Method and system for monitoring and controlling ring main unit based on multi-source data

Abstract

The invention discloses a method and a system for monitoring and controlling a ring main unit based on multi-source data. The method comprises the following steps: acquiring real-time monitoring data of electrical parameters, environmental parameters and equipment state parameters of the ring main unit through a multi-source sensor network; constructing an adaptive weight fusion algorithm with a data quality evaluation mechanism, performing reliability evaluation on the multi-source data, and outputting a fused multi-dimensional feature vector; establishing a deep convolutional neural network model based on a multi-scale attention mechanism, and performing time sequence and space double attention analysis on the multi-dimensional feature vector to realize fine recognition of the operation state; constructing a time sequence prediction model based on a state identification result, analyzing an operation trend, predicting a potential fault, and generating a grading early warning signal and a fault prediction report; and generating an optimal control strategy according to the early warning signal and executing a control instruction. Accurate identification, trend analysis and fault early warning of the operation state of the ring main unit are realized, and technical support is provided for safe and stable operation and intelligent operation and maintenance of a power distribution network.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of ring main unit monitoring, and in particular to a ring main unit monitoring control method and system based on multi-source data. BACKGROUND

[0002] With the rapid development of China's power industry and the continuous improvement of urban distribution network construction, as the core switching device in the 10kV distribution network, the ring main unit plays a crucial role in ensuring the reliability of urban power supply. Ring main units are widely used in load-intensive areas such as residential areas, commercial centers and industrial parks, and their operating status directly affects the safety of millions of households and the normal operation of the urban economy. However, in actual operation, ring main units face many challenges and technical problems.

[0003] In daily operation and maintenance practice, ring main units often face harsh environmental conditions. High temperature and humidity in summer lead to a decline in equipment insulation performance, and low temperature in winter causes mechanical operating mechanisms to move slowly. Lightning and overvoltage shocks in the rainy season pose a serious threat to equipment safety. At the same time, dust, smoke, chemical pollution and other factors in the urban environment accelerate the aging process of the equipment. The combined effects of these environmental factors make ring main units prone to failure. According to statistics, about 40% of distribution network failures are related to ring main units, of which insulation failure, mechanical failure and electrical connection failure are the most common types of failure.

[0004] The traditional ring main unit operation and maintenance mode mainly relies on regular manual inspection and planned maintenance. This passive maintenance method has obvious shortcomings in practical application. Maintenance personnel usually conduct routine inspections of ring main units once a month or once a quarter, mainly through visual observation, sound identification and simple electrical tests to determine the equipment status. This inspection method is limited by personnel experience and inspection time, and often fails to detect potential internal equipment problems. For example, early signs of failure such as trace leakage of sulfur hexafluoride gas, slight wear of switch contacts and gradual aging of insulation materials are difficult to detect through routine inspections, and when these problems develop to the point where they can be clearly detected, they often pose a serious threat to equipment safety.

[0005] Although some intelligent monitoring technologies have improved the shortcomings of traditional operation and maintenance mode to some extent, there are still many limitations in actual engineering applications. Taking patent CN202510761652.5 as an example, the technical solution proposes a full-cycle operation data monitoring system for ring network cabinets based on artificial intelligence, but its technical architecture faces multiple challenges in actual application. First, the scheme needs to inject a test signal of 0.5% rated current into the running ring network cabinet. This active injection method has safety risks in the actual operation of the distribution network, especially during peak load periods or system failures. Any additional current injection may cause protection devices to malfunction or increase the burden on the system. Second, the scheme is mainly based on single-time voltage and current complex vector analysis, lacking continuous tracking of the evolution process of device state, and has limited effect in dealing with the slow degradation process of the device.

[0006] In actual distribution network operation, ring network cabinet failures often show obvious progressive characteristics. For example, the process of increasing contact resistance caused by switch contact wear usually lasts for several months or even years, and the aging and degradation of insulation materials is also a long-term gradual process. Traditional instantaneous state evaluation methods are difficult to capture this gradual change and may cause fault warning lag. At the same time, the state of ring network cabinets under different operating conditions varies greatly, and changes in load, environmental temperature, and operation frequency will affect the operating characteristics of the device. The existing static evaluation model cannot adapt to such dynamic changes.

[0007] In addition, existing technologies also have obvious shortcomings in handling multi-source monitoring data. Actual ring network cabinet monitoring systems usually have temperature sensors, humidity sensors, barometric pressure sensors, current transformers, voltage transformers, and partial discharge sensors, among other types of monitoring devices. The data sampling frequency, accuracy level, and reliability level of these sensors vary. Traditional simple weighted fusion methods cannot effectively handle this heterogeneity, especially when a sensor fails or drifts, which can easily lead to false positives in the entire monitoring system. In actual operation, sensor failures are not uncommon, and problems such as temperature sensor aging, humidity sensor contamination, and current transformer saturation can affect data quality. Existing systems lack effective data quality evaluation and fault tolerance mechanisms.

[0008] More importantly, existing ring network cabinet monitoring technologies generally lack effective fault prediction capabilities and can only assess the current state and provide simple alarm functions. In actual operation and maintenance work, what maintenance personnel need is early warning of future fault risks to allow for reasonable scheduling of maintenance plans, preparation of spare parts, and development of emergency plans. Monitoring systems that lack prediction capabilities can only respond passively and cannot meet the needs of modern distribution network active operation and maintenance.

[0009] In view of the technical problems and engineering requirements existing in the above actual application, it is urgent to develop a more intelligent and practical monitoring and control technology for ring network cabinets, to realize comprehensive perception, accurate evaluation and early warning of the running state of the ring network cabinet through advanced multi-source data fusion algorithm, deep learning model and time series prediction technology, and to provide strong technical support for the safe and stable operation and lean operation and maintenance of the distribution network. SUMMARY

[0010] Therefore, the present application provides a method and system for monitoring and controlling ring network cabinets based on multi-source data, aiming to build an adaptive weight fusion algorithm with a data quality evaluation mechanism, establish a deep convolutional neural network model based on a multi-scale attention mechanism, and develop a time series prediction model based on a long short-term memory network, to realize accurate identification, trend analysis and fault warning of the running state of the ring network cabinet, and to provide technical support for the safe and stable operation and intelligent operation and maintenance of the distribution network.

