A new type of protective transformer voltage transformation method and device

Through intelligent sensor network and deep learning technology, multi-dimensional parameter acquisition and load prediction of transformers are carried out, and transformer switching decisions are made by combining multi-objective optimization and reinforcement learning methods, which solves the problem that traditional transformer control methods are difficult to adapt to complex load environments and improves the transformer's Operational safety and energy efficiency.

CN118801394BActive Publication Date: 2024-11-26DONGGUAN PANDA MAGNETIC FLUX ELECTRONICS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411281740.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-11-26
Estimated Expiration
2044-09-13

AI Technical Summary

Technical Problem

Traditional transformer control methods are difficult to adapt to complex and variable load environments, resulting in unstable voltage quality, shortened life and inefficient energy efficiency, and lack of real-time evaluation and prediction capabilities for transformer health status.

Method used

Intelligent sensor network and edge computing technology are used to collect multidimensional parameters in real time, combine timing analysis and deep learning technology to perform load prediction, evaluate the health status of the transformer through multidimensional evaluation and safety margin calculation, and use multi-objective optimization and reinforcement learning methods to transform Switch decision-making, combine model predictive control and adaptive control technology for precise control, and optimize decision-making solutions using online learning and knowledge graph technology.

Benefits of technology

It improves the operating safety and reliability of the transformer, optimizes energy efficiency, realizes rapid response and decision-making adjustments to complex scenarios, and enhances the flexibility and adaptability of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118801394B_ABST
    Figure CN118801394B_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of transformer switching, and discloses a transformer method and device for a novel protective transformer. The method includes: real-time acquisition of multi-dimensional parameters of the transformer to obtain an original monitoring data set; performing time series analysis and deep learning modeling to obtain an initial load prediction model and a short-term load prediction result; performing multi-dimensional evaluation and safety margin calculation to obtain a transformer health index and a dynamic safety threshold; performing multi-objective optimization and reinforcement learning to obtain a first transformer switching decision plan; performing segmented control and real-time parameter adjustment on the first transformer switching decision plan to obtain an adaptive switching control instruction and switching process data; performing online learning and knowledge graph construction on the switching process data, the adaptive switching control instruction and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision plan. The present application improves the transformer switching flexibility of the protective transformer.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of voltage transformation and switching technology, and in particular to a voltage transformation method and device for a novel protective transformer. Background Art

[0002] As a key device in the power system, transformers are facing increasingly severe operational challenges. Traditional transformer control methods often use fixed switching strategies, which are difficult to adapt to complex and changing load environments, resulting in unstable voltage quality, shortened transformer life, and low energy efficiency. In addition, due to the lack of real-time evaluation and prediction capabilities of the transformer's health status, traditional methods are difficult to achieve optimal transformer control while ensuring equipment safety.

[0003] In recent years, with the rapid development of artificial intelligence and big data technology, the advancement of smart substation and digital power grid construction has provided new opportunities for intelligent control of transformers. However, the existing intelligent control methods still have some shortcomings, such as insufficient data collection and processing, limited load prediction accuracy, and lack of comprehensive consideration of multiple objectives in the decision-making process. Summary of the invention

[0004] The present application provides a novel voltage transformation method and device for a protective transformer, which are used to improve the voltage transformation switching flexibility of the protective transformer.

[0005] In a first aspect, the present application provides a voltage transformation method of a novel protective transformer, the voltage transformation method of the novel protective transformer comprising:

[0006] Collect multi-dimensional parameters of transformers in real time to obtain original monitoring data sets;

[0007] Performing time series analysis and deep learning modeling on the original monitoring data set to obtain an initial load forecasting model and short-term load forecasting results;

[0008] Performing multi-dimensional evaluation and safety margin calculation on the original monitoring data set and the short-term load forecast result to obtain a transformer health index and a dynamic safety threshold;

[0009] Performing multi-objective optimization and reinforcement learning on the initial load prediction model, the transformer health index and the dynamic safety threshold to obtain a first transformer switching decision plan;

[0010] Performing segmented control and real-time parameter adjustment on the first transformer switching decision scheme to obtain adaptive switching control instructions and switching process data;

[0011] Online learning and knowledge graph construction are performed on the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision plan.

[0012] In a second aspect, the present application provides a new type of protective transformer transformer device, the new type of protective transformer transformer device comprising:

[0013] The acquisition module is used to collect multi-dimensional parameters of the transformer in real time to obtain the original monitoring data set;

[0014] A modeling module, used to perform time series analysis and deep learning modeling on the original monitoring data set to obtain an initial load forecasting model and a short-term load forecasting result;

[0015] A calculation module, used to perform multi-dimensional evaluation and safety margin calculation on the original monitoring data set and the short-term load forecast result to obtain a transformer health index and a dynamic safety threshold;

[0016] An optimization module, used for performing multi-objective optimization and reinforcement learning on the initial load prediction model, the transformer health index and the dynamic safety threshold to obtain a first transformer switching decision plan;

[0017] An adjustment module, used for performing segmented control and real-time parameter adjustment on the first transformer switching decision scheme to obtain adaptive switching control instructions and switching process data;

[0018] A construction module is used to perform online learning and knowledge graph construction on the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision plan.

[0019] The third aspect of the present application provides a new type of protective transformer transformation device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the new type of protective transformer transformation device executes the above-mentioned new type of protective transformer transformation method.

[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned voltage transformation method of the novel protective transformer.

[0021] In the technical solution provided by the present application, high-frequency sampling and real-time preprocessing of key parameters of transformers are realized through intelligent sensor networks and edge computing technologies. Combined with time series analysis and deep learning technologies, multi-scale and multi-model load prediction is realized, and prediction accuracy and robustness are improved. Through multi-dimensional evaluation and safety margin calculation, combined with expert knowledge and data-driven methods, dynamic evaluation of transformer health status is realized, and operational safety is improved. Multi-objective optimization and reinforcement learning methods are adopted to realize dynamic optimization of transformer switching decisions. Through segmented control and real-time parameter adjustment, combined with model predictive control and adaptive control technology, precise control of transformer switching process is realized, and the impact of switching on the system is reduced. Using online learning and knowledge graph technology, dynamic accumulation and optimization of transformer operation knowledge are realized, and the adaptive ability and intelligent decision-making level of the system are improved. Through the analysis and feedback of switching process data, continuous optimization of decision-making schemes is realized, and the long-term effect of transformer control is improved. Through the calculation and adjustment of dynamic safety thresholds, the safety and reliability of transformer operation are improved, while the operating efficiency is optimized, rapid response and decision adjustment to complex scenarios are realized, and the flexibility and adaptability of the system are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 A schematic diagram of an embodiment of a voltage transformation method of a novel protective transformer in an embodiment of the present application;

[0024] Figure 2 This is a schematic diagram of an embodiment of a transformer device of a new protective transformer in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiment of the present application provides a method and device for transforming a novel protective transformer. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims 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 described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the voltage transformation method of the novel protective transformer in the embodiment of the present application includes:

[0027] Step S101, real-time acquisition of multi-dimensional parameters of the transformer to obtain an original monitoring data set;

