highly nonlinear activation function (such as Sigmoid), which is suitable for situations where features are relatively dispersed.

[0101] Calculate the abstract value of input features: For each training cycle, calculate the abstract value H(X) of the input features to evaluate the distribution complexity of the data.

[0102] Calculate the rate of change of training error in the current cycle Used to judge the training effect of the model.

[0103] Adjust the model structure:

[0104]

[0105] when And H(X i )>Threshold 2 , increase the number of hidden layers.

[0106] when And H(X i ) <Threshold 4 , reduce the number of hidden layers.

[0107] According to the adjustment strategy, the skip connections or residual connections of the network are dynamically increased or decreased to adapt to the feature distribution of the input data.

[0108] It should be further explained that the traditional neural network structure adjustment method is usually based on a fixed training stage or preset conditions, and rarely considers the real-time feature distribution of the input data. In the present invention, a dynamic adjustment mechanism based on the extraction value of the input feature and the training error is introduced. This mechanism does not rely on a preset training cycle or fixed conditions, but dynamically adjusts the number of layers of the neural network based on the entropy value of the input feature and the real-time performance of the model. This method enables the model to adapt to different input data characteristics more flexibly, thereby improving the accuracy of training and prediction.

[0109] When the number of neural network layers is dynamically adjusted to achieve stability, a game theory optimization strategy is adopted inside the DNN model. In the later stage of model training, the utility function is used to further optimize the parameters of each layer of the DNN model. The expression is:

[0110] U l (w l )=-Loss(M l )+C·Complexity(M l )

[0111] Among them, U l (w l ) is the utility function of the lth layer, w l is the weight of the output of the lth layer, Loss(M l ) represents the contribution of the l-th layer output to the final prediction error, Complexity (M l ) indicates that the complexity C of the lth layer is an adjustment parameter.

[0112] According to the utility function of each layer, the parameters and structure of each layer of the DNN model are optimized. When optimizing the utility function of each layer, different layers are regarded as players in the game. Each layer maximizes its positive contribution to the final prediction result by adjusting its own output weights and parameters. The strategy of each layer is regarded as the setting of its output weights and parameters. The utility function U l (w l ) is the profit under the current strategy. The Nash equilibrium point in the game is solved through an iterative optimization algorithm so that each layer can reach the optimal state of the utility function under the current configuration.

[0113] After solving the Nash equilibrium point, the outputs of each layer of the DNN model are weighted and combined according to the optimal output weight of each layer to construct the final output feature, which is expressed as:

[0114]

[0115] Among them, hfinal is the final output feature of the DNN model, h l is the output feature of the lth layer, is the optimal weight of layer l.

[0116] After implementing the BoYun optimization strategy, the hierarchical structure and parameters of the DNN model are optimized. The optimized model is used to predict the heavy overload state. The final prediction result is based on the weighted combination of outputs at different levels to form a comprehensive prediction result, which is expressed as:

[0117]

[0118] in, is the final prediction result of the heavy overload state, W final and b final are the final optimized weights and biases.

[0119] Example 2, reference Figure 2 , which is an embodiment of the present invention, provides a system for a method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity, including: a data acquisition and preprocessing module, a multimodal data fusion and feature extraction module, an adaptive deep neural network prediction module, a game theory optimization and strategy integration module, and a prediction result fusion and output module;

[0120] The data acquisition and preprocessing module collects the operating data and environmental data of the distribution transformer through sensors, and performs denoising, normalization and outlier removal on the collected raw data;

[0121] The multimodal data fusion and feature extraction module conducts joint analysis on different types of data and uses multimodal data fusion technology to integrate data from different sources into advanced fusion features;

[0122] The adaptive deep neural network prediction module uses a deep neural network model to train and predict advanced fusion features, and dynamically adjusts the number of model layers according to the distribution characteristics of entropy;

[0123] The game theory optimization and strategy integration module regards each layer of the DNN model as a player in the game, uses game theory methods to solve the optimal weight and configuration of each layer, and optimizes the parameters and structure of each layer of the DNN model;

[0124] The prediction result fusion and output module is based on the output results of each layer after game theory optimization, and uses a weighted combination strategy to generate the final heavy overload state prediction result.

[0125] If the 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.

[0126] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0127] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0128] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0129] Embodiment 3: In this embodiment, in order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments. This embodiment experiments the existing traditional method and the method of this embodiment respectively.

[0130] The simulation experiment of the present invention is intended to verify its effectiveness in power system monitoring and distribution transformer heavy overload state prediction. The experimental environment is set in a simulated power distribution network, which contains multiple distribution transformers and environmental monitoring equipment for real-time collection of power data (such as current, voltage, power, etc.) and non-power data (such as temperature, humidity, etc.). The experiment uses a simulation platform built with Python programming language and TensorFlow deep learning framework, running on a computer with 16GB memory and 4-core CPU to ensure the accuracy and repeatability of the experimental results.

[0131] Method 1 predicts the heavy overload state based on a simple threshold judgment. When the collected current or voltage value exceeds the preset safety threshold, the system triggers an overload alarm. The advantage of this method is that it is simple to implement, but the disadvantage is that it does not consider the multidimensional characteristics and time series characteristics of the data, and is prone to false positives or false negatives.

