Voltage stabilizer intelligent detection and compensation method and system based on data deviation analysis

Abstract

The invention discloses a voltage stabilizer intelligent detection and compensation method and system based on data deviation analysis, and relates to the technical field of voltage stabilizer detection, and the method comprises the steps: obtaining multi-source data through the topology and working condition demands of a voltage detection system, and carrying out the collaborative preprocessing of the multi-source data, and constructing a full-working-condition effective data set; quantifying the total voltage deviation, establishing a full-working-condition deviation matrix, decomposing the full-working-condition deviation matrix into static and dynamic deviations, and constructing a deviation-influence factor dynamic correlation model; deviation characteristics are determined, and a fusion compensation model fusing the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model is constructed; real-time compensation is executed in combination with real-time collected data, and compensated data and a running state label are output; and on the basis of deviation data and trend prediction, constructing a closed-loop optimization link of passive correction and active pre-judgment. The method has the advantages that all-working-condition accurate adaptation is achieved, the dynamic response speed is high, working condition and element changes can be actively pre-judged, and fixed and real-time time-varying deviations are effectively offset.

Description

Technical Field

[0001] This invention relates to the field of voltage regulator testing technology, specifically to a method and system for intelligent detection and compensation of voltage regulators based on data deviation analysis. Background Technology

[0002] Voltage regulators are core voltage stabilizing components in power systems and industrial equipment, and their voltage stabilization accuracy directly affects the operational safety of downstream equipment. In actual operation, voltage regulators are prone to problems such as increased output voltage deviation and delayed dynamic adjustment response due to factors such as aging of internal components, input voltage fluctuations, sudden load changes, and changes in ambient temperature and humidity.

[0003] Existing intelligent voltage detection and compensation methods and systems based on data deviation analysis do not take into account the characteristics of voltage regulator input and output voltage coupling and steady-state / dynamic operating condition adaptation. They only compensate for general voltage data and cannot accurately reflect the actual operating status of the voltage regulator (such as the precursors of voltage regulator module failure and load adaptation deviation). This results in insufficient compensation specificity and poor long-term stability. Therefore, it is necessary to provide intelligent detection and compensation methods and systems for voltage regulators based on data deviation analysis to solve the above-mentioned problems. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a method and system for intelligent detection and compensation of voltage regulators based on data deviation analysis. This solution resolves the problems of existing intelligent voltage detection and compensation methods and systems based on data deviation analysis, which fail to consider the characteristics of voltage regulator input and output voltage coupling and steady-state / dynamic operating condition adaptation. They only compensate for general voltage data, making it difficult to accurately reflect the actual operating state of the voltage regulator (such as precursors to voltage regulator module failure and load adaptation deviations), resulting in insufficient compensation targeting and poor long-term stability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart detection and compensation method for voltage regulators based on data deviation analysis includes: By considering the voltage detection system topology and operating conditions, multi-source data is acquired, including raw voltage detection data, reference data, environmental interference data, and operating condition characteristic data. Collaborative preprocessing of multi-source data is performed to construct an effective dataset covering all operating conditions; Quantify the total voltage deviation, establish a full-condition deviation matrix, and decompose it into static deviation and dynamic deviation to construct a dynamic correlation model of deviation and influencing factors. Determine the static and dynamic bias characteristics, construct a fusion compensation model that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model, and achieve accurate cancellation of fixed bias and real-time time-varying bias. Acquire real-time data, combine it with the fusion compensation model, perform real-time compensation calculations on the voltage detection data, and output the compensated data and the regulator's operating status label; By combining real-time deviation data with trend prediction, a closed-loop optimization link of passive correction and active prediction is constructed to continuously update model parameters.

[0006] In an optional embodiment, the acquisition of multi-source data based on the voltage detection system topology and operating condition requirements specifically includes: Raw voltage detection data is collected at the deployed detection nodes at a frequency at least 10 times the power grid frequency to ensure that dynamic operating details are captured. The raw voltage detection data includes the instantaneous / RMS value of the voltage on the input side of the regulator, the instantaneous / RMS value of the voltage on the output side, the collection timestamp, the regulator number, the load condition information, and the regulator's operating status. Obtain the voltage range and accuracy requirements of the detection node, and select a standard voltage source that matches the rated output voltage range of the voltage regulator and has an accuracy level not lower than the preset accuracy level as the reference source. Under the same load conditions and acquisition frequency, reference voltage data is obtained through a reference source to ensure that the acquisition timing of the reference voltage data and the original voltage detection data is completely synchronized. The reference voltage data is a dynamic reference range of ±5% of the rated output voltage of the regulator, covering operating conditions including no load, steady-state conditions of 25% / 50% / 75% / rated load, and dynamic operating conditions of load change and input voltage fluctuation. During the raw data acquisition, environmental interference data, internal status data, and operating condition characteristic data are collected simultaneously. Perform time-series alignment and validity verification on multi-source data, remove data with time-series misalignment and failed acquisition, and establish a data ledger categorized by voltage regulator number, operating condition type, and acquisition time.

[0007] In an optional embodiment, the collaborative preprocessing of multi-source data to construct a valid dataset covering all operating conditions specifically includes: Based on the synchronous dataset, raw voltage data, reference data, and disturbance and operating condition characteristic data are extracted to construct the raw dataset, reference dataset, and auxiliary dataset, respectively. The significance level is preset to identify and remove outlier data points from the original dataset. The sliding window size is set to 50 data points. Data mutations within the window are detected in real time, marked as dynamic operating condition anomalies, and stored separately. The original dataset after removing outliers is subjected to four-level wavelet decomposition to separate high-frequency noise components from low-frequency effective components. High-frequency noise is further filtered out, and the preprocessed original detection data is reconstructed. The same outlier removal and noise reduction processes are performed on both the benchmark dataset and the auxiliary dataset to ensure data consistency and reconstruct the preprocessed benchmark data. The preprocessed raw test data and the reference data are verified a second time to ensure the consistency of the input and output voltage response timing of the voltage regulator. At the same time, the noise reduction efficiency of the raw test data and the reference data is determined. When the noise reduction efficiency is greater than or equal to the preset noise reduction efficiency threshold, the preprocessing is confirmed to be effective and the set of valid data under all operating conditions is output. Otherwise, the wavelet transform decomposition level is adjusted and the data is reprocessed.

[0008] In an optional embodiment, the quantification of total voltage deviation, the establishment of a full-condition deviation matrix, and the decomposition into static and dynamic deviations, and the construction of a dynamic correlation model between deviation and influencing factors, specifically include: Based on the preprocessed raw detection data and reference data, the total voltage deviation is determined point by point, and a full-condition deviation matrix of operating condition type-time-deviation value is constructed. The total voltage deviation is the difference between the detected voltage on the output side of the regulator and the rated output reference voltage. The deviation data of steady-state conditions in the full-condition deviation matrix are extracted, and the mean value of the fixed deviation data of output voltage under each steady-state condition is obtained. The static deviation is obtained by weighted averaging, which characterizes the inherent error of the component and the circuit drift. Based on the total voltage deviation and static deviation, the dynamic deviation is determined and a dynamic deviation dataset is constructed. The dataset is classified and labeled according to the dynamic operating condition type. The dynamic deviation is the time-varying deviation caused by input voltage fluctuation, load change, and internal temperature rise after removing the static deviation from the total deviation. Based on environmental disturbance data, internal state data, operating condition characteristic data and dynamic deviation dataset, the correlation coefficients between each environmental factor and dynamic deviation are determined. Based on the absolute value of the correlation coefficient, environmental factors with a correlation coefficient greater than or equal to the preset threshold are selected as key influencing factors. A time-series correlation model between key influencing factors and dynamic deviation is established to clarify the weight of each factor on dynamic deviation.