[0011] To achieve the above purpose, the present application provides a method for monitoring and controlling ring network cabinets based on multi-source data, comprising the following steps: A1: Obtain real-time monitoring data of each device of the ring network cabinet by establishing a multi-source sensor network, including electrical parameters, environmental parameters and device state parameters, to form multi-source raw data; A2: Build an adaptive weight fusion algorithm with a data quality evaluation mechanism to perform real-time reliability evaluation on the multi-source raw data, and output a fused multi-dimensional feature vector; A3: Establish a deep convolutional neural network model based on a multi-scale attention mechanism to perform time series and spatial dual attention analysis on the fused multi-dimensional feature vector, realize fine identification of the running state of the ring network cabinet, and obtain the state identification result; A4: Build a time series prediction model based on the state identification result, analyze the running trend of the ring network cabinet and predict potential faults, generate a graded warning signal and a fault prediction report; A5: According to the graded warning signal, combined with the running constraint conditions of the ring network cabinet, generate an optimal control strategy and execute the corresponding control instructions to realize intelligent monitoring and control of the ring network cabinet.

[0012] As a further improved method of the present application: Optionally, in the A1 step, the real-time monitoring data of each device of the ring network cabinet is obtained by establishing a multi-source sensor network, including electrical parameters, environmental parameters and device state parameters, to form multi-source raw data, including: A multi-source sensor network is established to connect the circuit breakers, load switches, voltage transformers, current transformers, temperature and humidity sensors, and partial discharge sensors in the ring network cabinet, and a unified data sampling frequency of 10 times per second is set to obtain real-time monitoring data of each device in the ring network cabinet; the monitoring data includes eight key parameter data, i.e., three-phase voltage effective value, three-phase current effective value, active power, environmental temperature, environmental humidity, partial discharge amount, circuit breaker position state, and load switch position state; the original monitoring data is preliminarily screened to eliminate abnormal values of three-phase voltage exceeding 200% of the rated value, three-phase current exceeding 300% of the rated value, temperature exceeding -40 degrees to 80 degrees, and humidity exceeding 0% to 100%, and the sensor communication state, signal quality, and calibration state are recorded as data quality identifiers to form multi-source original data containing the eight key parameter data and corresponding data quality identifiers.

[0013] Optionally, in the A2 step, an adaptive weight fusion algorithm with a data quality evaluation mechanism is constructed to perform real-time reliability evaluation on the multi-source original data, and a fused multi-dimensional feature vector is output, including: The adaptive weight fusion algorithm with the data quality evaluation mechanism is realized by using a data reliability evaluation module; the data reliability evaluation module evaluates the real-time reliability index of each data source by analyzing the time sequence consistency, physical rationality, and historical stability of each sensor data; The time sequence consistency is evaluated by calculating the standard deviation of the change rate of the current data and the data of the previous five sampling points: when the standard deviation of the change rate is less than 0.1, the time sequence consistency index is 1.0; when the standard deviation of the change rate is between 0.1 and 0.5, the time sequence consistency index is 0.5; and when the standard deviation of the change rate is greater than 0.5, the time sequence consistency index is 0.0; The physical rationality is evaluated by checking whether the data conforms to the basic laws of the power system: when the three-phase voltage symmetry deviation is less than 5% and the active power is between 0 and the rated power of the device, the physical rationality index is 1.0, otherwise it is 0.0; The historical stability is evaluated by calculating the deviation degree of the current data and the data of the same period in the past 24 hours: when the deviation is less than 20% of the historical mean, the historical stability index is 1.0; when the deviation is between 20% and 50% of the historical mean, the historical stability index is 0.5; and when the deviation is greater than 50% of the historical mean, the historical stability index is 0.0; Data sources in the multi-source original data At time The dynamic fusion weight The calculation formula is: Wherein, And Representing data sources respectively and data source At any moment The reliability index is calculated by weighted average of three dimensions: temporal consistency, physical rationality, and historical stability. and Indicates data source and data source At different times The correlation moderating factor is obtained by calculating the average cross-correlation coefficient between this data source and other data sources. The calculation process of the fused multidimensional feature vector includes three steps: data standardization, weight fusion, and feature dimensionality reduction. Data standardization uses the Z-score method to calculate the mean and standard deviation of each data source over the past 1000 sampling points, and then converts the original data at the current moment into standardized data to obtain 8 standardized data sources. The weight fusion step multiplies the 8 standardized data sources with their corresponding weights to form an 8-dimensional weighted vector. Feature dimensionality reduction uses principal component analysis to reduce the 8-dimensional weighted vector to 5 feature dimensions, forming the fused multidimensional feature vector.

[0014] It should be noted that this step establishes a multi-dimensional data quality assessment system. Through comprehensive analysis of three dimensions—temporal consistency, physical rationality, and historical stability—it can accurately identify and quantify the reliability level of each data source. Compared with traditional single-indicator assessment methods, this multi-dimensional assessment mechanism can more comprehensively reflect the true quality of sensor data and effectively avoid data fusion failures caused by errors in single-dimensional judgments. In particular, temporal consistency assessment can capture the dynamic characteristics of data changes, physical rationality assessment ensures that the data conforms to the basic laws of the power system, and historical stability assessment provides a reference benchmark for long-term trends. The three complement each other to form a complete quality assessment framework.