[0028] It is understandable that the execution subject of the present application may be a transformer device of the novel protective transformer, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0029] Specifically, intelligent sensor networks are deployed at multiple target locations of the transformer. These sensors are placed at key locations of the transformer to monitor parameters such as voltage, current, temperature, load, and on-load tap changer position, and generate multi-dimensional raw parameter data streams through high-frequency sampling. Edge computing preprocesses the multi-dimensional raw parameter data streams, and edge computing can process data close to the data source to reduce transmission delay and bandwidth consumption. Through this process, real-time parameter data is obtained after preliminary filtering, and adaptive threshold recognition and median replacement processing are performed on the outliers in these data to obtain corrected continuous parameter data. Timestamp alignment and multi-source data fusion are performed on the corrected continuous parameter data to generate a synchronized multi-dimensional collaborative data set. Timestamp alignment ensures that data from different sensors are compared on the same time scale, while multi-source data fusion combines these data together to form a complete collaborative data set. Wavelet transform denoising and Kalman filtering smoothing are performed on the synchronized multi-dimensional collaborative data set. Wavelet transform removes high-frequency noise in the data, while Kalman filtering further smoothes the data to generate a high signal-to-noise ratio data set, thereby improving the quality and reliability of the data. Perform time-frequency domain feature extraction on the high signal-to-noise ratio data set to generate a reduced-dimensionality feature vector set. Time-frequency domain feature extraction can capture the characteristics of data at different frequencies and time periods, while the reduced-dimensionality feature vector set simplifies the complexity of the data while retaining the most important information. Perform standardization on the reduced-dimensionality feature vector set to generate a standardized feature data set. Perform adaptive time window segmentation and overlapping sliding window processing on the standardized feature data set to generate a time series feature sample set. Adaptive time window segmentation can dynamically adjust the window size according to data changes, while overlapping sliding window processing can capture subtle changes in data by introducing overlapping parts between different windows. Perform compressed sensing encoding on the time series feature sample set to obtain the original monitoring data set. Compressed sensing encoding uses efficient coding technology to compress data into a smaller representation while retaining as much information as possible from the original data.

[0030] Step S102: Perform time series analysis and deep learning modeling on the original monitoring data set to obtain an initial load forecasting model and short-term load forecasting results;

[0031] Specifically, the original monitoring data set is subjected to wavelet transform time series decomposition, and the original data is decomposed into multi-scale trend items, seasonal items, and random items. Through wavelet transform, the long-term trend and periodic fluctuations in the data can be effectively captured, and short-term random noise can be eliminated, making the data smoother and more regular. According to the decomposed multi-scale trend items, seasonal items, and random items, the original monitoring data set is seasonally adjusted and fractional difference processed to obtain a stationary time series. By eliminating the periodic components and trend components in the time series, the data is made more stable and suitable for subsequent modeling and prediction. The autocorrelation function and partial autocorrelation function of the stationary time series are analyzed to determine the order parameters of the model. By analyzing the autocorrelation of the time series, the potential regularity and periodicity in the data are identified, and the optimal model order is determined. According to the model order parameters, the maximum likelihood estimation of the stationary time series is performed to obtain the preliminary load forecasting results. By statistical methods, the time series is fitted and predicted, and the load changes in the future period are preliminarily predicted. The original monitoring data set is subjected to adaptive time window segmentation and multi-dimensional feature extraction to obtain the deep learning input tensor. Adaptive time window segmentation can dynamically adjust the size of the time window to adapt to the changing characteristics of the data, while multi-dimensional feature extraction can extract more useful information from the original data and enrich the input of the model. The deep learning input tensor is input into the temporal convolutional network and long short-term memory network hybrid model architecture for multi-step prediction to obtain the deep learning prediction results. The temporal convolutional network can effectively capture the local patterns in the data, while the long short-term memory network is good at processing the long-term dependencies in time series data. The combination of the two can significantly improve the accuracy and stability of the prediction. The preliminary load forecast results and the deep learning forecast results are fused by dynamic weight ensemble learning to obtain the short-term load forecast results. Dynamic weight ensemble learning can make full use of the advantages of the preliminary forecast results and the deep learning forecast results by dynamically adjusting the weights of the preliminary forecast results and the deep learning forecast results to improve the accuracy and robustness of the prediction. According to the short-term load forecast results, historical forecast errors and transformer health index, online transfer learning and adaptive adjustment are performed to obtain the initial load forecast model. Online transfer learning improves the adaptability and prediction accuracy of the model by continuously learning new data and adapting to environmental changes, while adaptive adjustment dynamically adjusts the parameters and structure of the model according to real-time data to ensure that the model is always in the best state.

[0032] Step S103, performing multi-dimensional evaluation and safety margin calculation on the original monitoring data set and the short-term load forecast result to obtain the transformer health index and dynamic safety threshold;

[0033] Specifically, multi-dimensional feature extraction is performed on the original monitoring data set to generate a standardized feature matrix. By standardizing the data, the comparability between the features is ensured, and the preliminary health assessment results are calculated based on the preset weight coefficients. The weight coefficients are pre-set according to the different operating parameters of the transformer and the degree of their influence on the health status of the transformer. The preliminary health assessment results are classified with the historical fault data to obtain the health status classification results. By analyzing the historical data, the characteristic patterns under different health states are identified. According to the health status classification results, and combined with the pre-established expert knowledge base, adaptive neuro-fuzzy reasoning is performed to obtain the fuzzy health index. The expert knowledge base contains a large number of empirical rules and data. Through neuro-fuzzy reasoning, these rules are applied to the current health status assessment. Time series analysis is performed on the fuzzy health index and the short-term load forecast results to obtain the health trend forecast. Time series analysis can reveal the potential trends and periodic changes in the data, thereby predicting the health status of the transformer in the future. Based on the health trend forecast, transformer rated parameters and real-time load data, dynamic stress-strength interference analysis is performed to obtain the preliminary safety margin. Dynamic stress-strength interference analysis evaluates the safety margin by comparing the stress and strength of the transformer under different load conditions. Kernel density estimation is performed on the preliminary safety margin to obtain the probability distribution of the safety margin. Kernel density estimation is a non-parametric statistical method that can estimate the probability density function of data, thereby more accurately describing the distribution of the safety margin. According to the probability distribution of the safety margin, adaptive confidence interval calculation is performed to obtain the dynamic safety threshold. By adjusting the size of the confidence interval, the adaptive confidence interval calculation can more flexibly respond to the changes in the safety margin under different conditions and ensure the safe operation of the transformer under different load conditions. The fuzzy health index and the dynamic safety threshold are nonlinearly weighted fused to obtain a comprehensive health score. Nonlinear weighted fusion introduces nonlinear weights to more accurately comprehensively consider the influence of the fuzzy health index and the dynamic safety threshold, and obtain a more comprehensive health score. According to the comprehensive health score, historical operating data and the initial load prediction model, online correction and adaptive update are performed to finally obtain the transformer health index and dynamic safety threshold. Online correction and adaptive update can ensure that the transformer health index and dynamic safety threshold always reflect the current operating status by continuously learning and adjusting model parameters.

[0034] Step S104, performing multi-objective optimization and reinforcement learning on the initial load prediction model, the transformer health index and the dynamic safety threshold to obtain a first transformer switching decision plan;

[0035] Specifically, the short-term load forecasting results of the initial load forecasting model are kernel-density estimated to obtain the load forecasting probability density function. Kernel density estimation can effectively describe the probability distribution of load forecasting. Simulation is performed based on the load forecasting probability density function and the rated parameters of the transformer to generate multiple groups of load scenarios. The load scenarios cover various possible load conditions. A comprehensive evaluation is performed on multiple groups of load scenarios and the transformer health index to obtain a risk assessment matrix for each scenario. The risk assessment matrix quantitatively analyzes the health status and potential risks of the transformer under different load conditions. According to the risk assessment matrix and the dynamic safety threshold, an initial multi-objective optimization model is constructed, and the optimal solution of the model is calculated to obtain the optimal solution set. The multi-objective optimization model comprehensively considers the load forecast, health status and safety margin of the transformer, and is solved by the optimization algorithm to find the optimal solution set that meets multiple objectives. Multi-criteria decision analysis is performed based on the optimal solution set and the preset decision preference weights to generate candidate transformer switching decision schemes. Multi-criteria decision analysis selects the scheme that best meets the decision preference by comprehensively considering the priorities of different weights. At the same time, the state transition probability matrix and time-varying reward function of the candidate transformer switching decision scheme are calculated. The state transition probability matrix describes the possibility of transformer switching under different states, while the time-varying reward function measures the decision benefits at different time points. According to the state transition probability matrix and the time-varying reward function, the Q-value function of the continuous action space is generated. The Q-value function provides guidance for decision-making by measuring the value of different state-action pairs. On this basis, the Q-value function is explored by noise injection to avoid local optimal solutions and obtain a more comprehensive reinforcement learning strategy. According to the reinforcement learning strategy and the candidate transformer switching decision scheme, the solution is obtained to obtain the first transformer switching decision scheme. The reinforcement learning strategy optimizes the decision-making process through continuous trial and error and learning, and finally generates the optimal transformer switching decision scheme.