[0009] In an optional embodiment, the step of determining the static and dynamic deviation features and constructing a fusion compensation model that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model specifically includes: Based on static deviation and steady-state data under full operating conditions, the linear relationship between static deviation and preprocessed detection data is fitted to obtain static compensation coefficients. Furthermore, by solving for the static compensation coefficients and static constants that minimize the value of the quadratic linear function, the static compensation sub-model is determined. Based on the dataset of key influencing factors and dynamic deviations, a training set and a test set are divided, with the training set accounting for 70% and the test set accounting for 30%. The key influencing factors, working condition type codes, and deviation change rate are used as input features, and dynamic deviations are used as output labels to construct a dynamic compensation sub-model. Based on the static and dynamic compensation sub-models, a fusion compensation model is constructed, the final fusion compensation formula is determined, and the compensated data is obtained. From the preprocessed benchmark data, a validation set is selected to validate the fusion compensation model and obtain the validated fusion compensation model. The preprocessed detection data and key influencing factors are input into the validated fusion compensation model to determine the deviation between the compensated data and the preprocessed benchmark data. When the mean absolute value of the deviation is less than or equal to the preset benchmark threshold, the model construction is confirmed to be effective.

[0010] In an optional embodiment, the step of acquiring real-time collected data, combining it with a fusion compensation model, performing real-time compensation calculations on the voltage detection data, and outputting the compensated data and a regulator operating status label specifically includes: The sensors and data acquisition terminals at the detection nodes collect real-time data of the raw voltage and corresponding environmental interference. Outlier removal and wavelet denoising are performed on the raw real-time voltage data to obtain preprocessed raw real-time voltage data. At the same time, environmental interference real-time data is preprocessed synchronously to remove outliers and standardize. The original voltage real-time data is input into the static compensation sub-model to obtain the real-time static compensation value, and then the static compensation real-time coefficient is obtained. The preprocessed real-time environmental disturbance data is input into the dynamic compensation sub-model to obtain the real-time dynamic deviation prediction value. Based on the final fusion compensation formula, the real-time compensated voltage value is obtained, and the corresponding record information is recorded, including the compensation timestamp, sensor number, corresponding operating condition information and voltage regulator operating status label. The real-time compensated voltage value and associated recorded information are synchronously transmitted to the voltage monitoring platform. The monitoring platform generates a real-time compensation ledger, which simultaneously displays the original data, preprocessed data, compensation data, and deviation values.

[0011] In an optional embodiment, the comparison and compensation of the data with the benchmark reference data, and the iterative optimization of the compensation model parameters to form a continuously optimized compensation chain, specifically includes: Extract the real-time compensated voltage value from the real-time compensation ledger and synchronously obtain the reference voltage data under the same time sequence and the same operating condition by combining it with the reference source. Obtain the deviation after compensation, statistically analyze the absolute value distribution of the voltage value after real-time compensation, and set the compensation accuracy threshold. Determine if the absolute value of the post-compensation deviation exceeds the compensation accuracy threshold. If not, maintain the current fusion compensation model parameters unchanged and continue to perform real-time compensation. If such data exists and the proportion of data exceeding the threshold is greater than or equal to 5%, the model optimization process is triggered to extract the original real-time voltage data, environmental interference data, reference voltage data and deviation data under the same operating condition during the period exceeding the threshold, and supplement them into the effective data set and the total deviation data set. Retrain the static compensation sub-model, update the static compensation coefficients and static constant terms, and simultaneously retrain the dynamic compensation sub-model to optimize the model parameters, reduce the loss function value, and obtain the optimized fusion compensation model. The optimized fusion compensation model replaces the original model and is put into the real-time compensation process. At the same time, the model optimization time and the accuracy comparison data before and after optimization are recorded.

[0012] Furthermore, a voltage regulator intelligent detection and compensation system based on data deviation analysis is proposed to implement the compensation method described in any of the preceding claims, characterized in that it includes: A multi-source data acquisition module is used to acquire multi-source data, including raw voltage detection data, reference data, environmental interference data, and operating condition characteristic data, based on the voltage detection system topology and operating condition requirements. A multi-source data collaborative preprocessing module is used to perform collaborative preprocessing on multi-source data to construct a set of effective data for all operating conditions. The deviation analysis and dynamic correlation module is used to quantify the total voltage deviation, establish a full-condition deviation matrix, decompose it into static deviation and dynamic deviation, and construct a dynamic correlation model of deviation and influencing factors. A fusion compensation model construction module is used to determine static and dynamic deviation characteristics, construct a fusion compensation model that integrates a static compensation sub-model and an attention mechanism LSTM dynamic compensation sub-model, and achieve accurate cancellation of fixed deviation and real-time time-varying deviation. The real-time compensation and closed-loop optimization module is used to acquire real-time data, combine it with the fusion compensation model, perform real-time compensation calculations on the voltage detection data, and output the compensated data and the regulator's operating status label. At the same time, it combines real-time deviation data and trend prediction to construct a closed-loop optimization link of passive correction and active prediction, and continuously update the model parameters.

[0013] In an optional embodiment, the deviation analysis and dynamic correlation module includes: Total voltage deviation quantization unit, which is used to quantify the total voltage deviation and establish a full-condition deviation matrix; A static and dynamic deviation decomposition unit is used to decompose the full-condition deviation matrix into static and dynamic deviations. A dynamic correlation model construction unit is used to construct a dynamic correlation model of deviation and influencing factors. The deviation feature extraction unit is used to extract static deviation features and dynamic deviation features, providing data support for subsequent model construction.

[0014] In an optional embodiment, the fusion compensation model construction module includes: A static compensation sub-model construction unit is used to construct a static compensation sub-model based on static deviation characteristics. A dynamic compensation sub-model construction unit is used to construct an attention mechanism LSTM dynamic compensation sub-model based on dynamic deviation characteristics. A fusion model integration unit is used to fuse the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model to form a fusion compensation model. The model verification unit is used to verify the effectiveness of the fusion compensation model and ensure that it can accurately cancel out fixed deviations and real-time time-varying deviations.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The proposed intelligent detection and compensation method for voltage regulators based on data deviation analysis obtains original voltage detection data, reference data, environmental interference data, and operating condition characteristic data. It performs collaborative preprocessing on multi-source data, performs outlier removal, four-layer wavelet denoising, and secondary verification to construct an effective data set for all operating conditions. This achieves the integrity and accuracy of data under all operating conditions. The filtering ensures that deviation analysis covers various operating conditions such as steady-state load changes and input voltage fluctuations, avoiding the one-sidedness of compensation caused by the limitation of single-condition data. The proposed intelligent detection and compensation method for voltage regulators based on data deviation analysis quantifies the total voltage deviation to establish a full-condition deviation matrix, decomposes static and dynamic deviations to construct a dynamic correlation model of deviation influencing factors, and integrates a static compensation sub-model and an attention mechanism LSTM dynamic compensation sub-model to construct an intelligent fusion compensation model. This achieves precise targeted cancellation of fixed deviations such as component inherent errors and circuit drift, as well as real-time time-varying deviations caused by input voltage fluctuations and load mutations, significantly improving the voltage regulation accuracy of voltage regulators under complex operating conditions. The proposed intelligent detection and compensation method for voltage regulators based on data deviation analysis combines real-time data acquisition with a fusion compensation model to perform real-time compensation calculations, outputting compensated data and voltage regulator operating status labels. Based on real-time deviation data and trend prediction, a closed-loop optimization link of passive correction and active prediction is constructed to continuously update model parameters, achieving adaptive evolution of the compensation model. This effectively avoids the attenuation of compensation accuracy caused by long-term changes in component aging and operating conditions, ensuring the stability and reliability of the voltage regulator throughout its entire life cycle. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent detection and compensation method for voltage regulators based on data deviation analysis proposed in this invention; Figure 2 This is a flowchart illustrating the process of obtaining the effective data set in this invention. Figure 3 This is a flowchart illustrating the construction process of the fusion compensation model in this invention; Figure 4 This is a system framework diagram of the intelligent detection and compensation system for voltage regulators based on data deviation analysis proposed in this invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 - Figure 4 As shown, the intelligent detection and compensation method for voltage regulators based on data deviation analysis includes: By considering the voltage detection system topology and operating conditions, multi-source data is acquired, including raw voltage detection data, reference data, environmental interference data, and operating condition characteristic data. Collaborative preprocessing of multi-source data is performed to construct an effective dataset covering all operating conditions; Quantify the total voltage deviation, establish a full-condition deviation matrix, and decompose it into static deviation and dynamic deviation to construct a dynamic correlation model of deviation and influencing factors. Determine the static and dynamic bias characteristics, construct a fusion compensation model that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model, and achieve accurate cancellation of fixed bias and real-time time-varying bias. Acquire real-time data, combine it with the fusion compensation model, perform real-time compensation calculations on the voltage detection data, and output the compensated data and the regulator's operating status label; By combining real-time deviation data with trend prediction, a closed-loop optimization link of passive correction and active prediction is constructed to continuously update model parameters.

