A power distribution room remote control inspection and intelligent early warning method based on an internet of things

Abstract

The application discloses a power distribution room remote control inspection and intelligent early warning method based on an Internet of Things, which comprises the following steps: collecting environment and equipment data by arranging multiple types of sensors in a power distribution room, extracting features and detecting abnormalities by a lightweight artificial intelligence model after synchronization, filtering and normalization by an edge computing unit, uploading to the cloud through a 4G and LoRa dual-mode network, performing time series prediction in a long short-term memory network containing a physical constraint gate, dynamically optimizing the model structure in combination with a Beta-Bernoulli prior sparse learning, calculating the fault risk of each monitoring point to generate multi-level early warning and control instructions, and realizing intelligent inspection and remote control in cloud-edge collaboration. Through the fusion of edge lightweight intelligent analysis, cloud physical constraint gating and Bayesian sparse learning, the application realizes real-time sensing, time series prediction and multi-level intelligent early warning control of the operation state of the power distribution room.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system monitoring, and particularly relates to a power distribution room remote control inspection and intelligent early warning method based on Internet of Things. BACKGROUND

[0002] With the continuous expansion of the power distribution network scale and the rapid growth of power load, the power distribution room as an important node connecting the medium voltage distribution network and the end user, its operation state is directly related to the safety and reliability of power supply. The traditional monitoring and operation of the power distribution room mainly relies on manual inspection and periodic detection, which has problems such as long information collection period, lagging fault discovery, high risk of personnel operation, etc., and it is difficult to meet the requirements of real-time sensing and intelligent decision-making for the construction of new power system.

[0003] In the prior art, some power distribution rooms have introduced temperature and humidity sensors, gas detectors and video monitoring equipment, and realized basic remote data collection and state monitoring through the Internet of Things platform, but most systems still stay in the stage of data uploading and manual analysis, lacking intelligent fusion and prediction analysis capability of multi-source data. Especially in complex electromagnetic environment, sensors are easily disturbed, leading to data drift; there are blind areas in communication link, and the real-time performance and stability of data transmission cannot be guaranteed; at the same time, the protocols of monitoring equipment from different manufacturers are not unified, and the system compatibility is poor.

[0004] In the aspect of intelligent analysis, the existing schemes mostly use fixed structure deep learning models for state recognition or fault classification, but such models usually have large calculation amount and high energy consumption, which are not suitable for long-term operation on site of the power distribution room; on the other hand, the model parameters and structure cannot be adaptively adjusted according to the changes of equipment state or sensing nodes, resulting in inaccurate prediction results.

[0005] In addition, the current remote control and early warning mechanism is mainly based on threshold comparison or rule matching, which cannot dynamically adjust the threshold according to the changes of environment and operation conditions, and cannot form a closed-loop process of "monitoring-prediction-control-feedback". Due to the lack of collaborative mechanism of edge intelligence and cloud deep analysis, the system is slow in responding to sudden events or multi-point linkage faults, and it is difficult to realize active early warning and remote intervention in time.

[0006] Therefore, how to provide a power distribution room remote control inspection and intelligent early warning method based on Internet of Things is a problem to be solved by those skilled in the art. SUMMARY

[0007] One purpose of the present application is to propose a power distribution room remote control inspection and intelligent early warning method based on Internet of Things. The present application collects environmental and equipment operation data by arranging multiple types of sensors in the power distribution room, performs synchronization, filtering and feature extraction at the edge end using a lightweight artificial intelligence model based on the TinyML framework, transmits feature data to the cloud through a 4G and LoRa dual-mode communication network, introduces a physical constraint gating mechanism into the long short-term memory network in the cloud, and combines a Bayesian sparse self-organizing learning method to realize dynamic optimization of the model structure, thereby completing time series prediction, risk assessment and multi-level early warning control under the condition of meeting the energy conservation constraint, and finally forming a closed-loop operation system of cloud-edge collaboration to improve the real-time performance, prediction accuracy and operation reliability of power distribution room monitoring.

[0008] A power distribution room remote control inspection and intelligent early warning method based on Internet of Things according to an embodiment of the present application comprises the following steps:

[0009] Sensors are arranged in the power distribution room to collect environmental and equipment operation parameters to form an original data set.

[0010] The original data set is input into an edge computing unit for time synchronization, noise filtering and normalization processing, and a lightweight artificial intelligence model is used to extract features and perform initial screening of abnormalities on the processed data, and output a feature vector.

[0011] The feature vector is sent to a cloud server through a 4G and LoRa dual-mode fusion communication network, and the communication network dynamically switches transmission channels according to signal quality and task type, and outputs a feature data packet.

[0012] The feature data packet is input into a long short-term memory network model containing a physical constraint gating unit, and the gating unit dynamically modulates the update of cell state and hidden state, and outputs a time series feature prediction value.