[0015] Optionally, in step A3, a deep convolutional neural network model based on a multi-scale attention mechanism is established to perform temporal and spatial dual attention analysis on the fused multi-dimensional feature vectors, thereby achieving refined identification of the ring main unit's operating status and obtaining status identification results, including: A deep convolutional neural network model based on a multi-scale attention mechanism is established. The deep convolutional neural network model includes four parts: a multi-scale feature extraction layer, a temporal attention layer, a spatial attention layer, and a state classification layer. The multi-scale feature extraction layer receives a time sequence of a 5-dimensional fusion feature vector as input, and the input data dimension is 120*5, wherein 120 is the time step, and 5 is the feature dimension; the multi-scale feature extraction layer adopts three parallel one-dimensional convolution branches, respectively uses convolution layers with convolution kernel sizes of 3, 5 and 7 to extract short-term, medium-term and long-term time sequence features along the time dimension, each branch contains 64 convolution kernels, the activation function adopts a ReLU function, the step is set to 1, and the padding mode is same padding to keep the output sequence length as 120; the outputs of the three parallel branches are fused through a channel splicing operation to form a multi-scale feature map with a dimension of 120*192; The time sequence attention layer receives the multi-scale feature map with a dimension of 120*192 output by the multi-scale feature extraction layer as input, and identifies key time points within the 120 time steps through a self-attention mechanism; the time sequence attention layer adopts a Query-Key-Value structure to calculate attention weights, maps the input 192-dimensional features to a Query matrix, a Key matrix and a Value matrix through three independent linear transformation layers, respectively, and compresses the feature dimensions of each matrix to 64, to obtain three matrices with a dimension of 120*64; the attention weight calculation obtains an attention matrix with a dimension of 120*120 through a dot product operation of the Query matrix and the Key matrix after softmax normalization, and then multiplies the Value matrix to obtain a time sequence attention feature with a dimension of 120*64; The spatial attention layer receives the time sequence attention feature with a dimension of 120*64 output by the time sequence attention layer as input, first compresses the time dimension through global average pooling to obtain a 64-dimensional global feature vector; then identifies the correlation between the 64 feature dimensions through a cross-attention mechanism, adopts a two-layer fully connected network to calculate a correlation weight matrix between the features, the first layer of the fully connected network maps the 64-dimensional features to 32-dimensional intermediate features, and the second layer of the fully connected network maps the 32-dimensional intermediate features back to 64-dimensional features and obtains a feature weight vector through a sigmoid activation function; finally, the feature weight vector and the original 64-dimensional global feature are multiplied element by element to obtain a weighted 64-dimensional spatial attention feature; The state classification layer receives the 64-dimensional spatial attention feature vector output by the spatial attention layer as input, adopts a three-layer fully connected network for state classification, the first layer of the fully connected network maps the 64-dimensional input features to 128-dimensional hidden features and passes them through a ReLU activation function, the second layer of the fully connected network compresses the 128-dimensional hidden features to 64-dimensional features and passes them through a ReLU activation function, and the third layer of the fully connected network maps the 64-dimensional features to 4-dimensional outputs and outputs a probability distribution of 4 running states, including normal operation, slight abnormality, potential failure and emergency failure, through a softmax function.

[0016] It should be noted that the parallel processing architecture of the multi-scale feature extraction layer is adopted in this step, and the short-term, medium-term and long-term time sequence features are captured simultaneously through three parallel branches with different convolution kernel sizes. Compared with the traditional single-scale feature extraction method, this multi-scale design can more comprehensively capture the multi-level time patterns in the ring network cabinet operation data, and effectively identify abnormal changes in different time spans. Short-term features can capture instantaneous fluctuations in electrical parameters, medium-term features reflect periodic changes in device operation, and long-term features reveal gradual trends in device status.

[0017] Optionally, the A4 step constructs a time sequence prediction model based on the state recognition result, analyzes the operation trend of the ring network cabinet and predicts potential failures, generates a hierarchical warning signal and a failure prediction report, including: Based on the probability of the operating state, a time sequence prediction model containing an LSTM network is constructed to analyze the time sequence evolution law and trend change of the ring network cabinet operating state, and to predict the potential failure type and occurrence probability within the next 30 minutes; the LSTM network contains two hidden layers, each containing 128 LSTM units, and adopts a bidirectional LSTM structure to capture the forward and backward time sequence dependency relationship, and the outputs of the forward and backward LSTM are fused through a concatenation operation; the time sequence prediction model input is a state probability sequence of the past 120 time steps, each time step containing probability values of 4 state categories; the time sequence prediction model output is a failure state probability prediction sequence of the next 30 time steps; The failure risk assessment index is obtained by weighted summation of the predicted failure state probability sequence within the next 30 time steps, and the weight coefficients are determined according to the severity of the failure type, including slight abnormality, potential failure and emergency failure; when the failure risk assessment index is greater than 0.7, a red emergency warning is generated; when the failure risk assessment index is between 0.3 and 0.7, a yellow attention warning is generated; when the failure risk assessment index is less than 0.3, the green normal state is maintained, and a hierarchical warning signal is obtained; After the hierarchical warning signal is generated, the system automatically generates a failure prediction report, which includes the current device state, the predicted failure type, the predicted occurrence time, the failure probability, the impact range analysis and the recommended treatment measures. The report is stored in JSON format and pushed to the operation and maintenance management system through the HTTP interface, and the pushing frequency is determined according to the warning level, the red warning is pushed immediately, the yellow warning is pushed every 5 minutes, and the green state is pushed every hour.

[0018] It should be noted that this step adopts a bidirectional long short-term memory network (i.e., LSTM network) architecture, which can more comprehensively understand the complex patterns of state evolution through forward and backward time-dependent relationships. Compared with traditional unidirectional recurrent neural networks, the bidirectional LSTM structure can not only use historical information to predict future states, but also combine future information to verify the rationality of historical states. This bidirectional information flow mechanism significantly improves the model's learning ability for the inherent laws of time series data. Especially when dealing with ring network cabinet operation state data with strong time sequence correlation, the bidirectional structure can more accurately capture the precursor features and evolution trend of state transition, providing a solid technical foundation for accurate fault prediction.

[0019] Further, the hierarchical warning mechanism established in this step realizes intuitive expression and differentiated processing of risk levels. Through the red, yellow and green three-level warning system, operation and maintenance personnel can quickly understand the current risk state and take appropriate measures. This hierarchical mechanism not only conforms to human cognitive habits, but also facilitates the establishment of standardized operation processes, improving the practicality and operability of the warning system. In particular, the design of different push frequencies corresponding to different warning levels not only ensures the timeliness of information transmission in emergency situations, but also avoids information overload in normal situations.