[0019] Furthermore, by considering the voltage detection system topology and operating conditions, multi-source data is acquired, specifically including: Raw voltage detection data is collected at the deployed detection nodes at a frequency at least 10 times the power grid frequency to ensure that dynamic operating details are captured. The raw voltage detection data includes the instantaneous / RMS value of the voltage regulator input side, the instantaneous / RMS value of the voltage regulator output side, the collection timestamp, the voltage regulator number, the load condition information, and the voltage regulator operating status (standby / running / adjusting). Obtain the voltage range and accuracy requirements of the detection node, and select a standard voltage source that matches the rated output voltage range of the voltage regulator and has an accuracy level not lower than the preset accuracy level as the reference source. Under the same load conditions and acquisition frequency, reference voltage data is obtained through a reference source to ensure that the acquisition timing of the reference voltage data and the original voltage detection data is completely synchronized. The reference voltage data is a dynamic reference range of ±5% of the rated output voltage of the regulator, covering operating conditions including no load, steady-state conditions of 25% / 50% / 75% / rated load, and dynamic operating conditions such as load change (0-100% rated load / 100-0% rated load, switching time ≤10ms) and input voltage fluctuation (±10% rated voltage, fluctuation frequency 0.5Hz). During raw data acquisition, environmental interference data (sensor operating temperature, ambient humidity, power grid harmonic content), internal status data (voltage regulator core component temperature, voltage regulation module operating current), and operating condition characteristic data (load change rate, input voltage fluctuation amplitude, operating condition type label) are collected simultaneously. Perform time-series alignment and validity verification on multi-source data, remove data with time-series misalignment (difference greater than or equal to 3ms) and failed acquisition, and establish a data ledger classified by regulator number, operating condition type and acquisition time.

[0020] In the specific implementation process, the power grid frequency is set to 50Hz, therefore the acquisition frequency of the raw voltage detection data is set to 50Hz to ensure that the acquisition timing is completely matched with the power grid voltage change cycle. The acquisition period is 20ms / point, the duration of a single continuous acquisition is no less than 10 minutes, and the cumulative data acquisition volume of each detection node is no less than 3000 sets to ensure the statistical validity of the data. In the raw voltage detection data, the instantaneous voltage value acquisition range is 0-10kV, the acquisition accuracy is ±0.01kV, and the effective voltage value is calculated from the instantaneous value using the root mean square formula, i.e. ,in The voltage value is instantaneous, and N is the number of sampling points within one power frequency cycle. The sampling timestamp uses UTC time format, accurate to the millisecond level. Sensor numbers are compiled according to the "region-equipment-serial number" rule (e.g., "G1-D2-003"). Load condition information is obtained through PLC data associated with the industrial control system, including load type (e.g., inductive load, capacitive load) and real-time load power. The preset accuracy level is set to 0.05, therefore a 0.01-level standard voltage source (e.g., FLUKE5520A) is selected as the reference source. The output voltage range of this standard voltage source is consistent with the voltage range of the detection node (0-10kV), and the output voltage stability is ≤±0.001% / h. The reference voltage data is acquired through the synchronous output interface of the standard voltage source connected to the trigger interface of the data acquisition terminal (DTU), achieving hard synchronization between the raw voltage detection data and the reference voltage data. The synchronization error is controlled within ±10μs, ensuring accurate correspondence between the two types of data at the same time point.

[0021] Environmental interference data is acquired through an integrated sensor module. The temperature sensor uses a PT100 platinum resistance sensor with a range of -40℃ to 85℃ and an accuracy of ±0.1℃. The sampling frequency is 50Hz, consistent with the voltage data. The humidity sensor uses a capacitive humidity sensor with a range of 0-100%RH and an accuracy of ±1%RH. The power grid harmonic content is acquired through a harmonic analyzer, covering the 2nd to 50th harmonics, with a total harmonic distortion (THD) measurement accuracy of ±0.01%. The load current is acquired through a Rogowski coil sensor with a range of 0-500A and an accuracy of ±0.5A. All environmental interference data are synchronously transmitted to the data acquisition unit (DTU) via a 485 bus.

[0022] The timing alignment process is implemented using a timestamp matching algorithm. Based on the system clock of the data acquisition unit (DTU), the timestamp fields are extracted from the raw voltage data, reference data, and environmental data. The difference between the three timestamps is calculated, and data with a difference exceeding 5ms is considered misaligned and discarded. After timing alignment, a synchronization dataset is constructed and stored in CSV format. Fields include timestamp, sensor number, instantaneous voltage value, effective voltage value, reference voltage value, temperature, humidity, harmonic content, load current, and load condition. The data ledger is built on a MySQL database, establishing tables including raw data, reference data, environmental data, and synchronization dataset. Each table uses timestamps and sensor numbers as linked indexes, supporting data querying and filtering by sensor number, acquisition time, and operating condition type, facilitating data retrieval for subsequent deviation analysis and model training.

[0023] In multi-detection node scenarios, the data acquisition terminals (DTUs) of each detection node communicate with the cloud platform via the 5G network to achieve real-time data upload and storage. The cloud platform centrally manages the data of each node and supports data anomaly alarm function. When the data loss rate of a certain node exceeds 5%, the platform alarm is triggered and the operation and maintenance personnel are notified to check the equipment status to ensure the continuity and integrity of data acquisition.

[0024] Furthermore, collaborative preprocessing of multi-source data is performed to construct a valid dataset covering all operating conditions, specifically including: Based on the synchronous dataset, raw voltage data, reference data, and disturbance and operating condition characteristic data are extracted to construct the raw dataset, reference dataset, and auxiliary dataset, respectively. With a preset significance level (α=0.05), outlier data points in the original dataset are identified and removed. The sliding window size is set to 50 data points. Data mutations (change rate > 5% / ms) within the window are detected in real time and marked as dynamic working condition anomalies and stored separately. The original dataset after removing outliers is subjected to 4-level db4 wavelet decomposition to separate high-frequency noise from low-frequency effective components. An adaptive soft thresholding method (the threshold is dynamically adjusted according to the working conditions) is used to filter out noise and reconstruct the preprocessed original detection data. The same outlier removal and noise reduction processes are performed on both the benchmark dataset and the auxiliary dataset to ensure data consistency and reconstruct the preprocessed benchmark data. The preprocessed raw test data and the reference data are verified a second time to ensure the consistency of the input and output voltage response timing of the voltage regulator. At the same time, the noise reduction efficiency of the raw test data and the reference data is determined. When the noise reduction efficiency is greater than or equal to the preset noise reduction efficiency threshold, the preprocessing is confirmed to be effective and the set of valid data under all operating conditions is output. Otherwise, the wavelet transform decomposition level is adjusted and the data is reprocessed.