[0036] Step S105, performing segmented control and real-time parameter adjustment on the first voltage transformation switching decision scheme to obtain an adaptive switching control instruction and switching process data;

[0037] Specifically, the first transformer switching decision plan is decomposed in time sequence and divided into scenarios, and the entire decision process is divided into three sub-decision sequences: preparation, execution, and completion. According to the sub-decision sequence, transformer health index, and dynamic safety threshold, a nonlinear state space model is constructed to generate a segmented control state equation. The nonlinear state space model can describe the dynamic behavior of the system under different states. The segmented control state equation is subjected to model predictive control to obtain the initial control sequence. The model predictive control method formulates a preliminary control strategy by predicting the future system behavior. According to the initial control sequence, real-time load data, and short-term load forecast results, multi-time scale rolling optimization is performed to obtain the optimized control parameters. The multi-time scale rolling optimization method can adjust the control strategy at different time scales to cope with the changing load and system state, ensuring the real-time and effectiveness of the control strategy. The optimized control parameters are subjected to polynomial chaos expansion analysis to obtain the parameter adjustment range and probability distribution. Polynomial chaos expansion analysis is an efficient stochastic analysis method that determines the parameter adjustment range by analyzing the distribution of control parameters. According to the parameter adjustment range and the response characteristics of the transformer, a nonlinear adaptive control law is constructed. The nonlinear adaptive control law can adaptively adjust the control strategy according to the real-time response of the system to ensure the stable operation of the transformer under different load conditions. The nonlinear adaptive control law is subjected to random stability analysis and solution to obtain the probabilistic stability guarantee interval. The random stability analysis method evaluates the stability of the control strategy by considering the random disturbance of the system, providing a basis for the optimization of the control strategy. According to the probabilistic stability guarantee interval and the dynamic characteristic curve of the on-load tap changer, the multi-objective switching timing optimization is carried out to obtain the target switching time point. The multi-objective switching timing optimization method determines the optimal switching time point by comprehensively considering multiple objectives, including the stability, efficiency and safety of the system, to ensure the smooth operation of the system during the switching process. The target switching time point is verified by closed-loop simulation to obtain the adaptive switching control instruction. The closed-loop simulation verification verifies the effectiveness and reliability of the control strategy by simulating the actual system operation. According to the adaptive switching control instruction, the transformer is controlled in real time, and the voltage, current, switch state and vibration data during the switching process are collected to obtain the switching process data. Through real-time control and data acquisition, the operating state of the transformer during the switching process is monitored to ensure the safety and stability of the switching process.

[0038] Step S106: Perform online learning and knowledge graph construction on the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision plan.

[0039] Specifically, multi-source heterogeneous data fusion is performed on the switching process data, the original monitoring data set and the short-term load forecasting results to generate a spatiotemporal feature data set. Multi-source heterogeneous data fusion integrates data from different sources and different types to form a comprehensive and consistent data set. According to the spatiotemporal feature data set and the adaptive switching control instructions, a temporal causal graph and a semantic network are constructed to obtain an initial multi-layer knowledge graph. The temporal causal graph can reveal the causal relationship in the data, while the semantic network enriches the content of the knowledge graph through semantic association. Dynamic node embedding and edge attribute learning are performed on the initial multi-layer knowledge graph to obtain an evolving knowledge graph. Dynamic node embedding and edge attribute learning reflect the changes of the system at different time points by continuously updating the attributes of nodes and edges, making the knowledge graph more dynamic and real-time. According to the evolving knowledge graph, graph reasoning and relationship extraction are performed to obtain the dynamic operation law and potential fault evolution mode of the transformer. Graph reasoning and relationship extraction can extract important information and laws from the knowledge graph, revealing the behavior mode and potential fault path of the transformer under different operating conditions. Spatiotemporal sequence modeling is performed on the dynamic operation law and potential fault evolution mode of the transformer to generate an intermediate load prediction model. Spatiotemporal sequence modeling establishes a model that can predict future loads by analyzing the changes in data in time and space. Online ensemble learning and migration optimization are performed based on the intermediate load prediction model, the initial load prediction model and the historical prediction error to obtain the target load prediction model. Online ensemble learning improves the prediction accuracy by combining the prediction results of multiple models, while migration optimization ensures the real-time and reliability of the prediction model by continuously adjusting the model parameters to adapt it to the new data environment. Multi-agent collaborative learning is performed on the target load prediction model, the evolutionary knowledge graph and the transformer health index to obtain an optimized decision strategy. Multi-agent collaborative learning optimizes the decision-making process and improves the accuracy and efficiency of decision-making through the collaborative work of multiple agents. According to the optimized decision strategy and dynamic safety threshold, a search is performed to generate a multi-layer candidate decision tree, which is dynamically programmed to obtain a decision solution set. The multi-layer candidate decision tree is screened layer by layer to ensure that each decision point can obtain the optimal solution, while dynamic programming finds the optimal path of the entire decision tree through global optimization. According to the decision solution set, the first transformer switching decision solution and real-time operation data, adaptive selection and parameter adjustment are performed to finally obtain the second transformer switching decision solution. Adaptive selection and parameter adjustment dynamically adjust decision parameters by combining current operating data and historical decision-making experience to ensure safe and stable operation of the transformer under different operating conditions.

[0040] In the embodiment of the present application, high-frequency sampling and real-time preprocessing of key parameters of the transformer are realized through intelligent sensor networks and edge computing technologies. Combined with time series analysis and deep learning technologies, multi-scale and multi-model load prediction is realized, and prediction accuracy and robustness are improved. Through multi-dimensional evaluation and safety margin calculation, combined with expert knowledge and data-driven methods, dynamic evaluation of transformer health status is realized, and operational safety is improved. Multi-objective optimization and reinforcement learning methods are adopted to realize dynamic optimization of transformer switching decisions. Through segmented control and real-time parameter adjustment, combined with model predictive control and adaptive control technology, precise control of transformer switching process is realized, and the impact of switching on the system is reduced. Using online learning and knowledge graph technology, dynamic accumulation and optimization of transformer operation knowledge are realized, and the adaptive ability and intelligent decision-making level of the system are improved. Through the analysis and feedback of switching process data, continuous optimization of decision-making schemes is realized, and the long-term effect of transformer control is improved. Through the calculation and adjustment of dynamic safety thresholds, the safety and reliability of transformer operation are improved, while the operating efficiency is optimized, rapid response and decision adjustment to complex scenarios are realized, and the flexibility and adaptability of the system are improved.

