Condensation monitoring method for digital substation

Abstract

The application discloses a condensation monitoring method for a digitalized transformer substation, comprising the following steps: obtaining data information of a target digitalized transformer substation; determining a high-risk area of condensation occurrence and arranging a monitoring device to obtain monitoring data; selecting a monitoring frequency band, obtaining data of the monitoring frequency band and analyzing, and constructing a feature set related to a condensation degree; obtaining monitoring data of the target digitalized transformer substation in real time and selecting an electromagnetic wave monitoring frequency band; performing feature extraction and analysis to construct a condensation monitoring model and optimization; generating a condensation risk prediction result of the target digitalized transformer substation based on the optimized condensation monitoring model; giving optimal anti-condensation measures to the target digitalized transformer substation according to the prediction result, and completing condensation monitoring of the target digitalized transformer substation. Through obtaining, processing of environmental parameters of the digitalized transformer substation and establishment of a condensation prediction model, the condensation monitoring method not only realizes condensation monitoring of the digitalized transformer substation, but also has higher reliability and better accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electrical automation, and particularly relates to a condensation monitoring method for a digital substation. BACKGROUND

[0002] With the development of economy and technology and the improvement of people's living standards, electric energy has become an essential secondary energy in people's production and life, bringing endless convenience to people's production and life. Therefore, ensuring the stable and reliable supply of electric energy has become one of the most important tasks of the power system.

[0003] The substation is an important part of the power system, so the working state of the substation will directly affect the operating state of the power system. The internal environment of the substation is complex and changeable, and the changes of temperature, humidity and other factors are easy to cause condensation on the surface of the equipment, affecting the insulation performance and operation reliability of the equipment; therefore, the condensation monitoring in the substation is of great significance.

[0004] At present, the traditional condensation monitoring scheme adopted by the substation is still a simple sensor + threshold alarm scheme. Although this scheme is simple, the reliability of the monitoring data is poor, and the accuracy of the condensation monitoring result is also low. This undoubtedly makes the condensation monitoring effect of the substation poor, thereby bringing great hidden dangers to the safe and stable operation of the substation. SUMMARY

[0005] The purpose of the present application is to provide a condensation monitoring method for a digital substation with high reliability and good accuracy.

[0006] The condensation monitoring method for a digital substation provided by the present application comprises the following steps:

[0007] S1. obtaining data information of a target digital substation;

[0008] S2. based on the data information obtained in step S1, simulating the airflow movement characteristics in the target digital substation based on a numerical simulation scheme, determining a high-risk area of condensation occurrence, and arranging corresponding monitoring equipment to obtain monitoring data;

[0009] S3. selecting a monitoring frequency band according to the response characteristic differences of different frequency bands of electromagnetic waves to condensation, obtaining data of the monitoring frequency band and analyzing the data, thereby constructing a feature set related to the degree of condensation;

[0010] S4. obtaining monitoring data of the target digital substation in real time, and selecting an electromagnetic wave monitoring frequency band based on the current environment;

[0011] S5. performing feature extraction and analysis according to the electromagnetic wave monitoring frequency band obtained in step S4, to construct a condensation monitoring model based on the transmission characteristics of electromagnetic waves;

[0012] S6. The condensation monitoring model constructed in step S5 is optimized based on the fused environmental parameters and electromagnetic wave parameters;

[0013] S7. The condensation risk prediction result of the target digital substation is generated based on the optimized condensation monitoring model obtained in step S6;

[0014] S8. The optimal anti-condensation measures for the target digital substation are given according to the prediction result obtained in step S7, and the condensation monitoring of the target digital substation is completed.

[0015] The step S2 specifically comprises the following steps:

[0016] A three-dimensional model of the target digital substation is established, and a computational fluid dynamics scheme is used to numerically simulate the air flow movement in the target digital substation to obtain the air flow movement characteristic parameters of different regions in the target digital substation;

[0017] The obtained air flow movement characteristic parameters are correlated with temperature and humidity to determine the high-risk areas of condensation occurrence in the target digital substation;

[0018] In the determined high-risk areas, corresponding monitoring equipment is arranged to obtain monitoring data.

[0019] The step S3 specifically comprises the following steps:

[0020] Based on electromagnetic wave transmission theory and condensation physical mechanism analysis, a plurality of condensation-sensitive frequency bands are selected as candidate frequency bands;

[0021] A condensation simulation experiment platform is built, and experiments are carried out under different condensation conditions to obtain electromagnetic wave transmission characteristic test data of the candidate frequency bands;

[0022] The obtained electromagnetic wave transmission characteristic test data is fitted and regression analyzed to obtain a correlation scale model between the condensation degree and the electromagnetic wave frequency band and transmission characteristic parameter;

[0023] According to the correlation scale model, the monitoring frequency band is determined to obtain a monitoring frequency band set;

[0024] According to the obtained monitoring frequency band, electromagnetic wave transmission data is collected to obtain an electromagnetic wave transmission data set reflecting the condensation state;

[0025] The obtained electromagnetic wave transmission data reflecting the condensation state is subjected to video domain analysis to extract characteristic parameters reflecting the condensation degree;

[0026] The feature selection algorithm is used to screen and optimize the obtained feature parameters, so as to obtain an optimal feature subset related to the condensation degree, and a condensation degree prediction feature set is constructed;

[0027] The obtained condensation degree prediction feature set is used to train the support vector regression algorithm, so as to obtain a condensation degree prediction model;

[0028] The obtained condensation degree prediction model is used for condensation degree prediction of the target digital substation.

[0029] The step S4 specifically comprises the following steps:

[0030] Real-time acquisition of environmental monitoring data of the target digital substation; the environmental monitoring data includes temperature data, humidity data, pressure data and light intensity data of the monitoring point;

[0031] The clustering algorithm is used for clustering analysis of the obtained environmental monitoring data, and the target digital substation is divided into several environmental regions according to the clustering results;

[0032] Condensation sensors are installed in each environmental region to monitor the condensation data on the surface of the equipment in real time;

[0033] A monitoring frequency band is selected, and the selected monitoring frequency band is used for data transmission of the condensation data.

[0034] The step S5 specifically comprises the following steps:

[0035] The maximum relevance minimum redundancy algorithm and the Relief-F algorithm are used for feature selection of the electromagnetic wave attenuation feature data, so as to screen the most relevant attenuation features related to the condensation degree as a candidate feature set;

[0036] Feature extraction and fusion are performed on the candidate feature set; the principal component analysis method is used to extract the main component features of the attenuation features, the independent component analysis method is used to extract the independent component features of the attenuation features, and the auto-encoder is used to extract the deep nonlinear features of the attenuation features, so as to obtain the attenuation feature parameters;

[0037] Condensation image data and video data on the surface of the equipment are acquired;

[0038] Image feature parameters of the acquired image data and video data are extracted, and the condensation degree is scored according to the set requirements, so as to obtain a condensation observation sample data set;

[0039] The condensation observation sample data set is used to train the gradient boosting regression tree model, so as to obtain a condensation monitoring model;

[0040] Meteorological prediction data of a target digital substation is acquired, and internal environment monitoring data and electromagnetic wave attenuation data are acquired, and a long short-term memory neural network is trained using the acquired data to obtain a condensation trend prediction model;

[0041] The condensation trend prediction model is used to predict a condensation trend of the target substation.

[0042] The step S6 specifically comprises the following steps:

[0043] The acquired environment parameters and electromagnetic wave parameters are preprocessed by data alignment, time series interpolation and outlier rejection algorithm to obtain a fusion feature vector;

[0044] The fusion feature vector is used to optimize the condensation monitoring model constructed in step S5.

[0045] The step S7 specifically comprises the following steps:

[0046] The optimized condensation monitoring model is used to predict a condensation risk of the target digital substation;

[0047] According to the predicted condensation risk probability, a clustering algorithm is used for clustering grouping, and different risk levels are identified;

[0048] A monitoring grid of the target digital substation is set as an undirected weighted graph, each monitoring point corresponds to a node of the graph, there is a weighted edge between adjacent monitoring points, and the weight of the edge is set; a Node2Vec algorithm is used to obtain a low-dimensional embedding vector of the node by optimizing the node collinearity probability; an IDW algorithm is used to interpolate and predict the risk level of an unknown node; based on the interpolation prediction result, a risk distribution map is generated using a substation geographic information map as a base map and using monitoring points as control points; and different risk level regions are identified.

[0049] The step S8 specifically comprises the following steps:

[0050] According to the prediction result obtained in step S7, a region with a condensation risk greater than a set threshold is obtained, and combined with layout data of the target substation, a specific position and coverage range are determined to obtain a condensation high-risk region position list;

[0051] A condensation prevention measure knowledge base is constructed;

[0052] In the condensation prevention measure knowledge base, the environment characteristics of the condensation high-risk region are compared with the environment characteristic data in the condensation prevention measure knowledge base in terms of similarity, and a condensation prevention measure with a similarity higher than a set threshold is selected as a candidate measure;

[0053] Among the various candidate measures, the cost data of each candidate measure is calculated according to the installation position and installation mode of the candidate measure, and the regional environmental parameter change after implementation of each candidate measure is calculated through the three-dimensional model of the target digital substation, the anti-condensation effect of the candidate measure is evaluated, the final anti-condensation measure is determined according to the cost data and the anti-condensation effect, and the condensation monitoring of the target digital substation is completed.

[0054] The condensation monitoring method of the digital substation further includes the following steps:

[0055] S9. Establish a condensation monitoring protection platform to realize condensation monitoring, data management and protection measure optimization of the target digital substation.

[0056] The step S9 specifically includes the following steps:

[0057] The obtained data is preprocessed to generate a condensation monitoring data set;

[0058] Based on a support vector machine and a convolutional neural network, an initial condensation risk assessment model is constructed, and the generated condensation monitoring data set is used for training to obtain a condensation risk assessment model;

[0059] The obtained condensation risk assessment model is used to obtain a condensation risk assessment result, and a condensation risk distribution map is generated in combination with the three-dimensional model of the target digital substation;

[0060] According to the generated condensation risk distribution map, the anti-condensation measures are dynamically optimized based on a multi-objective planning scheme.

