Low-voltage power distribution cabinet intelligent control system applied to intelligent power grid

Abstract

The invention belongs to the technical field of power system automation control, and particularly relates to a low-voltage power distribution cabinet intelligent control system applied to an intelligent power grid, which comprises a multi-dimensional sensing module, an edge calculation and decision center and an execution and communication interface module. A multi-time-scale adaptive control mechanism adopted by the system separates protection, optimization and reconstruction decisions in different periods, the problems of response lag and target conflict of a single control loop are effectively solved, and the quick response capability and the adaptive adjustment capability of the system in a dynamic load and distributed energy access scene are remarkably improved. And the operation and maintenance mode is converted into predictive maintenance from post-maintenance through a deep fusion equipment health state evaluation and prediction function, so that the risk of unplanned shutdown is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of power system automation control technology, specifically a low-voltage distribution cabinet intelligent control system applied to smart grids. Background Technology

[0002] In the field of power system automation control, the smart grid, as a core component of the modern energy system, aims to achieve high efficiency, reliability, and intelligence in grid operation by integrating advanced sensing, communication, and decision-making technologies. Low-voltage switchgear, as a key device connecting the transmission network and end users, directly impacts regional power supply quality and energy management efficiency through the performance of its control system.

[0003] Among these, the intelligent control system of the low-voltage distribution cabinet is a crucial technical component for achieving the aforementioned goals. This system performs multiple functions, including protection, control, and optimization, through real-time monitoring and analysis of electrical parameters, equipment status, and environmental information within the cabinet. Existing technologies generally employ protection mechanisms based on fixed thresholds and control strategies with preset logic, which are ill-suited to complex and ever-changing operating conditions.

[0004] Existing technologies exhibit significant delays in control response and insufficient adaptive adjustment capabilities when facing high-proportion distributed energy integration and dynamic load fluctuations, failing to achieve multi-objective collaborative optimization. Traditional systems lack in-depth perception of equipment health status and predictive maintenance functions, resulting in low accuracy of fault warnings and poor information exchange between subsystems, hindering unified decision-making. Furthermore, existing solutions have limited room for optimization in energy efficiency management and cannot dynamically adjust strategies based on real-time electricity prices and load demand.

[0005] Therefore, the present invention provides an intelligent control system for low-voltage distribution cabinets applied to smart grids. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: An intelligent control system for low-voltage distribution cabinets applied to smart grids, comprising a multi-dimensional sensing module, an edge computing and decision-making center, and an execution and communication interface module deployed on the distribution cabinet body. The multi-dimensional sensing module is responsible for continuously collecting electrical parameters, equipment mechanical status signals, and micro-environment data inside the distribution cabinet. The edge computing and decision-making center integrates a data fusion and feature extraction unit, a multi-timescale adaptive control unit, an equipment health status assessment and prediction unit, and an energy efficiency optimization decision-making unit. This center processes, analyzes, and generates decisions based on the raw data uploaded by the sensing module. The execution and communication interface module drives the switching elements inside the cabinet to perform corresponding actions according to the instructions issued by the decision-making center, and realizes data interaction with the upper-level power grid dispatching system and adjacent distribution units.

[0008] Furthermore, the multi-dimensional sensing module specifically includes a high-precision electrical sensor array, a vibration and temperature monitoring submodule, and a partial discharge detection unit. The high-precision electrical sensor array synchronously measures three-phase voltage, three-phase current, power factor, and 2nd to 13th harmonic components with a 1-millisecond sampling period. The vibration and temperature monitoring submodule captures mechanical vibration characteristics and temperature rise gradient data in real time through accelerometers attached to key circuit breaker operating mechanisms and busbar connection points, and distributed fiber optic temperature measurement points. The partial discharge detection unit uses an ultra-high frequency sensor array to monitor electromagnetic discharge signals caused by insulation defects within the cabinet, with a detection frequency band covering 300 MHz to 1500 MHz.

[0009] Furthermore, the data fusion and feature extraction unit performs timestamp alignment and normalization preprocessing on the multi-source heterogeneous sensing data. This unit uses Fast Fourier Transform based on a 256-point sliding window to analyze current and voltage waveforms, extracting specific harmonic amplitude, phase, and total harmonic distortion (THD) indices. Simultaneously, this unit performs four-level wavelet packet decomposition on the vibration acceleration signal, extracting the energy ratio of the third-level fourth sub-band and the fourth-level seventh sub-band as mechanical loosening characteristics, and extracting the energy entropy of the fourth-level second sub-band as an early indication of bearing wear. The absolute value of the gradient change within a 10-minute time window and the Pearson correlation coefficient with the load current are calculated for the temperature sequence. The amplitude distribution histogram, pulse repetition frequency moving average, and phase resolution characteristics of the discharge pulses within each power frequency cycle are statistically analyzed for the partial discharge signal. All extracted 28-dimensional feature vectors and 5-dimensional key original electrical parameters together constitute a 33-dimensional high-dimensional state feature set.