[0041] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0042] (1) Deploy smart sensor networks at multiple target locations of the transformer and perform high-frequency sampling of voltage, current, temperature, load, and on-load tap changer position parameters to obtain multi-dimensional raw parameter data streams;

[0043] (2) Perform edge computing preprocessing on the multi-dimensional original parameter data stream to obtain real-time parameter data after preliminary filtering, and perform adaptive threshold recognition and median replacement processing on the outliers in the real-time parameter data after preliminary filtering to obtain corrected continuous parameter data;

[0044] (3) Perform timestamp alignment and multi-source data fusion on the corrected continuous parameter data to obtain a synchronized multidimensional collaborative data set, and perform wavelet transform denoising and Kalman filter smoothing on the synchronized multidimensional collaborative data set to obtain a high signal-to-noise ratio data set;

[0045] (4) Extract time-frequency domain features from the high signal-to-noise ratio data set to obtain a reduced-dimensional feature vector set, and standardize the reduced-dimensional feature vector set to obtain a standardized feature data set;

[0046] (5) Perform adaptive time window segmentation and overlapping sliding window processing on the standardized feature data set to obtain a time series feature sample set, and perform compressed sensing encoding on the time series feature sample set to obtain the original monitoring data set.

[0047] Specifically, various types of smart sensors are installed at key locations of the transformer. These sensors include voltage sensors, current sensors, temperature sensors, load sensors, and sensors for detecting the position of on-load tap changers. These sensors sample data at a high frequency to generate a multi-dimensional raw parameter data stream containing rich information. The multi-dimensional raw parameter data stream is pre-processed by edge computing. Edge computing transfers data processing tasks from the central server to devices close to the data source, thereby reducing data transmission latency and bandwidth requirements. In this process, preliminary data filtering is performed to remove obvious noise and outliers. Adaptive threshold identification and median replacement processing are performed on outliers in the real-time parameter data after preliminary filtering. Adaptive threshold identification identifies and replaces outliers by dynamically adjusting the threshold. For example, for the data stream of a certain voltage sensor , you can set a dynamic threshold ,in is the mean value of the voltage data, is the standard deviation, is the adjustment coefficient. When the threshold is exceeded, it is considered as an outlier and replaced with the median to obtain the corrected continuous parameter data. The corrected continuous parameter data is timestamp aligned and multi-source data fused. Timestamp alignment ensures that data from different sensors are compared on the same time scale, while multi-source data fusion combines these data to form a synchronized multi-dimensional collaborative data set. The synchronized multi-dimensional collaborative data set is subjected to wavelet transform denoising and Kalman filter smoothing. Wavelet transform can effectively remove high-frequency noise in the data, while Kalman filter smoothes the time series through a recursive algorithm to generate a high signal-to-noise ratio data set. Time-frequency domain feature extraction is performed on the high signal-to-noise ratio data set. By analyzing the characteristics of the data at different frequencies and times, a reduced dimension feature vector set is generated. For example, the time domain signal is converted into a frequency domain signal through Fourier transform, and the main frequency components are extracted from it. The reduced dimension feature vector set is standardized to generate a standardized feature data set. Standardization makes the data have the same scale in each dimension by subtracting the mean and dividing by the standard deviation, thereby improving the accuracy of subsequent analysis. Adaptive time window segmentation and overlapping sliding window processing are performed on the standardized feature data set. Adaptive time window segmentation dynamically adjusts the size of the time window according to the changes in the data, while overlapping sliding window processing captures subtle changes in the data by introducing overlapping parts between different windows. To standardize feature data, in the window Its adaptive time window segmentation can be expressed as ,in is the dynamically adjusted window width, and the overlapping sliding windows can be expressed as . Perform compressed sensing coding on the time series feature sample set to obtain the original monitoring data set. Compressed sensing coding reduces the burden of data storage and transmission by sparsely representing high-dimensional data. is the vector representation of the time series feature sample set, through the compressed sensing matrix , we can get the compressed representation ,in is the compressed data, It is a compressed sensing matrix, and the original data can be reconstructed through the optimization algorithm.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] (1) Perform wavelet transform time series decomposition on the original monitoring data set to obtain multi-scale trend terms, seasonal terms, and random terms;

[0050] (2) The original monitoring data set is seasonally adjusted and fractional-order difference processed according to the multi-scale trend term, seasonal term, and random term to obtain a stationary time series;

[0051] (3) Perform autocorrelation function and partial autocorrelation function analysis on the stationary time series to obtain the model order parameters, and perform maximum likelihood estimation on the stationary time series based on the model order parameters to obtain preliminary load forecast results;

[0052] (4) Adaptively segment the original monitoring data set into time windows and extract multi-dimensional features to obtain a deep learning input tensor, and then input the deep learning input tensor into a temporal convolutional network and long short-term memory network hybrid model architecture for multi-step prediction to obtain a deep learning prediction result;

[0053] (5) Perform dynamic weight integration learning on the preliminary load forecast results and the deep learning forecast results to obtain the short-term load forecast results;

[0054] (6) Online transfer learning and adaptive adjustment are performed based on the short-term load forecast results, historical prediction errors and transformer health index to obtain the initial load forecast model.

[0055] Specifically, the original monitoring data set is decomposed by wavelet transform time series. Wavelet transform is a mathematical tool that can decompose signals into different frequency components. Through this decomposition, various characteristics in the data are captured. Wavelet transform is used to decompose the data into multi-scale trend terms, seasonal terms, and random terms. For example, suppose is the original monitoring data, and the decomposition form obtained by wavelet transform can be expressed as:

[0056] ;

[0057] in, It is a multi-scale trend term, reflecting the long-term trend of data; is the seasonal term, reflecting cyclical fluctuations; is a random term, representing noise and short-term fluctuations. According to the multi-scale trend term, seasonal term and random term, the original monitoring data set is seasonally adjusted and fractional-order difference processed to obtain a stationary time series. Seasonal adjustment makes the data more stable by eliminating the periodic components in the data. Fractional-order difference processing is an improved difference method that more effectively removes trend components and makes the time series more stable by performing non-integer-order difference operations on the data. Assume For seasonally adjusted data, the fractional difference processing can be expressed as:

[0058] ;

[0059] in, is the backshift operator, is the order of fractional differences, by adjusting The value of can control the effect of the difference. Perform autocorrelation function and partial autocorrelation function analysis on the stationary time series to determine the order parameter of the model. The autocorrelation function is used to analyze the correlation of the time series at different lags, while the partial autocorrelation function is used to analyze the correlation after excluding the influence of the intermediate lag terms. The order parameter of the time series model is determined by the graphs of the autocorrelation function and the partial autocorrelation function, such as the autoregressive order of the ARIMA model. and the moving average order . After determining the model order parameters, the maximum likelihood estimation is performed on the stationary time series to obtain the preliminary load forecast results. The maximum likelihood estimation achieves the fitting and prediction of the time series model by finding the parameter value that maximizes the probability of the observed data. The original monitoring data set is adaptively segmented in time windows and multi-dimensional feature extracted to obtain the deep learning input tensor. Adaptive time window segmentation captures the changing characteristics of the data by dynamically adjusting the size of the time window, while multi-dimensional feature extraction extracts rich feature information by analyzing and processing the data in multiple dimensions. Assume For the original monitoring data, the adaptive time window segmentation can be expressed as:

[0060] ;

[0061] in, is the length of the adaptive time window. By performing multi-dimensional feature extraction on the segmented data, the input tensor for deep learning is obtained. The deep learning input tensor is input into the temporal convolutional network and long short-term memory network hybrid model architecture for multi-step prediction to obtain the deep learning prediction result. The temporal convolutional network captures the local features of the data through convolution operations, while the long short-term memory network handles the long-term dependencies in the time series through memory and forgetting mechanisms. The preliminary load forecast results and the deep learning forecast results are fused by dynamic weight ensemble learning to obtain the short-term load forecast results. Dynamic weight ensemble learning improves the accuracy of the prediction by assigning weights to different prediction results and combining the advantages of different models. Assume This is the preliminary load forecast result. Deep learning prediction results, short-term load prediction results It can be expressed as:

[0062] ;

[0063] in, and is the dynamically adjusted weight coefficient. Online transfer learning and adaptive adjustment are performed based on the short-term load forecast results, historical forecast errors, and transformer health index to obtain the initial load forecast model. Online transfer learning continuously learns new data and adjusts model parameters so that the model can adapt to the new data environment. Adaptive adjustment improves the accuracy and robustness of the forecast model through real-time feedback and dynamic adjustment. For model parameters, through online transfer learning and adaptive adjustment, the model parameters can be continuously updated to obtain a more accurate load prediction model.