[0013] The time series feature prediction value is taken as a training sample, a Beta-Bernoulli process prior is introduced in the input channel and gating connection of the model, and online variational inference is performed to dynamically determine the activation probability of the connection, and output the network structure and parameter set after sparse optimization.

[0014] According to the time series feature prediction value, the network structure and parameter set after sparse optimization output by the fusion model, the fault risk level of each monitoring point in the power distribution room is calculated, and when the risk value exceeds the dynamic threshold, a multi-level early warning signal and the corresponding control suggestion are output.

[0015] Control instructions are generated according to the control suggestions, and are issued to the edge control device through the communication network for execution, and the early warning event and control execution result are input into the cloud model as feedback data.

[0016] Further, the sensors include temperature and humidity sensors, gas detection sensors, current sensors, voltage sensors, partial discharge sensors, door magnetic sensors, water immersion sensors, and smoke sensors,

[0017] Each sensor is arranged at a transformer, a switch cabinet, a bus duct, and an environmental area in the power distribution room, for simultaneously collecting environmental parameters and equipment operation parameters and forming a multi-source time sequence original data set.

[0018] Further, the generation of the feature vector includes:

[0019] The data output by each type of sensor is time-stamped according to a unified sampling period to generate multi-dimensional time sequence data at equal time intervals;

[0020] The aligned data is sequentially subjected to moving average filtering and median filtering to remove abnormal mutation points and random noise;

[0021] The mean and standard deviation of each dimension of the filtered data are calculated, and the normalized data matrix is obtained by subtracting the mean and dividing by the standard deviation of each dimension of the data;

[0022] The data matrix is input into a lightweight artificial intelligence model built based on a TinyML framework, and the following operations are sequentially performed:

[0023] Multi-layer depth separable convolution operation is performed on the input data to extract local features;

[0024] The convolution output is input into a gated recurrent unit structure to extract time-dependent features;

[0025] The extracted features are compared with a set abnormality judgment threshold to identify abnormal samples;

[0026] The results after feature extraction and abnormality identification are combined into a feature vector, and the feature vector is output.

[0027] Further, the generation of the feature data packet includes:

[0028] The feature vector output by the edge computing unit is data packed and time labeled to form a feature data frame with a unique identification number;

[0029] The signal strength, network delay, and bandwidth parameters of the current communication environment are detected, and the detection results are compared with preset thresholds;

[0030] When the signal strength is detected to be higher than the threshold and the task type is a high-bandwidth transmission task, a 4G communication channel is selected for uplink transmission of the feature data frame;

[0031] When the signal strength is detected to be lower than the threshold value or the task type is a low-rate transmission task, a LoRa communication channel is selected for uplink transmission of the feature data frame;

[0032] Each transmitted data frame is sequentially numbered and acknowledged during transmission, and if no acknowledgement signal is received within a preset time, the corresponding data frame is automatically retransmitted;

[0033] After completing the transmission of all data frames, the data frames are reorganized into a feature data packet according to the sequence number.

[0034] Further, the generation step of the time sequence feature prediction value comprises:

[0035] The feature data packets transmitted through the communication network are arranged in timestamp order to form a continuous time sequence input sequence;

[0036] The time sequence input sequence is gradually input to the input layer of the long short-term memory network, and the node power imbalance, current overrun and voltage deviation of the power distribution room operation are simultaneously obtained at each time step;

[0037] According to the obtained power imbalance, current overrun and voltage deviation, the corresponding constraint coefficient is calculated, which is used as the input quantity of the constraint gate for correcting the gated output after the input gate, the forget gate and the output gate are calculated;

[0038] During the cell state update process, the cell state of the previous time step is weighted and adjusted by the constraint coefficient, and the adjusted cell state is input as the cell state of the current time step;

[0039] During the implicit state calculation process, the implicit state activation value of the current time step is weighted and corrected by the constraint coefficient, so that the update of the implicit state is simultaneously controlled by the time sequence input and the constraint coefficient;

[0040] After completing the recursive calculation of all time steps, the final implicit state output is output as the time sequence feature prediction value of the corresponding time sequence of the power distribution room.