[0025] Specifically, the secondary verification verifies that the consistency error of the time-domain characteristics (peak value, valley value, response time) between the original detection data and the benchmark data is ≤0.3%, and the noise reduction efficiency is ≥92%. At the same time, it verifies the correlation between the operating condition characteristic data and the voltage data (such as the correlation coefficient between the load change rate and the voltage fluctuation is ≥0.6). After the verification is passed, the valid data set of the entire operating condition is output. In the specific implementation process, when extracting raw voltage detection data and reference voltage data based on the synchronous dataset, the data field filtering algorithm is used to extract "instantaneous voltage value and effective voltage value" as core data fields, and "collection timestamp" as the associated field to construct raw dataset D_1 and reference dataset D_2 respectively. The datasets are stored in a two-dimensional array format, with row index as time series number and column index as data field identifier to ensure high efficiency when retrieving data.

[0026] With a preset significance level of α=0.05, for the original dataset D_1, the formula is used... Calculate the mean μ (where For a single data point in D_1, (where D_1 is the sample size), using the formula Calculate the standard deviation σ; based on the Grubbs criterion, using the formula... Identify outlier data points (of which) The Grubbs threshold is defined as G(0.05, 3000) = 3.29 when n = 3000. Data points meeting this condition are identified as outliers and removed from D_1. If the data missing rate exceeds 3% after removal, data for the corresponding time period is re-collected. The original dataset after outlier removal... Wavelet decomposition was performed, using the db4 wavelet basis function and setting the decomposition level to 3. The wavelet transform formula was then applied. in This is the voltage signal in D1'. Signal decomposition was performed using the db4 wavelet basis function, resulting in one low-frequency component A3 and three high-frequency components D_1, D_2, and D_3. A soft thresholding method was then used to process the high-frequency components, setting a threshold value. (σ is the noise standard deviation of the high-frequency components). High-frequency coefficients with absolute values ​​less than λ are set to zero, while high-frequency coefficients with absolute values ​​greater than or equal to λ are shrunk. Finally, the preprocessed original detection data is reconstructed using inverse wavelet transform. Based on the same significance level α=0.05, outlier data points were removed using the Grubbs criterion, resulting in... The db4 wavelet basis function was used for three-level wavelet decomposition. After processing the high-frequency components using the same soft thresholding method, the preprocessed reference data was reconstructed. This ensures that the preprocessing standards for the two types of data are consistent.

[0027] The raw detection data after preprocessing Compared with benchmark data A secondary verification is performed, comparing the time synchronization of the two data points one by one based on the acquisition timestamp, with the synchronization error controlled within ±1ms; by calculating the time-domain characteristic parameters (including peak, valley, and period) of the two types of data, the consistency error of the time-domain characteristics is ensured to be ≤0.5%; at the same time, the formula is used... Calculate the noise reduction efficiency η (where For the original dataset standard deviation For the preprocessed dataset (Standard deviation), with a preset noise reduction efficiency threshold of 90%. When η≥90%, the preprocessing is confirmed to be effective, and a valid data set is output. , If η < 90%, adjust the wavelet decomposition layer to 4 layers and repeat the above noise reduction and verification process until the threshold requirement is met.

[0028] In multi-condition scenarios, the above preprocessing process is executed for three steady-state conditions: no load, 50% rated load, and rated load, respectively, to obtain the effective data set for the corresponding condition. At the same time, the preprocessing parameters for each condition (including significance level, wavelet basis function, number of decomposition layers, and threshold size) are recorded and stored in the preprocessing parameter configuration library. The configuration parameters can be directly called for subsequent similar conditions to improve preprocessing efficiency. In addition, the preprocessing process is accelerated by FPGA chip, and the data processing latency is controlled within 50ms to meet the timing requirements of real-time compensation.

[0029] Furthermore, the total voltage deviation is quantified, a full-condition deviation matrix is ​​established, and decomposed into static and dynamic deviations. A dynamic correlation model of deviation and influencing factors is constructed, specifically including: Based on the preprocessed raw detection data and reference data, the total voltage deviation is determined point by point, and a full-condition deviation matrix of operating condition type-time-deviation value is constructed. The total voltage deviation is the difference between the detected voltage on the output side of the regulator and the rated output reference voltage. The deviation data of steady-state conditions in the full-condition deviation matrix are extracted, and the mean value of the fixed deviation data of output voltage under each steady-state condition is obtained. The static deviation is obtained by weighted averaging, which characterizes the inherent error of the component and the circuit drift. Based on the total voltage deviation and static deviation, the dynamic deviation is determined and a dynamic deviation dataset is constructed. The dataset is classified and labeled according to the dynamic operating condition type. The dynamic deviation is the time-varying deviation caused by input voltage fluctuation, load change, and internal temperature rise after removing the static deviation from the total deviation. Based on environmental disturbance data, internal state data, operating condition characteristic data and dynamic deviation dataset, the correlation coefficients between each environmental factor and dynamic deviation are determined. Based on the absolute value of the correlation coefficient, environmental factors with a correlation coefficient greater than or equal to the preset threshold are selected as key influencing factors. A time-series correlation model between key influencing factors and dynamic deviation is established to clarify the weight of each factor on dynamic deviation.

[0030] In the specific implementation process, based on the preprocessed raw detection data Compared with benchmark data Through formula The total voltage deviation is calculated for each time series point, with the timing point matching accuracy consistent with the time-domain consistency verification standard in the preprocessing stage (synchronization error ≤ ±1ms). The total voltage deviation of all time series points is stored in the order of the acquisition timestamps, constructing a total deviation dataset D_Δtotal. The dataset is stored in array format, with the index being the time series point number and the element being the corresponding total deviation value. The operating condition label for each time series point is also stored in association for subsequent operating condition selection. Three steady-state operating conditions are selected: no-load (load power of 0kW), 50% rated load (load power of 50% of rated power), and rated load (load power of the equipment's rated power, such as 100kW). The deviation data for the corresponding operating condition is extracted from the total deviation dataset D_Δtotal using the operating condition label, resulting in no-load deviation subset D_Δs1, 50% rated load deviation subset D_Δs2, and rated load deviation subset D_Δs3, respectively. The formula is then used to... Calculate the mean deviation of the no-load condition μ_Δs1 (where Let D_Δs1 be the sample size. (For a single deviation value in D_Δs1), μ_Δs2 and μ_Δs3 are calculated similarly, and then obtained through the formula. The static deviation ΔU_s is calculated with an accuracy controlled within ±0.001kV. Based on the total voltage deviation ΔU_total and the static deviation ΔU_s, the dynamic deviation is calculated point by point using the formula ΔU_d=ΔU_total−ΔU_s. All dynamic deviation values ​​are stored according to their corresponding time points to construct a dynamic deviation dataset D_Δd. This dataset maintains the same time-series index as the total deviation dataset D_Δtotal to facilitate subsequent correlation analysis with environmental data.

[0031] Environmental interference data includes sensor operating temperature T, ambient humidity H, power grid harmonic content THD, and load current I. Based on the dynamic deviation dataset D_Δd and environmental interference data, the Pearson correlation coefficient formula is used. (Where x represents single environmental factor data, y represents dynamic deviation data, and n represents the synchronous sample size), calculate the correlation coefficients r_T, r_H, r_THD, and r_I between T, H, THD, I and ΔU_d respectively.

