Multi-source data acquisition and fusion method for power distribution network

Abstract

The invention discloses a power distribution network multi-source data acquisition and fusion method, and the method comprises the following steps: S1, collecting multi-source data in a power distribution network, including power equipment operation data, user power consumption data and environment monitoring data; s2, preprocessing the collected data, including noise removal, missing value filling, abnormal value detection and processing, data dimension reduction and data coding; s3, performing feature extraction on the preprocessed data; and S4, performing data fusion based on machine learning. According to the multi-source data acquisition and fusion method for the power distribution network, the multi-source data is comprehensively acquired by means of various professional devices, the data basis can be tamped, the data quality can be improved, the redundancy can be reduced, the operation state of the power distribution network can be accurately analyzed, a scientific basis is provided for optimal management of the power distribution network, and the method is suitable for popularization and application. And reliability and stability are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network data processing, and particularly relates to a power distribution network multi-source data acquisition and fusion method. BACKGROUND

[0002] With the continuous expansion of the scale of the power distribution network and the deepening of the intelligent degree, the data generated in the operation process of the power distribution network presents an explosive growth trend. The source data is widely sourced, covering power equipment, user electricity consumption and environmental monitoring data. In the power equipment operation data, the change of transformer oil temperature can reflect internal faults, the line current and voltage amplitude affect power transmission, the distributed power supply data show the power generation capacity, the switch device state and the power distribution network harmonic data are related to the equipment operating condition. User electricity consumption data can help analyze electricity consumption behavior and facilitate the development of strategies by the power department. In the environmental monitoring data, temperature, humidity and wind speed affect the power distribution network equipment, high temperature, high humidity and strong wind will increase the risk of equipment failure, which is of great significance to the equipment operating condition.

[0003] However, these massive multi-source data have many problems. The data is scattered in different systems and devices, and the formats are different, some are stored in text form and some are in binary format, which greatly increases the difficulty of data integration. Noise exists in the data, which may be caused by sensor measurement error, communication interference and other factors; the occurrence of missing values may be due to data transmission interruption, equipment failure and other reasons; the generation of abnormal values, such as sudden values in power equipment operation data, will seriously interfere with the accuracy and reliability of data analysis. Moreover, the data dimension is very high, and there are many redundant features, a large number of features that are not valuable for subsequent analysis not only occupy storage space, but also make the data processing computation grow exponentially.

[0004] Therefore, there is an urgent need for an effective method to comprehensively and systematically collect, finely preprocess, accurately extract features and efficiently fuse these multi-source data. Through this method, the data quality is improved, noise is removed, missing values are filled, abnormal values are corrected, key features are selected, and data dimension is reduced, making the data more accurate, complete and usable. In this way, it can provide solid data support for the optimized operation and management of the power distribution network, and help the power industry move towards a more efficient and intelligent development stage. SUMMARY

[0005] The purpose of the present application is to provide a power distribution network multi-source data acquisition and fusion method, which solves the problems of dispersion, different formats, noise, missing values, abnormal values and high-dimensional redundant features in existing power distribution network multi-source data, can systematically collect and fuse data, improve quality, remove noise, select key features and reduce dimension, accurately analyze the operating state, provide support for optimized management, and improve the reliability and stability of the power distribution network.

[0006] To achieve the above object, the application provides a power distribution network multi-source data acquisition and fusion method, comprising the following steps:

[0007] S1, collecting multi-source data in the power distribution network, including power equipment operation data, user power consumption data and environmental monitoring data;

[0008] S2, preprocessing the collected data, including removing noise, filling missing values, detecting and processing outliers, data dimensionality reduction and data encoding;

[0009] S3, feature extraction is performed on the preprocessed data;

[0010] S4, data fusion based on machine learning.

[0011] Preferably, in S1, the power equipment operation data in the power distribution network is collected in real time by sensors, including transformer oil temperature T t , line current I l , voltage amplitude U a , distributed power data, switch device state data, power distribution network harmonic data and power distribution network fault data;

[0012] The transformer oil temperature T t reflects the thermal state of the transformer, and the oil temperature is too high, indicating that there is a fault inside the transformer, the line current I l and the voltage amplitude U a measure the power transmission and power quality of the power distribution network respectively;

[0013] The distributed power data includes generated power P g and generated power cumulative value E g , the generated power P g represents the electric power injected by the distributed power source to the power distribution network at a certain time, reflecting the power generation capacity and real-time power generation state of the distributed power source;

[0014] The calculation formula of the generated power cumulative value is:

[0015]

[0016] Where t1 and t2 are the starting time and ending time of the time interval, used to calculate the total amount of power generation of the distributed power source in a certain time period.

[0017] Preferably, in S2, the switch device state data includes switch opening and closing state S S and switch action number N S , wherein the switch opening and closing state S S takes value 0 or 1, 0 represents the switch is in the open state, 1 represents the switch is in the closed state; the switch action number N SUsed to assess the lifespan and health status of switching equipment;

[0018] Harmonic data for distribution networks includes total harmonic voltage distortion (THD). U Total Harmonic Distortion (THD) I The formula for calculating the total harmonic voltage distortion rate is:

[0019]

[0020] Among them, U h U1 is the effective value of the h-th harmonic voltage, and U2 is the effective value of the fundamental voltage. h For the highest harmonic order considered, THD U Used to measure the degree of distortion of voltage waveforms in a distribution network, expressed in %;

[0021] The formula for calculating the total distortion rate of harmonic currents is:

[0022]

[0023] Among them, I h Ih is the effective value of the h-th harmonic current, I1 is the effective value of the fundamental current, and n is the effective value of the fundamental current. h For the highest harmonic order considered, THD I Used to measure the degree of distortion of current waveforms in a distribution network, expressed in %;

[0024] Distribution network fault data includes fault type F type and the time of failure T type Fault type F type This includes short-circuit faults denoted as S, open-circuit faults as O, and ground faults as G, with the fault occurrence time T. type Used to record the exact moment a fault occurs.

[0025] Preferably, in S1, user electricity consumption data, including user electricity consumption E, is collected through a smart meter. u This represents the user's total electricity consumption and electricity consumption time t. u Used to analyze users' electricity consumption habits and power consumption P u The user's electricity consumption at different times;

[0026] Environmental monitoring data is collected through weather stations, including ambient temperature (T). e Humidity (H) and wind speed (V) data.