[0020] Optionally, according to the hierarchical warning signal, the A5 step generates an optimal control strategy and executes corresponding control instructions in combination with the ring network cabinet operation constraints, realizing intelligent monitoring and control of the ring network cabinet, including: Based on the hierarchical warning signal, a control strategy generation mechanism is established. When receiving a green normal state signal, an optimal operation mode is adopted; when receiving a yellow attention warning signal, a preventive adjustment mode is adopted; and when receiving a red emergency warning signal, an emergency disposal mode is adopted. The optimal operation mode mainly targets economy, optimizing system operation efficiency by adjusting controllable load distribution and switch state. The preventive adjustment mode mainly targets safety, preferentially adjusting abnormal risk equipment load and protection setting value. The emergency disposal mode targets safety only, immediately executing load shedding, switch tripping and protection action emergency measures; The control strategy includes three levels of load adjustment strategy, switch operation strategy and protection action strategy. The load adjustment strategy optimizes the load distribution of the ring network cabinet by adjusting the switching state of controllable loads, including air conditioning load, charging pile load and industrial load. The switch operation strategy changes the network topology by controlling the on-off state of circuit breakers and load switches in the ring network cabinet. The protection action strategy ensures system safety by setting the action parameters of protection devices; The particle swarm optimization algorithm is used for control parameter optimization, and three targets of system safety, economy and reliability are considered; the safety target function is the weighted sum of voltage deviation square sum and current exceeding square sum; the economy target function is the sum of load regulation cost and switch operation cost; and the reliability target function is the reciprocal of system average outage time; The constraint conditions of the control strategy include voltage constraint, current constraint, power constraint and device capacity constraint; the voltage constraint requires that the voltage of each node is kept between 95% and 105% of the rated voltage; the current constraint requires that the current of each branch is not more than 90% of the rated current; the power constraint requires that the total load power is not more than 80% of the total capacity of the system; and the device capacity constraint requires that the operating power of each device is not more than its rated capacity; After the particle swarm optimization algorithm outputs the optimal control parameters, the optimal control strategy is generated and the corresponding control instructions are executed; the control instruction execution process includes four links of instruction generation, safety check, execution confirmation and result feedback; the instruction generation link generates specific control instructions according to the optimal control parameter optimization results; the safety check link verifies whether the system state after instruction execution meets the safety constraint conditions of the control strategy through power flow calculation; the execution confirmation link sends the control instructions to the field devices and waits for the execution confirmation signal, and the timeout time is set to 5 seconds; and the result feedback module collects the instruction execution results and updates the system state.

[0021] The application further discloses a system for monitoring and controlling the ring network cabinet based on multi-source data, which comprises: A data acquisition module is configured to acquire real-time monitoring data of each device of the ring network cabinet through a multi-source sensor network, including electrical parameters, environmental parameters and device state parameters. A data fusion module is configured to construct an adaptive weight fusion algorithm with a data quality evaluation mechanism, to perform real-time reliability evaluation on multi-source original data, and to form multi-source original data. A state recognition module is configured to construct a deep convolutional neural network model based on a multi-scale attention mechanism, to perform time sequence and spatial double attention analysis on the fused multi-dimensional feature vector, and to realize fine identification of the running state of the ring network cabinet. An early warning module is configured to construct a time sequence prediction model based on the state recognition result, to analyze the running trend of the ring network cabinet and to predict potential faults, to generate a graded early warning signal and a fault prediction report. A control module is configured to generate an optimal control strategy and execute corresponding control instructions according to the early warning signal and in combination with the running constraint conditions of the ring network cabinet, to realize intelligent monitoring and control of the ring network cabinet.

[0022] Compared with the prior art, the application has at least the following beneficial effects: The application realizes intelligent processing and reliability evaluation of multi-source heterogeneous data by constructing an adaptive weight fusion algorithm with a data quality evaluation mechanism. Through comprehensive analysis of three dimensions of time sequence consistency, physical rationality and historical stability, the algorithm can accurately identify and quantify the reliability level of each data source, and dynamically adjust the fusion weight according to the data quality change. When a sensor fails or the data quality decreases, the system can automatically reduce its weight and correspondingly increase the weight of other reliable sensors, ensuring the stability and accuracy of the fusion result.

[0023] The deep convolutional neural network model based on the multi-scale attention mechanism established by the application realizes fine identification and deep feature extraction of the operating state of the ring network cabinet. The model captures short-term, medium-term and long-term time sequence features through a multi-scale feature extraction layer, identifies key time points combined with a time sequence attention mechanism, and excavates the correlation between feature dimensions combined with a spatial attention mechanism, which can accurately identify four operating states of normal operation, slight abnormality, potential failure and emergency failure.

[0024] The time sequence prediction model based on the bidirectional long short-term memory network constructed by the application realizes accurate prediction of the operating trend of the ring network cabinet and early identification of potential failures. The model captures forward and backward time sequence dependencies through a bidirectional LSTM structure, which can predict the failure type and occurrence probability within 30 minutes in the future, providing sufficient response time for operation and maintenance personnel. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flowchart of a method for monitoring and controlling a ring network cabinet based on multi-source data according to an embodiment of the application; Figure 2 A schematic diagram of the weight dynamic adjustment effect of the multi-source data fusion algorithm; Figure 3 A schematic diagram of the state recognition accuracy convergence process of the deep convolutional neural network. DETAILED DESCRIPTION

[0026] The application will be further described below with reference to the accompanying drawings, but the application is not limited in any way by the application. Any transformation or replacement based on the teaching of the application is within the protection scope of the application.

[0027] Example 1: A method for monitoring and controlling a ring network cabinet based on multi-source data, as shown in Figure 1 The method comprises the following steps: A1: Obtain real-time monitoring data of each device of the ring network cabinet through the establishment of a multi-source sensor network, including electrical parameters, environmental parameters and device state parameters, to form multi-source raw data: A multi-source sensor network is established to connect the circuit breakers, load switches, voltage transformers, current transformers, temperature and humidity sensors, and partial discharge sensors in the ring network cabinet, and a unified data sampling frequency of 10 times per second is set to obtain real-time monitoring data of each device in the ring network cabinet; the monitoring data includes eight key parameter data, i.e., three-phase voltage effective value, three-phase current effective value, active power, environmental temperature, environmental humidity, partial discharge amount, circuit breaker position state, and load switch position state; The original monitoring data is preliminarily screened to eliminate abnormal values of three-phase voltage exceeding 200% of the rated value, three-phase current exceeding 300% of the rated value, temperature exceeding -40 degrees to 80 degrees, and humidity exceeding 0% to 100%, and the sensor communication state, signal quality, and calibration state are recorded as data quality identifiers to form multi-source original data including the eight key parameter data and the corresponding data quality identifiers.