[0064] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0065] (1) Perform multi-dimensional feature extraction on the original monitoring data set to obtain a standardized feature matrix, and generate preliminary health assessment results based on the standardized feature matrix and preset weight coefficients;

[0066] (2) Classify the preliminary health assessment results and historical fault data to obtain the health status classification results, and perform adaptive neuro-fuzzy reasoning based on the health status classification results and the pre-established expert knowledge base to obtain the fuzzy health index;

[0067] (3) Perform time series analysis on the fuzzy health index and short-term load forecast results to obtain health trend forecast;

[0068] (4) Perform dynamic stress-strength interference analysis based on health trend prediction, transformer rated parameters and real-time load data to obtain preliminary safety margin;

[0069] (5) Perform kernel density estimation on the preliminary safety margin to obtain the safety margin probability distribution, and perform adaptive confidence interval calculation based on the safety margin probability distribution to obtain the dynamic safety threshold;

[0070] (6) Perform nonlinear weighted fusion of the fuzzy health index and dynamic safety threshold to obtain a comprehensive health score;

[0071] (7) Online correction and adaptive update are performed based on the comprehensive health score, historical operating data and initial load prediction model to obtain the transformer health index and dynamic safety threshold.

[0072] Specifically, multidimensional feature extraction is performed on the original monitoring data set. The original monitoring data set includes parameters such as voltage, current, temperature, load, and on-load tap changer position. Through multidimensional feature extraction, these parameters are converted into feature matrices. Principal component analysis or other feature extraction methods are used to reduce the dimensionality of high-dimensional data. Assume that there is an original data matrix , whose size is ,in is the number of time steps, is the number of parameters. Through feature extraction, the standardized feature matrix is ​​obtained , whose size is ,in is the number of features after extraction. Standardized feature matrix Each column of represents a standardized feature, which is standardized by subtracting the mean and dividing by the standard deviation. Using the preset weight coefficient The standardized feature matrix is ​​weighted summed to generate the preliminary health assessment results. is a weight vector whose size is , the initial health assessment results It can be expressed as: .in, is a size of The vector of represents the health assessment result at each time step. The preliminary health assessment results and historical fault data are classified to obtain the health status classification results. By analyzing the historical fault data, the assessment results are divided into different health states, such as "normal", "warning" and "fault". This classification task is achieved by using machine learning classification algorithms, such as support vector machines or random forests. According to the health status classification results and the pre-established expert knowledge base, adaptive neuro-fuzzy reasoning is performed to obtain the fuzzy health index. The expert knowledge base contains a large number of rules and experience. Through the neuro-fuzzy reasoning system, these rules are applied to the current data to generate the fuzzy health index. Time series analysis is performed on the fuzzy health index and short-term load forecast results to predict future health trends. Time series analysis can use an autoregressive integral moving average model or other prediction models. The health trend prediction is obtained by modeling and predicting the historical data. Based on the health trend prediction, the rated parameters of the transformer and the real-time load data, a dynamic stress-strength interference analysis is performed to obtain the preliminary safety margin. The dynamic stress-strength interference analysis evaluates the safety of the transformer under different load conditions and calculates the safety margin by comparing the real-time load with the rated parameters of the transformer. The preliminary safety margin is subjected to kernel density estimation to obtain the probability distribution of the safety margin. Kernel density estimation is a nonparametric statistical method used to estimate the probability density function of data. The probability distribution function of the safety margin is obtained through kernel density estimation. According to the probability distribution of the safety margin, an adaptive confidence interval is calculated to obtain the dynamic safety threshold. The adaptive confidence interval calculates the threshold interval of the safety margin according to different confidence levels, thereby evaluating the safety of the transformer under different load conditions. The fuzzy health index and the dynamic safety threshold are nonlinearly weighted fused to obtain a comprehensive health score. Nonlinear weighted fusion combines different indicators through nonlinear functions to generate a unified score. Online correction and adaptive update are performed based on the comprehensive health score, historical operating data and initial load prediction model to obtain the transformer health index and dynamic safety threshold. Online correction improves the accuracy and robustness of the prediction model by adjusting the model parameters in real time. Adaptive update optimizes the performance of the model by continuously learning new data to ensure its accuracy under different operating conditions.

[0073] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0074] (1) Perform kernel density estimation on the short-term load forecasting results of the initial load forecasting model to obtain the load forecasting probability density function;

[0075] (2) Perform simulation based on the load prediction probability density function and transformer rated parameters to obtain multiple groups of load scenarios;

[0076] (3) Comprehensively evaluate multiple groups of load scenarios and transformer health indexes to obtain a risk assessment matrix for each scenario;

[0077] (4) Construct an initial multi-objective optimization model based on the risk assessment matrix and the dynamic safety threshold, and calculate the optimal solution of the initial multi-objective optimization model to obtain the optimal solution set;

[0078] (5) Perform multi-criteria decision analysis based on the optimal solution set and the preset decision preference weights to obtain candidate transformer switching decision schemes, and calculate the state transition probability matrix and time-varying reward function of the candidate transformer switching decision schemes;

[0079] (6) Generate the Q-value function of the continuous action space based on the state transition probability matrix and the time-varying reward function, and perform noise injection exploration on the Q-value function to obtain the reinforcement learning strategy;

[0080] (7) According to the reinforcement learning strategy and the candidate transformer switching decision schemes, the solution is obtained to obtain the first transformer switching decision scheme.

[0081] Specifically, kernel density estimation is performed on the short-term load forecasting results of the initial load forecasting model. Kernel density estimation is a non-parametric statistical method used to estimate the probability density function of a random variable. is the short-term load forecast result of the initial load forecast model, then the load forecast probability density function It can be obtained by kernel density estimation:

[0082] ;

[0083] in, It is a kernel function. Commonly used kernel functions include Gaussian kernel, uniform kernel, etc. is the bandwidth parameter, which controls the smoothness of the estimation. The simulation is performed based on the load prediction probability density function and the rated parameters of the transformer to generate multiple sets of load scenarios. The simulation simulates different load conditions by randomly sampling on the load prediction probability density function. Assume that the rated load of the transformer is , multiple groups of load scenarios can be obtained through simulation Comprehensively evaluate multiple groups of load scenarios and transformer health indexes to generate a risk assessment matrix for each scenario. Transformer health index Can be combined with load scenarios Conduct risk assessment and generate a risk assessment matrix , whose elements In the load scenario Next, the health index For example, risk assessment can calculate overload risk and health decline risk by comparing the thresholds of load and health index. and dynamic safety thresholds , construct the initial multi-objective optimization model. The goal of the multi-objective optimization model is to minimize the risk and ensure the safety of the system at the same time. The optimization problem can be expressed as:

[0084] ;

[0085] in, is the decision preference weight, which indicates the importance attached to different risks. The optimal solution of the initial multi-objective optimization model is calculated to obtain the optimal solution set , each optimal solution Represents an optimized transformer strategy. Based on the optimal solution set and the preset decision preference weights, a multi-criteria decision analysis is performed to obtain candidate transformer switching decision schemes. The multi-criteria decision analysis selects the optimal transformer switching strategy by comprehensively considering the weights of different criteria. Calculate the state transition probability matrix of the candidate transformer switching decision scheme and time-varying reward function . State transition probability matrix Describes the probability of voltage switching in different states, time-varying reward function The benefits of making decisions at different times are measured. According to the state transition probability matrix and time-varying reward function , generating the Q-value function of the continuous action space Q-value function By measuring the status Take Action The long-term return provides a basis for decision-making. Noise injection exploration is performed on the Q-value function to avoid falling into the local optimal solution, thereby obtaining a reinforcement learning strategy. Noise injection can be done by adding random noise To achieve this, the formula is:

[0086] ;

[0087] in, is the Q value function after noise injection, is the noise term. According to the reinforcement learning strategy and the candidate transformer switching decision scheme, the solution is obtained to obtain the first transformer switching decision scheme. The reinforcement learning strategy selects the optimal transformer switching action under different conditions through continuous learning and optimization, thereby ensuring the safe and efficient operation of the transformer.