[0027] Preferably, in S2, the collected power equipment operation data, user electricity consumption data, and environmental monitoring data are subjected to noise removal using a mean filtering algorithm;

[0028] For transformer oil temperature, line current and voltage amplitude data in power equipment operation data, if there are missing values, linear interpolation is used to fill them in.

[0029] If there are missing time periods in the user's electricity consumption data, the electricity consumption pattern of adjacent time periods will be used to fill them in.

[0030] The standard deviation-based method is used to detect outliers in the data. Let the data sequence be c = c1, c2, ... c q First, calculate the mean μ and standard deviation σ of the data sequence. If a certain data point c k Satisfy |c k If -μ∣>3σ, then c is considered k For outliers, the detected outliers are corrected by linear interpolation of the adjacent non-outliers before and after the data point.

[0031] Preferably, in S2, the data dimensionality reduction operation for multi-source data of the distribution network includes principal component analysis and linear discriminant analysis. The principal component analysis (PCA) process is as follows:

[0032] Let the original data matrix be X. m'×n' Where m' is the number of samples and n' is the number of features, the data is first centered to obtain... Calculate the covariance matrix so that the mean of each column of data is 0. Then, solve for the eigenvalues ​​λ1≥λ2≥...≥λ of the covariance matrix S. n′ and the corresponding feature vectors v1, v2, ..., v n′ ;

[0033] Select the eigenvectors v1, v2, ..., v corresponding to the first k' largest eigenvalues. k' Forming the projection matrix V n'×k' Projecting the original data matrix X onto these k' principal components yields the dimensionality-reduced data matrix Y. m'×n' =XV, where k' is the number of features after dimensionality reduction. An appropriate value needs to be selected according to the actual situation in order to reduce the number of features and reduce the complexity of the data without losing too much information.

[0034] The linear discriminant analysis (LDA) process is as follows:

[0035] Assume the data has C' categories, and the i'-th category has m" i' Number of samples, total number of samples Calculate the mean vector u′ for each class. i Given the global mean vector u', calculate the within-class scatter matrix:

[0036]

[0037] Calculate the inter-class scatter matrix:

[0038]

[0039] Solving the generalized eigenvalue problem S′ b v′=τ'S′ w v′, we obtain the eigenvalues ​​τ'1≥τ'2≥...≥τ' n′-1 and the corresponding feature vectors v'1, v'2, ..., v' n′-1 Select the eigenvectors v'1, v'2, ..., v' corresponding to the first k largest eigenvalues. k" Forming the projection matrix w′ n′×k" The original data is projected onto k directions to achieve data dimensionality reduction.

[0040] Preferably, in S2, the data encoding uses one-hot encoding and label encoding. The one-hot encoding process is as follows: For a categorical variable X with n" different values, its value set is x′1, x′2, ..., x′ n Convert the categorical variable into an n-dimensional binary vector. If the sample value is x′ j′ Then the encoded vector is e j′ e j′ It is an n"-dimensional vector, where the j'-th element is 1 and the rest are 0;

[0041] The label encoding process is as follows: for each different value x′ of the categorical variable X”. i" Assign a unique integer code l′ i" , i" = 1, 2, ..., n".

[0042] Preferably, in S3, feature extraction of the preprocessed data includes extracting voltage fluctuation features and current change rate features from power equipment operation data; extracting peak-valley ratio features and load curve slope features from user electricity consumption data; and extracting trend features and correlation features from environmental monitoring data.

[0043] The formula for calculating voltage fluctuation is:

[0044]

[0045] Wherein, ΔU is the voltage fluctuation value, reflecting the degree of voltage fluctuation, presented as a percentage, used to measure the relative change in voltage deviation from the rated value; U max It represents the maximum voltage amplitude within a statistical time period, indicating the highest voltage state; U min It represents the minimum voltage amplitude within the statistical time period, indicating the lowest voltage value; U ratedThe rated voltage of the power system is the benchmark value for measuring voltage fluctuations.

[0046] The formula for calculating the rate of change of current is:

[0047]

[0048] in, I represents the rate of change of current, measured in amperes per second, reflecting how quickly the current changes over time. t2 It is the line current value at time t2, representing the magnitude of the current at that moment; I t1 It is the line current value at time t1, used to correlate with I. t2 The difference is used to calculate the change in current; t2-t1 is the time interval in seconds, representing the length of the time period selected for calculating the rate of change of current.

[0049] Preferably, in S3, the formula for calculating the peak-to-valley ratio of electricity consumption is:

[0050]

[0051] Among them, P ratio The peak-to-valley ratio reflects the relative magnitude of electricity consumption by users during peak and off-peak hours; P peak It is the peak power consumption of a user within the statistical period, representing the user's maximum load state; P valley It is the lowest value of the user's electricity consumption within the statistical period;

[0052] The formula for calculating the slope of the load curve is:

[0053]

[0054] Where k is the slope of the load curve, in kilowatts per hour, representing the change in user power consumption per unit time, used to measure the rate of increase or decrease in user power load; P t2 P represents the power consumption of a user at time t2, reflecting the user's power load at that time. t1 It is the power consumption value of user at time t2, and P t2 In conjunction with the calculation of power change; t2-t1 is also a time interval in hours, used to determine the time interval corresponding to the calculated power change rate;

[0055] Linear trend formula: Assuming that the data y has a linear relationship with time t:

[0056] y = β0 + β1t + ∈;

[0057] Where β0 is the intercept, β1 is the slope, and ∈ is the error term. The least squares method is used to estimate β0 and β1, i.e.

[0058] The calculation of seasonal trends uses seasonal decomposition methods, including additive models:

[0059] y t =T t +S t +R t ; or multiplication model: y t =T t ×S t ×R t ;

[0060] Among them, y t It is the observation value at time t, T t It is a trend component, S t It is a seasonal ingredient, R t It is a residual component;

[0061] Temperature T is measured using the Pearson correlation coefficient. ei and humidity Linear correlation between them:

[0062]

[0063] The value of r ranges from [-1, 1], and the closer the absolute value is to 1, the stronger the correlation.

[0064] Preferably, S4 includes the following steps:

[0065] S41. Constructing a fusion model:

[0066] Artificial neural network (ANN) is selected as the fusion model to construct a multilayer perceptron (MLP) structure with input layer, hidden layer and output layer. The number of nodes in the input layer corresponds to the total number of features after feature extraction from various data sources.