[0061] The condensation monitoring method of the digital substation provided by the application, through the acquisition, processing of the environmental parameters of the digital substation and the establishment of the condensation prediction model, not only realizes the condensation monitoring of the digital substation, but also has higher reliability and better accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a method flowchart of the method of the application. DETAILED DESCRIPTION

[0063] As Figure 1 It is a method flowchart of the method of the application: the condensation monitoring method of the digital substation disclosed by the application includes the following steps:

[0064] S1. Obtain data information of a target digital substation;

[0065] S2. Based on the data information obtained in step S1, the airflow movement characteristics in the target digitalized substation are simulated based on a numerical simulation scheme, high-risk areas of condensation occurrence are determined, and corresponding monitoring devices are arranged to obtain monitoring data; specifically including the following steps:

[0066] A three-dimensional model of the target digitalized substation is established, a computational fluid dynamics scheme is used to numerically simulate the airflow movement in the target digitalized substation, and airflow movement characteristic parameters of different regions in the target digitalized substation are obtained;

[0067] The obtained airflow movement characteristic parameters are correlated and analyzed with temperature and humidity to determine high-risk areas of condensation occurrence in the target digitalized substation;

[0068] In the determined high-risk areas, corresponding monitoring devices are arranged to obtain monitoring data;

[0069] In specific implementation, the computational fluid dynamics method is used to numerically simulate the airflow movement in the substation, the finite volume method is used to discretize the control equation, and the SIMPLE algorithm is used to solve the velocity field and pressure field. The calculation domain is divided into 500,000 hexahedral grids, the boundary conditions are set as an inlet velocity of 5 m / s, an outlet pressure of standard atmospheric pressure, and a wall surface of adiabatic no-slip boundary. The simulation results show that the airflow velocity distribution in the substation is uneven, the airflow velocity is low near the equipment, and the airflow velocity is high in the channel, with a maximum speed of 2 m / s. The airflow velocity and direction are correlated and analyzed with temperature and humidity, and the least square method is used to fit the airflow velocity and temperature, which are negatively correlated, and the airflow velocity and humidity, which are positively correlated, with correlation coefficients of -78 and 69, respectively. According to the correlation analysis results, the area near the equipment is determined as the high-risk area of condensation, 10 temperature and humidity sensors and 5 electromagnetic wave monitoring points are arranged in this area, the sampling frequency is 1 minute, and the environmental parameter and electromagnetic wave intensity data are collected in real time. Through statistical analysis of the monitoring data, the average temperature of the condensation high-risk area is 25℃, the average humidity is 85%, and the average electromagnetic wave intensity is 50V / m, which provides data support for condensation early warning;

[0070] S3. According to the response characteristic differences of different frequency bands of electromagnetic waves to condensation, the monitoring frequency band is selected, the data of the monitoring frequency band are obtained and analyzed, and a feature set related to the degree of condensation is constructed; specifically including the following steps:

[0071] Based on electromagnetic wave transmission theory and condensation physical mechanism analysis, a number of frequency bands sensitive to condensation are selected as candidate frequency bands;

[0072] A condensation simulation experiment platform is built, experiments are carried out under different condensation conditions, and electromagnetic wave transmission characteristic test data of the candidate frequency bands are obtained;

[0073] Fitting and regression analysis are performed on the obtained electromagnetic wave transmission characteristic test data to obtain a correlation scale model between the condensation degree and the electromagnetic wave frequency band and the transmission characteristic parameter;

[0074] According to the correlation scale model, a monitoring frequency band is determined to obtain a monitoring frequency band set;

[0075] According to the obtained monitoring frequency band, electromagnetic wave transmission data are collected to obtain an electromagnetic wave transmission data set reflecting the condensation state;

[0076] Video domain analysis is performed on the obtained electromagnetic wave transmission data reflecting the condensation state to extract characteristic parameters capable of reflecting the condensation degree;

[0077] A feature selection algorithm is used to screen and optimize the obtained characteristic parameters to obtain an optimal feature subset related to the condensation degree, and a condensation degree prediction feature set is constructed;

[0078] The obtained condensation degree prediction feature set is used to train a support vector regression algorithm to obtain a condensation degree prediction model;

[0079] The obtained condensation degree prediction model is used to predict the condensation degree of a target digital substation;

[0080] In specific implementation, MATLAB and other tools are used to perform curve fitting and regression analysis on experimental data, to obtain a mathematical model between the degree of condensation and the frequency band of electromagnetic waves, transmission characteristic parameters, and to quantitatively describe the correlation between them, such as the negative correlation between the degree of condensation and the frequency of electromagnetic waves, and the positive correlation between the degree of condensation and the transmission loss. According to the correlation model, the optimal monitoring frequency band set that is most sensitive to condensation and least affected by environmental interference is determined, providing guidance for subsequent data collection. The multi-band electromagnetic wave transmission data collection technology is used to synchronously collect electromagnetic wave transmission data in each frequency band in the selected monitoring frequency band set, and to obtain a multi-band electromagnetic wave transmission data set reflecting the condensation state. Wavelet transform, Fourier transform and other signal processing algorithms are used to perform time-frequency domain analysis on the preprocessed multi-band electromagnetic wave transmission data, and to extract characteristic parameters such as attenuation characteristics, phase characteristics and spectral characteristics that can reflect the degree of condensation. A feature selection algorithm based on recursive feature elimination is used to screen and optimize the extracted multiple characteristic parameters, remove redundant and irrelevant features, and obtain an optimal feature subset highly correlated with the degree of condensation, thereby constructing a feature set for condensation degree prediction. A machine learning algorithm such as support vector regression is selected, the optimal feature subset is taken as the input, and the degree of condensation is taken as the output, to construct a condensation degree prediction model. The kernel function, penalty coefficient, error tolerance and other hyperparameters of SVR are optimized and selected by using grid search and other methods, to improve the fitting ability of the model. The monitoring data set is randomly divided into a training set and a test set according to a certain proportion, the SVR model is trained on the training set, and a nonlinear mapping relationship between the degree of condensation and the characteristic parameters is obtained; the performance is evaluated on the test set, the generalization ability of the model is verified through mean square error, determination coefficient and other indicators, the model is fine-tuned and optimized, and finally a condensation degree prediction model with superior performance is obtained. The trained condensation degree prediction model is integrated into the substation condensation monitoring system, multi-band electromagnetic wave transmission data are collected in real time, and characteristic parameters are extracted, which are input into the prediction model for online evaluation of the degree of condensation, to monitor the condensation risk in the substation in real time. When the predicted degree of condensation exceeds the preset threshold, the early warning mechanism is triggered, and warning is given to the on-duty personnel through sound and light alarm, short message notification and other ways, and the condensation suppression device such as heating, ventilation and the like is started, to timely relieve the condensation risk and ensure the safe operation of the substation. The condensation monitoring data and warning records are stored in the database to form condensation big data, which provides data support for subsequent condensation rule analysis and prediction model optimization, and realizes the intelligentization and digitization of substation condensation monitoring.

[0081] In the construction of the condensation simulation experiment platform, the following equipment and parameters can be used: the temperature range of the environmental simulation cabin is -20℃-50℃, the humidity range is 30%-100%, and the pressure range is 0.5-1.5 atm; the frequency range of the multi-band electromagnetic wave emission module covers 1 GHz-1000 GHz, the emission power is adjustable, and the maximum power is not less than 1 W; the receiving sensitivity of the multi-band electromagnetic wave receiving module is better than -100 dBm, the receiving frequency is synchronized with the emission frequency, and wideband reception can be achieved through difference frequency technology; the sampling rate of the data acquisition module is not less than 10 GSPS, the bit depth is not less than 12 bits, and real-time data transmission and storage functions are supported. By controlling the temperature and humidity parameters of the environmental simulation cabin, different condensation levels are simulated, and the transmission characteristics of electromagnetic waves of different frequency bands are tested by changing the emission frequency of the electromagnetic wave emission module. By statistical analysis, key indicators such as transmission attenuation and phase delay are extracted, and the correlation between the electromagnetic wave frequency band and the condensation response is preliminarily judged. Through mathematical methods such as multiple regression and principal component analysis, a quantitative relationship model between the frequency band, environmental parameters, and transmission characteristics indicators is established, and the best monitoring frequency band is finally selected. In this process, the selection of environmental parameters, the collection and analysis of experimental data are the key, and the technical settings in related condensation experiments and electromagnetic wave transmission research can be referred to. For the preprocessing of multi-band electromagnetic wave transmission data, the wavelet transform combined with EMD empirical mode decomposition method can be used. First, the original data is decomposed into wavelet coefficients at different frequency scales using wavelet transform. At each scale, the wavelet coefficients are decomposed into several intrinsic mode functions (IMFs) and a residual component using EMD method. According to the frequency characteristics and energy distribution of IMFs, the effective IMFs reflecting condensation information and the ineffective IMFs representing noise interference are distinguished. The ineffective IMFs are removed, the effective IMFs are retained, and the denoised wavelet coefficients are reconstructed. Finally, the preprocessed electromagnetic wave transmission data is obtained using wavelet inverse transform. On this basis, the Hilbert transform can be used to calculate the instantaneous frequency and instantaneous amplitude, or the singular value decomposition can be used to extract the main feature vector of the data as the input of the subsequent feature index extraction. In the construction of the condensation level prediction model, the kernel function of the SVR model can be selected as the Gaussian kernel function, i.e. where x, y are input feature vectors, and s is the kernel function width parameter. The penalty coefficient C can be adjusted in the range of 0.1-100, and the error tolerance e can be adjusted in the range of 0.01-1. In the grid search process, a rough value grid can be determined first, such as C in {0.1, 1, 10, 100}, s in {0.01, 0.1, 1, 10}, and e in {0.01, 0.05, 0.1, 0.5}. The model performance under each parameter combination is determined by 5-fold cross-validation to select the optimal parameter range. Then a fine value grid is constructed near the optimal parameters, such as C in {5, 7, 9, 11, 13}, s in {0.5, 0.7, 0.9, 1.1, 1.3}, and the above process is repeated to find the optimal parameter combination. On this basis, the learning curve can also be used to diagnose whether the model is overfitting or underfitting, and to improve it through regularization terms, sample size, etc. The model performance evaluation indicators can be selected from mean square error (MSE), mean absolute error (MAE), and determination coefficient (R 2 etc. to comprehensively evaluate the fitting effect and prediction ability of the model. In the condensation early warning and mitigation mechanism, multiple warning thresholds can be set, such as triggering yellow warning at 30% condensation level, orange warning at 50%, and red warning at 70%. Different warning levels correspond to different mitigation measures, such as starting low-power heating and natural ventilation at yellow warning, medium-power heating and mechanical ventilation at orange warning, and high-power heating and forced convection at red warning. At the same time, different frequency and content of warning messages are sent to the on-duty personnel according to the warning level, such as once an hour, once every 30 minutes, once every 10 minutes, etc. In practical application, the warning thresholds and mitigation measures can also be optimized according to the environmental characteristics, equipment layout, etc. of the substation to improve the sensitivity of the warning and the effectiveness of the mitigation. For data from multiple monitoring points, prediction models and warning mechanisms can be established respectively, and comprehensive warning can be achieved through data fusion, logical judgment, etc. to improve the reliability of the system;

[0082] S4. Real-time acquisition of monitoring data of the target digital substation, and selection of electromagnetic wave monitoring frequency band based on the current environment; specifically including the following steps:

[0083] Real-time acquisition of environmental monitoring data of the target digital substation; the environmental monitoring data includes temperature data, humidity data, pressure data, and light intensity data of each monitoring point;

[0084] For the acquired environmental monitoring data, clustering algorithm is used for clustering analysis, and the target digital substation is divided into several environmental regions according to the clustering results;

[0085] Condensation sensors are installed in each environmental region to monitor the condensation data on the surface of the equipment in real time;

[0086] The monitoring frequency band is selected, and the condensation data is transmitted using the selected monitoring frequency band.