[0088] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0089] (1) The first transformer switching decision plan is decomposed into time series and divided into scenarios to obtain a sub-decision sequence of three stages: preparation, execution, and completion;

[0090] (2) A nonlinear state space model is constructed based on the sub-decision sequence, transformer health index, and dynamic safety threshold to obtain the piecewise control state equation;

[0091] (3) Perform model predictive control on the segmented control state equation to obtain the initial control sequence, and perform multi-time scale rolling optimization based on the initial control sequence, real-time load data, and short-term load forecast results to obtain the optimized control parameters;

[0092] (4) Perform polynomial chaos expansion analysis on the optimized control parameters to obtain the parameter adjustment range and probability distribution, and construct a nonlinear adaptive control law based on the parameter adjustment range and transformer response characteristics;

[0093] (5) Perform random stability analysis and solve the nonlinear adaptive control law to obtain the probabilistic stability guarantee interval, and perform multi-objective switching timing optimization based on the probabilistic stability guarantee interval and the dynamic characteristic curve of the on-load tap changer to obtain the target switching time point;

[0094] (6) Perform closed-loop simulation verification on the target switching time point to obtain an adaptive switching control instruction. According to the adaptive switching control instruction, the transformer is controlled in real time and the voltage, current, switch status and vibration data during the switching process are collected to obtain the switching process data.

[0095] Specifically, the entire decision-making process is divided into three sub-decision sequences: preparation, execution, and completion. In the preparation stage, the operating data and status information of the transformer are prepared, the execution stage is the actual process of transformer operation, and the completion stage is the evaluation and feedback of the transformer operation effect. According to the sub-decision sequence, transformer health index, and dynamic safety threshold, a nonlinear state space model is constructed to describe the state changes of the transformer in different operation stages. The nonlinear state space model can be expressed by the following equation:

[0096] ;

[0097] in, is the system state vector, indicating that the transformer is The state of the moment; is the control input vector, indicating Control operation at all times; and are process noise and measurement noise respectively; function and The state transition and observation model of the system are described respectively. The initial control sequence is obtained by performing model predictive control on the piecewise control state equation. Model predictive control generates a set of optimized control operations by predicting the behavior of the system in a period of time in the future. Using the rolling optimization technology, the initial control sequence is subjected to multi-time scale rolling optimization based on real-time load data and short-term load forecast results. The rolling optimization method obtains the optimized control parameters by recalculating the optimal control strategy at each time step to adapt to the dynamic changes of the system. The optimized control parameters are subjected to polynomial chaos expansion analysis to obtain the parameter adjustment range and probability distribution. Polynomial chaos expansion analysis is a method to model system uncertainty as a polynomial chaotic sequence. This method can estimate the range and probability distribution of system parameters under different conditions. Assumptions is the optimized control parameter, which can be expanded by polynomial chaos using the following formula:

[0098] ;

[0099] in, is the chaotic basis function, is the expansion coefficient, is a random variable. According to the parameter adjustment range and the transformer response characteristics, a nonlinear adaptive control law is constructed. The nonlinear adaptive control law enables the system to maintain stable and efficient operation under different conditions by adjusting the control parameters in real time. The nonlinear adaptive control law is subjected to random stability analysis and solution to determine the stability range of the system under different operating conditions and obtain the probabilistic stability guarantee interval. The random stability analysis evaluates the robustness and stability of the system by introducing random disturbances. According to the probabilistic stability guarantee interval and the dynamic characteristic curve of the on-load tap changer, a multi-objective switching timing optimization is performed to determine the optimal target switching time point. The multi-objective switching timing optimization determines the optimal switching time and maximizes the operating performance of the system by comprehensively considering the stability, efficiency and safety of the system. The target switching time point is verified by closed-loop simulation to ensure the effectiveness of the optimization strategy. The closed-loop simulation verifies the control strategy in the simulation environment, evaluates its performance in actual operation, and obtains the adaptive switching control instruction. The adaptive switching control instruction is adjusted according to the simulation results to ensure its effectiveness and reliability in actual operation. The transformer is controlled in real time according to the adaptive switching control instruction, and the voltage, current, switch state and vibration data during the switching process are collected to obtain the switching process data. Through real-time monitoring and data collection, the operating status of the transformer during the switching process can be fully understood.

[0100] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0101] (1) Multi-source heterogeneous data fusion is performed on the switching process data, original monitoring data set, and short-term load forecasting results to obtain a spatiotemporal feature data set;

[0102] (2) Based on the spatiotemporal feature data set and the adaptive switching control instructions, a temporal causal graph and a semantic network are constructed to obtain an initial multi-layer knowledge graph. Dynamic node embedding and edge attribute learning are performed on the initial multi-layer knowledge graph to obtain an evolving knowledge graph.

[0103] (3) Perform graph reasoning and relationship extraction based on the evolving knowledge graph to obtain the dynamic operation rules and potential fault evolution patterns of the transformer;

[0104] (4) Model the dynamic operation rules and potential fault evolution patterns of the transformer in time and space to obtain an intermediate load prediction model;

[0105] (5) Perform online ensemble learning and migration optimization based on the intermediate load prediction model, the initial load prediction model, and the historical prediction error to obtain the target load prediction model;

[0106] (6) Conduct multi-agent collaborative learning on the target load prediction model, evolutionary knowledge graph, and transformer health index to obtain an optimized decision-making strategy;

[0107] (7) Searching based on the optimized decision strategy and dynamic safety threshold to obtain a multi-layer candidate decision tree, and dynamically programming the multi-layer candidate decision tree to obtain a decision solution set;

[0108] (8) Adaptively select and adjust parameters according to the decision plan set, the first transformer switching decision plan, and the real-time operation data to obtain a second transformer switching decision plan.