[0067] S42. Training the model:

[0068] The multi-source data after S1 acquisition, S2 preprocessing and S3 feature extraction are divided into training set, validation set and test set. Taking the training set as an example, the processed power equipment operation data feature vector D1, the processed user electricity consumption data feature vector D2 and the processed environmental monitoring data feature vector combination D3 are used as input data.

[0069] The backpropagation algorithm is used to minimize the error between the model's prediction and the labeled data by continuously adjusting the weights and biases of the neural network. During the training process, the model is evaluated using a validation set to prevent overfitting.

[0070] S43, Data Fusion Computation:

[0071] Once the model training is complete and it passes the test set validation, the feature vectors of the multi-source data collected and processed in real time are input into the trained neural network model. The model performs calculations on the input data based on the learned feature relationships and weights, and finally outputs the fused result F.

[0072] Therefore, the present invention employs the above-mentioned method for multi-source data acquisition and fusion in power distribution networks, and the beneficial effects are as follows:

[0073] (1) This invention uses a variety of professional equipment to comprehensively collect multi-source data such as power equipment operation, user electricity consumption and environmental monitoring, including rich indicators such as transformer oil temperature and line current, which lays a solid data foundation for subsequent analysis.

[0074] (2) The present invention uses a mean filtering algorithm to effectively filter out noise such as sensor error and communication interference, making the data smoother and more accurate, and reducing the interference of noise on the analysis.

[0075] (3) The present invention uses linear interpolation for power equipment operation data and fills user electricity consumption data according to the electricity consumption pattern of adjacent time periods to ensure data integrity and prevent analysis deviation due to missing values.

[0076] (4) This invention uses a standard deviation-based method to accurately detect and correct outliers, ensuring data reliability and making the analysis results close to the actual operating state of the power distribution network.

[0077] (5) This invention utilizes principal component analysis (PCA) and linear discriminant analysis (LDA) to reduce dimensionality, remove redundant features, reduce the number and complexity of data features, improve processing efficiency, reduce computational resource consumption, and accelerate analysis speed.

[0078] (6) This invention extracts specific features for different types of data, such as voltage fluctuations of power equipment, peak-valley ratio of user electricity consumption, and correlation features of environmental monitoring, to help accurately analyze the operation of the power distribution network.

[0079] (7) This invention uses the multilayer perceptron (MLP) structure of artificial neural network (ANN) to fuse data, comprehensively consider the complex relationship of multi-source data, output the fusion result, and use it to accurately predict the operating status of the distribution network, provide a scientific basis for optimizing management, and improve the reliability and stability of the distribution network.

[0080] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0081] Figure 1 This is an overall flowchart of an embodiment of a multi-source data acquisition and fusion method for power distribution networks according to the present invention. Detailed Implementation

[0082] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0084] like Figure 1 As shown, a method for multi-source data acquisition and fusion in a power distribution network includes the following steps:

[0085] S1. Collect multi-source data in the power distribution network, including power equipment operation data, user electricity consumption data, and environmental monitoring data;

[0086] Real-time data collection of power equipment operation data in the power distribution network using sensors, including transformer oil temperature T. t Line current I l Voltage amplitude U a Distributed power source data, switchgear status data, distribution network harmonic data, and distribution network fault data;

[0087] Transformer oil temperature T t Reflecting the thermal state of the transformer, excessively high oil temperature indicates an internal fault in the transformer, and the line current I... l and voltage amplitude U a These measures power transmission and power quality in the distribution network, respectively; distributed generation data includes generation power P. g and cumulative power generation value E g Power generation P g It represents the electrical power injected into the distribution network by distributed generation at a certain moment, reflecting the generation capacity and real-time generation status of distributed generation.

[0088] The formula for calculating the cumulative value of electricity generation is:

[0089]

[0090] Where t1 and t2 are the start and end times of the time interval, this formula is derived by applying the power generation P... g Integrating over the time interval [t1, t2] yields the total power generation of the distributed power source within this time interval, which is used to statistically analyze the total power generation of the distributed power source over a certain period of time.

[0091] In addition, smart meters collect user electricity consumption data, including user electricity consumption E. u This represents the user's total electricity consumption and electricity consumption time t. u Used to analyze users' electricity consumption habits and power consumption P u User electricity consumption at different times; environmental monitoring data collected from weather stations, including ambient temperature T. eHumidity (H) and wind speed (V) data.

[0092] Switchgear status data includes switch open / closed status S S and the number of switching actions N S Among them, the switch open / closed state S S The value is either 0 or 1, where 0 indicates the switch is in the open state and 1 indicates the switch is in the closed state; the number of switch actions N S Used to assess the lifespan and health status of switchgear.

[0093] Harmonic data for distribution networks includes total harmonic voltage distortion (THD). U Total Harmonic Distortion (THD) I Total Harmonic Voltage Distortion (THD) U The calculation formula is:

[0094]

[0095] Among them, U h U1 is the effective value of the h-th harmonic voltage, and U2 is the effective value of the fundamental voltage. h For the highest harmonic order considered, THD U Used to measure the degree of distortion of voltage waveforms in a distribution network, expressed in %;

[0096] The formula for calculating the total distortion rate of harmonic currents is:

[0097]

[0098] Among them, I h Ih is the effective value of the h-th harmonic current, I1 is the effective value of the fundamental current, and n is the effective value of the fundamental current. h For the highest harmonic order considered, THD I Used to measure the degree of distortion of current waveforms in a distribution network, measured in percent.

[0099] Distribution network fault data includes fault type F type and the time of failure T type Fault type F type This includes short-circuit faults denoted as S, open-circuit faults as O, and ground faults as G, with the fault occurrence time T. type Used to record the exact moment a fault occurs.

[0100] S2. Preprocess the collected data, including noise removal, missing value filling, outlier detection and handling, data dimensionality reduction and data encoding;

[0101] For the collected power equipment operation data, user electricity consumption data, and environmental monitoring data, a mean filtering algorithm is used for noise removal, with a one-dimensional data sequence x = x1, x2, ... xm For example, suppose the filter window size is k (k is an odd number), and the data sequence after mean filtering is y = y1, y2, ... y m , where y i The calculation formula is:

[0102]

[0103] In the formula, i = 1, 2, ..., m; when j < 1 or j > m, boundary value filling is used; x i Let y be the i-th data point in the original data sequence. i Let be the i-th data point in the data sequence after mean filtering, and k be the size of the filtering window, which controls the number of data points participating in the mean calculation. The data is smoothed and noise interference is removed by averaging the data within the window.