[0041] Further, the generation of the network structure and the parameter set comprises:

[0042] The time sequence feature prediction values are divided into multiple training batches according to time sequence, and are input to the input channels of the long short-term memory network in turn;

[0043] An independent activation probability parameter is set for each input channel and gate connection, and a sparse prior distribution is assigned to each activation probability parameter based on the Beta-Bernoulli process;

[0044] In an initial training stage, the activation states of each connection are randomly initialized according to the prior distribution, and the connection index is recorded;

[0045] In each training iteration, the likelihood function value of each connection is calculated according to the time series feature prediction value of the current batch, and the posterior distribution of the activation probability is determined according to the joint of the likelihood function and the prior distribution;

[0046] The online variational inference method is used to update the parameters of the posterior distribution to obtain new activation probability values;

[0047] When the activation probability is lower than a preset threshold, the corresponding connection is closed and its weight is set to zero;

[0048] When the activation probability is higher than the threshold, the connection is maintained and the weight parameter is updated;

[0049] In a plurality of consecutive training iterations, when it is detected that the activation probability is stable between batches, the corresponding connection structure is fixed and the final activation set is recorded;

[0050] After all training batches are completed, the weight parameters, connection indexes and activation sets of the retained connections are combined into a sparsely optimized network structure and parameter set.

[0051] Further, the calculation step of the failure risk level comprises:

[0052] The sparsely optimized network structure and parameter set are imported into the risk calculation unit, and the time series feature prediction values corresponding to each monitoring point are read as input variables;

[0053] The change rates of temperature, current, voltage and gas concentration and other operating indicators are calculated on the prediction sequence of each monitoring point, and a state vector is generated according to the change rates;

[0054] The conditional probability distribution of the state vector is calculated according to the Bayesian inference method to obtain the failure occurrence probability corresponding to each operating indicator;

[0055] According to the failure occurrence probability of each operating indicator and its weight coefficient in the network structure, the comprehensive risk value of the monitoring point is obtained;

[0056] The comprehensive risk value is compared with a dynamic threshold, and when the comprehensive risk value is in different threshold intervals, it is respectively identified as a first, second and third risk state;

[0057] According to the level of the risk state, the corresponding warning signal and control suggestion are output, and the warning signal and control suggestion are input to the remote control step.

[0058] Further, the step of generating a control instruction according to the control suggestion comprises:

[0059] According to the control suggestion corresponding to the multi-level early warning signal, a matched control scheme is selected from a preset control strategy library;

[0060] The selected control scheme is parsed into an executable control instruction sequence, and each control instruction contains a control target, an execution parameter and an execution duration;

[0061] The control instruction sequence is encoded and time-stamped to generate a control data packet with a unique identification number;

[0062] The control data packet is sent to the edge control device through the current effective channel of the communication network, and the response detection and timeout retransmission operation are performed during the sending process;

[0063] After the edge control device receives the control data packet, the ventilation, cooling, circuit breaking or isolation operation is performed in sequence;

[0064] After the operation is completed, the execution result data and the corresponding early warning event identification are bound to form a feedback data packet, which is uploaded to the cloud model through the communication network for updating the model parameters and the control strategy.

[0065] The beneficial effects of the present application are:

[0066] The present application arranges temperature and humidity, gas, current, voltage, partial discharge, water immersion, smoke and door magnetic sensors in the power distribution room, establishes a multi-source synchronous acquisition mechanism, so that the environmental parameters and equipment operation parameters are collected and aggregated under a unified time reference, providing a high-precision data basis for subsequent intelligent analysis. The lightweight artificial intelligence model based on the TinyML framework is deployed in the edge computing unit to perform synchronization, filtering and normalization processing on the original data, realize feature extraction and abnormal preliminary screening, and complete real-time calculation under low power consumption conditions, significantly reducing communication bandwidth and cloud load.

[0067] Through the 4G and LoRa dual-mode fusion communication network, the transmission channel is dynamically selected according to the signal quality and task type, which can not only guarantee the high-speed upload of large bandwidth data, but also maintain the continuous transmission of low-rate data in weak signal or remote scenes, realizing stable data upload and control instruction download.

[0068] The application introduces a physical constraint gating unit in a cloud long short-term memory network, takes node power imbalance, current overrun and voltage deviation as constraint inputs, so that the model meets the energy conservation and power balance conditions of the power distribution system during the learning process, and fundamentally improves the stability and physical consistency of time series prediction. The Beta-Bernoulli process prior is introduced in the model input channel and gating connection, and online variational inference is performed to dynamically determine the connection activation probability, realize automatic pruning of the model structure and adaptive growth of new connections, so that the model can automatically adjust the structure with the change of the monitoring node and the difference of the working condition, has the ability of continuous learning and evolution.