[0087] In implementation, environmental monitoring points are arranged in different areas inside the substation to collect environmental parameter data such as temperature, humidity, pressure, and light intensity, and the collected data is transmitted to the data processing center of the monitoring system in real time. In the data processing center, the collected environmental parameter data is preprocessed through data cleaning, data normalization, and other operations to remove outliers and invalid values, convert the data to a unified dimension and scale, and form a standardized environmental parameter dataset. According to the environmental characteristics of the substation, the range of candidate frequency bands is determined, such as extremely low frequency, ultra-low frequency, high frequency, and microwave, and the electromagnetic wave transmission characteristics of each candidate frequency band are tested under different environmental conditions to obtain transmission attenuation, phase delay, signal-to-noise ratio, and other index data. The obtained index data is analyzed, and the comprehensive performance score of each frequency band under different environmental conditions is calculated. The higher the score, the less the frequency band is affected by the environment. According to the comprehensive performance score, several optimal frequency bands are selected for each environmental condition to establish a frequency band-environment condition mapping table and form an electromagnetic wave monitoring frequency band library. The frequency band library can be established offline through field testing, simulation calculation, or real-time optimized through online learning and dynamic updating. The standardized environmental parameter dataset is analyzed using clustering algorithms such as K-means and DBSCAN, and the internal environment of the substation is divided into several environmental regions based on the clustering results. The environmental conditions in each region are relatively uniform, and there are significant differences between different regions. For each environmental region, several candidate frequency bands with the least environmental condition interference are selected from the electromagnetic wave monitoring frequency band library based on the environmental condition characteristics of the region to form an optimal monitoring frequency band set for the region. Condensation sensors are installed at each monitoring point to detect the condensation amount on the surface of the equipment in real time through capacitive, resistive, and optical sensing principles, and the detection value is converted to standardized condensation data. At the same time, the current electromagnetic wave monitoring frequency band of each monitoring point is recorded, and the condensation data and frequency band information are associated to form a frequency band-condensation data pair. The collected frequency band-condensation data pairs are summarized and stored in the data processing center of the monitoring system to construct a frequency band-condensation mapping dataset. The type, installation location, and sampling frequency of the condensation sensor need to be optimized based on the characteristics of the equipment and the monitoring requirements to improve the timeliness and reliability of the condensation monitoring data. When new environmental parameter data is collected at the environmental monitoring points, the environmental region of each monitoring point is determined through data processing and feature extraction, and the optimal monitoring frequency band with the least environmental condition interference is dynamically selected from the optimal monitoring frequency band set of the region for condensation monitoring of the monitoring point. The frequency band-condensation mapping dataset is grouped according to environmental conditions and monitoring frequency bands, and the mean, variance, maximum, and minimum of the condensation degree in each group are calculated to evaluate the condensation monitoring performance of different frequency bands under different environmental conditions.The statistical methods such as hypothesis testing and variance analysis are used to determine whether there is a significant difference in the condensation monitoring performance between different frequency bands, and to analyze the reasons for the performance difference. According to the results of performance evaluation and difference analysis, the original frequency band selection strategy is dynamically adjusted and optimized, such as removing the frequency bands with poor performance, adding new frequency bands with excellent performance, adjusting the priority of each frequency band, etc. The optimized frequency band selection strategy is applied to the dynamic frequency band selection process to realize the adaptive optimization of the monitoring frequency band. The parameters such as the cycle of statistical analysis, the threshold of indicators, and the update trigger condition need to be flexibly set according to the real-time and accuracy requirements of condensation monitoring to balance the system performance and computing cost. At the same time, machine learning algorithms are fully utilized to realize the automatic optimization and intelligent evolution of the frequency band selection strategy, and continuously improve the accuracy and reliability of condensation monitoring. When arranging environmental monitoring points inside the substation, a grid-based layout can be used to divide the substation into several 20m x 20m grids, and an environmental monitoring point is installed at the center of each grid. Each monitoring point is equipped with temperature, humidity, pressure, and light intensity sensors with a range of -40℃ to 80℃, 0% to 100%, 500hPa to 1100hPa, and 0lux to 100000lux, respectively, and a resolution of 0.1℃, 0.1%, 1hPa, and 1lux, respectively. The sampling frequency is 1 time per minute. The environmental monitoring points are connected to the data processing center through the RS485 bus and use the Modbus RTU protocol for data transmission. Each data frame includes the monitoring point ID, sensor type, data timestamp, and measurement value fields, and the communication baud rate is 9600bps. The data processing center uses the FP-growth algorithm to perform frequent pattern mining on the received environmental parameter data, discovers the association rules between parameters, and uses the box plot and Z-score methods for anomaly detection and noise reduction to generate a standardized environmental parameter dataset. The test of electromagnetic wave transmission characteristics can choose an omnidirectional radiating monopole antenna as the transmitting antenna and a logarithmic periodic antenna as the receiving antenna with antenna gains of 2dBi and 10dBi, respectively, and a standing wave ratio less than 1.5. The spectrum analyzer and vector network analyzer are used to measure the frequency spectrum characteristics and S parameters of the transmission signal, respectively. The scanning frequency range is 10kHz to 10GHz, the scanning point number is 1601, the intermediate frequency bandwidth is 10kHz, and the dynamic range is greater than 120dB. The antennas are placed in different environmental areas of the substation, and the antenna spacing, incident angle, and other parameters are changed to measure the transmission attenuation, phase delay, and signal-to-noise ratio of electromagnetic waves in different frequency bands. Each frequency band is tested 10 times, and the average value is taken. The weighted sum method is used to comprehensively score the transmission characteristic indicators of each frequency band, and the weight coefficients can be determined by the AHP hierarchical analysis method. The comprehensive performance score of each frequency band under different environmental conditions is obtained, and the top 5 frequency bands with the highest score are selected as the optimal monitoring frequency bands under the environmental conditions.In the clustering analysis of environmental parameter data, the K-means algorithm can be used, the number of clusters K is set to 5, the Euclidean distance is used as the distance measure, the initial cluster center is selected by the K-means++ algorithm, and the iteration number is 100 times. The Silhouette index is evaluated for the clustering results, and if the index is greater than 0.7, it is considered that the clustering effect is good, and the substation environment can be divided into 5 regions. In each region, select the 3 frequency bands with the highest comprehensive performance score as the preferred monitoring frequency bands of the region, and construct a frequency band-environment region matrix. When collecting condensation data, a capacitive condensation sensor can be selected, with a measurement range of 0-100%, a resolution of 0.1%, and a response time of less than 10s. Condensation sensors are arranged on the surfaces of key equipment such as insulators, busbars, and transformers in the substation, with a sensor spacing of not more than 1m and a sampling frequency of 1 / 10min. The condensation monitoring data and frequency band information are aligned by timestamp, and stored as a frequency band-condensation time series dataset. When performing statistical analysis on the frequency band-condensation dataset, the following methods can be used: single-factor variance analysis is used for hypothesis testing, the significance level a = 0.5, and F test is used to determine whether there is a significant difference in the condensation mean value of different frequency bands; Tukey HSD post-hoc test is used to determine whether the condensation mean value difference between frequency bands is significant, and the frequency bands are clustered and sorted. The ARIMA model is used to predict the trend of the condensation time series, with a prediction step of 1 day, and the autocorrelation coefficient and partial autocorrelation coefficient of the residual series are determined by LB test to determine their stationarity. Compare the prediction results with the historical data, calculate the MAPE and RMSE indicators, and evaluate the condensation monitoring performance and prediction accuracy ranking of each frequency band. Update the condensation prediction model once a day, and trigger the update of the frequency band selection strategy when the performance ranking changes. The improved particle swarm algorithm is used to optimize the frequency band selection strategy, the fitness function includes condensation detection accuracy, energy consumption, cost, etc., the particle number is 50, the maximum iteration number is 200, the learning factor c1 = c2 = 2, and the inertia weight w decays from 0.9 to 0.4 according to the cosine function. The optimized frequency band selection strategy is sent to each monitoring point to dynamically update the monitoring frequency band.

[0088] S5. According to the electromagnetic wave monitoring frequency band obtained in step S4, feature extraction and analysis are performed to construct a condensation monitoring model based on electromagnetic wave transmission characteristics; specifically including the following steps:

[0089] Based on the maximum correlation minimum redundancy algorithm and the Relief-F algorithm, the electromagnetic wave attenuation feature data is selected, and the attenuation feature with the largest correlation with the condensation degree is selected as the candidate feature set;

[0090] Feature extraction and fusion are performed on the candidate feature set, principal component analysis is used to extract the main component features of the attenuation features, independent component analysis is used to extract the independent component features of the attenuation features, and a self-encoder is used to extract deep nonlinear features of the attenuation features to obtain attenuation feature parameters;

[0091] Obtain the condensation image data and video data of the surface of the device;

[0092] Extract the image feature parameters of the obtained image data and video data, score the condensation degree according to the set requirements, and obtain a condensation observation sample data set;

[0093] Use the condensation observation sample data set to train a gradient boosting regression tree model to obtain a condensation monitoring model;

[0094] Obtain the weather forecast data, internal environment monitoring data and electromagnetic wave attenuation data of the target digital substation, and use the obtained data to train a long short-term memory neural network to obtain a condensation trend prediction model;

[0095] Use the condensation trend prediction model to predict the condensation trend of the target substation;