[0104] For transformer oil temperature, line current, and voltage amplitude data in power equipment operation data, if missing values ​​exist, linear interpolation is used to fill them in.

[0105] Suppose the data sequence z = z1, z2, ... z j , ..., z n In the middle, z j The missing value is z, and its two adjacent non-missing values ​​are z and z. j-1 and z j+1 Then z j The formula for calculating the fill value is:

[0106]

[0107] After simplification, it becomes:

[0108]

[0109] In the formula, j is the index of the missing value in the data sequence, and z j-1 z is the non-missing data point preceding the missing value. j+1 For each non-missing data point following a missing value, the missing value is estimated using a linear relationship to ensure data continuity.

[0110] For the electricity consumption time in the user's electricity consumption data, if there are missing time periods, the electricity consumption pattern of adjacent time periods will be used to fill in the gaps:

[0111] Assuming the power consumption of adjacent time periods t3 and t5 is known to be P3 and P5 respectively, the power consumption of the missing time period t4, P4, can be determined through a linear relationship:

[0112]

[0113] Where t3, t4, and t5 are time points, and P3, P4, and P5 are the power consumption at the corresponding time points.

[0114] The standard deviation-based method is used to detect outliers in the data. Let the data sequence be c = c1, c2, ... c q First, calculate the mean μ and standard deviation σ of the data series.

[0115] The calculation formulas are as follows:

[0116]

[0117] If a certain data point c k Satisfy |c k If -μ∣>3σ, then c is considered k For outliers, a linear interpolation of the adjacent non-outliers before and after the detected outlier is used for correction. Let's assume the outlier is c. k The adjacent non-outlier values ​​are c k-s′ and c k+t' (where s' and t' are positive integers), then the corrected outlier c k' The calculation formula is:

[0118]

[0119] In the formula, q is the length of the data sequence, μ is the mean of the data sequence, σ is the standard deviation of the data sequence, and c i Let c be the i-th data point in the data sequence. k c represents the potentially abnormal data points to be detected. k-s' and c k+t' These are outlier values ​​c. k Adjacent non-outlier values, c k' These are the corrected outliers.

[0120] Distribution network multi-source data may contain a large number of features, some of which may be redundant or have little impact on subsequent analysis. Dimensionality reduction of distribution network multi-source data can improve data processing efficiency and reduce computational load.

[0121] Data dimensionality reduction operations for multi-source data in power distribution networks include principal component analysis (PCA) and linear discriminant analysis (LCA). The PCA process is as follows:

[0122] Let the original data matrix be X. m'×n' Where m' is the number of samples and n' is the number of features, the data is first centered to obtain... Calculate the covariance matrix so that the mean of each column of data is 0. Then, solve for the eigenvalues ​​λ1≥λ2≥...≥λ of the covariance matrix S. n'and the corresponding feature vectors v1, v2, ..., v n' .

[0123] Select the eigenvectors v1, v2, ..., v corresponding to the first k' largest eigenvalues. k' Forming the projection matrix V n'×k' Projecting the original data matrix X onto these k' principal components yields the dimensionality-reduced data matrix Y. m'×n' =XV, where k' is the number of features after dimensionality reduction. An appropriate value needs to be selected according to the actual situation to reduce the number of features and reduce the complexity of the data without losing too much information.

[0124] The linear discriminant analysis (LDA) process is as follows:

[0125] Assume the data has C' categories, and the i'-th category has m" i' Number of samples, total number of samples Calculate the mean vector u' of each class. i Calculate the within-class scatter matrix using 'and the global mean vector u':

[0126]

[0127] Calculate the inter-class scatter matrix:

[0128]

[0129] Solving the generalized eigenvalue problem S' bv '=τ'S' wv ', thus obtaining the eigenvalues ​​τ'1≥τ'2≥...≥τ' n'-1 and the corresponding feature vectors v'1, v'2, ..., v' n'-1 Select the eigenvectors v'1, v'2, ..., v' corresponding to the first k largest eigenvalues. k" Forming the projection matrix w' n'×k" The original data is projected onto k" directions to achieve data dimensionality reduction. Here, k" ≤ C'-1. By finding projection directions, the distance between different categories is maximized, and the distance within the same category is minimized.

[0130] Data encoding employs one-hot encoding and label encoding. The one-hot encoding process is as follows: For a categorical variable X with n" distinct values, its value set is x'1, x'2, ..., x'. n" Convert the categorical variable into an n"-dimensional binary vector. If the sample value is x′ j′ Then the encoded vector is e j′ e j′It is an n"-dimensional vector, where the j′-th element is 1 and the rest are 0. For example, if there are three types of faults: short circuit, open circuit, and ground fault, then the short circuit code is 1, 0, 0, the open circuit code is 0, 1, 0, and the ground fault code is 0, 1, 0.

[0131] The label encoding process is as follows: for each different value x' of the categorical variable X”. i" Assign a unique integer code l' i" , i"=1,2,...,n". For example, in the fault types, the short circuit code is 0, the open circuit code is 1, and the grounding code is 2.

[0132] S3. Extract features from the preprocessed data, including extracting voltage fluctuation features and current change rate features from power equipment operation data; extracting peak-valley ratio features and load curve slope features from user electricity consumption data; and extracting trend features and correlation features from environmental monitoring data.

[0133] Voltage change rate characteristics

[0134] The formula for calculating voltage fluctuation is:

[0135]

[0136] Wherein, ΔU is the voltage fluctuation value, reflecting the degree of voltage fluctuation, presented as a percentage, used to measure the relative change in voltage deviation from the rated value; U max It represents the maximum voltage amplitude within a statistical time period, indicating the highest voltage state; U min It represents the minimum voltage amplitude within the statistical time period, indicating the lowest voltage value; U rated The rated voltage of a power system is a benchmark value for measuring voltage fluctuations. Different voltage levels of distribution networks have corresponding standard rated voltage values.