[0109] Specifically, multi-source heterogeneous data fusion is performed on the switching process data, the original monitoring data set and the short-term load prediction results. The switching process data includes voltage, current, switch state and vibration data, the original monitoring data set includes parameters such as voltage, current, temperature, load and on-load tap changer position, and the short-term load prediction result is the predicted value of the transformer load in the future. Through data preprocessing, these data from different sources are converted into a unified format, and then multi-source heterogeneous data fusion is performed to generate a spatiotemporal feature data set. According to the spatiotemporal feature data set and the adaptive switching control instructions, a temporal causal graph and a semantic network are constructed to reveal the causal relationship and semantic association between the data. The temporal causal graph is used to capture the causal relationship in the time series data, while the semantic network constructs a multi-level knowledge structure through the semantic association of nodes and edges. The initial multi-layer knowledge graph can be constructed by the following steps: according to each feature in the spatiotemporal feature data set, a corresponding node is created; according to the causal relationship and semantic association between the features, edges between nodes are added; and through iterative optimization, the initial multi-layer knowledge graph is generated. In order to further improve the dynamic and real-time performance of the knowledge graph, dynamic node embedding and edge attribute learning are performed on the initial multi-layer knowledge graph. Dynamic node embedding reflects the dynamic relationship between nodes by embedding nodes into a high-dimensional space and continuously updating the positions of nodes according to real-time data. Edge attribute learning learns the attributes of edges, such as weight and direction, to more accurately describe the relationship between nodes. Through these steps, an evolving knowledge graph is obtained. Based on the evolving knowledge graph, graph reasoning and relationship extraction are performed to identify the dynamic operation laws and potential fault evolution patterns of the transformer. Graph reasoning uses the causal relationships and semantic associations in the knowledge graph to reason about nodes and edges and identify the behavior patterns of the transformer under different operating conditions. Relationship extraction identifies potential fault modes and evolution paths by analyzing the attributes of nodes and edges. Spatiotemporal sequence modeling is performed on the dynamic operation laws and potential fault evolution patterns of the transformer to obtain an intermediate load prediction model. Spatiotemporal sequence modeling establishes a model that can predict future loads by analyzing the temporal and spatial change patterns in the spatiotemporal feature data set. Assumptions is a spatiotemporal feature dataset, and the spatiotemporal sequence model can be expressed as:

[0110] ;

[0111] in, is the load forecast result, is the spatiotemporal sequence model function, is a model parameter. Based on the intermediate load prediction model, the initial load prediction model and the historical prediction error, online ensemble learning and migration optimization are performed to obtain the target load prediction model. Online ensemble learning improves the prediction accuracy by combining the results of multiple prediction models. Migration optimization continuously adjusts the model parameters to adapt it to the new data environment, thereby improving the real-time and robustness of the model. Assume and They are the prediction results of the intermediate load prediction model and the initial load prediction model, and the target load prediction model It can be expressed as:

[0112] ;

[0113] in, and is the weight coefficient, is the historical prediction error. The target load prediction model is subjected to multi-agent collaborative learning with the evolutionary knowledge graph and the transformer health index to obtain an optimized decision strategy. Multi-agent collaborative learning optimizes the decision process and improves the accuracy and efficiency of decision-making through the collaborative work of multiple agents. Agents can make more comprehensive analysis and decisions by sharing information and collaborative learning. Search is performed based on the optimized decision strategy and dynamic safety threshold to generate a multi-layer candidate decision tree. The multi-layer candidate decision tree is screened layer by layer to ensure that each decision point can get the optimal solution. Dynamic programming is performed on the multi-layer candidate decision tree to obtain the final decision solution set. Dynamic programming finds the optimal path of the entire decision tree through global optimization, thereby maximizing the operating performance of the system. According to the decision solution set, the first transformer switching decision solution and real-time operating data, adaptive selection and parameter adjustment are performed to obtain the second transformer switching decision solution. Adaptive selection dynamically adjusts decision parameters by combining current operating data and historical decision experience to ensure safe and stable operation of the transformer under different operating conditions.

[0114] The above describes the voltage transformation method of the novel protective transformer in the embodiment of the present application. The following describes the voltage transformation device of the novel protective transformer in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the transformer device of the novel protective transformer includes:

[0115] The acquisition module 201 is used to collect multi-dimensional parameters of the transformer in real time to obtain an original monitoring data set;

[0116] Modeling module 202, used to perform time series analysis and deep learning modeling on the original monitoring data set to obtain an initial load forecasting model and short-term load forecasting results;

[0117] The calculation module 203 is used to perform multi-dimensional evaluation and safety margin calculation on the original monitoring data set and the short-term load forecast result to obtain the transformer health index and dynamic safety threshold;

[0118] The optimization module 204 is used to perform multi-objective optimization and reinforcement learning on the initial load prediction model, the transformer health index and the dynamic safety threshold to obtain a first transformer switching decision plan;

[0119] An adjustment module 205 is used to perform segmented control and real-time parameter adjustment on the first voltage transformation switching decision scheme to obtain an adaptive switching control instruction and switching process data;

[0120] The construction module 206 is used to perform online learning and knowledge graph construction on the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain the target load prediction model and the second transformer switching decision plan.

[0121] Through the collaboration of the above components, the high-frequency sampling and real-time preprocessing of the key parameters of the transformer are realized through the intelligent sensor network and edge computing technology. Combined with time series analysis and deep learning technology, multi-scale and multi-model load prediction is realized, and the prediction accuracy and robustness are improved. Through multi-dimensional evaluation and safety margin calculation, combined with expert knowledge and data-driven methods, the dynamic evaluation of the health status of the transformer is realized, and the operation safety is improved. The dynamic optimization of transformer switching decision is realized by adopting multi-objective optimization and reinforcement learning methods. Through segmented control and real-time parameter adjustment, combined with model predictive control and adaptive control technology, the precise control of the transformer switching process is realized, and the impact of switching on the system is reduced. Using online learning and knowledge graph technology, the dynamic accumulation and optimization of transformer operation knowledge is realized, and the adaptive ability and intelligent decision-making level of the system are improved. Through the analysis and feedback of the switching process data, the continuous optimization of the decision-making plan is realized, and the long-term effect of transformer control is improved. Through the calculation and adjustment of dynamic safety thresholds, the safety and reliability of transformer operation are improved, and the operation efficiency is optimized at the same time, and the rapid response and decision adjustment to complex scenarios are realized, which improves the flexibility and adaptability of the system.

[0122] The present application also provides a new type of protective transformer transformation device, the new type of protective transformer transformation device includes a memory and a processor, the memory stores computer-readable instructions, when the computer-readable instructions are executed by the processor, the processor executes the steps of the new type of protective transformer transformation method in the above-mentioned embodiments.

[0123] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the voltage transformation method of the novel protective transformer.