[0096] In implementation, through the electromagnetic wave transmission test system, the electromagnetic wave attenuation data of selected frequency bands under different condensation environments are collected, including amplitude attenuation, phase delay, spectral change and other multi-dimensional time series data, and the sampling frequency is not less than 10 kHz, and the number of samples collected for each frequency band is not less than 10,000. The collected electromagnetic wave attenuation data is preprocessed, the data is denoised and smoothed by wavelet transform, the data is scaled and distributed by Z-score normalization, and the missing data is repaired by interpolation method to form a standardized attenuation data set. The MRMR maximum correlation minimum redundancy algorithm and the Relief-F algorithm are used to select the features of the standardized attenuation data set, and the 10 attenuation features with the highest correlation with the condensation degree are selected as the candidate feature set. The feature selection process ensures the stability and reliability of the results through 10-fold cross-validation. Further feature extraction and fusion are performed on the candidate feature set, the principal component analysis PCA is used to extract the main component features of the attenuation data, the independent component analysis ICA is used to extract the independent component features of the attenuation data, and the self-encoder is used to extract the deep nonlinear features of the attenuation data. Finally, a set of comprehensive attenuation feature parameters is obtained. In the transformer substation, several key positions prone to condensation, such as insulators, busbars, transformers, etc., are selected as condensation observation points, and high-definition cameras and visual sensors are installed at each observation point to collect image and video data on the surface of the equipment in real time. Computer vision algorithms are used to analyze the image and video data, extract the pixel number and gray value of the condensation area, and obtain physical quantity indicators such as condensation amount and coverage area. According to the condensation physical quantity indicators, combined with the insulation performance of electrical equipment, meteorological conditions and other factors, the condensation degree at each sampling time is scored, with a score range of 0-10 points. The higher the score, the more serious the condensation degree. The physical quantity indicators and scores are corresponded one by one to generate a condensation observation sample data set, each sample including fields such as collection time, collection location, physical quantity indicators and condensation degree score. The gradient boosting regression tree GBRT algorithm is used to train the condensation monitoring model, with the attenuation feature parameters as input and the condensation degree score as output. The model hyperparameters such as the number of trees, depth, learning rate, etc. are optimized through grid search, the model performance is evaluated through 5-fold cross-validation, and the optimal model is selected for deployment and application according to the actual application requirements. The trained condensation monitoring model is integrated into the online condensation monitoring system of the transformer substation to collect electromagnetic wave attenuation data of selected frequency bands in real time, extract attenuation feature parameters as input to the model, and continuously predict the condensation degree score. The meteorological forecast data of the area where the transformer substation is located, including temperature, humidity, pressure, wind speed, rainfall and other meteorological elements in the future period of time, are collected as external input features for condensation prediction. The meteorological forecast data is spatio-temporally aligned and fused with the internal environmental monitoring data and electromagnetic wave attenuation data of the transformer substation to form a multi-source heterogeneous condensation prediction sample set.A time series prediction model such as a long short-term memory (LSTM) neural network is used to train a multi-step and multi-variable condensation trend prediction model, with historical condensation levels and multi-source physical quantities as inputs and future condensation levels as outputs. The input data of the model is updated in real time using a sliding window method to dynamically predict the condensation level trend in the future period of time, and the alarm threshold and shutdown threshold are adjusted in a timely manner according to the prediction results. The condensation level prediction results are compared with the preset threshold, and if the warning threshold is exceeded, an audible and light alarm is triggered to prompt the operation and maintenance personnel, and if the shutdown threshold is exceeded, the device is forced to shut down for protection. The condensation trend prediction results are published to related business systems and decision makers through Web services to assist them in formulating condensation prevention measures and operation and maintenance strategies. In practical applications, attention should be paid to key technical issues such as spatio-temporal data fusion, multi-step prediction, and model updating to improve the timeliness and accuracy of the prediction. At the same time, visualization of the prediction results and human-computer interaction design should be done well to improve the usability and user experience of the system. When collecting electromagnetic wave attenuation data, a spectrum analyzer with a frequency range of 1 MHz-10 GHz, a dynamic range greater than 120 dB, and a sensitivity better than -130 dBm can be selected, and a horn antenna with a gain greater than 30 dBi and a standing wave ratio less than 1.2 can be used. More than 10 collection points are arranged in different areas of the substation, and data is collected every 10 minutes at each collection point. Each candidate frequency band is swept 100 times during each collection, and the average value is taken as the attenuation data for that frequency band. In the data preprocessing stage, the Daubechies wavelet basis function is used for 6-level discrete wavelet transform of the attenuation data, then the wavelet coefficients are denoised using the Donoho threshold, and the denoised attenuation data is obtained by wavelet reconstruction. The data is normalized using the Z-score method, and the linear interpolation method is used to fill in the data with a missing rate less than 10%, and the data with a missing rate greater than 10% is directly excluded. In the feature selection stage, the MRMR algorithm is used to calculate the mutual information of each attenuation feature and the condensation level, and the top 10 features with the largest mutual information are selected as the candidate feature set. Then the Relief-F algorithm is used to calculate the correlation statistics of each candidate feature, and the candidate features are sorted according to the correlation statistics. The top 5 features are selected as the final feature selection results. In the feature extraction stage, the PCA algorithm is used to reduce the dimension of the candidate feature set, and the principal components with a cumulative contribution rate greater than 95% are selected as the main component features. The FastICA algorithm is used to extract the independent components of the candidate feature set, and the independent components with a kurtosis greater than 5 are selected as the independent component features. Finally, the self-encoder is used for nonlinear transformation of the candidate feature set, and the hidden layer node output with a reconstruction error less than 0.01 is selected as the deep feature.In the process of condensation observation data collection, high-definition digital camera with resolution greater than 1000 million pixels and laser condensation sensor with measurement range of 0-100% and resolution better than 0.1% can be selected. Two cameras and four sensors are installed at key positions such as insulators, busbars and transformers in the substation. Image and condensation data are collected synchronously every 30 minutes. After pre-processing, image data is input into faster R-CNN target detection algorithm to extract the position and size information of condensation area. After outlier rejection and smoothing filtering, condensation data is used as condensation parameter. Image parameter and condensation parameter at the same time are combined into an observation sample. In the process of condensation degree scoring, each condensation observation sample is scored independently. According to the condensation image and condensation data in the sample, combined with the equipment layout, meteorological conditions, operation conditions and other factors of the substation, the score of 0-10 is given. The median of 10 score results is taken as the final score of the sample. Then, the image parameter, condensation parameter and score result of all samples are combined into a condensation observation sample data set. The data set is randomly divided into training set and test set according to the ratio of 8:2. In the training stage of condensation monitoring model, GBRT algorithm is used to fit the training set data. The model hyperparameters are optimized by grid search method. The parameter search range is: number of trees [50, 100, 200, 500], maximum depth of tree [3, 5, 7, 9], learning rate [0.01, 0.05, 0.1, 0.2]. The loss function is mean square error, and the evaluation index is determination coefficient R. 2 Then, the model is evaluated by 5-fold cross-validation method. The average R 2The group of hyperparameters with the maximum value is taken as the optimal parameters, and a trained GBRT condensation monitoring model is finally obtained. In the model application stage, the trained GBRT model is used to predict the newly collected electromagnetic wave attenuation data. First, the feature parameters of the attenuation data are extracted, and then input into the model to obtain the condensation degree score. Then, the score is compared with the pre-set threshold value. If the score exceeds the early warning threshold value, an early warning short message is sent to the on-duty personnel. If the score still exceeds the shutdown threshold value after 3 consecutive early warnings, the device is immediately triggered to shut down. At the same time, a multivariate time series prediction model is constructed using the LSTM network. The input variables include the condensation degree score in the past 24 hours, electromagnetic wave attenuation features, meteorological elements, etc. The output variable is the condensation degree score in the future 1 hour, 2 hours and 4 hours. The input data of the model is updated every 1 hour and a prediction is made once every 1 hour. The prediction result is used to assist in adjusting the early warning and shutdown threshold values. The main parameter settings of the LSTM model are as follows: input time step 24, hidden layer neuron number 128, output neuron number 3, learning rate 0.001, batch size 32, training round number 100, activation function tanh, loss function mean square error, and optimization algorithm Adam. In the deployment application, a RESTful API service is built using the Flask Web framework to receive real-time data from the monitoring device, call the prediction model for calculation, and push the prediction result to the Web front end in real time through the WebSocket protocol. The front end uses the Echarts component to visualize the prediction result and sets a human-computer interaction function to allow users to adjust the threshold settings and query historical data.

[0097] S6. The condensation monitoring model constructed in step S5 is optimized based on the fused environmental parameters and electromagnetic wave parameters. The specific steps include the following steps:

[0098] The obtained environmental parameters and electromagnetic wave parameters are preprocessed through data alignment, time series interpolation and outlier elimination algorithm to obtain a fused feature vector.

[0099] The fused feature vector is used to optimize the condensation monitoring model constructed in step S5.