[0032] The correlation coefficient threshold is preset to 0.7. The correlation coefficients are sorted by absolute value, and environmental factors with an absolute value ≥ 0.7 are selected as key influencing factors (e.g., if...). The key influencing factors are temperature (T) and harmonic content (THD). Based on the key influencing factors and dynamic deviation dataset, a correlation model is established using a multiple linear regression algorithm. in As a key influencing factor, Let be the weighting coefficients of each factor (c is a constant term). The weighting coefficients that minimize the mean square error between the model's predicted values ​​and the actual dynamic deviations are obtained using the least squares method. The weights of each key influencing factor on the dynamic deviation are defined (e.g., temperature weighting coefficient is 0.62, harmonic content weighting coefficient is 0.38). The goodness-of-fit R-squared of the correlation model is then calculated. 2 A value ≥ 0.85 is required to ensure a good fit of the model.

[0033] In multi-detection-node scenarios, the above-mentioned deviation quantification and correlation analysis process is performed on each node. For the differences in key influencing factors of different nodes, a node-specific correlation model is established to improve the targeting of compensation. At the same time, the correlation coefficient is recalculated and the key influencing factors and correlation model parameters are updated regularly (e.g., every 7 days) to adapt to long-term changes in the environment and equipment status and ensure the timeliness and accuracy of deviation analysis.

[0034] Furthermore, static and dynamic bias characteristics are identified, and a fusion compensation model is constructed that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model. Specifically, this includes: Based on static deviation and steady-state data under full operating conditions, the linear relationship between static deviation and preprocessed detection data is fitted to obtain static compensation coefficients. Furthermore, by solving for the static compensation coefficients and static constants that minimize the value of the quadratic linear function, the static compensation sub-model is determined. Based on the dataset of key influencing factors and dynamic deviations, a training set and a test set are divided, with the training set accounting for 70% and the test set accounting for 30%. The key influencing factors, working condition type codes, and deviation change rate are used as input features, and dynamic deviations are used as output labels to construct a dynamic compensation sub-model. Based on the static and dynamic compensation sub-models, a fusion compensation model is constructed, the final fusion compensation formula is determined, and the compensated data is obtained. From the preprocessed benchmark data, a validation set is selected to validate the fusion compensation model and obtain the validated fusion compensation model. The preprocessed detection data and key influencing factors are input into the validated fusion compensation model to determine the deviation between the compensated data and the preprocessed benchmark data. When the mean absolute value of the deviation is less than or equal to the preset benchmark threshold, the model construction is confirmed to be effective.

[0035] In the specific implementation process, based on the static deviation ΔU_s, the preprocessed original detection data U_d1 and the label data for three steady-state conditions—no load, 50% rated load, and rated load—are selected to construct a static fitting dataset. The least squares method is used to fit the linear relationship between the static deviation and U_d1, and the fitting formula is as follows: (Where k_s is the static compensation coefficient, and b is the static constant term). This is achieved by solving a quadratic linear function. (n is the sample size of the static fitting dataset, Let be the static bias of the i-th sample. To determine the minimum value of the preprocessed detection data for the i-th sample, the optimal static compensation coefficient k_s and constant term b are determined. Finally, a static compensation sub-model is constructed, with the compensation formula U_c1=U_d1-k_s, and the goodness-of-fit Ri is calculated. 2 A value ≥ 0.9 is required to ensure a good linear fit.

[0036] Based on key influencing factors (such as temperature T and power grid harmonic content THD) and the dynamic deviation dataset DΔd, a dynamic model training dataset was constructed. The dataset was randomly divided into a 70% training set and a 30% test set, using stratified sampling to ensure consistent data proportions for each operating condition across the training and test sets. The dynamic compensation sub-model was constructed using an LSTM neural network. The number of neurons in the input layer matched the number of key influencing factors (e.g., two neurons for two key factors). Two hidden layers were set, each with 32 neurons, using the ReLU activation function. The output layer consisted of one neuron, corresponding to the predicted dynamic deviation value. The mean squared error was used as the basis for the calculation. (m is the sample size) is used as the loss function. The Adam optimizer is employed, with an initial learning rate of 0.001 and 100 iterations. The condition is met when the MSE on the test set ≤ 0.001V. 2 Training is stopped when the time is right, and the dynamic bias prediction model is obtained, and the predicted dynamic bias ΔU_d(pred) is output.

[0037] Based on the static and dynamic compensation sub-models, a fusion compensation model is constructed, and the final fusion compensation formula is determined as follows: (in (For the compensated data). From the preprocessed baseline data U_r1, samples are randomly selected at a ratio of 20% to serve as the validation set. The validation set samples must cover all steady-state conditions and typical dynamic conditions. The preprocessed detection data U_d1 and key influencing factors are input into the fusion compensation model to obtain the compensated data U_comp. The values ​​of U_comp and the corresponding baseline data in the validation set are then calculated. deviation And calculate the mean absolute value of the deviation. .

[0038] The preset reference threshold is 0.1% × U_r1 (i.e., 0.1% of the reference voltage data). At that time, the fusion compensation model was confirmed to be effective; if If the threshold is not met, return to the dynamic compensation sub-model training stage, adjust the number of hidden layer neurons in the LSTM neural network (e.g., adjust to 48) or the learning rate (e.g., adjust to 0.0005), retrain the model and perform validation again until the threshold requirement is met.

[0039] In multi-detection node scenarios, each node executes the aforementioned model building process to generate a node-specific fusion compensation model, which is then stored in the model library and associated with the sensor number. Simultaneously, a model parameter configuration file is created to record the static compensation coefficients k_s, constant term b, LSTM neural network structure parameters, and training hyperparameters for each node, facilitating subsequent model calls and optimization.

[0040] Furthermore, real-time data is acquired, and combined with the fusion compensation model, real-time compensation calculations are performed on the voltage detection data, outputting the compensated data and the regulator's operating status label. Specifically, this includes: The sensors and data acquisition terminals at the detection nodes collect real-time data of the raw voltage and corresponding environmental interference. Outlier removal and wavelet denoising are performed on the raw real-time voltage data to obtain preprocessed raw real-time voltage data. At the same time, environmental interference real-time data is preprocessed synchronously to remove outliers and standardize. The original voltage real-time data is input into the static compensation sub-model to obtain the real-time static compensation value, and then the static compensation real-time coefficient is obtained. The preprocessed real-time environmental disturbance data is input into the dynamic compensation sub-model to obtain the real-time dynamic deviation prediction value. Based on the final fusion compensation formula, the real-time compensated voltage value is obtained, and the corresponding record information is recorded, including the compensation timestamp, sensor number, corresponding operating condition information and voltage regulator operating status label. The real-time compensated voltage value and associated recorded information are synchronously transmitted to the voltage monitoring platform. The monitoring platform generates a real-time compensation ledger, which simultaneously displays the original data, preprocessed data, compensation data, and deviation values.

[0041] In the specific implementation process, voltage sensors (such as Hall voltage sensors with a range of 0-10kV and an accuracy of ±0.1%FS) deployed at the detection nodes and industrial-grade data acquisition terminals (DTUs, supporting 4G / 5G communication) are used to collect real-time raw voltage data. The acquisition frequency is consistent with the power grid frequency at 50Hz, and the single acquisition delay is controlled within 10ms. Real-time environmental interference data, including sensor operating temperature, ambient humidity, power grid harmonic content, and load current, are collected synchronously at the same frequency of 50Hz. Parallel acquisition is achieved through the multi-channel interface of the DTU to ensure timing synchronization with the raw voltage data.