[0137] By calculating voltage fluctuation characteristics, one can intuitively understand the stability of voltage in the distribution network. Excessive voltage fluctuations may affect the normal operation of power equipment or even cause equipment damage. This characteristic is crucial for assessing the power quality of the distribution network.

[0138] Current change rate characteristics

[0139] The formula for calculating the rate of change of current is:

[0140]

[0141] in, It represents the rate of change of current, with units of amperes per second (A / s), reflecting how quickly the current changes over time; I t2 It is the line current value at time t2, representing the magnitude of the current at that moment; I t1It is the line current value at time t1, used to correlate with I. t2 The difference is used to calculate the change in current; t2-t1 is the time interval in seconds (s), which represents the length of the time interval selected for calculating the rate of change of current.

[0142] The rate of change of current can be used to monitor the operating status of power equipment. For example, when equipment such as transformers fails, the rate of change of current may become abnormal. Analyzing this characteristic can help to detect potential faults in a timely manner.

[0143] Feature extraction of user electricity consumption data

[0144] Electricity peak-valley ratio characteristics

[0145] Formula for calculating peak-to-valley electricity consumption ratio:

[0146]

[0147] Among them, P ratio P represents the peak-to-valley ratio, which reflects the relative magnitude of electricity consumption by users during peak and off-peak hours. peak It is the peak power consumption of a user within the statistical period, representing the user's maximum load state; P valley It represents the lowest power consumption of a user within a statistical period, reflecting the user's minimum power load. The statistical period can typically be one day, one week, or one month, depending on the specific analysis requirements.

[0148] Peak-valley ratio helps power companies understand users' electricity consumption patterns. For users with high peak-valley ratios, targeted peak-valley pricing strategies can be developed to encourage them to consume more electricity during off-peak hours, thereby balancing grid load and improving the efficiency of power resource utilization.

[0149] Load curve slope characteristics

[0150] Formula for calculating the slope of the load curve (taking linear fitting as an example to approximate the slope over a certain period):

[0151]

[0152] Where k is the slope of the load curve, in kilowatts per hour (kW / h), representing the change in user power consumption per unit time, used to measure the rate of increase or decrease in user power load; P t2 P represents the power consumption of a user at time t2, reflecting the user's power load at that time. t1 It is the power consumption value of user at time t2, and P t2 In conjunction with the calculation of power change, t2-t1 is also a time interval in hours (h), used to determine the time interval corresponding to the calculated power change rate.

[0153] The slope characteristics of the load curve can help analyze the changing trends of users' electricity consumption behavior. For example, a larger slope may mean that users turn their electrical equipment on or off more frequently during that period, which may have a certain impact on the stability of the power system. By monitoring and analyzing this characteristic, the power sector can take countermeasures in advance.

[0154] Trend characteristics

[0155] Linear trend analysis uses linear regression to determine whether environmental factors are trending upward, downward, or stable over time. For example, a sustained rise in temperature may affect the heat dissipation of electrical equipment, increasing the risk of equipment failure.

[0156] Linear trend formula: Assuming that the data y has a linear relationship with time t:

[0157] y = β0 + β1t + ∈;

[0158] Where β0 is the intercept, β1 is the slope, and ∈ is the error term. The least squares method is used to estimate β0 and β1, i.e.

[0159] Seasonal trends

[0160] Environmental data often exhibits seasonal variations; for example, temperature changes cyclically throughout the year. Seasonal trends, obtained by decomposing time-series data and extracting seasonal components, help predict environmental conditions in different seasons and allow for proactive maintenance and management of power distribution networks.

[0161] The calculation of seasonal trends uses seasonal decomposition methods, including additive models:

[0162] y t =T t +S t +R t ; or multiplication model: y t =T t ×S t ×R t ;

[0163] Among them, y t It is the observation value at time t, T t It is a trend component, S t It is a seasonal ingredient, R t It is a residual component.

[0164] Correlation characteristics

[0165] The correlation between temperature and humidity studies the relationship between temperature and humidity. The combination of high temperature and high humidity has a greater impact on the insulation performance of electrical equipment.

[0166] Temperature T is measured using the Pearson correlation coefficient. ei and humidity Linear correlation between them:

[0167]

[0168] The value of r ranges from [-1, 1], and the closer the absolute value is to 1, the stronger the correlation.

[0169] The correlation between wind speed and other factors involves analyzing the relationship between wind speed and factors such as temperature and humidity. For example, strong winds may accelerate heat loss and affect equipment temperature. The Pearson correlation coefficient is used to calculate the correlation between wind speed and other environmental factors. S4. Data fusion based on machine learning includes the following steps:

[0170] S41. Constructing a fusion model:

[0171] Artificial neural network (ANN) is selected as the fusion model to construct a multilayer perceptron (MLP) structure with input layer, hidden layer and output layer. The number of nodes in the input layer corresponds to the total number of features after feature extraction from various data sources.

[0172] If 5 features are extracted from power equipment operation data, 3 features are extracted from user electricity consumption data, and 4 features are extracted from environmental monitoring data, then the number of input layer nodes is 5 + 3 + 4 = 12.

[0173] The number of hidden layer nodes can be determined based on empirical formulas or through experimental debugging, such as using...

[0174]

[0175] Where, n h n is the number of hidden layer nodes. i n is the number of nodes in the input layer. o The number of output layer nodes is determined based on the specific analysis objective. If the objective is to predict the operating status of a certain type of distribution network, the output layer may have only one node. 'a' is the adjustment constant (generally a value between 1 and 10; here we assume a = 5).

[0176] S42. Training the model:

[0177] The multi-source data, after being collected in S1, preprocessed in S2, and feature extracted in S3, is divided into training set, validation set, and test set. Taking the training set as an example, the processed power equipment operation data feature vector D1, the processed user electricity consumption data feature vector D2, and the processed environmental monitoring data feature vector combination D3 are used as input data. At the same time, corresponding label data is prepared. For example, for the task of predicting the operation status of the distribution network, the label data can be status indicators such as normal operation, slight abnormality, and severe abnormality.

[0178] The backpropagation algorithm is used to minimize the error between the model's prediction and the labeled data by continuously adjusting the weights and biases of the neural network. During the training process, the model is evaluated using a validation set to prevent overfitting.

[0179] S43, Data Fusion Computation:

[0180] Once the model training is complete and it passes the test set validation, the feature vectors of the multi-source data collected and processed in real time are input into the trained neural network model. The model performs calculations on the input data based on the learned feature relationships and weights, and finally outputs the fused result F.