[0069] Based on the time series prediction value and sparse structure parameter output by the fusion model, the fault probability distribution of each monitoring point is calculated, and the graded warning signal is generated through dynamic threshold determination to automatically match the control strategy and issue control instructions, realizing remote operation such as ventilation, cooling, circuit breaking or isolation. BRIEF DESCRIPTION OF DRAWINGS

[0070] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0071] Fig. 1 A flowchart of a power distribution room remote control inspection and intelligent early warning method based on the Internet of Things is provided for the application;

[0072] Fig. 2 A communication transmission flowchart of a power distribution room remote control inspection and intelligent early warning method based on the Internet of Things is provided for the application;

[0073] Fig. 3 A cloud model reasoning flowchart of a power distribution room remote control inspection and intelligent early warning method based on the Internet of Things is provided for the application. DETAILED DESCRIPTION

[0074] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0075] REFERENCE Figs. 1-3 A power distribution room remote control inspection and intelligent early warning method based on the Internet of Things, comprising the following steps:

[0076] Temperature and humidity sensors, gas detection sensors, current sensors, voltage sensors, partial discharge sensors, door magnets, water immersion and smoke sensors are arranged in the power distribution room to collect environmental and equipment operation parameters to form an original data set;

[0077] The original data set is input into an edge computing unit for time synchronization, noise filtering and normalization processing, and a lightweight artificial intelligence model is used to extract features and perform initial screening of abnormalities on the processed data, outputting a feature vector containing key feature quantities and abnormal markers;

[0078] The feature vector and abnormal markers are sent to a cloud server through a 4G and LoRa dual-mode fusion communication network, which dynamically switches transmission channels according to signal quality and task type, and outputs feature data packets;

[0079] The feature data packets are input into a long short-term memory network model containing a physical constraint gate unit, which uses node power imbalance, current overrun and voltage deviation as input parameters to dynamically modulate the update of cell state and hidden state, and outputs time series feature prediction values that satisfy the energy conservation constraint;

[0080] The time series feature prediction values are used as training samples, and a Beta-Bernoulli process prior is introduced into the input channel and gate connection of the model and online variational inference is performed to dynamically determine the activation probability of the connection, and output the network structure and parameter set after sparse optimization;

[0081] According to the time series feature prediction values and the network structure and parameter set after sparse optimization output by the fusion model, the fault risk level of each monitoring point in the power distribution room is calculated, and when the risk value exceeds the dynamic threshold, a multi-level early warning signal and the corresponding control suggestion are output;

[0082] Control instructions are generated according to the control suggestions and are issued to edge control devices through the communication network to perform ventilation, cooling, circuit breaking or isolation operations, and the warning event and control execution result are input as feedback data into the cloud model to update the sparse structure parameters and periodically issued to the edge lightweight model.

[0083] In this embodiment, the sensors include temperature and humidity sensors, gas detection sensors, current sensors, voltage sensors, partial discharge sensors, door magnetic sensors, water immersion sensors, and smoke sensors,

[0084] Each sensor is arranged in a transformer, switch cabinet, bus duct and environment area in the power distribution room for simultaneously collecting environmental parameters and equipment operating parameters and forming a multi-source time series original data set;

[0085] During data acquisition, the sampling frequencies of various sensors are uniformly set so that the collected data is recorded at the same time interval, and all sensor signals are aggregated to generate a complete original data frame after each sampling period ends.

[0086] In this embodiment, the generation of the feature vector includes:

[0087] The data output by various sensors is time-stamped according to a unified sampling period to align the time intervals of the multi-dimensional time-series data;

[0088] The aligned data is sequentially subjected to moving average filtering and median filtering, a fixed window length is set in the moving average process, and the median value is taken in the order of adjacent sampling points in the median filtering process to remove abnormal mutation points and random noise;

[0089] The mean and standard deviation of each dimension of the filtered data are calculated, and the normalized data matrix is obtained by subtracting the mean and dividing by the standard deviation of each dimension of the data;

[0090] The data matrix is input into a lightweight artificial intelligence model based on the TinyML framework, and the following operations are sequentially performed:

[0091] Multi-layer depth separable convolution operation is performed on the input data to extract local features, the convolution kernel size and step length of each layer of convolution are preset according to the type of sensor data, and the convolution result is passed to the next layer after being processed by an activation function;

[0092] The convolution output is input into a gated recurrent unit structure to extract time-related features, and the gated structure updates the hidden state in a time step recursive manner and maintains the time sequence continuity of the features;

[0093] The extracted features are compared with the set abnormality judgment threshold to identify abnormal samples, and when the feature value exceeds the threshold interval, it is marked as an abnormal sample, and when it does not exceed, it is marked as a normal sample;

[0094] The results after feature extraction and abnormality identification are combined into a feature vector, the feature vector is arranged in time sequence and attached with a sampling time stamp, and the feature vector is output.