[0028] A2: An adaptive weight fusion algorithm with a data quality evaluation mechanism is constructed to perform real-time reliability evaluation on the multi-source original data, and a fused multi-dimensional feature vector is output: The adaptive weight fusion algorithm with the data quality evaluation mechanism is implemented by using a data reliability evaluation module; specifically: The data reliability evaluation module calculates the real-time reliability index of each data source by analyzing the time sequence consistency, physical rationality, and historical stability of each sensor data; The time sequence consistency is evaluated by calculating the standard deviation of the change rate of the current data and the data of the previous five sampling points: when the standard deviation of the change rate is less than 0.1, the time sequence consistency index is 1.0; when the standard deviation of the change rate is between 0.1 and 0.5, the time sequence consistency index is 0.5; and when the standard deviation of the change rate is greater than 0.5, the time sequence consistency index is 0.0; The physical rationality is evaluated by checking whether the data conforms to the basic laws of the power system: when the three-phase voltage symmetry deviation is less than 5% and the active power is between 0 and the rated power of the device, the physical rationality index is 1.0, otherwise it is 0.0; The historical stability is evaluated by calculating the deviation degree of the current data and the data of the same period in the past 24 hours: when the deviation is less than 20% of the historical mean, the historical stability index is 1.0; when the deviation is between 20% and 50% of the historical mean, the historical stability index is 0.5; and when the deviation is greater than 50% of the historical mean, the historical stability index is 0.0; For the case where the historical data is less than 24 hours at the initial stage of system operation, a gradual historical window mechanism is adopted, and the historical window length is gradually expanded from a minimum of 6 hours to 24 hours; Data sources in the multi-source original data At time Dynamic fusion weights ,like Figure 2 As shown, the calculation formula is: in, and Representing data sources respectively and data source At any moment The reliability index is calculated by weighted average of three dimensions: temporal consistency, physical rationality, and historical stability. In this embodiment, the weights are 0.3, 0.4, and 0.3, respectively. and Indicates data source and data source At different times The correlation moderating factor is obtained by calculating the average cross-correlation coefficient between this data source and other data sources; in this embodiment, the correlation moderating factor... The Pearson correlation coefficient is used for calculation, specifically: in , Indicates data source and Pearson correlation coefficient, For data source and Covariance over the past 100 sampling points and These are the corresponding standard deviations; when the correlation between data sources is extremely low, such as If the value is less than 0.1, a correlation measurement method based on information entropy is used as a supplement. The calculation process of the fused multidimensional feature vector includes three steps: data standardization, weight fusion, and feature dimensionality reduction. Data standardization uses the Z-score method to calculate the mean and standard deviation of each data source over the past 1000 sampling points, and then converts the original data at the current moment into standardized data to obtain 8 standardized data sources. The weight fusion step multiplies the 8 standardized data sources with their corresponding weights to form an 8-dimensional weighted vector. Feature dimensionality reduction uses principal component analysis to reduce the 8-dimensional weighted vector to 5 feature dimensions, forming the fused multidimensional feature vector.

[0029] A3: Establish a deep convolutional neural network model based on a multi-scale attention mechanism, perform temporal and spatial dual attention analysis on the fused multi-dimensional feature vectors, realize refined identification of the operating status of the ring main unit, and obtain the status identification results: A deep convolutional neural network model based on a multi-scale attention mechanism is established, which includes four parts of a multi-scale feature extraction layer, a time sequence attention layer, a spatial attention layer and a state classification layer; The multi-scale feature extraction layer receives a time sequence of a 5-dimensional fusion feature vector as input, and the input data dimension is 120x5, wherein 120 is the time step and 5 is the feature dimension; the multi-scale feature extraction layer adopts three parallel one-dimensional convolution branches, which respectively use convolution layers with convolution kernel sizes of 3, 5 and 7 to extract short-term, medium-term and long-term time sequence features along the time dimension, each branch contains 64 convolution kernels, the activation function uses the ReLU function, the step size is set to 1, and the padding mode is same padding to keep the output sequence length as 120; the outputs of the three parallel branches are fused through a channel concatenation operation to form a multi-scale feature map with a dimension of 120x192; The time sequence attention layer receives the 120x192-dimensional multi-scale feature map output by the multi-scale feature extraction layer as input, and identifies the key time points within the 120 time steps through a self-attention mechanism; the time sequence attention layer adopts a Query-Key-Value structure to calculate the attention weight, and maps the input 192-dimensional features to Query matrix, Key matrix and Value matrix through three independent linear transformation layers, and the feature dimension of each matrix is compressed to 64, obtaining three matrices with a dimension of 120x64; the attention weight is calculated by the dot product operation of the Query matrix and the Key matrix, and then normalized by the softmax function to obtain a 120x120 attention matrix, and then multiplied by the Value matrix to obtain a time sequence attention feature with a dimension of 120x64; The spatial attention layer receives the 120x64-dimensional time sequence attention feature output by the time sequence attention layer as input, first compresses the time sequence dimension through global average pooling to obtain a 64-dimensional global feature vector; then identifies the correlation between the 64 feature dimensions through a cross-attention mechanism, and calculates the correlation weight matrix between the features through a two-layer fully connected network, the first fully connected network maps the 64-dimensional feature to a 32-dimensional intermediate feature, and the second fully connected network maps the 32-dimensional intermediate feature back to 64-dimensional and obtains a feature weight vector through a sigmoid activation function; finally, the feature weight vector and the original 64-dimensional global feature are multiplied element by element to obtain a weighted 64-dimensional spatial attention feature; The state classification layer receives the 64-dimensional spatial attention feature vector output by the spatial attention layer as input, and uses a three-layer fully connected network for state classification. The first fully connected network maps the 64-dimensional input feature to a 128-dimensional hidden feature and passes it through a ReLU activation function. The second fully connected network compresses the 128-dimensional hidden feature to 64 dimensions and passes it through a ReLU activation function. The third fully connected network maps the 64-dimensional feature to a 4-dimensional output and outputs the probability distribution of the four operating states, including normal operation, slight abnormality, potential failure, and emergency failure, through a softmax function. The deep convolutional neural network model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training rounds. The loss function uses the cross-entropy loss function with an L2 regularization term to prevent overfitting, with a regularization coefficient of 0.0001. The ratio of training data to validation data is 8:2. The state recognition accuracy convergence process of the deep convolutional neural network model during training is shown in Figure 3 .