[0181] If input D1 contains voltage fluctuation characteristic value ΔU and current change rate characteristic value The vector D2, composed of the peak-to-valley ratio of electricity consumption, contains the characteristic value P. ratio D3 is a vector composed of load curve slope characteristic value k, etc., and D3 is a vector composed of ambient temperature trend characteristic value t1, humidity and temperature correlation characteristic value r, etc. After internal neuron calculation and weight adjustment, the model outputs a result F that comprehensively reflects the operation status of the distribution network. This result can be a numerical value (such as the predicted load value), a category (such as the distribution network operation status category), or a vector (containing multiple comprehensive indicators).

[0182] Example

[0183] This embodiment focuses on a typical power distribution network area in a medium-sized city, which includes 10 substations and 50 distribution lines, serving approximately 10,000 residential users and 500 commercial users. Sensors, smart meters, and weather stations have been deployed to collect multi-source data for collecting operational data from power equipment.

[0184] Multi-source data acquisition

[0185] (I) Power Equipment Operation Data

[0186] Transformer oil temperature: Each transformer is equipped with a high-precision temperature sensor, which collects data every 15 minutes and transmits it to the data acquisition center in real time via a wireless module.

[0187] Line current and voltage amplitude: Current transformers and voltage sensors are installed at key line locations, and data is collected once per second. After signal conditioning, the data is transmitted to the acquisition system via fiber optic network.

[0188] Distributed power data: Establish communication with the equipment to obtain the power generation in real time, and record the cumulative power generation value every 10 minutes.

[0189] Switchgear status data: The switch is equipped with an intelligent monitoring module that collects the opening and closing status (0 for open, 1 for closed) and the number of actions every 30 minutes, and transmits the data to the acquisition center via power line carrier communication.

[0190] Harmonic data of power distribution network: The professional harmonic monitoring instrument collects the total harmonic voltage distortion rate and total harmonic current distortion rate every hour, stores the data and sends it to the management platform remotely on a timed basis.

[0191] Distribution network fault data: When the fault monitoring device detects a fault, it immediately records the fault type (short circuit 0, open circuit 1, grounding 2) and the time of occurrence, and uploads it to the dispatch center via the emergency communication channel.

[0192] (II) User electricity consumption data

[0193] All users in the area are equipped with smart meters, which collect electricity consumption, time of use, and power consumption every hour, and then aggregate the data to the power company's data acquisition server via a wireless communication network.

[0194] (III) Environmental Monitoring Data

[0195] Three meteorological stations are set up in the area to collect ambient temperature, humidity and wind speed every 30 minutes. The data is then transmitted to the environmental data monitoring platform via GPRS communication and interacts with the power distribution network data acquisition system.

[0196] III. Data Preprocessing

[0197] (I) Noise Removal

[0198] For data such as transformer oil temperature, line current, voltage amplitude, ambient temperature, humidity, and wind speed, a mean filtering algorithm is used. For example, the oil temperature data is calculated by averaging the values ​​at 5 consecutive time points, and boundary value filling is performed.

[0199] (ii) Missing value imputation

[0200] 1. Power equipment operation data: Missing values ​​for transformer oil temperature, line current, and voltage amplitude are calculated using linear interpolation. For example, if the j-th value of the line current is missing, the sum of the adjacent non-missing values ​​is used to fill in the missing value.

[0201] 2. User electricity consumption data: For missing time periods, fill in the missing time periods according to the electricity consumption patterns of adjacent time periods. For example, if the power consumption of adjacent time periods is known, the power consumption of the missing period is _____.

[0202] (III) Outlier Detection and Handling

[0203] For all data sequences, outliers are detected using a standard deviation-based method. If a line current data point meets the criteria for an outlier, it is corrected by linear interpolation of adjacent non-outlier values.

[0204] (iv) Data dimensionality reduction

[0205] Principal Component Analysis (PCA): PCA is performed on multi-source data, such as the original power equipment operation data matrix X. m'×n' With sample size m' = 1000 and feature size n' = 10, after centering, calculate the covariance matrix, and then take the eigenvectors corresponding to the top k' = 5 largest eigenvalues ​​to form the projection matrix V. n'×k' The dimensionality reduction matrix Y is obtained. m'×n' =XV.

[0206] Linear Discriminant Analysis (LDA): Assuming the data is divided into three classes: normal operation, early warning, and fault, the mean vector, within-class and between-class scatter matrices are calculated. The generalized eigenvalue problem is solved, and the eigenvectors corresponding to the first k largest eigenvalues ​​are used to form the projection matrix w'. n'×k" Achieve dimensionality reduction.

[0207] (v) Data Encoding

[0208] Unique hot coding: Encodes categorized variables such as switch opening and closing status and fault type, such as short circuit [1,0,0], open circuit [0,1,0], and grounding [0,0,1] in fault types.

[0209] Label coding: Fault type: short circuit code 0, open circuit code 1, grounding code 2.

[0210] IV. Feature Extraction

[0211] (I) Power Equipment Operation Data

[0212] Voltage fluctuation characteristics: Calculated according to the formula, for example, if the maximum voltage is 105V and the minimum voltage is 95V during a certain period, and the rated voltage is 100V, the voltage fluctuation value is 10%.

[0213] Current change rate characteristics: Calculated using a formula, for example, when t1 = 10s the current is 50A, and when t2 = 15s the current is 60A, the current change rate is 2A / s.

[0214] (II) User electricity consumption data

[0215] Peak-valley ratio characteristics of electricity consumption: Calculated according to the formula, if a user's peak power consumption is 5kW and the valley power consumption is 1kW in a day, the peak-valley ratio of electricity consumption is 5.

[0216] Load curve slope characteristics: Calculated using a formula, for example, when t1 = 2h the power is 2kW, and when t2 = 3h the power is 2.5kW, the load curve slope is 0.5kW / h.

[0217] (III) Environmental Monitoring Data

[0218] Linear trend: Assuming the ambient temperature is y = β0 + β1t + ∈, the least squares method is used to estimate it. For example, if β0 = 20 and β1 = 0.5, the temperature rises by 0.5℃ per hour.

[0219] Seasonal trends: Use an additive model to decompose temperature data and analyze the seasonal variation patterns.