[0095] In this embodiment, the generation of the feature data packet includes:

[0096] The feature vector output by the edge computing unit is data packed and time stamped to form a feature data frame with a unique identification number, each feature data frame contains a time stamp, a serial number and an integrity check code;

[0097] Detect the signal strength, network delay and bandwidth parameters of the current communication environment, and compare the detection results with the preset threshold;

[0098] When the signal strength is detected to be higher than the threshold and the task type is a high-bandwidth transmission task, a 4G communication channel is selected for the uplink transmission of the feature data frame, and the corresponding time stamp is recorded before each transmission for synchronization confirmation;

[0099] When the signal strength is detected to be lower than the threshold value or the task type is a low-rate transmission task, a LoRa communication channel is selected for uplink transmission of the feature data frame, and a fixed-interval time slot mode is adopted to sequentially transmit the data frame during transmission to reduce signal congestion;

[0100] During transmission, each transmitted data frame is sequentially numbered and acknowledged, and when no response signal from the cloud is received within a preset acknowledgement time, the unacknowledged data frame is automatically retransmitted until an acknowledgement is received or the maximum number of retransmissions is reached.

[0101] After completing the transmission of all data frames, the acknowledged data frames are reorganized in sequence number order, duplicate frames are removed, and integrity is verified, and the data frames are reorganized into feature data packets in sequence number.

[0102] In the embodiment, the generation step of the timing feature prediction value comprises:

[0103] The feature data packets transmitted through the communication network are arranged in timestamp order, and linear interpolation is performed on samples with inconsistent timestamp intervals during the arrangement process, and duplicate timestamp samples are processed to form a continuous timing input sequence;

[0104] The timing input sequence is gradually input to the input layer of the long short-term memory network, the corresponding input vector is read at each time step, and the hidden state and cell state of the previous time step are loaded, and the node power imbalance, current overrun and voltage deviation of the power distribution room are obtained at each time step.

[0105] According to the obtained power imbalance, current overrun and voltage deviation, the corresponding constraint coefficient is calculated, the constraint coefficient is used as the input quantity of the constraint gate, and the constraint coefficient is obtained by weighted summation of the three types of input quantities, which is used to modify the gated output after input gate, forget gate and output gate calculation;

[0106] During cell state updating, the cell state of the previous time step is multiplied by the input gate output of the current time step, the cell state of the previous time step is weighted and adjusted by the constraint coefficient, and then added to the candidate state of the current time step, and the adjusted cell state is input as the cell state of the current time step.

[0107] During the calculation of the hidden state, the cell state of the current time step is processed by an activation function and multiplied by the output gate result, and the hidden state activation value of the current time step is weighted and corrected by the constraint coefficient, so that the update of the hidden state is controlled by the timing input and the constraint coefficient.

[0108] After completing the recursive calculation of all time steps, the implicit state vector of the last recursive output is extracted, and the final implicit state output is taken as the time series feature prediction value of the corresponding time sequence of the power distribution room.

[0109] In the embodiment, the generation of the network structure and the parameter set comprises:

[0110] The time series feature prediction value is divided into multiple training batches in chronological order, and time segments of equal length are extracted in each training batch to ensure that the sample length of each batch is consistent and is input into the input channel of the long short-term memory network in turn;

[0111] An independent activation probability parameter is set for each input channel and gate connection, and when initialized, an activation probability between zero and one is randomly assigned to each connection, and a sparse prior distribution is assigned to each activation probability parameter based on the Beta-Bernoulli process;

[0112] In the initial stage of training, the connection state vector is generated by sampling from the prior distribution according to the activation probability parameter, the connection with a state value of one is kept enabled, the connection with a state value of zero is disabled, and the connection index is recorded;

[0113] In each training iteration, the likelihood function value of each connection is calculated according to the time series feature prediction value of the current batch, and the posterior distribution of the activation probability is determined according to the joint of the likelihood function and the prior distribution;

[0114] The online variational inference method is used to update the parameters of the posterior distribution, and the activation probability of each connection is recalculated after each update, and the new activation probability replaces the old value;

[0115] When the activation probability is lower than the preset threshold, the corresponding connection is closed and its weight is set to zero;

[0116] When the activation probability is higher than the threshold, the connection is kept and the weight parameter is updated, and the update process calculates the correction value of the weight according to the standard gradient descent algorithm;

[0117] In a plurality of consecutive training iterations, when it is detected that the activation probability of each connection changes by less than a set threshold between batches, it is determined that the activation state is stable, the corresponding connection structure is fixed, and the final activation set is recorded;

[0118] After all training batches are completed, the weight parameters, connection indexes and activation sets of all retained connections are reordered according to the connection number to form a sparsely optimized network structure and parameter set.