[0030] A4: Based on the state recognition result, a time series prediction model is constructed to analyze the operation trend of the ring network cabinet and predict potential failures, generating a hierarchical warning signal and a failure prediction report: Based on the probability of the operating state, a time series prediction model based on the long short-term memory network (LSTM network) is constructed to analyze the time series evolution law and trend change of the ring network cabinet operating state, and to predict the potential failure type and occurrence probability within the next 30 minutes. The LSTM network contains two hidden layers, each containing 128 LSTM units. A bidirectional LSTM structure is used to capture both forward and backward time series dependencies. The outputs of the forward and backward LSTMs are fused through concatenation. The time series prediction model input is a state probability sequence of the past 120 time steps, each containing 4 probability values for each state category. The time series prediction model output is a fault state probability prediction sequence for the next 30 time steps. The training process of the time series prediction model uses a sliding window mechanism with a window size of 120 time steps and a sliding step of 1 time step to construct training samples from historical data. The time series prediction model is trained using the mean square error loss function with the Adam optimizer, a learning rate of 0.0005, a training batch size of 64, and 200 training rounds. An early stopping mechanism is used to prevent overfitting, and training is stopped when the validation set loss does not decrease for 10 consecutive rounds. To prevent gradient explosion, gradient clipping is used with a maximum gradient value of 1.0. When encountering a sudden change in data distribution, the system enables an online learning mechanism to update the model parameters using an exponential weighted moving average. The calculation method of the fault risk assessment index is obtained by weighted summation of the predicted future 30 time step fault state probability sequence, and the weight coefficient is determined according to the severity of the fault type. In the embodiment, the weight of the slight anomaly is 0.3, the weight of the potential fault is 0.6, and the weight of the emergency fault is 1.0; when the fault risk assessment index is greater than 0.7, a red emergency warning is generated, when the fault risk assessment index is between 0.3 and 0.7, a yellow attention warning is generated, and when the fault risk assessment index is less than 0.3, a green normal state is maintained, to obtain a graded warning signal. It should be noted that, in order to avoid frequent changes of the warning signal, the system adopts a hysteresis judgment mechanism, and the rising threshold and the falling threshold are set to have a deviation of ±0.05, that is, the risk index needs to be greater than 0.35 for the green to yellow transition, and the risk index needs to be less than 0.25 for the yellow to green transition; at the same time, time persistence verification is introduced, and the warning level change needs to be maintained for 3 consecutive prediction periods to take effect, to improve the warning stability.

[0031] After the graded warning signal is generated, the system automatically generates a fault prediction report, and the report content includes the current device state, the predicted fault type, the predicted occurrence time, the fault probability, the influence range analysis and the recommended treatment measures. The report is stored in JSON format and pushed to the operation and maintenance management system through the HTTP interface, and the pushing frequency is determined according to the warning level, the red warning is pushed immediately, the yellow warning is pushed every 5 minutes, and the green state is pushed every hour.

[0032] A5: According to the warning signal, combined with the operating constraints of the ring network cabinet, an optimal control strategy is generated and corresponding control instructions are executed to realize intelligent monitoring and control of the ring network cabinet: Based on the graded warning signal, a control strategy generation mechanism is established; when receiving a green normal state signal, an optimal operation mode is adopted, when receiving a yellow attention warning signal, a preventive adjustment mode is adopted, and when receiving a red emergency warning signal, an emergency disposal mode is adopted; the optimal operation mode takes economy as the main target, and adjusts the controllable load distribution and switch state to optimize the system operation efficiency, and in the embodiment, the target function weight is set to safety 0.3, economy 0.5 and reliability 0.2; the preventive adjustment mode takes safety as the main target, and preferentially adjusts the abnormal risk device load and protection setting value, and in the embodiment, the target function weight is set to safety 0.5, economy 0.3 and reliability 0.2; the emergency disposal mode takes safety as the only target, and immediately executes emergency measures such as load shedding, switch tripping and protection action, and in the embodiment, the target function weight is set to safety 0.7, economy 0.1 and reliability 0.2; The control strategy comprises three levels of load regulation strategy, switch operation strategy and protection action strategy; the load regulation strategy optimizes the load distribution of the ring network cabinet by adjusting the switching state of controllable loads, the controllable loads comprising air conditioning loads, charging pile loads and industrial loads; the switch operation strategy changes the network topology by controlling the on-off state of circuit breakers and load switches in the ring network cabinet; and the protection action strategy ensures system safety by setting the action parameters of protection devices; The particle swarm optimization algorithm is used for control parameter optimization, while considering three objectives of system safety, economy and reliability; the safety objective function is a weighted sum of the square sum of voltage deviation and the square sum of current exceeding; the economy objective function is the total of load regulation cost and switch operation cost; and the reliability objective function is the reciprocal of system average outage time; in the embodiment, the particle number of the particle swarm optimization algorithm is set to 50, the maximum iteration number is set to 100, the inertia weight is linearly decreased from 0.9 to 0.4, and the learning factor and is set to 2.0; The constraint conditions of the control strategy comprise voltage constraint, current constraint, power constraint and device capacity constraint; the voltage constraint requires that the voltage of each node is kept between 95% and 105% of the rated voltage; the current constraint requires that the current of each branch is not more than 90% of the rated current; the power constraint requires that the total load power is not more than 80% of the total capacity of the system; and the device capacity constraint requires that the operating power of each device is not more than the rated capacity thereof; After the particle swarm optimization algorithm outputs the optimal control parameters, the optimal control strategy is generated and the corresponding control instructions are executed; the control instruction execution process comprises four links of instruction generation, safety check, execution confirmation and result feedback; the instruction generation module generates specific control instructions according to the optimal control parameter optimization results; the safety check module verifies whether the system state after instruction execution meets the safety constraint conditions of the control strategy through power flow calculation; the execution confirmation module sends the control instructions to the field devices and waits for the execution confirmation signal, and the timeout time is set to 5 seconds; and the result feedback module collects the instruction execution results and updates the system state.

[0033] Embodiment 2 The application further discloses a ring network cabinet monitoring and control system based on multi-source data, comprising the following five modules: A data acquisition module acquires real-time monitoring data of each device of the ring network cabinet through a multi-source sensor network, including electrical parameters, environmental parameters and device state parameters; A data fusion module constructs a self-adaptive weight fusion algorithm with a data quality evaluation mechanism, performs real-time reliability evaluation on multi-source original data, and forms multi-source original data; State recognition module: a deep convolutional neural network model based on multi-scale attention mechanism is established to perform time series and spatial double attention analysis on the fused multi-dimensional feature vector, and realize fine identification of the running state of the ring network cabinet. Early warning module: a time series prediction model is constructed based on the state recognition result, the running trend of the ring network cabinet is analyzed, potential faults are predicted, and a graded early warning signal and a fault prediction report are generated; Control module: according to the early warning signal, combined with the running constraint condition of the ring network cabinet, an optimal control strategy is generated and corresponding control instructions are executed to realize intelligent monitoring and control of the ring network cabinet.