[0220] Correlation characteristics

[0221] Temperature and humidity: Calculated using the Pearson correlation coefficient, assuming a correlation coefficient r = -0.6, indicating a negative correlation between the two.

[0222] Wind speed and other factors: The calculated correlation coefficient between wind speed and temperature is r = -0.3, showing a weak negative correlation.

[0223] V. Data Fusion Based on Machine Learning

[0224] (I) Constructing a Fusion Model

[0225] A multilayer perceptron (MLP) structure based on artificial neural networks (ANNs) is selected. Assume that 5 features are extracted from power equipment operation data, 3 from user electricity consumption data, and 4 from environmental monitoring data, with 12 nodes in the input layer. The number of hidden layer nodes is calculated using an empirical formula (output layer node count is 1, adjustment constant a = 5), approximately 8. The number of output layer nodes is set according to the analysis objective, such as predicting the operating status of the distribution network, with output values ​​ranging from [0,1], where 0 represents normal and 1 represents severe anomaly.

[0226] (II) Training Model

[0227] The collected, preprocessed, and feature-extracted data were divided into a training set (70%), a validation set (15%), and a test set (15%). The training set used the feature vectors of the three types of processed data as input, with the distribution network operating status as the label (normal 0, slightly abnormal 0.3, severely abnormal 0.8). It was trained using the backpropagation algorithm and evaluated using the validation set to prevent overfitting.

[0228] (III) Data Fusion Computing

[0229] The trained model receives feature vectors from multi-source data processed in real time, including power equipment operation [0.2, 0.5, 0.3, 0.1, 0.4], user electricity consumption [0.8, 0.6, 0.3], and environmental monitoring [0.5, 0.7, 0.6, 0.4]. The model outputs a fusion result F = 0.4. Based on the threshold (0.3 is the boundary between normal and slightly abnormal, and 0.6 is the boundary between slightly abnormal and severely abnormal), the distribution network is judged to be in a slightly abnormal state.

[0230] Therefore, the present invention adopts the above-mentioned method for multi-source data acquisition and fusion of distribution networks. By comprehensively collecting multi-source data of distribution networks, using various preprocessing and feature extraction methods, and combining machine learning fusion, the operating status of distribution networks can be accurately analyzed, providing a strong basis for optimized management and effectively improving the reliability and stability of distribution networks.

[0231] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for multi-source data acquisition and fusion in a power distribution network, characterized in that, Includes the following steps: S1. Collect multi-source data in the power distribution network, including power equipment operation data, user electricity consumption data, and environmental monitoring data; S2. Preprocess the collected data, including noise removal, missing value filling, outlier detection and handling, data dimensionality reduction and data encoding; S3. Extract features from the preprocessed data; S4. Data fusion based on machine learning.

2. The method for multi-source data acquisition and fusion in a power distribution network according to claim 1, characterized in that, In S1, sensors are used to collect real-time operating data of power equipment in the distribution network, including transformer oil temperature T. t Line current I l Voltage amplitude U a Distributed power source data, switchgear status data, distribution network harmonic data, and distribution network fault data; Transformer oil temperature T t Reflecting the thermal state of the transformer, excessively high oil temperature indicates an internal fault in the transformer, and the line current I... l and voltage amplitude U a These measures power transmission and power quality in the distribution network, respectively. Distributed power data includes power generation P g and cumulative power generation value E g Power generation P g It represents the electrical power injected into the distribution network by distributed generation at a certain moment, reflecting the generation capacity and real-time generation status of distributed generation. The formula for calculating the cumulative value of power generation is: Where t1 and t2 are the start and end times of the time interval, used to count the total power generation of distributed power sources within a certain time period.

3. The method for multi-source data acquisition and fusion in a power distribution network according to claim 2, characterized in that, In S2, the switchgear status data includes the switch open / closed state S. S and the number of switching actions N S Among them, the switch open / closed state S S The value is either 0 or 1, where 0 indicates the switch is in the open state and 1 indicates the switch is in the closed state; the number of switch actions N S Used to assess the lifespan and health status of switching equipment; Harmonic data for distribution networks includes total harmonic voltage distortion (THD). U Total Harmonic Distortion (THD) I The formula for calculating the total harmonic voltage distortion rate is: Among them, U h U1 is the effective value of the h-th harmonic voltage, and U2 is the effective value of the fundamental voltage. h For the highest harmonic order considered, THD U Used to measure the degree of distortion of voltage waveforms in a distribution network, expressed in %; The formula for calculating the total distortion rate of harmonic currents is: Among them, I h Ih is the effective value of the h-th harmonic current, I1 is the effective value of the fundamental current, and n is the effective value of the fundamental current. h For the highest harmonic order considered, THD I Used to measure the degree of distortion of current waveforms in a distribution network, expressed in %; Distribution network fault data includes fault type F type and the time of failure T type Fault type F type This includes short-circuit faults denoted as S, open-circuit faults as O, and ground faults as G, with the fault occurrence time T. type Used to record the exact moment a fault occurs.

4. The method for multi-source data acquisition and fusion in a power distribution network according to claim 3, characterized in that, In S1, user electricity consumption data, including user electricity consumption E, is collected through smart meters. u This represents the user's total electricity consumption and electricity consumption time t. u Used to analyze users' electricity consumption habits and power consumption P u The user's electricity consumption at different times; Environmental monitoring data is collected through weather stations, including ambient temperature (T). e Humidity (H) and wind speed (V) data.

5. The method for multi-source data acquisition and fusion in a power distribution network according to claim 4, characterized in that, In S2, the mean filtering algorithm is used to remove noise from the collected power equipment operation data, user electricity consumption data and environmental monitoring data. For transformer oil temperature, line current and voltage amplitude data in power equipment operation data, if there are missing values, linear interpolation is used to fill them in. If there are missing time periods in the user's electricity consumption data, the electricity consumption pattern of adjacent time periods will be used to fill them in. The standard deviation-based method is used to detect outliers in the data. Let the data sequence be c = c1, c2, ... c q First, calculate the mean μ and standard deviation σ of the data sequence. If a certain data point c k Satisfy |c k If -μ∣>3σ, then c is considered k For outliers, the detected outliers are corrected by linear interpolation of the adjacent non-outliers before and after the data point.