[0119] In the embodiment, the calculation step of the fault risk level comprises:

[0120] The network structure and parameter set optimized by sparseness are imported into the risk calculation unit, the time series characteristic prediction values corresponding to each monitoring point are read as input variables, and an input data table is established according to the monitoring point number sequence to ensure one-to-one correspondence between the characteristics and the network parameters;

[0121] The change rates of the operation indexes such as temperature, current, voltage and gas concentration are calculated on the prediction sequence of each monitoring point, the change rate sequence is obtained by the difference between the adjacent time step values divided by the time interval, and the state vector is generated according to the change rate;

[0122] The conditional probability distribution of the state vector is calculated according to the Bayesian inference method, the prior probability of each operation index is determined first in the calculation process, and then the posterior probability is updated according to the observation result of the prediction value to obtain the fault occurrence probability corresponding to each operation index;

[0123] According to the fault occurrence probability of each operation index and its weight coefficient in the network structure, the index probability is multiplied by the corresponding weight and summed to obtain the comprehensive risk value of the monitoring point;

[0124] The comprehensive risk value is compared with the dynamic threshold, the dynamic threshold is updated in real time according to the operating environment temperature, humidity and load condition, and when the comprehensive risk value is in different threshold intervals, it is respectively identified as first, second and third risk states;

[0125] According to the level of risk state, the corresponding warning signal and control suggestion are output, the first risk corresponds to the circuit breaking or isolation control, the second risk corresponds to the cooling or ventilation control, and the third risk corresponds to the state observation and data review, and the warning signal and control suggestion are input to the remote control step.

[0126] In the embodiment, the step of generating a control instruction according to the control suggestion comprises:

[0127] According to the control suggestion output by the risk calculation step, the control scheme corresponding to the risk level is searched in the preset control strategy table, the circuit breaking or isolation scheme is selected when the first risk is detected, the ventilation or cooling scheme is selected when the second risk is detected, and the state observation scheme is selected when the third risk is detected;

[0128] The selected control scheme is parsed into an executable control instruction sequence, each instruction contains target device number, control action, execution duration and safety confirmation mark, and a time stamp and a unique identification number are added to the instruction sequence after generation;

[0129] The connection with the edge control device is established through the communication network, the 4G or LoRa channel is selected according to the current communication quality to send the control instruction sequence, the instruction data is encoded and checked before sending, and the receiving end confirmation information is waited after sending;

[0130] When the edge control device receives the control instruction, it sequentially executes the corresponding operation according to the instruction content and records the start time and end time of the execution; if communication interruption or incomplete instruction occurs during the execution process, the edge device automatically retries and stores the execution log locally;

[0131] After the operation is completed, the execution log is bound to the corresponding early warning event number to generate a feedback data packet, and the feedback data includes the execution result, the actual duration and the change of the environmental parameters;

[0132] The feedback data packet is uploaded to the cloud model through the communication network, the cloud model updates the model parameters according to the feedback result, and records the latest risk event and control execution corresponding relationship for subsequent model retraining and strategy correction.

[0133] Embodiment 1:

[0134] In order to verify the feasibility of the application in implementation, the application is applied to a certain 110kV power distribution room intelligent operation and maintenance system to monitor and intelligently warn the equipment operating environment and electrical parameters in real time.

[0135] The power distribution room is provided with 8 temperature and humidity sensors, 4 SF6 gas concentration sensors, 12 current sensors, 8 voltage sensors, 4 partial discharge sensors, 2 door magnets and water immersion sensors, and 2 smoke sensors. All sensors synchronously collect data at a sampling frequency of 1Hz, the edge side uses an embedded TinyML model based on ESP32 for feature extraction, and the cloud side deploys a long short-term memory network containing physical constraint gating and a Bayesian sparse self-organizing optimization mechanism.

[0136] During the test, 604,800 groups of operation data were collected for 7 consecutive days. After the edge side executes the moving average (window=5) and median filtering (window=3) combined filtering, the data are standardized by Z-score, and the feature vector dimension is 32. The dynamic switching threshold of 4G and LoRa communication mode is set to RSSI=-85dBm. The cloud model sets the number of hidden units of the LSTM layer to 128, the constraint input includes node power imbalance ΔP, current overrun ΔI and voltage deviation ΔV; the Beta-Bernoulli prior parameter is set to α=1.0, β=0.5, and the pruning threshold is set to 0.25. The training uses an online variational inference algorithm, the batch size is 64 per iteration, and the learning rate is 0.001.

[0137] Table 1 Operation results

[0138]

[0139] According to the data in the table, the system detects that the node power imbalance ΔP rises to 8.4kW, the current exceeds ΔI to 15.8A, and the voltage deviation ΔV is 15.2V in the 12-16 hour interval. At this time, the risk value R calculated by the fusion model is 0.82, which exceeds the dynamic threshold 0.7, and the system determines a first-level warning and automatically issues a circuit breaking instruction to the corresponding feeder. After execution, the current and power quickly fall back to the safe range. This process verifies the real-time constraint adjustment function of the KCLSTM structure: as ΔP, ΔI and ΔV change, the gating coefficient dynamically adjusts the network output, so that the predicted risk is synchronized with the actual fault event.