[0100] In implementation, environmental sensors and electromagnetic wave sensors are arranged in different areas inside the substation to collect environmental parameters such as temperature, humidity, and pressure, as well as electromagnetic wave amplitude, phase, and spectrum characteristic parameters. The sampling frequency of the sensors is not less than 1 time per minute, the range covers the actual working condition range of the substation, the data resolution is better than 0.1 level, and the collected environmental parameters and electromagnetic wave characteristic parameters are uploaded to the edge computing gateway in real time through MQTT and other Internet of Things communication protocols. In the edge computing gateway, the collected environmental parameters and electromagnetic wave characteristic parameters are subjected to data fusion processing, and parameters of different sources and different spatiotemporal granularities are integrated into a unified time series data set to form a fusion feature vector containing environmental features and electromagnetic wave features. The fusion feature vector is encoded using the UUID algorithm to generate a globally unique feature identification code. The local fusion feature data set is uploaded to the cloud big data platform through a secure communication tunnel, and massive fusion feature data is stored, cleaned, indexed, and calculated using Hadoop, Spark, and other big data processing frameworks. Specifically, the fusion feature data is stored in the Parquet columnar storage format on the HDFS distributed file system, and the compression and encoding mechanism of Parquet is used to improve the storage efficiency and reading performance of the data. SparkSQL is used to perform ETL processing on the fusion feature data, and through data filtering, data conversion, data aggregation, and other operations, data cleaning and conversion are realized. The query optimizer and in-memory computing engine of SparkSQL are used to speed up data processing. SparkMLlib machine learning library is used for feature engineering and statistical analysis of fusion feature data, such as feature selection, feature scaling, correlation analysis, etc., to mine the internal rules and correlations of the data. SparkStreaming is used to process newly collected fusion feature data in real time, and through the setting of sliding window and trigger interval, real-time cleaning and calculation of data are realized. The processing results are written into a distributed in-memory database such as Redis for subsequent model training and prediction. Through data visualization technology, multi-dimensional display and interactive analysis of fusion features are realized to discover the correlation rules and abnormal patterns between environmental parameters and electromagnetic wave characteristic parameters. Based on the cloud fusion feature data set, a machine learning sample set for condensation monitoring is constructed, and the sample features include fusion feature vectors, feature identification codes, and condensation levels. According to the 8:2 ratio, the sample set is divided into training set and test set. Gradient boosting decision tree algorithms such as XGBoost and LightGBM are used to train the condensation monitoring model with the training set samples as input.During the model training process, data augmentation and transfer learning mechanisms are added. Through data augmentation operations such as random rotation, translation, and noise addition on training samples, the adaptability of the model to environmental changes is improved. Through transfer learning technology, the condensation monitoring model trained in other substation scenarios is used as a pre-trained model for fine-tuning and retraining in the new substation environment, reducing the demand for training samples, accelerating the convergence speed of the model, and improving the robustness of the model. Through feature engineering and parameter optimization, the feature expression ability and generalization performance of the model are further improved. The model evaluation indicators include precision, recall, F1 value, etc., which comprehensively balance the fitting ability and prediction ability of the model. The robustness of the condensation monitoring model is analyzed and the generalization performance is evaluated. Through multiple repeated experiments, the mean, variance, and confidence interval of the model output results are evaluated to judge the anti-interference ability and prediction reliability of the model. ROC curve and confusion matrix are used to evaluate the classification performance of the model, and the model structure and parameters are optimized accordingly. The trained and optimized condensation monitoring model is deployed to the edge computing gateway to realize real-time monitoring and early warning locally. When the edge gateway collects new environmental parameters and electromagnetic wave feature parameters, it automatically performs data fusion and feature extraction, inputs the fused features into the condensation monitoring model, and calculates the condensation degree prediction value online. According to the comparison result of the prediction value and the preset threshold, the corresponding level of condensation warning is triggered, and the warning information is pushed to the on-duty personnel through SMS, email, voice call, etc. At the same time, the monitoring data, prediction results, and warning state information are transmitted back to the cloud to realize cloud-edge collaborative closed-loop optimization. A data analysis and model optimization platform is built in the cloud to receive feedback information such as monitoring data, prediction results, and warning state from the edge, and store it in the time series database and object storage to form a historical data set. Time series data analysis and visualization tools are used to analyze trends, detect anomalies, and discover patterns in historical data to uncover underlying rules and issues. AutoML tools are used to automatically optimize the condensation monitoring model through hyperparameter search, model integration, and automatic feature engineering to find the optimal model structure and parameter configuration. Online learning algorithms are used to incrementally train and update the model to adapt to changing data distributions and environmental conditions. The optimized model and configuration parameters are downloaded to the edge to update the monitoring model and warning thresholds, enabling online model updates and performance improvements. A / B testing and multi-armed bandit experiments are used to evaluate the effectiveness and benefits of the new model, forming a data-driven continuous optimization loop. In the edge gateway, linear interpolation is used to align environmental parameter data over time, Tukey algorithm is used to remove outliers from electromagnetic wave feature parameter data, and standardization is used to normalize both types of data to a standard normal distribution with a mean of 0 and a variance of 1.The environmental parameter data and electromagnetic wave characteristic parameter data are merged by timestamp according to a 1-minute time window to generate a fusion feature vector, each vector containing 10 environmental features and 50 electromagnetic wave features. A 128-bit feature identification code is generated using the UUID algorithm to form the final fusion feature dataset. The QoS level of the MQTT message is set to 2 to ensure the reliability of data transmission. The fusion feature dataset is stored in Parquet format on HDFS, partitioned by collection time and feature identification code, and compressed using the Snappy algorithm. The fusion features are processed using SparkSQL, filtering out abnormal data, selecting key features, and aggregating data using the agg function to generate statistical indicators such as mean, variance, and correlation coefficient for environmental parameters and electromagnetic wave parameters. The random forest algorithm is used to evaluate the importance of the fusion feature data, calculate the Gini index of each feature, and select the top 20% of important features as model inputs. Pearson correlation coefficient, Spearman rank correlation coefficient, and other methods are used to analyze the correlation between environmental parameters and electromagnetic wave parameters, and heat maps and scatter plots are used to visualize the results. The Word2Vec algorithm based on SparkMLlib is used to convert the fusion feature vector into a 256-dimensional word vector for subsequent model training and similarity calculation. The sliding window mechanism is used with a window length of 10 minutes and a sliding step of 5 minutes to accumulate statistics and detect anomalies in real-time data streams, and the results are written to the Redis cluster with a data expiration time of 24 hours. The training sample set for the condensation monitoring model contains 10,000 positive samples and 50,000 negative samples. Positive samples are fusion feature vectors when condensation occurs, and negative samples are fusion feature vectors when condensation does not occur. The samples are arranged in chronological order, with the first 80% as the training set and the last 20% as the test set. The XGBoost algorithm is used to train a binary classification model, using gbtree as the base classifier, setting the maximum tree depth to 6, the learning rate to 0.1, the regularization coefficient to 1, the objective function to cross-entropy loss function, and the evaluation indicators to AUC value, F1 score, precision, and recall. The grid search method is used to optimize the number of trees, feature sampling ratio, and other hyperparameters. Each set of hyperparameter combinations is trained for 10 rounds, and the average AUC value is used as the scoring criterion to select the hyperparameter combination with the highest average AUC value as the optimal parameters. Data augmentation operations such as 5-degree random rotation, 10% random scaling, and 0.01 Gaussian noise are performed on the training samples to improve the model's generalization ability. The condensation monitoring model trained in other 10 substations is used as a pre-trained model, and 800 samples are used to fine-tune the model to optimize the weight parameters of the fully connected layer. Transfer learning can improve the model's generalization performance by 3-5 percentage points.The condensation monitoring model is cross-validated 10 times with 10-fold cross-validation. Each time, the training set and the validation set are randomly divided, the model is trained on the training set, and the model is evaluated on the validation set. The evaluation indicators of the 10 models are obtained, and the mean and standard deviation of the indicators are calculated. The mean reflects the average performance of the model, and the standard deviation reflects the stability of the model. When the mean of AUC reaches 0.95 and the standard deviation is less than 0.01, it indicates that the model performance is stable. Using the Monte Carlo simulation method, 10000 groups of random data are generated for the environmental parameters, electromagnetic wave parameters, and condensation labels of the model input. The data distribution is the same as the real distribution. The performance of the model on random data is evaluated. If the generalization performance decreases by no more than 5%, it indicates that the model has good robustness. The model is deployed on the edge gateway and loaded and predicted using the FlinkML library. The average prediction latency is less than 50ms, and the resource utilization rate is not more than 50%. It can meet the performance requirements of real-time early warning. When the condensation occurrence probability exceeds 70%, trigger level one warning, send warning message to the on-duty personnel, and control the equipment heating and ventilation at the same time. When the condensation occurrence probability exceeds 90%, trigger level two warning, send voice call to the operation responsible person, and control the equipment shutdown and isolation at the same time. After each warning trigger, upload the warning data to the cloud, and update the warning state and processing record on the cloud. Use InfluxDB time series database to store condensation monitoring and warning data on the cloud. The data retention policy is 1 year, and the data sampling granularity is 1 minute. Use Grafana to display the time series curve of condensation occurrence, the spatio-temporal distribution of warning, the statistical charts of environmental parameters and electromagnetic wave parameters, etc. Configure the Grafana alarm plugin. When the edge data delay exceeds 5 minutes and the warning times exceed 10, send an alarm email to the system administrator. Use Auto-Sklearn automatic machine learning tool to input historical data and optimize model performance. Automatically search for data preprocessing, feature engineering, algorithm selection, hyperparameter tuning, and a series of modeling steps to get an optimal model Pipeline. Export the model in ONNX format and deploy it on the edge for prediction. Use TensorflowServing to build a model service and use Flask to implement a web service interface to receive prediction requests from the edge and return prediction results. The average service latency is controlled within 100ms. Use A / B testing to randomly split the edge requests in a 9:1 ratio to request the current production model and the new training model. Record the prediction results and prediction time of the two models. Use t-test to compare the performance difference of the two models. When the new model performance is significantly better than the current model and the prediction time is not more than 1.5 times of the current model, start the model replacement process, deploy the new model online, and update the edge configuration parameters.A cloud-edge monitoring platform is built using Prometheus and Grafana to collect and aggregate various monitoring indicators of edge nodes in the cloud, including CPU utilization, memory utilization, disk IO, network bandwidth, data delay, model prediction time consumption, etc. The LSTM algorithm is used to predict the trend of the monitoring indicators, and the Isolation Forest algorithm is used for anomaly detection. When the indicators exceed the threshold or the predicted trend is abnormal, the cloud alarm is triggered to notify the on-duty personnel for inspection and processing, ensuring the high reliability and high availability of the condensation monitoring service.

[0101] S7. Based on the optimized condensation monitoring model obtained in step S6, generate the condensation risk prediction result of the target digital substation; specifically including the following steps:

[0102] Using the optimized condensation monitoring model, predict the condensation risk of the target digital substation;

[0103] According to the predicted condensation risk probability, use clustering algorithm for clustering grouping, and identify different risk levels;

[0104] Set the monitoring grid of the target digital substation as an undirected weighted graph, each monitoring point corresponds to a node of the graph, there is a weighted edge between adjacent monitoring points, and the weight of the edge is set. Using Node2Vec algorithm, the low-dimensional embedding vector of the node is obtained by optimizing the node collinearity probability; using IDW algorithm to interpolate and predict the risk level of unknown nodes; based on the interpolation prediction result, taking the substation geographic information map as the base map and the monitoring point as the control point, generating the risk distribution map; for different risk level areas, identify them;