[0034] It should be noted that the above-mentioned embodiment numbers of the application are only for description, and do not represent the advantages and disadvantages of the embodiments. And the terms "include", "contain" or any other variant in this paper are intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, device, article or method including the element.

[0035] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of software product, which is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc) as described above, including a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server or network device) execute the method described in each embodiment of the present application.

[0036] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method for monitoring and controlling ring main units based on multi-source data, characterized in that, Includes the following steps: A1: Real-time monitoring data of each device in the ring main unit is obtained by establishing a multi-source sensor network, including electrical parameters, environmental parameters and equipment status parameters, forming multi-source raw data; A2: Construct an adaptive weight fusion algorithm with a data quality assessment mechanism to perform real-time reliability assessment on multi-source raw data and output the fused multi-dimensional feature vector; A3: Establish a deep convolutional neural network model based on a multi-scale attention mechanism, perform temporal and spatial dual attention analysis on the fused multi-dimensional feature vectors, realize refined identification of the operating status of the ring main unit, and obtain the status identification results; A4: Based on the state recognition results, construct a time-series prediction model, analyze the ring main unit's operating trend and predict potential faults, and generate graded early warning signals and fault prediction reports; A5: Based on the graded early warning signals and combined with the operating constraints of the ring main unit, the optimal control strategy is generated and the corresponding control commands are executed to realize the intelligent monitoring and control of the ring main unit.

2. The method for monitoring and controlling ring main units based on multi-source data according to claim 1, characterized in that, Step A1 includes: A multi-source sensor network is established to connect circuit breakers, load switches, voltage transformers, current transformers, temperature and humidity sensors, and partial discharge sensors in the ring main unit. A unified data sampling frequency of 10 times per second is set to acquire real-time monitoring data of each device in the ring main unit. The monitoring data includes eight key parameters: three-phase voltage RMS value, three-phase current RMS value, active power, ambient temperature, ambient humidity, partial discharge quantity, circuit breaker position status, and load switch position status. The monitoring data is initially screened to remove abnormal values ​​such as three-phase voltage exceeding 200% of the rated value, three-phase current exceeding 300% of the rated value, temperature exceeding -40°C to 80°C, and humidity exceeding 0% to 100%. At the same time, the sensor communication status, signal quality, and calibration status are recorded as data quality identifiers to form multi-source raw data containing 8 key parameter data and corresponding data quality identifiers.

3. The method for monitoring and controlling ring main units based on multi-source data according to claim 2, characterized in that, Step A2 includes: An adaptive weight fusion algorithm with a data quality assessment mechanism is constructed, and the adaptive weight fusion algorithm is implemented using a data reliability assessment module; specifically: The data reliability assessment module calculates the real-time reliability index of each data source by analyzing three dimensions: temporal consistency, physical rationality, and historical stability of the data from each sensor. Temporal consistency is evaluated by calculating the standard deviation of the rate of change of the current data compared to the data of the previous 5 sampling points: when the standard deviation of the rate of change is less than 0.1, the temporal consistency index is 1.0; when the standard deviation of the rate of change is between 0.1 and 0.5, the temporal consistency index is 0.5; when the standard deviation of the rate of change is greater than 0.5, the temporal consistency index is 0.

0. Physical rationality is assessed by verifying whether the data conforms to the basic laws of the power system: when the three-phase voltage symmetry deviation is less than 5% and the active power is between 0 and the rated power of the equipment, the physical rationality index is 1.0; otherwise, it is 0.

0. Historical stability is assessed by calculating the degree of deviation between the current data and the data from the same period in the past 24 hours: when the deviation is less than 20% of the historical mean, the historical stability index is 1.0; when the deviation is between 20% and 50% of the historical mean, the historical stability index is 0.5; when the deviation is greater than 50% of the historical mean, the historical stability index is 0.

0. Data sources in multi-source raw data At any moment Dynamic fusion weights The calculation formula is: in, and Representing data sources respectively and data source At any moment Reliability indicators and Indicates data source and data source At different times Correlation regulators; The calculation process of the fused multidimensional feature vector includes three steps: data standardization, weight fusion, and feature dimensionality reduction. Data standardization uses the Z-score method to calculate the mean and standard deviation of each data source over the past 1000 sampling points, and then converts the original data at the current moment into standardized data to obtain 8 standardized data sources. The weight fusion step multiplies the 8 standardized data sources with their corresponding weights to form an 8-dimensional weighted vector. Feature dimensionality reduction uses principal component analysis to reduce the 8-dimensional weighted vector to 5 feature dimensions, forming the fused multidimensional feature vector.

4. The method for monitoring and controlling ring main units based on multi-source data according to claim 3, characterized in that, Step A3 includes: A deep convolutional neural network model based on a multi-scale attention mechanism is established. The deep convolutional neural network model includes four parts: a multi-scale feature extraction layer, a temporal attention layer, a spatial attention layer, and a state classification layer. The multi-scale feature extraction layer receives a temporal sequence of 5-dimensional fused feature vectors as input. The input data dimension is 120×5, where 120 is the time step and 5 is the feature dimension. The multi-scale feature extraction layer uses three parallel one-dimensional convolutional branches, which use convolutional layers with kernel sizes of 3, 5, and 7 to extract short-term, medium-term, and long-term temporal features along the time dimension, respectively. Each branch contains 64 convolutional kernels, the activation function is ReLU, the stride is set to 1, and the padding method is same padding to maintain the output sequence length of 120. The outputs of the three parallel branches are fused through channel concatenation to form a multi-scale feature map with a dimension of 120×192. The temporal attention layer receives a 120×192-dimensional multi-scale feature map output from the multi-scale feature extraction layer as input, and identifies key time points within 120 time steps through a self-attention mechanism. The temporal attention layer uses a Query-Key-Value structure to calculate attention weights, and maps the input 192-dimensional features into a Query matrix, a Key matrix, and a Value matrix through three independent linear transformation layers. The feature dimension of each matrix is ​​compressed to 64, resulting in three matrices with a dimension of 120×64. The attention weights are calculated by performing a dot product operation between the Query and Key matrices, followed by softmax normalization to obtain a 120×120 attention matrix, which is then multiplied by the Value matrix to obtain a temporal attention feature with a dimension of 120×64. The spatial attention layer receives the 120×64-dimensional temporal attention features output from the temporal attention layer as input. First, it compresses the temporal dimension through global average pooling to obtain a 64-dimensional global feature vector. Then, it identifies the correlation between the 64 feature dimensions through a cross-attention mechanism. A two-layer fully connected network is used to calculate the correlation weight matrix between the features. The first fully connected network maps the 64-dimensional features to 32-dimensional intermediate features, and the second fully connected network maps the 32-dimensional intermediate features back to 64 dimensions and obtains the feature weight vector through a sigmoid activation function. Finally, the feature weight vector is multiplied element-wise with the original 64-dimensional global features to obtain the weighted 64-dimensional spatial attention features. The state classification layer receives 64-dimensional spatial attention features from the spatial attention layer as input and uses a three-layer fully connected network for state classification. The first fully connected network maps the 64-dimensional input features to 128-dimensional hidden features and activates them using the ReLU function. The second fully connected network compresses the 128-dimensional hidden features to 64-dimensional features and activates them using the ReLU function. The third fully connected network maps the 64-dimensional features to 4-dimensional output and outputs the probability distribution of four operating states using the softmax function, including normal operation, minor anomaly, potential fault, and emergency fault.