[0140] The change of risk value from 4-8 hour interval to 8-12 hour interval (0.37→0.61) corresponds to the synchronous increase of equipment temperature and current fluctuation. The IBP sparse learning mechanism automatically closes the input channels with low contribution in iterative training, so that the number of model parameters is reduced from the initial 1.02×10 6 to 6.7×10 5 , and the calculation load is reduced by about 34%, while the prediction accuracy is improved by about 8.3%.

[0141] Further analysis of the matching between model prediction results and actual warning events: during the test period, there were 5 valid warning events (including 2 second-level, 2 second-level and 1 first-level), the system accurately identified 4 times, and misreported 1 time, with a correct rate of 80%, which can be stabilized at more than 85% after subsequent parameter adjustment. Compared with the traditional fixed threshold method (correct rate 62%), the model of the present application improves by about 23 percentage points under the same environment.

[0142] In addition, the cloud-edge collaborative feedback mechanism realizes adaptive optimization. After the edge executes the control instruction, the temperature recovery curve and the execution time are uploaded to the cloud, the cloud updates and reestimates the risk weight, and automatically selects the delayed circuit breaking strategy for similar abnormal events next time, avoiding overprotection. This verification shows that the constraint gating and sparse self-learning mechanism of the model can operate stably under real working conditions, and the algorithm calculation resources can meet the requirements of embedded deployment.

[0143] The test results verify the feasibility and stability of the present application in multi-source data collaborative processing, dynamic risk identification and automatic control closed loop, and prove that the method can realize accurate prediction and active warning control of the equipment operating state in the actual power distribution room scene.

[0144] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes within the technical scope disclosed by the present application according to the technical solutions and inventive concepts of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for remote control inspection and intelligent early warning of power distribution rooms based on the Internet of Things, characterized in that, Includes the following steps: Sensors are deployed in the power distribution room to collect environmental and equipment operating parameters, forming a raw dataset. The original dataset is input into the edge computing unit for time synchronization, noise filtering and normalization. A lightweight artificial intelligence model is used to extract features and perform initial anomaly screening on the processed data, and output feature vectors. The feature vector is sent to the cloud server through a 4G and LoRa dual-mode converged communication network. The communication network dynamically switches the transmission channel according to the signal quality and task type, and outputs feature data packets. The feature data packet is input into a long short-term memory network model containing a physical constraint gating unit. The gating unit dynamically modulates the updates of cell state and hidden state and outputs time-series feature prediction values. Using the predicted values ​​of the time series features as training samples, the Beta-Bernoulli process prior is introduced into the input channel and gated connections of the model, and online variational inference is performed to dynamically determine the activation probability of the connection, and output the sparsely optimized network structure and parameter set. Based on the time-series feature predictions output by the fusion model and the sparsely optimized network structure and parameter set, the fault risk level of each monitoring point in the power distribution room is calculated. When the risk value exceeds the dynamic threshold, multi-level early warning signals and corresponding control suggestions are output. Control commands are generated based on the control recommendations and sent to the edge control device for execution via the communication network. At the same time, the warning events and control execution results are input into the cloud model as feedback data. The steps for generating the predicted time-series feature values ​​include: The characteristic data packets transmitted through the communication network are arranged in timestamp order to form a continuous time-series input sequence; The time-series input sequence is gradually input into the input layer of the long short-term memory network, and at each time step, the node power imbalance, current over-limit and voltage deviation of the power distribution room operation are simultaneously obtained. Based on the obtained power imbalance, current over-limit, and voltage deviation, the corresponding constraint coefficients are calculated. These constraint coefficients serve as the input to the constraint gate and are used to correct the gated output after calculation by the input gate, forget gate, and output gate. During the cell state update process, the cell state of the previous time step is weighted and adjusted using the constraint coefficient, and the adjusted cell state is used as the cell state input for the current time step. During the hidden state calculation process, the activation value of the hidden state at the current time step is weighted and corrected using the constraint coefficients, so that the update of the hidden state is controlled by both the time-series input and the constraint coefficients. After completing the recursive calculations for all time steps, the final implicit state output is used as the predicted time series characteristic value of the corresponding time series of the power distribution room. The generation of the network structure and parameter set includes: The predicted values ​​of temporal features are divided into multiple training batches according to time order and then sequentially input into the input channel of the Long Short-Term Memory Network. Set independent activation probability parameters for each input channel and gating connection, and assign sparse prior distributions to each activation probability parameter based on the Beta–Bernoulli process; In the initial training phase, the activation state of each connection is randomly initialized according to the prior distribution, and the connection index is recorded. In each training iteration, the likelihood function value of each connection is calculated based on the temporal feature prediction value of the current batch, and the posterior distribution of the activation probability is determined by combining the likelihood function with the prior distribution. The posterior distribution is updated using an online variational inference method to obtain new activation probability values; When the activation probability is lower than a preset threshold, close the corresponding connection and reset its weight to zero. When the activation probability is higher than the threshold, the connection is maintained and the weight parameters are updated; In several consecutive training iterations, once the activation probability is detected to be stable across multiple batches, the corresponding connection structure is fixed and the final activation set is recorded. After all training batches are completed, the weight parameters of the retained connections, connection indices, and activation sets are combined into a sparsely optimized network structure and parameter set.