[0042] The preprocessing procedure for the raw real-time voltage data is the same as that for the offline stage: the significance level is preset to α=0.05, and the mean is calculated in real time based on a sliding window (window size set to 100 data points). with standard deviation Through the Grubbs criterion formula Identify and remove outlier data points; perform 3-level wavelet decomposition using the db4 wavelet basis function, and apply a soft thresholding method (threshold). The high-frequency noise components are processed, and then reconstructed using inverse wavelet transform to obtain the preprocessed original real-time voltage data U_d1_real. Real-time environmental interference data are preprocessed synchronously: temperature, humidity, harmonic content, and load current data are all processed using the same Grubbs criterion to remove outliers, and then standardized using the Z-score formula. (in The moving window mean of environmental data. The data is standardized to the [-1,1] interval (where the standard deviation is 1) to obtain the preprocessed real-time environmental disturbance data X_std.

[0043] The preprocessed raw voltage real-time data U_d1_real is input into the static compensation sub-model, based on the static compensation formula. ( (The static compensation coefficients obtained from offline training) are used to calculate the real-time static compensation value. Synchronously output static compensation real-time coefficients (Maintain consistency with the offline model; no real-time updates required). The standardized environmental disturbance data will be updated in real-time. The input dynamic compensation sub-model (LSTM neural network) outputs real-time dynamic bias predictions based on real-time input features. The model inference latency is controlled within 20ms to ensure real-time requirements.

[0044] Based on the final fusion compensation formula The real-time compensated voltage value is calculated with an accuracy controlled within ±0.001kV. Simultaneously, supporting information is recorded, with the compensation timestamp using UTC time format, accurate to the millisecond level; sensor numbering follows the "area-device-serial number" rule; operating condition information is obtained in real-time through data from the associated industrial control system PLC, including the current load type, load power, and grid operating status (such as steady state and transient state).

[0045] The real-time compensated voltage value U_comp_real and its associated recording information are synchronously transmitted to the voltage monitoring platform via a 5G communication module. The transmission protocol uses MQTT, with a transmission latency of ≤50ms, ensuring that the monitoring platform receives data in real time. A real-time compensation ledger is established on the monitoring platform, and a visual interface is used to synchronously display the raw voltage data, preprocessed data, static compensation value, dynamic deviation prediction value, compensated data, and real-time deviation value of each detection node. ,in (This serves as the real-time output data for the reference source). The ledger supports data filtering and export by sensor number, time range, and operating condition type. It also includes a deviation warning function, which alerts the system when the real-time deviation value... Exceeding the preset warning threshold (e.g., 0.15% × When this happens, the platform will automatically trigger an audible and visual alarm to notify maintenance personnel to investigate.

[0046] In multi-detection node cluster scenarios, the monitoring platform supports distributed data reception and centralized management, and uses a load balancing algorithm to allocate data processing resources to each node. A single platform can simultaneously access data from no less than 100 detection nodes, and the data processing of each node does not interfere with each other. The platform has a built-in data caching mechanism with a caching time of 72 hours. The cached data supports offline querying and tracing, providing sufficient real-time data samples for subsequent closed-loop optimization modules.

[0047] Furthermore, by comparing the compensated data with the benchmark reference data, the parameters of the compensation model are iteratively optimized to form a continuously optimized compensation chain, which specifically includes: Extract the real-time compensated voltage value from the real-time compensation ledger and synchronously obtain the reference voltage data under the same time sequence and the same operating condition by combining it with the reference source. Obtain the deviation after compensation, statistically analyze the absolute value distribution of the voltage value after real-time compensation, and set the compensation accuracy threshold. Determine if the absolute value of the post-compensation deviation exceeds the compensation accuracy threshold. If not, maintain the current fusion compensation model parameters unchanged and continue to perform real-time compensation. If such data exists and the proportion of data exceeding the threshold is greater than or equal to 5%, the model optimization process is triggered to extract the original real-time voltage data, environmental interference data, reference voltage data and deviation data under the same operating condition during the period exceeding the threshold, and supplement them into the effective data set and the total deviation data set. Retrain the static compensation sub-model, update the static compensation coefficients and static constant terms, and simultaneously retrain the dynamic compensation sub-model to optimize the model parameters, reduce the loss function value, and obtain the optimized fusion compensation model. The optimized fusion compensation model replaces the original model and is put into the real-time compensation process. At the same time, the model optimization time and the accuracy comparison data before and after optimization are recorded.

[0048] In the specific implementation process, when extracting data from the real-time compensation ledger, a timed extraction mechanism is adopted, with an extraction cycle of 24 hours. The real-time compensated voltage values ​​U_comp_real of all detection nodes within the 24 hours prior to each extraction are included, and the extraction is associated with tags such as "collection timestamp, sensor number, and operating condition type" to ensure data traceability. Simultaneously, reference voltage data Ur_real under the same time series and operating condition is synchronously acquired through a reference source (0.01-level standard voltage source). The time series synchronization error is controlled within ±10μs to ensure the validity of the comparison between the two types of data. Based on the extracted U_comp_real and Ur_real, the compensated deviation is calculated point by point using the formula ΔU_comp = |U_comp_real - U_r_real|. The absolute value distribution of all deviation values ​​is statistically analyzed, and a deviation distribution histogram is plotted to determine statistical parameters such as the median and 95th quantile of the deviation. Set the compensation accuracy threshold ε = 0.1% × U_r_real (i.e., 0.1% of the reference voltage data). This threshold can be adjusted within the range of 0.05%-0.2% according to the accuracy requirements of the actual application scenario. Iterate through all the deviation data extracted at this node and calculate the ratio r = n_out / n_total of the number of data points exceeding the threshold ε to the total number of extracted data points. If r < 5%, the current fusion compensation model parameters are considered to be adapted, the parameters are kept unchanged, and real-time compensation continues; if r ≥ 5%, the model optimization process is triggered.

[0049] After triggering optimization, the original real-time voltage data U_d_real, real-time environmental interference data X_real, reference voltage data U_r_real under the same operating condition, and compensated deviation data ΔU_comp are extracted from the time periods exceeding the threshold (i.e., all time periods corresponding to deviation ΔU_comp>ε). After outlier removal, noise reduction, and standardization of these data according to offline preprocessing standards, they are added to the original valid dataset and the total deviation dataset. The sample size of the supplemented dataset is increased by at least 10% to ensure that the model training has sufficient new feature data.

[0050] Based on the supplemented effective dataset, steady-state operating condition data were selected, and the linear relationship between the static deviation and the preprocessed detection data was refitted. By solving the quadratic linear function The minimum value is used to update the static compensation coefficient. and constant term Goodness of fit R 2The value must be ≥0.9. Simultaneously, the dynamic compensation model is retrained based on the supplemented key impact factor data and dynamic bias dataset. The LSTM neural network structure (number of neurons in the input layer = number of key impact factors, 2 hidden layers with 32 neurons each) remains unchanged. The learning rate is adjusted to 0.0008, and the number of iterations is set to 80 rounds, using the mean squared error... The loss function is used, and training is performed until the MSE on the test set is ≤ 0.0008V. 2 Thus, the optimized dynamic compensation sub-model is obtained.

[0051] The updated static and dynamic compensation sub-models are integrated to form an optimized fusion compensation model, whose compensation formula remains the same. (in This is the updated static bias. (This refers to the predicted dynamic bias of the optimized model). The optimized model replaces the original model and is put into the real-time compensation process. At the same time, the model optimization time (accurate to the minute), the MSE of the test set before optimization, the MSE of the test set after optimization, the percentage of data exceeding the threshold before optimization (r_old), and the percentage of data exceeding the threshold after optimization (r_new) are recorded as accuracy comparison data. These data are stored in the model optimization log, which is named "sensor number - optimization date" for easy traceability and analysis later.

[0052] In multi-node cluster scenarios, each node independently triggers the optimization process without interfering with others. The optimized model parameters are synchronized to the cloud platform model library, enabling distributed management and rapid deployment of the model. In addition, a mandatory optimization cycle of 30 days is set for the model. Even if the percentage of data exceeding the threshold does not reach 5%, the model will be retrained according to the above process to ensure that the model can adapt to long-term, slow changes in equipment and environment, further guaranteeing the stability of the compensation link.