5. The method for monitoring and controlling ring main units based on multi-source data according to claim 4, characterized in that, Step A4 includes: Based on the probability of operating states, a time-series prediction model incorporating an LSTM network is constructed to analyze the temporal evolution and trend changes of the ring main unit's operating states, predicting potential fault types and their probabilities within the next 30 minutes. The LSTM network contains two hidden layers, each with 128 LSTM units. A bidirectional LSTM structure is used to simultaneously capture forward and backward temporal dependencies. The outputs of the forward and backward LSTMs are fused through a concatenation operation. The input to the time-series prediction model is the state probability sequence over the past 120 time steps, with each time step containing probability values ​​for four state categories. The output of the time-series prediction model is the fault state probability prediction sequence for the next 30 time steps. A fault risk assessment index is obtained by weighted summation of the predicted fault state probability sequences over the next 30 time steps. Weighting coefficients are determined based on the severity of the fault type, which includes minor anomalies, potential faults, and emergency faults. A red emergency warning is generated when the fault risk assessment index is greater than 0.7; a yellow warning is generated when the fault risk assessment index is between 0.3 and 0.7; and a green normal state is maintained when the fault risk assessment index is less than 0.3, thus obtaining a graded warning signal. After a graded early warning signal is generated, the system automatically generates a fault prediction report. The report includes the current equipment status, predicted fault type, predicted occurrence time, fault probability, impact range analysis, and suggested handling measures. The report is stored in JSON format and pushed to the operation and maintenance management system via an HTTP interface. The push frequency is determined according to the early warning level: red warnings are pushed immediately, yellow warnings are pushed every 5 minutes, and green warnings are pushed every hour.

6. The method for monitoring and controlling ring main units based on multi-source data according to claim 5, characterized in that, Step A5 includes: A control strategy generation mechanism is established based on tiered early warning signals. When a green normal status signal is received, an optimized operation mode is adopted; when a yellow warning signal is received, a preventative adjustment mode is adopted; and when a red emergency warning signal is received, an emergency response mode is adopted. The optimized operation mode prioritizes economy, adjusting controllable load distribution and switch states to optimize system operating efficiency. The preventative adjustment mode prioritizes safety, prioritizing adjustments to equipment loads and protection settings with abnormal risks. The emergency response mode prioritizes safety, immediately implementing emergency measures such as load shedding, switch tripping, and protection actions. The control strategy comprises three levels: load regulation strategy, switch operation strategy, and protection action strategy. The load regulation strategy optimizes the load distribution of the ring main unit by adjusting the switching status of controllable loads, including air conditioning loads, charging pile loads, and industrial loads. The switch operation strategy changes the network topology by controlling the opening and closing status of circuit breakers and load switches within the ring main unit. The protection action strategy ensures system safety by setting the action parameters of protection devices. The particle swarm optimization algorithm is used to optimize the control parameters, while considering three objectives: system safety, economy, and reliability. The safety objective function is the weighted sum of the squares of voltage deviation and the squares of current exceedance. The economy objective function is the sum of load regulation cost and switching operation cost. The reliability objective function is the reciprocal of the system's average outage time. The constraints of the control strategy include voltage constraints, current constraints, power constraints, and equipment capacity constraints. Voltage constraints require that the voltage of each node be maintained between 95% and 105% of the rated voltage. Current constraints require that the current of each branch does not exceed 90% of the rated current. Power constraints require that the total load power does not exceed 80% of the total system capacity. Equipment capacity constraints require that the operating power of each device does not exceed its rated capacity. After the particle swarm optimization algorithm outputs the optimal control parameters, it generates the optimal control strategy and executes the corresponding control commands. The control command execution process includes four stages: command generation, safety verification, execution confirmation, and result feedback. The command generation stage generates specific control commands based on the optimal control parameters. The safety verification stage verifies whether the system state after command execution meets the constraints of the control strategy through power flow calculation. The execution confirmation stage sends control commands to field devices and waits for execution confirmation signals, with a timeout of 5 seconds. The result feedback stage collects the command execution results and updates the system state.

7. A system for monitoring and controlling ring main units based on multi-source data, characterized in that, include: Data acquisition module: Acquires real-time monitoring data of each device in the ring main unit by establishing a multi-source sensor network, including electrical parameters, environmental parameters and equipment status parameters; Data fusion module: Constructs an adaptive weighted fusion algorithm with a data quality assessment mechanism to perform real-time reliability assessment on multi-source raw data and form multi-source raw data; Status recognition module: Establish a deep convolutional neural network model based on multi-scale attention mechanism, perform temporal and spatial dual attention analysis on the fused multi-dimensional feature vector, and realize refined recognition of the operating status of the ring main unit; Early warning module: Based on the status recognition results, a time-series prediction model is constructed to analyze the operating trend of the ring main unit and predict potential faults, generating graded early warning signals and fault prediction reports; Control module: Based on the early warning signal and the operating constraints of the ring main unit, it generates the optimal control strategy and executes the corresponding control commands to realize intelligent monitoring and control of the ring main unit; To achieve the method for monitoring and controlling ring main units based on multi-source data as described in any one of claims 1-6.

Images