2. The method for remote control inspection and intelligent early warning of power distribution rooms based on the Internet of Things as described in claim 1, characterized in that, The sensors include temperature and humidity sensors, gas detection sensors, current sensors, voltage sensors, partial discharge sensors, door magnetic sensors, water immersion sensors, and smoke sensors. Each sensor is installed in the transformer, switch cabinet, busbar trunking, and environmental area of ​​the power distribution room to simultaneously collect environmental parameters and equipment operating parameters and form a multi-source time-series raw dataset.

3. The method for remote control inspection and intelligent early warning of power distribution rooms based on the Internet of Things as described in claim 1, characterized in that, The generation of the feature vector includes: The data output from various sensors are timestamped according to a unified sampling period to generate multidimensional time-series data with equal time intervals. The aligned data is then subjected to moving average filtering and median filtering in sequence to remove abnormal abrupt changes and random noise. The mean and standard deviation of each dimension are calculated for the filtered data, and the data is standardized by subtracting the mean from each dimension and then dividing by the standard deviation to obtain the normalized data matrix. Input the data matrix into a lightweight artificial intelligence model built on the TinyML framework, and perform the following operations in sequence: Multi-layer depthwise separable convolution operations are performed on the input data to extract local features; The convolutional output is input into a gated recurrent unit structure to extract time-related features; The extracted features are compared with the set anomaly detection threshold to identify abnormal samples; The results of feature extraction and anomaly identification are combined to form a feature vector, and the feature vector is output.

4. The method for remote control inspection and intelligent early warning of power distribution rooms based on the Internet of Things according to claim 1, characterized in that, The generation of the feature data packet includes: The feature vectors output by the edge computing unit are packaged and time-stamped to form a feature data frame with a unique identifier. Detect the signal strength, network latency, and bandwidth parameters of the current communication environment, and compare the detection results with preset thresholds; When the detected signal strength is higher than the threshold and the task type is a high-bandwidth transmission task, the 4G communication channel is selected for uplink transmission of the feature data frame. When the signal strength is detected to be lower than the threshold or the task type is a low-rate transmission task, the LoRa communication channel is selected for uplink transmission of the characteristic data frame. During transmission, each sent data frame is sequenced and acknowledged. If no acknowledgment signal is received within a preset time, the corresponding data frame is automatically retransmitted. After all data frames have been transmitted, the data frames are reassembled into feature data packets according to their sequence numbers.

5. The method for remote control inspection and intelligent early warning of power distribution rooms based on the Internet of Things according to claim 1, characterized in that, The calculation steps for the fault risk level include: The sparsely optimized network structure and parameter set are imported into the risk calculation unit, and the time series feature prediction values ​​corresponding to each monitoring point are read as input variables. The rate of change of operating indicators such as temperature, current, voltage and gas concentration is calculated on the predicted sequence of each monitoring point, and a state vector is generated based on the rate of change. The conditional probability distribution of the state vector is calculated using the Bayesian inference method to obtain the probability of failure for each operating indicator. The comprehensive risk value of the monitoring point is obtained by calculating the failure probability of each operational indicator and its weight coefficient in the network structure. The comprehensive risk value is compared with the dynamic threshold. When the comprehensive risk value is in different threshold ranges, it is identified as a level 1, level 2, and level 3 risk status, respectively. Based on the level of risk, corresponding warning signals and control suggestions are output, and these warning signals and control suggestions are input into the remote control process.

6. The method for remote control inspection and intelligent early warning of power distribution rooms based on the Internet of Things according to claim 1, characterized in that, The step of generating control commands based on the control recommendations includes: Based on the control recommendations corresponding to the multi-level early warning signals, a matching control scheme is selected from the preset control strategy library; The selected control scheme is parsed into an executable sequence of control instructions, each of which includes the control objective, execution parameters, and execution duration. The control command sequence is encoded and time-stamped to generate a control data packet with a unique identifier; The control data packet is sent to the edge control device through the current valid channel of the communication network, and response detection and timeout retransmission are performed during the transmission process; After receiving the control data packet, the edge control device performs ventilation, cooling, circuit breaking, or isolation operations in sequence; After the operation is completed, the execution result data is bound with the corresponding early warning event identifier to form a feedback data packet, which is then uploaded to the cloud model through the communication network for updating model parameters and control strategies.

Images