[0132] Method 2 uses a linear regression model to analyze the operating data of the distribution transformer and predicts the future load state through regression analysis of historical data. This method has some improvements over the threshold method and can handle more data features, but it still lacks the characterization of nonlinear relationships in the data and has limited prediction accuracy.

[0133] Table 1 Data comparison table

[0134]

[0135] As can be seen from Table 1, the method of the present invention performs significantly better than the traditional method in predicting the heavy overload state of the distribution transformer.

[0136] The method of the present invention achieves a prediction accuracy of 93.5%, which is significantly higher than the 75.2% of the threshold method and the 82.7% of the linear regression method. In addition, the false alarm rate and the missed alarm rate of the method of the present invention are 3.2% and 3.3% respectively, which are much lower than the 15.3% and 9.5% of the threshold method and the 10.8% and 6.5% of the linear regression method. In terms of response time, the method of the present invention is also faster, only 5ms, which is significantly better than the 10ms and 20ms of the threshold method and the linear regression method.

[0137] The method of the present invention has high adaptability and can dynamically adjust the structure and parameters of the model to adapt to different load characteristics and environmental changes, which is impossible to achieve in traditional methods. Overall, the present invention has significant advantages in prediction accuracy, response speed and adaptability, fully demonstrating its beneficial effects in power system monitoring and optimization.

[0138] Table 2 Effect comparison table

[0139]

[0140]

[0141] Referring to Table 2, although the method of the present invention requires a longer training time (120 seconds), this matches the complexity of the deep neural network and the large amount of data required for training. In contrast, the linear regression method only takes 5 seconds, and the threshold method has no model training time because it does not require training.

[0142] In terms of model complexity, the method of the present invention has 1,000,000 parameters, which is significantly higher than the 200 parameters of the linear regression method. This high complexity enables the model to capture more data features and provide higher prediction accuracy.

[0143] In terms of data processing efficiency, the method of the present invention only takes 0.3 ms to process each piece of data, which is better than the 0.5 ms of the threshold method and the 1.2 ms of the linear regression method, demonstrating its advantage in efficiently processing large-scale data.

[0144] In terms of the advance time of fault prediction, the method of the present invention can successfully predict the actual overload event 15 minutes before it occurs, which is earlier than 5 minutes of the linear regression method and 1 minute of the threshold method, greatly improving the response capability of the power system.

[0145] Finally, in terms of model stability, the method of the present invention maintains high stability under different data conditions, while the traditional method performs poorly. The threshold method has low stability when the data changes, and the linear regression method is medium. However, the method of the present invention still performs well in complex and changing environments through an adaptive adjustment mechanism.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity, characterized in that: include: Collect distribution transformer data through sensors and perform preprocessing; Based on the preprocessing results, multimodal data fusion technology is used to jointly analyze power data and non-power data, extract time series features and static features, evaluate and output advanced fusion features; Use high-level fusion features as input to deep neural networks for prediction and dynamically adjust model parameters and structures; A game theory optimization strategy is adopted within the DNN model to dynamically adjust the weights of different layers of the model and output the optimal heavy overload state prediction results.

2. The method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity according to claim 1, characterized in that: The distribution transformer data includes operation data and environmental data; The preprocessing is to use a filtering algorithm to denoise the collected distribution transformer operation data and environmental data, normalize the denoised data, convert them to the same scale, and detect and eliminate abnormal values ​​through statistical analysis methods.

3. The method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity according to claim 2, characterized in that: The joint analysis is performed by using an improved dynamic time warping algorithm combined with fast Fourier transform to perform time alignment of data, inputting power data and non-power data into an adaptive convolutional neural network respectively, designing an adaptive convolution kernel mechanism in the convolution layer, dynamically adjusting the size and shape of the convolution kernel according to the local time-frequency characteristics of the input data, and extracting the time series characteristics of power data and the spatial characteristics of non-power data respectively through a multi-level feature extraction strategy; A dual attention mechanism is introduced in the fusion layer, and the fusion ratio of power and environmental features is dynamically adjusted through the feature weighted combination method to generate high-level fusion features.

4. The method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity according to claim 3, characterized in that: The improved dynamic time warping algorithm introduces weights into the traditional dynamic time warping algorithm and combines the features of time domain and frequency domain to align data; the power data and non-power data are respectively subjected to fast Fourier transform to obtain their respective frequency domain representations, the difference between the time domain and frequency domain features is calculated, and the weighted cumulative distance matrix is ​​constructed, which is expressed as: C(i,j)=w t ·|x i -y j |+w f ·|FFT(x i )-FFT(y j )| Among them, x i and j are the sampling points of power data and non-power data in the time domain, FFT(x i ) and FFT(y j ) are their representations in the frequency domain, w t and w f are the weight coefficients in time domain and frequency domain; The path that minimizes the weighted cumulative distance is calculated through a dynamic programming algorithm, and a cumulative distance matrix is ​​constructed. The optimal alignment path is obtained by reverse tracing the matrix, so that power data and non-power data are synchronously aligned in the time dimension and frequency domain characteristics. The dual attention mechanism is to use the dual attention mechanism in the fusion layer after the convolution layer extracts the features, calculate the weights of the features in the same modality, use the cross-modal attention mechanism to calculate the correlation between power data and non-power data, and output high-level fusion features.