[0124] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0125] 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 application is essentially 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 to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0126] 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 new type of protective transformer transformation method, characterized in that: The method comprises: The transformer is subjected to real-time multi-dimensional parameter acquisition to obtain an original monitoring data set; specifically, the method comprises: deploying an intelligent sensor network at multiple target locations of the transformer, and performing high-frequency sampling on the voltage, current, temperature, load and on-load tap changer position parameters to obtain a multi-dimensional original parameter data stream; performing edge computing preprocessing on the multi-dimensional original parameter data stream to obtain real-time parameter data after preliminary filtering, and performing adaptive threshold recognition and median replacement processing on the abnormal values ​​in the real-time parameter data after preliminary filtering to obtain corrected continuous parameter data; performing timestamp alignment and multi-source data fusion on the corrected continuous parameter data to obtain a synchronous multi-dimensional collaborative data set, and performing wavelet transform denoising and Kalman filter smoothing processing on the synchronous multi-dimensional collaborative data set to obtain a high signal-to-noise ratio data set; performing time-frequency domain feature extraction on the high signal-to-noise ratio data set to obtain a reduced dimension feature vector set, and performing standardization processing on the reduced dimension feature vector set to obtain a standardized feature data set; performing adaptive time window segmentation and overlapping sliding window processing on the standardized feature data set to obtain a time series feature sample set, and performing compressed sensing encoding on the time series feature sample set to obtain an original monitoring data set; Performing time series analysis and deep learning modeling on the original monitoring data set to obtain an initial load forecasting model and short-term load forecasting results; specifically including: performing wavelet transform time series decomposition on the original monitoring data set to obtain multi-scale trend terms, seasonal terms and random terms; performing seasonal adjustment and fractional-order difference processing on the original monitoring data set according to the multi-scale trend terms, the seasonal terms and the random terms to obtain a stationary time series; performing autocorrelation function and partial autocorrelation function analysis on the stationary time series to obtain model order parameters, and performing maximum likelihood estimation on the stationary time series according to the model order parameters to obtain preliminary load forecasting results; performing adaptive time window segmentation and multi-dimensional feature extraction on the original monitoring data set to obtain a deep learning input tensor, and inputting the deep learning input tensor into a temporal convolutional network and a long short-term memory network hybrid model architecture for multi-step prediction to obtain a deep learning prediction result; performing dynamic weight integration learning fusion on the preliminary load forecasting results and the deep learning prediction results to obtain a short-term load forecasting result; performing online transfer learning and adaptive adjustment based on the short-term load forecasting results, historical prediction errors and transformer health index to obtain an initial load forecasting model; The original monitoring data set and the short-term load forecast result are subjected to multi-dimensional evaluation and safety margin calculation to obtain the transformer health index and dynamic safety threshold; specifically, the method comprises: performing multi-dimensional feature extraction on the original monitoring data set to obtain a standardized feature matrix, and generating a preliminary health assessment result according to the standardized feature matrix and a preset weight coefficient; classifying the preliminary health assessment result and historical fault data to obtain a health status classification result, and performing adaptive neural fuzzy reasoning according to the health status classification result and a pre-established expert knowledge base to obtain a fuzzy health index; performing time-domain neural network reasoning on the fuzzy health index and the short-term load forecast result; Sequence analysis is performed to obtain a health trend prediction; dynamic stress-strength interference analysis is performed based on the health trend prediction, transformer rated parameters and real-time load data to obtain a preliminary safety margin; kernel density estimation is performed on the preliminary safety margin to obtain a safety margin probability distribution, and adaptive confidence interval calculation is performed based on the safety margin probability distribution to obtain a dynamic safety threshold; nonlinear weighted fusion is performed on the fuzzy health index and the dynamic safety threshold to obtain a comprehensive health score; online correction and adaptive update are performed based on the comprehensive health score, historical operating data and the initial load prediction model to obtain a transformer health index and a dynamic safety threshold; Performing multi-objective optimization and reinforcement learning on the initial load prediction model, the transformer health index and the dynamic safety threshold to obtain a first transformer switching decision plan; Performing segmented control and real-time parameter adjustment on the first transformer switching decision scheme to obtain adaptive switching control instructions and switching process data; Online learning and knowledge graph construction are performed on the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision plan.

2. The voltage transformation method of the novel protective transformer according to claim 1 is characterized in that: The multi-objective optimization and reinforcement learning of the initial load prediction model, the transformer health index and the dynamic safety threshold are performed to obtain a first transformer switching decision scheme, including: Performing kernel density estimation on the short-term load forecasting result of the initial load forecasting model to obtain a load forecasting probability density function; Perform simulation according to the load prediction probability density function and the transformer rated parameters to obtain multiple groups of load scenarios; Comprehensively evaluate the multiple groups of load scenarios and the transformer health index to obtain a risk assessment matrix for each scenario; Constructing an initial multi-objective optimization model according to the risk assessment matrix and the dynamic safety threshold, and calculating an optimal solution for the initial multi-objective optimization model to obtain an optimal solution set; Performing a multi-criteria decision analysis based on the optimal solution set and the preset decision preference weights to obtain a candidate transformer switching decision scheme, and calculating a state transition probability matrix and a time-varying reward function of the candidate transformer switching decision scheme; Generate a Q-value function of a continuous action space according to the state transition probability matrix and the time-varying reward function, and perform noise injection exploration on the Q-value function to obtain a reinforcement learning strategy; A solution is performed according to the reinforcement learning strategy and the candidate transformer switching decision schemes to obtain a first transformer switching decision scheme.

3. The voltage transformation method of the novel protective transformer according to claim 2 is characterized in that: The step of performing segmented control and real-time parameter adjustment on the first transformer switching decision scheme to obtain an adaptive switching control instruction and switching process data includes: Performing time sequence decomposition and scenario division on the first transformer switching decision scheme to obtain sub-decision sequences of three stages: preparation, execution and completion; Constructing a nonlinear state space model according to the sub-decision sequence, the transformer health index and the dynamic safety threshold to obtain a segmented control state equation; Performing model predictive control on the segmented control state equation to obtain an initial control sequence, and performing multi-time scale rolling optimization based on the initial control sequence, real-time load data and the short-term load forecast result to obtain optimized control parameters; Performing polynomial chaos expansion analysis on the optimized control parameters to obtain parameter adjustment range and probability distribution, and constructing a nonlinear adaptive control law according to the parameter adjustment range and transformer response characteristics; Performing random stability analysis and solving the nonlinear adaptive control law to obtain a probabilistic stability guarantee interval, and performing multi-objective switching timing optimization according to the probabilistic stability guarantee interval and the dynamic characteristic curve of the on-load tap changer to obtain a target switching time point; A closed-loop simulation is performed on the target switching time point to obtain an adaptive switching control instruction, and the transformer is controlled in real time according to the adaptive switching control instruction, and the voltage, current, switch state and vibration data during the switching process are collected to obtain the switching process data.

4. The voltage transformation method of the novel protective transformer according to claim 3 is characterized in that: The online learning and knowledge graph construction of the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision scheme include: Performing multi-source heterogeneous data fusion on the switching process data, the original monitoring data set and the short-term load forecast result to obtain a spatiotemporal feature data set; Constructing a temporal causal graph and a semantic network according to the spatiotemporal feature data set and the adaptive switching control instruction to obtain an initial multi-layer knowledge graph, and performing dynamic node embedding and edge attribute learning on the initial multi-layer knowledge graph to obtain an evolving knowledge graph; Performing graph reasoning and relationship extraction based on the evolutionary knowledge graph to obtain the transformer dynamic operation law and potential fault evolution mode; Performing spatiotemporal sequence modeling on the dynamic operation law of the transformer and the potential fault evolution mode to obtain an intermediate load prediction model; Perform online integrated learning and migration optimization according to the intermediate load prediction model, the initial load prediction model and the historical prediction error to obtain a target load prediction model; Performing multi-agent collaborative learning on the target load prediction model, the evolution knowledge graph and the transformer health index to obtain an optimized decision-making strategy; Searching according to the optimized decision strategy and the dynamic safety threshold to obtain a multi-layer candidate decision tree, and dynamically programming the multi-layer candidate decision tree to obtain a decision solution set; Adaptive selection and parameter adjustment are performed according to the decision plan set, the first voltage transformation switching decision plan and real-time operation data to obtain a second voltage transformation switching decision plan.

5. A new type of protective transformer transformer device, characterized in that: The device is used to perform the voltage transformation method of the novel protective transformer according to any one of claims 1 to 4, comprising: The acquisition module is used to collect multi-dimensional parameters of the transformer in real time to obtain the original monitoring data set; A modeling module, used to perform time series analysis and deep learning modeling on the original monitoring data set to obtain an initial load forecasting model and a short-term load forecasting result; A calculation module, used to perform multi-dimensional evaluation and safety margin calculation on the original monitoring data set and the short-term load forecast result to obtain a transformer health index and a dynamic safety threshold; An optimization module, used for performing multi-objective optimization and reinforcement learning on the initial load prediction model, the transformer health index and the dynamic safety threshold to obtain a first transformer switching decision plan; An adjustment module, used for performing segmented control and real-time parameter adjustment on the first transformer switching decision scheme to obtain adaptive switching control instructions and switching process data; A construction module is used to perform online learning and knowledge graph construction on the switching process data, the adaptive switching control instructions and the original monitoring data set to obtain a target load prediction model and a second transformer switching decision plan.

6. A new type of protective transformer transformer equipment, characterized in that: The transformer device of the novel protective transformer comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the voltage transformation device of the novel protective transformer to execute the voltage transformation method of the novel protective transformer according to any one of claims 1 to 4.

7. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, a voltage transformation method of a novel protective transformer as described in any one of claims 1 to 4 is implemented.