[0053] Furthermore, a voltage regulator intelligent detection and compensation system based on data deviation analysis is proposed to implement any of the compensation methods described above, characterized by comprising: The multi-source data acquisition module is used to acquire multi-source data based on the voltage detection system topology and operating conditions, including raw voltage detection data, reference data, environmental interference data, and operating condition characteristic data. Multi-source data collaborative preprocessing module: This module is used to collaboratively preprocess multi-source data to construct an effective data set for all operating conditions. The deviation analysis and dynamic correlation module is used to quantify the total voltage deviation, establish a full-condition deviation matrix, decompose it into static deviation and dynamic deviation, and construct a dynamic correlation model of deviation and influencing factors. The fusion compensation model construction module is used to determine the static and dynamic deviation characteristics, and construct a fusion compensation model that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model to achieve accurate cancellation of fixed deviation and real-time time-varying deviation. The real-time compensation and closed-loop optimization module is used to acquire real-time data, combine it with the fusion compensation model, perform real-time compensation calculations on the voltage detection data, and output the compensated data and the regulator's operating status label. At the same time, it combines real-time deviation data and trend prediction to build a closed-loop optimization link of passive correction and active prediction, and continuously update the model parameters.

[0054] In an optional embodiment, the deviation analysis and dynamic correlation module includes: Total voltage deviation quantization unit: The total voltage deviation quantization unit is used to quantify the total voltage deviation and establish the deviation matrix under all operating conditions. The static and dynamic deviation decomposition unit is used to decompose the full-condition deviation matrix into static and dynamic deviations. The dynamic correlation model building unit is used to construct a dynamic correlation model between deviation and influencing factors. The deviation feature extraction unit is used to extract static and dynamic deviation features, providing data support for subsequent model construction.

[0055] In an optional embodiment, the fusion compensation model construction module includes: The static compensation sub-model building unit is used to build a static compensation sub-model based on static deviation characteristics. The dynamic compensation sub-model building unit is used to build an attention mechanism LSTM dynamic compensation sub-model based on dynamic deviation features. The fusion model integration unit is used to fuse the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model to form a fusion compensation model. The model validation unit is used to verify the effectiveness of the fusion compensation model and ensure that it can accurately cancel out fixed biases and real-time time-varying biases.

[0056] The advantages of this invention are: a refined and intelligent design for the entire process of voltage regulator detection and compensation, enabling precise adaptation to complex operating scenarios and ensuring long-term stable performance. It utilizes multi-source data acquisition covering original voltage, reference, environmental interference, and operating condition characteristics, combined with a high-frequency acquisition frequency 10 times the power grid frequency and full operating condition coverage, ensuring no details of dynamic operating conditions are missed, providing comprehensive data support for subsequent analysis. Through collaborative preprocessing including outlier removal, four-layer wavelet denoising, and secondary verification, data accuracy is effectively guaranteed, avoiding deviations caused by single data processing. In the deviation analysis and compensation stage, the total deviation is decomposed into static and dynamic categories, and a dynamic correlation model is established. This integrates a static compensation sub-model and an attention-based LSTM dynamic compensation sub-model, achieving precise cancellation of fixed deviations such as inherent component errors and circuit drift with time-varying deviations such as input voltage fluctuations and load mutations, significantly improving voltage regulation accuracy under complex operating conditions. Meanwhile, the closed-loop optimization chain of "passive correction + active prediction" not only triggers model updates when deviations exceed the standard, but also optimizes parameters in advance based on trend prediction, effectively avoiding accuracy decay caused by component aging and changes in operating conditions, and ensuring the stability and reliability of the voltage regulator throughout its entire life cycle.

[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for intelligent detection and compensation of voltage regulators based on data deviation analysis, characterized in that, include: By considering the voltage detection system topology and operating conditions, multi-source data is acquired, including raw voltage detection data, reference data, environmental interference data, and operating condition characteristic data. Collaborative preprocessing of multi-source data is performed to construct an effective dataset covering all operating conditions; Quantify the total voltage deviation, establish a full-condition deviation matrix, and decompose it into static deviation and dynamic deviation to construct a dynamic correlation model of deviation and influencing factors. Determine the static and dynamic bias characteristics, construct a fusion compensation model that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model, and achieve accurate cancellation of fixed bias and real-time time-varying bias. Acquire real-time data, combine it with the fusion compensation model, perform real-time compensation calculations on the voltage detection data, and output the compensated data and the regulator's operating status label; By combining real-time deviation data with trend prediction, a closed-loop optimization link of passive correction and active prediction is constructed to continuously update model parameters.

2. The intelligent detection and compensation method for voltage regulators based on data deviation analysis according to claim 1, characterized in that, The acquisition of multi-source data through the voltage detection system topology and operating condition requirements specifically includes: Raw voltage detection data is collected at the deployed detection nodes at a frequency at least 10 times the power grid frequency to ensure that dynamic operating details are captured. The raw voltage detection data includes the instantaneous / RMS value of the voltage on the input side of the regulator, the instantaneous / RMS value of the voltage on the output side, the collection timestamp, the regulator number, the load condition information, and the regulator's operating status. Obtain the voltage range and accuracy requirements of the detection node, and select a standard voltage source that matches the rated output voltage range of the voltage regulator and has an accuracy level not lower than the preset accuracy level as the reference source. Under the same load conditions and acquisition frequency, reference voltage data is obtained through a reference source to ensure that the acquisition timing of the reference voltage data and the original voltage detection data is completely synchronized. The reference voltage data is a dynamic reference range of ±5% of the rated output voltage of the regulator, covering operating conditions including no load, steady-state conditions of 25% / 50% / 75% / rated load, and dynamic operating conditions of load change and input voltage fluctuation. During the raw data acquisition, environmental interference data, internal status data, and operating condition characteristic data are collected simultaneously. Perform time-series alignment and validity verification on multi-source data, remove data with time-series misalignment and failed acquisition, and establish a data ledger categorized by voltage regulator number, operating condition type, and acquisition time.

3. The intelligent detection and compensation method for voltage regulators based on data deviation analysis according to claim 2, characterized in that, The collaborative preprocessing of multi-source data to construct a valid dataset covering all operating conditions specifically includes: Based on the synchronous dataset, raw voltage data, reference data, and disturbance and operating condition characteristic data are extracted to construct the raw dataset, reference dataset, and auxiliary dataset, respectively. The significance level is preset to identify and remove outlier data points from the original dataset. The sliding window size is set to 50 data points. Data mutations within the window are detected in real time, marked as dynamic operating condition anomalies, and stored separately. The original dataset after removing outliers is subjected to four-level wavelet decomposition to separate high-frequency noise components from low-frequency effective components. High-frequency noise is further filtered out, and the preprocessed original detection data is reconstructed. The same outlier removal and noise reduction processes are performed on both the benchmark dataset and the auxiliary dataset to ensure data consistency and reconstruct the preprocessed benchmark data. The preprocessed raw test data and the reference data are verified a second time to ensure the consistency of the input and output voltage response timing of the voltage regulator. At the same time, the noise reduction efficiency of the raw test data and the reference data is determined. When the noise reduction efficiency is greater than or equal to the preset noise reduction efficiency threshold, the preprocessing is confirmed to be effective and the set of valid data under all operating conditions is output. Otherwise, the wavelet transform decomposition level is adjusted and the data is reprocessed.