[0105] In implementation, by dividing the monitoring grid inside the substation, the grid size is set to 10m*10m or 20m*20m according to the actual area and equipment distribution of the substation, one condensation monitoring point is arranged in each grid, the condensation state of electrical equipment in the grid is monitored, the data collected by each monitoring point is transmitted back to the condensation monitoring cloud platform in real time through the wireless sensor network, and the monitoring data is stored and managed in the cloud platform using the HDFS distributed file system and the HBase distributed database. The massive monitoring data returned by each monitoring point is processed in parallel using the MapReduce distributed computing framework, the data quality is improved through data preprocessing operations such as data cleaning, anomaly detection, and missing value filling. Then the condensation monitoring data is divided into fixed-length time windows, such as 30 minutes, 1 hour, etc., each time window contains multiple time steps, forming a two-dimensional matrix, the two-dimensional matrix is converted into a grayscale image, each time step corresponds to a column of pixels in the image, and different monitoring indicators correspond to different channels in the image. The converted image is input into a pre-trained CNN model to extract high-level features such as pooling layer output and fully connected layer output, and a fixed-length feature vector is obtained. The CNN model can select classic structures such as LeNet, AlexNet, and VGGNet, or design a customized network structure according to the specific data characteristics. The model training can use the Adam optimizer, the learning rate is set to 0.001, the batch size is set to 32, and the number of training rounds is set to 100. After each round, the model's accuracy and loss are evaluated on the validation set to prevent overfitting. The model's effectiveness can be evaluated using the leave-one-out cross-validation method, dividing the dataset into n parts, selecting one part as the test set and the remaining n-1 parts as the training set, repeating the experiment n times, and taking the average as the final evaluation indicator. Finally, the feature vector is used as the input of the condensation monitoring model to judge the condensation risk probability of each monitoring point in real time. According to the level of condensation risk probability, the clustering algorithm is used to automatically group the risk probability and use different colors such as green, yellow, orange, and red to identify different risk levels, generating a mapping table of condensation risk levels and colors. The substation monitoring grid is abstracted as an undirected weighted graph, each monitoring point corresponds to a node in the graph, and there is a weighted edge between adjacent monitoring points. The weight of the edge can be Euclidean distance, correlation coefficient, etc. The Node2Vec algorithm is used to learn the embedding vector of the node. Node2Vec learns the low-dimensional vector representation of the node by optimizing the probability of node co-occurrence, which can capture the structural similarity and homogeneity between nodes.Then the risk level of unknown nodes is interpolated and predicted using the IDW algorithm, which is a local interpolation algorithm based on spatial correlation. The main parameters of the Node2Vec algorithm include embedding dimension, walk length, and walk number, which can be optimized using grid search. The main parameters of the IDW algorithm include the number of nearest neighbors and distance power, which can be optimized using cross-validation to evaluate the accuracy and robustness of the interpolation. Finally, the interpolation results are mapped to a grid map to obtain the continuous condensation risk distribution of the substation. Based on the interpolation prediction results, open-source GIS tool libraries such as Leaflet are used to generate a heat map of the substation condensation risk distribution using the substation geographic information map as the base map and the monitoring points as the control points. Different risk levels are rendered using different colors and transparencies. The deeper the color and the lower the transparency, the higher the risk level. The trend of risk level change and key risk points are marked using contour lines, wind direction indicators, and other graphical symbols. On the basis of the heat map, thematic layers such as building layout and equipment arrangement within the substation are superimposed, and key areas and critical equipment in the substation are marked. The Shiny package of R language is used to develop an interactive WebGIS system, providing zooming in and out, panning, layer control, and other interactive functions. Users can customize the condensation risk level threshold and color scheme to dynamically adjust the display effect of the risk distribution map, achieving intuitive display of condensation risk. In addition, multi-source data such as risk distribution map, environmental monitoring data, and weather forecast data can be cleaned and converted to extract key fields such as time, location, and indicators, and convert them into standard spatiotemporal data cube format. Apache Kylin and other pre-aggregated OLAP engines are used to precompute multi-dimensional and multi-granular data cubes to generate aggregated indicators and statistical reports, such as average risk level at different time scales and number of high-risk monitoring points in different regions. Seaborn and other statistical plotting libraries are used to visualize and analyze aggregated indicators, generating line charts, bar charts, heat maps, and other charts to reveal the correlation and trend between risk level and environmental factors and weather conditions, providing a more comprehensive basis for condensation prevention. The design of data cubes needs to consider business requirements and data characteristics, select appropriate dimensions, measures, and granularities, and balance query performance and storage costs. The association of multi-source data needs to be based on spatiotemporal attributes such as timestamps and geographic coordinates, which can be implemented using SQL operations such as JOIN and GROUPBY. Correlation analysis can use Pearson correlation coefficient, mutual information, and other indicators, and trend analysis can use Mann-Kendall rank correlation test, Theil-Sen estimation, and other non-parametric methods.

[0106] S8. According to the prediction result obtained in step S7, the optimal anti-condensation measure is given to the target digitalized transformer substation, and the condensation monitoring of the target digitalized transformer substation is completed; specifically including the following steps:

[0107] According to the prediction result obtained in step S7, the area with a condensation risk greater than a set threshold is obtained, and combined with the layout data of the target transformer substation, the specific position and coverage range are determined to obtain a high-risk condensation area position list;

[0108] A knowledge base of anti-condensation measures is constructed;

[0109] In the anti-condensation measure knowledge base, the environmental characteristics of the high-risk condensation area are compared with the environmental characteristic data in the anti-condensation measure knowledge base in terms of similarity, and the anti-condensation measures with a similarity higher than a set threshold are selected as candidate measures;

[0110] Among the various candidate measures, the cost data of each candidate measure is calculated according to the installation position and installation method of the candidate measure, and through the three-dimensional model of the target digitalized transformer substation, the change of the regional environmental parameters after the implementation of each candidate measure is calculated, the anti-condensation effect of the candidate measure is evaluated, and the final anti-condensation measure is determined according to the cost data and the anti-condensation effect, and the condensation monitoring of the target digitalized transformer substation is completed;

[0111] In implementation, the list attributes include region number, center coordinates, boundary coordinates, area, perimeter, and covered equipment. Based on the high-risk area location list, environmental monitoring data within the region, including temperature, humidity, pressure, and wind speed, are extracted. Using historical data over five years, statistical indicators such as annual average, seasonal average, and monthly average of each environmental parameter are calculated. Combined with the meteorological knowledge base, the environmental characteristics of the region are summarized to form a high-risk area environmental characteristics description. The description attributes include region number, temperature characteristics, humidity characteristics, pressure characteristics, wind speed characteristics, and condensation mode. The measure attributes include measure number, measure name, applicable conditions, technical principle, installation location, investment cost, operation and maintenance cost, and service life. When calculating the similarity between the condensation prevention measure applicable conditions and the regional environmental characteristics, both the measure applicable conditions and the regional environmental characteristics are represented as a multi-dimensional feature vector, with each dimension corresponding to an environmental parameter. The distance between the feature values of the measure applicable conditions and the regional environmental characteristics is calculated for each dimension. Common distance metrics such as Euclidean distance and Manhattan distance are used to obtain a comprehensive distance by weighted averaging of the distances in each dimension. The weights are determined based on the importance of each environmental parameter in condensation. The comprehensive distance is converted to similarity through a Gaussian kernel function, and by setting different parameter values, the similarity decay rate is controlled to adjust the matching sensitivity. According to actual requirements and experience, appropriate parameter values are selected, and a similarity threshold is set. Measures with a similarity greater than the threshold are considered as candidate measures, sorted from high to low according to similarity, and a list of available measures in the region is formed. For each measure in the available measure list, its specific installation location and installation method are determined in the substation 3D model based on its installation location attribute. Computational fluid dynamics software is used to simulate the temperature and humidity distribution and airflow field in the region after the implementation of the measure. In CFD simulation, boundary conditions, initial conditions, turbulence models, and grid division of the calculation domain are set reasonably based on the actual layout and equipment parameters of the substation. Appropriate fluid property parameters and solution settings are selected to ensure the convergence and accuracy of the simulation results. By comparing the temperature and humidity distribution and airflow streamline before and after the implementation of the measure, the anti-condensation effect of the measure is evaluated, and a measure effect evaluation report is formed. The report attributes include measure number, installation location, temperature effect, humidity effect, airflow effect, and comprehensive effect. Based on the measure effect evaluation report and the cost attributes of the measures, the cost-effectiveness of each measure is calculated, sorted from high to low according to cost-effectiveness, and a measure recommendation list is formed. The list attributes include region number, measure name, cost-effectiveness, investment cost, operation and maintenance cost, and comprehensive effect. The measure with the highest cost-effectiveness is selected as the optimal measure for the region. For the optimal measure belonging to the dehumidification device type, the operation control strategy of the dehumidification device is optimized based on the regional environmental characteristics and the technical parameters of the dehumidification device, including start and stop time, dehumidification capacity, and regeneration period.An intelligent control algorithm such as fuzzy control and predictive control is used to design a control model for the dehumidification device, establish a dynamic mapping relationship between temperature and humidity, dehumidification capacity, and energy consumption, and adaptively adjust the control parameters through offline simulation and online learning to minimize dehumidification energy consumption and improve dehumidification efficiency. For the optimal measures that belong to the ventilation optimization type, the operation control strategy of the ventilation system is optimized according to the regional environmental characteristics and the design parameters of the ventilation system, including ventilation time, ventilation volume, fan power, etc. An optimization model of the ventilation system is established, the ventilation control parameters are taken as decision variables, the condensation occurrence probability, ventilation energy consumption, and equipment life are taken as optimization objectives, and the environmental parameter range, fan rated power, and ventilation volume range are taken as constraint conditions to form a multi-objective optimization problem. A heuristic search algorithm such as particle swarm optimization and genetic algorithm is used to search for the optimal combination of ventilation control parameters. In the particle swarm optimization algorithm, the position vector of each particle is encoded as a set of ventilation control parameters, a set of particles is randomly initialized, the fitness function value of each particle is calculated, the individual optimal position and global optimal position are updated, the particle speed and position are adjusted, and the iteration is repeated until convergence, and the global optimal solution is output as the best ventilation control strategy. The optimal anti-condensation measures for each region are formed into a decision report and reported to the substation management personnel, and the report attributes include region number, condensation risk level, environmental characteristic description, optimal measure name, measure cost performance, energy saving and emission reduction benefit, etc. After the management personnel approve, the optimal measures are included in the substation condensation control scheme library for subsequent implementation. According to the real-time update of the condensation risk distribution map, the above process is periodically triggered to re-evaluate the changes in risk level and environmental characteristics of each region and adjust the existing anti-condensation measures. Combined with factors such as equipment operating conditions and weather forecasts, when the condensation risk level rises by more than 5% or the environmental parameter fluctuation exceeds 10%, the optimization process is triggered, the latest monitoring data and external conditions are used to update the anti-condensation measures, and a closed-loop control of condensation risk is formed. Through the cooperation of cloud big data platforms and edge intelligent gateways, the continuous optimization and adaptive update of anti-condensation measures are realized, and the timeliness and pertinence of the measures are improved. In terms of deployment and implementation, microservice architecture and containerization technology are used to decouple each functional module into independent microservices, Kubernetes and other container orchestration platforms are used for elastic scheduling and management to improve the scalability and maintainability of the system. At the same time, a unified data standard and interface specification are established to realize seamless integration and data sharing between subsystems, and to provide intelligent decision support for the whole life cycle of substation condensation control.In identifying high-risk areas, the following method can be used for image segmentation of the condensation risk distribution map. A region growing-based segmentation algorithm, such as region merging algorithm, is used. The pixel points with risk level greater than 75% are taken as seed points. The risk level difference less than 5% is taken as the growth criterion. The adjacent high-risk pixel points are merged into a region to obtain the segmentation result of the high-risk area. Then, multi-scale geometric features are extracted for each segmentation area, including area, perimeter, rectangularity, circularity, etc. According to the empirical threshold, the areas with area greater than 100 square meters and rectangularity greater than 0.8 are selected as candidate areas. Finally, the candidate areas are superimposed and analyzed with the substation geographic information map and the equipment layout map to extract the position information of the area center coordinates, boundary coordinates, etc. and the attribute information of the main equipment type and quantity covered, etc. to form a high-risk area position list. In calculating the statistical indicators of the environmental parameters, the incremental learning method can be used. It does not need to store a large amount of historical data. Only a set of statistics, including sample number, total, sum of squares, etc. needs to be maintained. Whenever new monitoring data arrives, the statistics are dynamically updated using recursive formulas. The recursive formula for the mean is: where μ_n is the mean after the nth sample, μ_{n-1} is the mean after the n-1th sample, and x_n is the nth sample value. There are similar recursive formulas for statistical quantities such as variance and standard deviation. When performing monthly, quarterly, and annual statistics, only the current statistical quantity needs to be calculated at the corresponding time node to obtain the corresponding index value. In matching anti-condensation measures, there are many measures in the knowledge base, and the applicable conditions are different. In order to improve the matching efficiency and accuracy, a multi-stage filtering method based on semantics can be used. First, the text attributes of the measures such as measure name and technical principle are segmented and semantically vectorized, and a pre-trained word vector model such as Word2Vec is used to map each word to a fixed-length real-valued vector. Then, the semantic vectors of the measures are clustered, and measures with similar semantics are clustered into a class. When optimizing the control strategy of the dehumidification device, a model predictive control method can be used to establish a dynamic model of the dehumidification process to describe the nonlinear relationship between temperature, humidity, dehumidification capacity, and energy consumption, such as a partial differential equation model based on the laws of energy and mass conservation. Then, in each control period, based on the current temperature and humidity measurements and the predicted temperature and humidity values in the future period, an optimal control problem in a finite time domain is solved, such as minimizing the cumulative energy consumption in the prediction period, with the constraint that the dehumidification capacity is not less than a certain threshold, to obtain the optimal dehumidification capacity sequence in the future period. Finally, the first value of the sequence is set as the dehumidification capacity set value in the current control period, and the dehumidification device is controlled to run. In the next control period, the above calculation is repeated to realize rolling optimization. During the implementation of the anti-condensation measures, the performance of the system needs to be continuously monitored and evaluated to timely discover and solve problems. For this purpose, a monitoring system based on microservice architecture can be deployed. The system consists of multiple independent monitoring services, including data collection services, data storage services, data analysis services, and alarm services, each running in an independent Docker container and communicating through REST APIs. The data collection service is responsible for collecting performance indicator data from various subsystems, such as the energy consumption of the dehumidification device, temperature and humidity, the wind speed and current of the ventilation system, and pushing the data to the message queue. The data storage service is responsible for persisting the data to the time series database and providing query and aggregation APIs. The data analysis service is responsible for periodically reading data from the database, calculating key performance indicators such as energy efficiency and humidity compliance rate, identifying abnormal situations using anomaly detection algorithms, and pushing the analysis results to the message queue. The alarm service is responsible for subscribing to the message queue and automatically sending alarm notifications such as emails and SMS messages when abnormalities are detected according to pre-set alarm rules. The services are decoupled through the message queue and can be independently scaled and upgraded, making the entire system highly scalable and fault-tolerant. The operation and maintenance personnel can view the running status and performance indicators of the system through the Web interface to timely discover and locate problems;