6. The method for multi-source data acquisition and fusion in a power distribution network according to claim 5, characterized in that, In S2, the dimensionality reduction operation for multi-source data of the distribution network includes principal component analysis and linear discriminant analysis. The principal component analysis (PCA) process is as follows: Let the original data matrix be X. m'×n' Where m' is the number of samples and n' is the number of features, the data is first centered to obtain... Calculate the covariance matrix so that the mean of each column of data is 0. Then, solve for the eigenvalues ​​λ1≥λ2≥...≥λ of the covariance matrix S. n′ and the corresponding feature vectors v1, v2, ..., v n′ ; Select the eigenvectors v1, v2, ..., v corresponding to the first k′ largest eigenvalues. k′ Forming the projection matrix V n′×k′ Projecting the original data matrix X onto these k′ principal components yields the dimensionality-reduced data matrix Y. m′×n′ =XV, where k′ is the number of features after dimensionality reduction. An appropriate value needs to be selected according to the actual situation in order to reduce the number of features and reduce the complexity of the data without losing too much information. The linear discriminant analysis (LDA) process is as follows: Assume the data has C′ categories, and the i′-th category has m" i′ Number of samples, total number of samples Calculate the mean vector u′ for each class. i’ Given the global mean vector u′, calculate the within-class scatter matrix: Calculate the inter-class scatter matrix: Solving the generalized eigenvalue problem S' b v'=τ'S' w v', we obtain the eigenvalues ​​τ'1≥τ'2≥...≥τ' n'-1 and the corresponding feature vectors v'1, v'2, ..., v' n'-1 Select the eigenvectors v'1, v'2, ..., v' corresponding to the first k" largest eigenvalues. k" Forming the projection matrix w' n'×k" The original data is projected onto k" directions to achieve data dimensionality reduction.

7. The method for multi-source data acquisition and fusion in a power distribution network according to claim 6, characterized in that, In S2, data encoding employs one-hot encoding and label encoding. The one-hot encoding process is as follows: For a categorical variable X with n" distinct values, its value set is x′1, x′2, ..., x′ n" Convert the categorical variable into an n"-dimensional binary vector. If the sample value is x' j' Then the encoded vector is e j′ e j′ It is an n"-dimensional vector, where the j'-th element is 1 and the rest are 0; The label encoding process is as follows: for each different value x′ of the categorical variable X”. i" Assign a unique integer code l′ i" , i" = 1, 2, ..., n".

8. The method for multi-source data acquisition and fusion in a power distribution network according to claim 7, characterized in that, In S3, feature extraction of the preprocessed data includes extracting voltage fluctuation features and current change rate features from the power equipment operation data; Extract peak-to-valley ratio and load curve slope characteristics from user electricity consumption data; Extract trend and correlation features from environmental monitoring data; The formula for calculating voltage fluctuation is: Wherein, ΔU is the voltage fluctuation value, reflecting the degree of voltage fluctuation, presented as a percentage, used to measure the relative change in voltage deviation from the rated value; U max It represents the maximum voltage amplitude within a statistical time period, indicating the highest voltage state; U min It represents the minimum voltage amplitude within the statistical time period, indicating the lowest voltage value; U rated The rated voltage of the power system is the benchmark value for measuring voltage fluctuations. The formula for calculating the rate of change of current is: in, I represents the rate of change of current, measured in amperes per second, reflecting how quickly the current changes over time. t2 It is the line current value at time t2, representing the magnitude of the current at that moment; I t1 It is the line current value at time t1, used to correlate with I. t2 The difference is used to calculate the change in current; t2-t1 is the time interval in seconds, representing the length of the time period selected for calculating the rate of change of current.

9. The method for multi-source data acquisition and fusion in a power distribution network according to claim 8, characterized in that, In S3, the formula for calculating the peak-to-valley ratio of electricity consumption is: Among them, P ratio The peak-to-valley ratio reflects the relative magnitude of electricity consumption by users during peak and off-peak hours; P peak It is the peak power consumption of a user within the statistical period, representing the user's maximum load state; P valley It is the lowest value of the user's electricity consumption within the statistical period; The formula for calculating the slope of the load curve is: Where k is the slope of the load curve, in kilowatts per hour, representing the change in user power consumption per unit time, used to measure the rate of increase or decrease in user power load; P t2 P represents the power consumption of a user at time t2, reflecting the user's power load at that time. t1 The power consumption of user at time t2 is related to P. t2 In conjunction with the calculation of power change; t2-t1 is also a time interval in hours, used to determine the time interval corresponding to the calculated power change rate; Linear trend formula: Assuming that the data y has a linear relationship with time t: y = β0 + β1t + ∈; Where β0 is the intercept, β1 is the slope, and ∈ is the error term. The least squares method is used to estimate β0 and β1, i.e. The calculation of seasonal trends uses seasonal decomposition methods, including additive models: y t =T t +S t +R t ; or multiplication model: y t =T t ×S t ×R t ; Among them, y t It is the observation value at time t, T t It is a trend component, S t It is a seasonal ingredient, R t It is a residual component; Temperature T is measured using the Pearson correlation coefficient. ei and humidity Linear correlation between them: The value of r ranges from [-1, 1], and the closer the absolute value is to 1, the stronger the correlation.

10. The method for multi-source data acquisition and fusion in a power distribution network according to claim 9, characterized in that, S4 includes the following steps: S41. Constructing a fusion model: Artificial neural network (ANN) is selected as the fusion model to construct a multilayer perceptron (MLP) structure with input layer, hidden layer and output layer. The number of nodes in the input layer corresponds to the total number of features after feature extraction from various data sources. S42. Training the model: The multi-source data after S1 acquisition, S2 preprocessing and S3 feature extraction are divided into training set, validation set and test set. Taking the training set as an example, the processed power equipment operation data feature vector D1, the processed user electricity consumption data feature vector D2 and the processed environmental monitoring data feature vector combination D3 are used as input data. The backpropagation algorithm is used to minimize the error between the model's prediction and the labeled data by continuously adjusting the weights and biases of the neural network. During the training process, the model is evaluated using a validation set to prevent overfitting. S43, Data Fusion Computation: Once the model training is complete and it passes the test set validation, the feature vectors of the multi-source data collected and processed in real time are input into the trained neural network model. The model performs calculations on the input data based on the learned feature relationships and weights, and finally outputs the fused result F.

Images