4. The intelligent detection and compensation method for voltage regulators based on data deviation analysis according to claim 3, characterized in that, The total voltage deviation is quantified, a full-condition deviation matrix is ​​established, and it is decomposed into static deviation and dynamic deviation. A dynamic correlation model of deviation and influencing factors is constructed, specifically including: Based on the preprocessed raw detection data and reference data, the total voltage deviation is determined point by point, and a full-condition deviation matrix of operating condition type-time-deviation value is constructed. The total voltage deviation is the difference between the detected voltage on the output side of the regulator and the rated output reference voltage. The deviation data of steady-state conditions in the full-condition deviation matrix are extracted, and the mean value of the fixed deviation data of output voltage under each steady-state condition is obtained. The static deviation is obtained by weighted averaging, which characterizes the inherent error of the component and the circuit drift. Based on the total voltage deviation and static deviation, the dynamic deviation is determined and a dynamic deviation dataset is constructed. The dataset is classified and labeled according to the dynamic operating condition type. The dynamic deviation is the time-varying deviation caused by input voltage fluctuation, load change, and internal temperature rise after removing the static deviation from the total deviation. Based on environmental disturbance data, internal state data, operating condition characteristic data and dynamic deviation dataset, the correlation coefficients between each environmental factor and dynamic deviation are determined. Based on the absolute value of the correlation coefficient, environmental factors with a correlation coefficient greater than or equal to the preset threshold are selected as key influencing factors. A time-series correlation model between key influencing factors and dynamic deviation is established to clarify the weight of each factor on dynamic deviation.

5. The intelligent detection and compensation method for voltage regulators based on data deviation analysis according to claim 4, characterized in that, The process of determining static and dynamic deviation characteristics and constructing a fusion compensation model that integrates the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model specifically includes: Based on static deviation and steady-state data under full operating conditions, the linear relationship between static deviation and preprocessed detection data is fitted to obtain static compensation coefficients. Furthermore, by solving for the static compensation coefficients and static constants that minimize the value of the quadratic linear function, the static compensation sub-model is determined. Based on the dataset of key influencing factors and dynamic deviations, a training set and a test set are divided, with the training set accounting for 70% and the test set accounting for 30%. The key influencing factors, working condition type codes, and deviation change rate are used as input features, and dynamic deviations are used as output labels to construct a dynamic compensation sub-model. Based on the static and dynamic compensation sub-models, a fusion compensation model is constructed, the final fusion compensation formula is determined, and the compensated data is obtained. From the preprocessed benchmark data, a validation set is selected to validate the fusion compensation model and obtain the validated fusion compensation model. The preprocessed detection data and key influencing factors are input into the validated fusion compensation model to determine the deviation between the compensated data and the preprocessed benchmark data. When the mean absolute value of the deviation is less than or equal to the preset benchmark threshold, the model construction is confirmed to be effective.

6. The intelligent detection and compensation method for voltage regulators based on data deviation analysis according to claim 5, characterized in that, The process of acquiring real-time data, combining it with a fusion compensation model, performing real-time compensation calculations on the voltage detection data, and outputting the compensated data and the regulator's operating status label specifically includes: The sensors and data acquisition terminals at the detection nodes collect real-time data of the raw voltage and corresponding environmental interference. Outlier removal and wavelet denoising are performed on the raw real-time voltage data to obtain preprocessed raw real-time voltage data. At the same time, environmental interference real-time data is preprocessed synchronously to remove outliers and standardize. The original voltage real-time data is input into the static compensation sub-model to obtain the real-time static compensation value, and then the static compensation real-time coefficient is obtained. The preprocessed real-time environmental disturbance data is input into the dynamic compensation sub-model to obtain the real-time dynamic deviation prediction value. Based on the final fusion compensation formula, the real-time compensated voltage value is obtained, and the corresponding record information is recorded, including the compensation timestamp, sensor number, corresponding operating condition information and voltage regulator operating status label. The real-time compensated voltage value and associated recorded information are synchronously transmitted to the voltage monitoring platform. The monitoring platform generates a real-time compensation ledger, which simultaneously displays the original data, preprocessed data, compensation data, and deviation values.

7. The intelligent detection and compensation method for voltage regulators based on data deviation analysis according to claim 6, characterized in that, The comparison and compensation data are compared with the benchmark reference data, and the compensation model parameters are iteratively optimized to form a continuously optimized compensation chain, specifically including: Extract the real-time compensated voltage value from the real-time compensation ledger and synchronously obtain the reference voltage data under the same time sequence and the same operating condition by combining it with the reference source. Obtain the deviation after compensation, statistically analyze the absolute value distribution of the voltage value after real-time compensation, and set the compensation accuracy threshold. Determine if the absolute value of the post-compensation deviation exceeds the compensation accuracy threshold. If not, maintain the current fusion compensation model parameters unchanged and continue to perform real-time compensation. If such data exists and the proportion of data exceeding the threshold is greater than or equal to 5%, the model optimization process is triggered to extract the original real-time voltage data, environmental interference data, reference voltage data and deviation data under the same operating condition during the period exceeding the threshold, and supplement them into the effective data set and the total deviation data set. Retrain the static compensation sub-model, update the static compensation coefficients and static constant terms, and simultaneously retrain the dynamic compensation sub-model to optimize the model parameters, reduce the loss function value, and obtain the optimized fusion compensation model. The optimized fusion compensation model replaces the original model and is put into the real-time compensation process. At the same time, the model optimization time and the accuracy comparison data before and after optimization are recorded.

8. A voltage regulator intelligent detection and compensation system based on data deviation analysis, used to implement the compensation method as described in any one of claims 1-7, characterized in that, include: A multi-source data acquisition module is used to acquire multi-source data, including raw voltage detection data, reference data, environmental interference data, and operating condition characteristic data, based on the voltage detection system topology and operating condition requirements. A multi-source data collaborative preprocessing module is used to perform collaborative preprocessing on multi-source data to construct a set of effective data for all operating conditions. The deviation analysis and dynamic correlation module is used to quantify the total voltage deviation, establish a full-condition deviation matrix, decompose it into static deviation and dynamic deviation, and construct a dynamic correlation model of deviation and influencing factors. A fusion compensation model construction module is used to determine static and dynamic deviation characteristics, construct a fusion compensation model that integrates a static compensation sub-model and an attention mechanism LSTM dynamic compensation sub-model, and achieve accurate cancellation of fixed deviation and real-time time-varying deviation. The real-time compensation and closed-loop optimization module is used to acquire real-time data, combine it with the fusion compensation model, perform real-time compensation calculations on the voltage detection data, and output the compensated data and the regulator's operating status label. At the same time, it combines real-time deviation data and trend prediction to construct a closed-loop optimization link of passive correction and active prediction, and continuously update the model parameters.

9. The intelligent detection and compensation system for voltage regulators based on data deviation analysis according to claim 8, characterized in that, The deviation analysis and dynamic correlation module includes: Total voltage deviation quantization unit, which is used to quantify the total voltage deviation and establish a full-condition deviation matrix; A static and dynamic deviation decomposition unit is used to decompose the full-condition deviation matrix into static and dynamic deviations. A dynamic correlation model construction unit is used to construct a dynamic correlation model of deviation and influencing factors. The deviation feature extraction unit is used to extract static deviation features and dynamic deviation features, providing data support for subsequent model construction.

10. The intelligent detection and compensation system for voltage regulators based on data deviation analysis according to claim 8, characterized in that, The fusion compensation model construction module includes: A static compensation sub-model construction unit is used to construct a static compensation sub-model based on static deviation characteristics. A dynamic compensation sub-model construction unit is used to construct an attention mechanism LSTM dynamic compensation sub-model based on dynamic deviation characteristics. A fusion model integration unit is used to fuse the static compensation sub-model and the attention mechanism LSTM dynamic compensation sub-model to form a fusion compensation model. The model verification unit is used to verify the effectiveness of the fusion compensation model and ensure that it can accurately cancel out fixed deviations and real-time time-varying deviations.

Images