[0112] S9. Establish a condensation monitoring and protection platform to realize condensation monitoring, data management and protection measure optimization of the target digital substation; specifically including the following steps:

[0113] Data preprocessing is performed on the obtained data to generate a condensation monitoring dataset;

[0114] Based on support vector machines and convolutional neural networks, an initial condensation risk assessment model is constructed, and the generated condensation monitoring dataset is used for training to obtain a condensation risk assessment model;

[0115] The obtained condensation risk assessment model is used to obtain condensation risk assessment results, and combined with the three-dimensional model of the target digital substation, a condensation risk distribution map is generated;

[0116] According to the generated condensation risk distribution map, the condensation prevention measures are dynamically optimized based on a multi-objective planning scheme;

[0117] In implementation, through the Internet of Things technology, the condensation monitoring equipment such as temperature and humidity sensor, dew point instrument and conductivity instrument in the transformer substation is connected to the intelligent management platform to realize the unified management and remote control of the equipment. The OneM2M, OMA LwM2M general Internet of Things standard and protocol are used to improve the interoperability and scalability of the system. For different types of monitoring equipment, the secure communication protocol and lightweight data serialization format are used to perform data preprocessing, compression and encryption at the device end through the edge computing gateway, thereby reducing the cloud computing and storage burden. The management platform groups and encodes the equipment according to the equipment type, location and function attribute, forms the equipment account, and uses the MQTT lightweight communication protocol to establish a real-time bidirectional communication link with the equipment to regularly obtain the working state and monitoring data of the equipment. The device metadata, configuration parameters and alarm events are modeled and described by referring to the LWM2M information model and resource definition, so as to realize the functions of remote configuration, firmware upgrade and fault diagnosis management of the equipment. On the basis of the condensation monitoring data set, a condensation risk assessment model is constructed. The support vector machine, random forest machine learning algorithm, LSTM and convolutional neural network deep learning algorithm are used to train and optimize the historical monitoring data, so as to establish the nonlinear relationship between the condensation risk level and the monitoring indexes such as temperature, humidity, dew point temperature and water vapor pressure. The model takes the real-time monitoring data of each monitoring point as input and outputs the condensation risk level of the point. The risk assessment accuracy and real-time performance are improved through regular retraining and fine-tuning. According to the output result of the condensation risk assessment model, a condensation risk distribution map of the transformer substation is dynamically generated to intuitively display the risk level and distribution of each monitoring point. The risk level is encoded using red, yellow and green colors, and the risk distribution is visualized using the contour line and heat map methods. The transformer substation geographic information layer and equipment layout layer are superimposed to identify and analyze the high-risk areas, predict the occurrence time, location and severity of condensation, and provide decision support for subsequent protection strategy optimization. Combined with the risk distribution map, environmental conditions and equipment performance factors, and combined with the safety and stability operation requirements of the power system, the existing condensation protection strategy of the transformer substation is dynamically optimized using the multi-objective optimization and reinforcement learning method. For the identified high-risk objects, the environmental parameters, equipment parameters and protection cost multi-dimensional attributes are extracted to construct protection strategy optimization models such as multi-objective linear programming and nonlinear programming. The minimum condensation risk and protection cost, the maximum equipment reliability objective function and weight coefficient are set, the environmental parameter, equipment parameter and protection cost budget constraint conditions are set, the NSGA-II and MOEA / D multi-objective optimization algorithm is used for solution, the Pareto optimal solution set is obtained, the optimal protection strategy parameters are selected from the solution set according to the decision preference, and the sensitivity analysis and robustness evaluation are performed.The optimized protection strategy is issued to the dehumidification device, ice flashover prevention coating spraying equipment, insulator cleaning equipment, and system, and through the edge computing gateway intelligent hardware, the strategy is converted into specific control instructions and action sequences, the device working parameters and operating state are adjusted in real time, and according to the device execution feedback and monitoring data, the strategy is closed loop corrected and self-adaptively modified. A condensation monitoring and protection intelligent operation and maintenance platform is constructed, big data analysis, digital twinning, virtual reality emerging technologies are used, and the condensation monitoring, risk assessment, protection optimization, and device scheduling links are simulated, monitored, warned, diagnosed, and optimized in the whole process, whole scene, and whole period. Using big data analysis technology, massive monitoring data, environmental data, and device data are associated, mined, and deeply analyzed, and the condensation law, risk mode, and protection experience implicit knowledge are refined. Using digital twinning technology, a virtual simulation model of the substation in whole elements, whole process, and whole life cycle is constructed, the monitoring scene is realistically reproduced, the risk event is preformed and deduced, and the protection strategy is virtually verified. Using virtual reality technology, an immersive and interactive three-dimensional visualization interface is generated, the management personnel can perceive the condensation risk in situ, and the protection operation is commanded in real time. In the realization of the intelligent perception and monitoring of the substation, intelligent sensors and acquisition devices conforming to the IEC61850 standard can be selected, such as intelligent temperature and humidity sensors, gas monitors, and partial discharge monitors, and through Modbus-TCP, IEC60870-5-104, and other standard protocols, the sensor data is uploaded to the substation monitoring system. For traditional analog sensors, intelligent terminals based on ARMCortex-M series chips and RS485 bus can be used to realize the digitization and networking transformation of analog signals. For different manufacturers' equipment, the IEC61970 / 61968-CIM public information model is used to model and describe the attributes, telemetry, and remote signaling of the equipment, form a unified information exchange format, and eliminate information islands. The data of each monitoring terminal is converged to the edge computing server of the substation control layer, and through MQTT message queues and Kafka data pipelines, the cross-protocol forwarding, shunting storage, and multi-level caching of data are realized. Using the time series database InfluxDB and the containerized deployment tool Docker, a high-availability and scalable data processing and storage node is built, and using the Kubernetes container orchestration platform, the elastic scheduling and fault self-healing of services are realized. The collected multi-source heterogeneous data is updated once every 15 minutes, through data fidelity compression, data quality exploration, consistency checking, and other means, the integrity and accuracy of the data are ensured to be more than 99.9%.In the process of condensation risk assessment, the random forest algorithm is used to select features of environmental factors and equipment factors that affect condensation. Through grid search and cross-validation methods, the parameters of random forest are optimized, including the number of decision trees, the maximum depth of each tree, and the maximum number of features considered when splitting nodes. The feature importance ranking is obtained, and the top 10 features with a cumulative contribution rate of 95% are selected as the model input. A long short-term memory network is used to build an input-output mapping model. The past 24 hours of temperature, humidity, wind direction, wind speed, air pressure, and equipment operating status are used as input, and the future 1 hour of condensation risk level, such as low, medium, and high, is used as output, forming a many-to-one time series prediction problem. A network structure containing 2 LSTM hidden layers and 1 Softmax output layer is built, using Adam optimizer and cross-entropy loss function, with a learning rate of 0.001, a batch size of 64, and an iteration number of 500. Model evaluation is performed every 50 iterations. The accuracy on the training set reaches 98%, and the accuracy on the test set reaches 95%, with an AUC value of 0.99, indicating that the model can well fit and predict the condensation risk. The model is retrained daily to continuously improve the timeliness and robustness of the prediction. In optimizing the anti-condensation strategy, the insulation reliability is improved, and the protection cost is reduced as the optimization objective. Environmental factor thresholds, protection device operating parameters, and operation plan frequencies are used as optimization variables. The condensation characteristic curve provided by the equipment manufacturer and the protection experience of the operation and maintenance personnel are used as constraint conditions to establish a multi-objective mixed integer programming model. An improved non-dominated sorting genetic algorithm is used to solve the model, with a population size of 100, a crossover probability of 0.8, a mutation probability of 0.1, an iteration number of 500, and a number of Pareto optimal solutions of 20. The algorithm is implemented in the Gurobi optimization library and accelerated using a high-performance computing cluster, reducing the solution time from hours to minutes. The protection cost and reliability level are used as evaluation indicators to quantitatively evaluate and sort the Pareto optimal solution set, and the solution with the best comprehensive performance is automatically selected as the final protection strategy. Simulation experiments show that the optimized strategy can reduce the device failure rate from 2% to 0.5% and the annual protection cost from 500,000 yuan to 300,000 yuan, achieving a double improvement in reliability and economy. In the construction of the intelligent operation and maintenance platform, a micro-service architecture is adopted, using mainstream frameworks such as SpringBoot and Vue.js to develop core functional modules such as monitoring data management, condensation risk assessment, protection strategy optimization, intelligent alarm decision-making, device account management, and three-dimensional visualization. REST interfaces are used to realize data interaction between modules. Stream computing frameworks such as Storm and SparkStreaming are used to perform real-time cleaning, filtering, aggregation, and analysis of monitoring data. Redis caching and RabbitMQ message queues are used to realize high-speed data access and cross-platform sharing.Using visualization engines such as ECharts and BabylonJS, build cool 2D and 3D interfaces, and use immersive technologies such as VR / AR to create a virtual operation experience that feels like being there. Using Neo4j graph database and Cypher query language, build semantic networks of device, environment, and policy knowledge, and realize knowledge reasoning and intelligent question answering based on graph. Using Drools rule engine and BPMN workflow engine, solidify business processes into computer executable rules and processes, and realize the automation and standardization of decision-making. Using digital twin simulation platform, build multi-physical field simulation model of substation, carry out fine simulation of typical scenarios such as equipment aging, material corrosion and condensation accumulation, and optimize sensor layout scheme and inspection and maintenance strategy. Using OneMapGIS platform, integrate condensation monitoring data with geographic spatial data to generate substation condensation risk map and protection resource map, and assist management personnel in visual analysis and scientific decision-making.