5. The method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity according to claim 4, characterized in that: The method of using the high-level fusion features as the input of the deep neural network for prediction is to construct the high-level fusion features into a feature set F fused =[f1, f2, ..., f n ], regularized LASSO regression with dynamic weighting strategy is used to screen the output weights of advanced fusion features, and the expression is: Among them, δ(f i ) represents the feature f i The change range, △f i Represents feature f i The rate of change of , λ is the adjustment parameter; The high-level fusion feature set is adjusted according to the output feature weights. The expression is: F weighted =[ω(f1)·f1,ω(f2)·f2,…,ω(f n )·f n ] Among them, F weighted is the optimized feature vector; The optimized high-level fusion feature F weighted Discretize into several intervals, calculate the probability distribution of eigenvalues ​​in each interval, and output the corresponding entropy value. The expression is: Among them, X i represents the i-th feature interval, P(x ik ) is the interval X i Medium feature x ik The probability distribution of Based on the entropy value of each interval, a suitable activation function is dynamically selected for each neuron in the neural network, and the adjusted high-level fusion feature set F weighted As the input of the deep neural network model, it is trained using the selected activation function, dynamically adjusting the structure and parameters according to the distribution of features in different intervals, monitoring the training performance of the model in real time, and dynamically adjusting the model parameters and structure.

6. A method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity as claimed in claim 5, characterized in that: The dynamic adjustment of model parameters and structure is to dynamically adjust the number of layers of the neural network by calculating the entropy value of the input feature and the rate of change of the training error in real time; when When H(x)>O2, increase the number of hidden layers; When and H(x) < O4, the number of hidden layers is reduced; Among them, O1, O2, O3, and O4 are training thresholds.

7. The method for reconstructing the response of a distribution transformer structure with adaptive resampling intensity according to claim 6, characterized in that: The game theory optimization strategy adopted in the DNN model is to dynamically adjust the number of neural network layers to achieve stability, enter the later stage of model training, and use the utility function to further optimize each layer of the DNN model. The expression is: U l (w l )=-Loss(M l )+λ·Complexity(M l ) Among them, U l (w l ) is the utility function of the lth layer, w l is the weight of the output of the lth layer, Loss(M l ) represents the contribution of the l-th layer output to the final prediction error, Complexity (M l ) indicates that the complexity λ of the lth layer is a tuning parameter; According to the utility function of each layer, the parameters and structure of each layer of the DNN model are optimized. When optimizing the utility function of each layer, different layers are regarded as players in the game. Each layer maximizes its positive contribution to the final prediction result by adjusting its own output weights and parameters. The strategy of each layer is regarded as the setting of its output weights and parameters. The utility function U l (w l ) is the profit under the current strategy. The Nash equilibrium point in the game is solved through an iterative optimization algorithm so that each layer can reach the optimal state of the utility function under the current configuration; After solving the Nash equilibrium point, the outputs of each layer of the DNN model are weighted and combined according to the optimal output weight of each layer to construct the final output feature, which is expressed as: Among them, h final is the final output feature of the DNN model, h l is the output feature of the first layer, is the optimal weight of layer 1; After implementing the Bobian optimization strategy, the hierarchical structure and parameters of the DNN model are optimized. The optimized model is used to predict the heavy overload state. The final prediction result is based on the weighted combination of outputs at different levels to form a comprehensive prediction result, which is expressed as: in, is the final prediction result of the heavy overload state, W final and b final are the final optimized weights and biases.

8. A system using a method for reconstructing the structural response of a distribution transformer using an adaptive resampling force as claimed in any one of claims 1 to 7, characterized in that: It includes data acquisition and preprocessing module, multimodal data fusion and feature extraction module, adaptive deep neural network prediction module, game theory optimization and strategy integration module, and prediction result fusion and output module; The data acquisition and preprocessing module collects the operating data and environmental data of the distribution transformer through sensors, and performs denoising, normalization and outlier removal on the collected raw data; The multimodal data fusion and feature extraction module performs joint analysis on different types of data and uses multimodal data fusion technology to integrate data from different sources into advanced fusion features; The adaptive deep neural network prediction module uses a deep neural network model to train and predict advanced fusion features, and dynamically adjusts the number of layers of the model according to the distribution characteristics of entropy; The game theory optimization and strategy integration module regards each layer of the DNN model as a player in the game, uses game theory methods to solve the optimal weight and configuration of each layer, and optimizes the parameters and structure of each layer of the DNN model; The prediction result fusion and output module generates the final heavy overload state prediction result using a weighted combination strategy based on the output results of each layer after game theory optimization.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a method for reconstructing the response of a matched variable structure with adaptive resampling intensity according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for reconstructing the response of a matched transformer structure with adaptive resampling force as described in any one of claims 1 to 7 are implemented.

Images