Claims

1. A method for monitoring condensation in a digital substation, comprising the following steps: S1. Obtain data information from the target digital substation; S2. Based on the data obtained in step S1, simulate the airflow characteristics in the target digital substation using a numerical simulation scheme, determine the high-risk areas for condensation, and deploy corresponding monitoring equipment to obtain monitoring data. S3. Based on the differences in the response characteristics of electromagnetic waves of different frequency bands to condensation, select the monitoring frequency band, acquire and analyze the data of the monitoring frequency band, and thus construct a feature set related to the degree of condensation; S4. Acquire real-time monitoring data of the target digital substation and select the electromagnetic wave monitoring frequency band based on the current environment; S5. Based on the electromagnetic wave monitoring frequency band obtained in step S4, perform feature extraction and analysis to construct a condensation monitoring model based on electromagnetic wave transmission characteristics; specifically including the following steps: Based on the maximum correlation minimum redundancy algorithm and the Relief-F algorithm, feature selection is performed on the electromagnetic wave attenuation characteristic data, and the attenuation features with the highest correlation to the degree of condensation are selected as candidate feature sets. Feature extraction and fusion are performed on the candidate feature set. Principal component analysis is used to extract the main component features of the attenuation feature, independent component analysis is used to extract the independent component features of the attenuation feature, and an autoencoder is used to extract the deep nonlinear features of the attenuation feature to obtain the attenuation feature parameters. Acquire condensation image and video data on the equipment surface; Extract image feature parameters from the acquired image and video data, score the degree of condensation according to the set requirements, and obtain a condensation observation sample dataset; A condensation monitoring model was obtained by training a gradient boosting regression tree model using a condensation observation sample dataset. Meteorological forecast data, internal environmental monitoring data, and electromagnetic wave attenuation data of the target digital substation are obtained, and the obtained data are used to train a long short-term memory neural network to obtain a condensation trend prediction model. A condensation trend prediction model is used to predict the condensation trend of the target substation. S6. Optimize the condensation monitoring model constructed in step S5 based on the fused environmental parameters and electromagnetic wave parameters; S7. Based on the optimized condensation monitoring model obtained in step S6, generate condensation risk prediction results for the target digital substation; S8. Based on the prediction results obtained in step S7, provide the optimal anti-condensation measures for the target digital substation and complete the condensation monitoring of the target digital substation.

2. The condensation monitoring method for a digital substation according to claim 1, characterized in that... Step S2 specifically includes the following steps: A three-dimensional model of the target digital substation was established, and a computational fluid dynamics scheme was used to numerically simulate the airflow motion within the target digital substation, thereby obtaining the airflow motion characteristic parameters of different regions within the target digital substation. The obtained airflow motion characteristic parameters are correlated with temperature and humidity to identify high-risk areas for condensation within the target digital substation. In identified high-risk areas, corresponding monitoring equipment is deployed to obtain monitoring data.

3. The condensation monitoring method for a digital substation according to claim 2, characterized in that... Step S3 specifically includes the following steps: Based on the analysis of electromagnetic wave transmission theory and condensation physical mechanism, several frequency bands sensitive to condensation were selected as candidate frequency bands. A condensation simulation experimental platform was built, and experiments were conducted under different condensation conditions to obtain test data on the electromagnetic wave transmission characteristics of candidate frequency bands. By fitting and regression analysis of the obtained electromagnetic wave transmission characteristic test data, a scale model of the correlation between condensation degree and electromagnetic wave frequency band and transmission characteristic parameters is obtained. Based on the correlation scale model, the monitoring frequency bands are determined, and the set of monitoring frequency bands is obtained; Based on the obtained monitoring frequency band, electromagnetic wave transmission data is collected to obtain a set of electromagnetic wave transmission data reflecting the condensation state. Video domain analysis was performed on the electromagnetic wave transmission data that reflects the condensation state to extract characteristic parameters that reflect the degree of condensation. A feature selection algorithm is used to filter and optimize the obtained feature parameters to obtain the optimal feature subset related to the degree of condensation, and to construct a feature set for predicting the degree of condensation. The obtained set of features for predicting condensation level is used to train the support vector regression algorithm to obtain the condensation level prediction model. The obtained condensation level prediction model is used to predict the condensation level of the target digital substation.

4. The condensation monitoring method for a digital substation according to claim 3, characterized in that... Step S4 specifically includes the following steps: Real-time acquisition of environmental monitoring data of the target digital substation; the environmental monitoring data includes temperature data, humidity data, pressure data and light intensity data of the monitoring points; Based on the acquired environmental monitoring data, a clustering algorithm was used for cluster analysis, and the target digital substation was divided into several environmental zones according to the clustering results. Install condensation sensors in various environmental areas to monitor condensation data on the equipment surface in real time; Select a monitoring frequency band and use the selected monitoring frequency band to transmit condensation data.

5. The condensation monitoring method for a digital substation according to claim 4, characterized in that... Step S6 specifically includes the following steps: The acquired environmental and electromagnetic wave parameters are preprocessed using data alignment, time series interpolation, and outlier removal algorithms to obtain a fused feature vector. The condensation monitoring model constructed in step S5 is optimized by using fused feature vectors.

6. The condensation monitoring method for a digital substation according to claim 5, characterized in that... Step S7 specifically includes the following steps: An optimized condensation monitoring model is used to predict the condensation risk of the target digital substation. Based on the predicted condensation risk probability, a clustering algorithm is used to cluster and group the data, and different risk levels are identified. The monitoring grid of the target digital substation is set as an undirected weighted graph, with each monitoring point corresponding to a node in the graph, and a weighted edge between adjacent monitoring points, and the weight of the edge is set. The Node2Vec algorithm is used to obtain the low-dimensional embedding vector of the node by optimizing the probability of node collinearity; the IDW algorithm is used to interpolate and predict the risk level of unknown nodes; based on the interpolation prediction results, a risk distribution map is generated with the substation geographic information map as the base map and the monitoring points as control points. Areas with different risk levels are marked accordingly.

7. The condensation monitoring method for a digital substation according to claim 6, characterized in that... Step S8 specifically includes the following steps: Based on the prediction results obtained in step S7, areas with condensation risk greater than the set threshold are identified. Combined with the layout data of the target substation, the specific locations and coverage areas are determined, resulting in a list of high-risk condensation areas. Build a knowledge base for anti-condensation measures; In the knowledge base of anti-condensation measures, the environmental characteristics of high-risk condensation areas are compared with the environmental characteristics data in the knowledge base, and anti-condensation measures with similarity higher than the set threshold are selected as candidate measures. Among the candidate measures, the cost data of each candidate measure is calculated according to the installation location and installation method. Then, the changes in regional environmental parameters after the implementation of each candidate measure are calculated using the three-dimensional model of the target digital substation. The anti-condensation effect of the candidate measures is evaluated. Based on the cost data and the anti-condensation effect, the final anti-condensation measure is determined, and the condensation monitoring of the target digital substation is completed.

8. The condensation monitoring method for a digital substation according to any one of claims 1 to 7, characterized in that, It also includes the following steps: S9. Establish a condensation monitoring and protection platform to achieve condensation monitoring, data management, and optimization of protection measures for the target digital substation.

9. The condensation monitoring method for a digital substation according to claim 8, characterized in that, Step S9 specifically includes the following steps: The acquired data is preprocessed to generate a condensation monitoring dataset; Based on support vector machines and convolutional neural networks, an initial model for condensation risk assessment was constructed, and the generated condensation monitoring dataset was used for training to obtain the condensation risk assessment model. The obtained condensation risk assessment model is used to obtain the condensation risk assessment results, and combined with the three-dimensional model of the target digital substation, a condensation risk distribution map is generated. Based on the generated condensation risk distribution map, the anti-condensation measures are dynamically optimized using a multi-objective programming approach.

Images