[0010] Furthermore, the multi-timescale adaptive control unit operates in three parallel decision cycles. The first decision cycle is millisecond-level fast protection. During this cycle, the unit continuously reads the instantaneous values ​​of the three-phase current at 50-microsecond intervals and compares them in real time with the instantaneous overcurrent threshold dynamically calculated based on the historical data of the effective load current over the most recent 100 milliseconds. This threshold is adaptively adjusted within the range of 1.2 to 1.8 times the rated current according to the load curve type of the day. Once an over-limit is detected, the highest priority trip command is immediately triggered. The second decision cycle is second-level optimized control. During this cycle, the unit is activated once every 500 milliseconds. Using a pre-trained deep neural network model as input, it predicts the trend of total active power, total reactive power, and power factor changes within the next 2 seconds using the state feature sequence of the previous 5 seconds. Based on the prediction results, it dynamically adjusts the switching strategy of the reactive power compensation device, with the control objective being to maintain the power factor above 0.95. The third decision cycle is a minute-level reconfiguration decision. During this cycle, the unit performs a comprehensive evaluation every 5 minutes and calculates the load rate variance of each outgoing circuit as an equilibrium index. When the variance exceeds the threshold of 0.15, the improved ant colony optimization algorithm is launched. This algorithm models the distribution network topology as a weighted graph, initializes 50 artificial ants for parallel search, sets the pheromone evaporation coefficient to 0.1, and sets the heuristic factor weight to 2.0. After a maximum of 100 iterations, it outputs the switching operation sequence that minimizes the load rate variance.

[0011] Furthermore, the equipment health status assessment and prediction unit constructs a comprehensive health index model that includes electrical stress factors, mechanical wear factors, and thermal aging factors. This unit utilizes the feature set output by the data fusion and feature extraction unit to recursively predict the future 7-day change trajectory of the contact resistance of critical circuit breakers and the remaining life of the insulation material through a long short-term memory network. When the predicted daily increase rate of contact resistance exceeds a baseline value of 5% or the predicted remaining life is less than a 30-day safety margin, the unit generates warning signals at three levels: attention level, early warning level, and emergency level. Based on the inflection point analysis of the prediction curve, it recommends maintenance time windows accurate to a 6-hour interval.

[0012] Furthermore, the energy efficiency optimization decision-making unit accesses real-time electricity price signals released by the grid side and local load forecast data. This unit establishes a mixed-integer programming model with 22 inequality constraints, aiming to minimize total electricity costs and maximize local distributed energy consumption. The model's decision variables include eight types of controllable load switching times, the 15-minute interval charging and discharging power setting value of the energy storage system, and three types of interruptible load contract execution flags. The unit employs a non-dominated sorting genetic algorithm with an elite retention strategy to solve this multi-objective optimization problem. The algorithm initializes the population size to 200, with a crossover probability of 0.85, a mutation probability of 0.02, and a maximum evolutionary generation of 500. It ultimately outputs a set of recommended operating schemes from the Pareto optimal frontier, and the edge computing and decision-making center selects the final execution strategy based on preset preference weights.

[0013] Furthermore, the various units within the edge computing and decision-making hub interact with each other via a shared memory data bus. The data fusion and feature extraction unit publishes a 33-dimensional high-dimensional state feature set to the designated address area of ​​the bus every 20 milliseconds. The multi-timescale adaptive control unit subscribes to all electrical and vibration / temperature features, the equipment health status assessment and prediction unit subscribes to vibration, temperature, partial discharge, and contact resistance-related features, and the energy efficiency optimization decision-making unit subscribes to load power, electricity price, and energy storage status-related features. The output instructions from each decision-making unit are aggregated to a central arbitrator, which scans the output queue every 10 milliseconds, performing conflict detection and priority sorting based on a predefined rule base. Protection instructions have a priority of 9, control instructions have a priority of 6, warning instructions have a priority of 3, and energy efficiency instructions have a priority of 1, ensuring that only the highest priority instruction is output in the event of a resource conflict.

[0014] Furthermore, the execution and communication interface module includes an intelligent circuit breaker drive circuit, a communication protocol conversion gateway, and a local human-machine interface. The intelligent circuit breaker drive circuit converts 24-volt digital commands into control pulses with a width of 50 to 200 milliseconds to drive specific circuit breaker tripping coils or energy storage motors. The communication protocol conversion gateway supports three industrial communication protocols: IEC61850 manufacturing message specification, Modbus TCP application protocol data unit, and DNP3 transmission protocol data unit, achieving second-level data synchronization with the station control layer system and cloud management platform. The local human-machine interface uses a 7-inch color LCD touchscreen, dynamically displaying real-time waveforms of three-phase voltage and current, power factor curves, key circuit breaker health status indices, and a list of the latest warning information.

[0015] Furthermore, the system possesses a continuously self-optimizing adaptive learning mechanism. The edge computing and decision-making hub incorporates an online learning engine that continuously collects feedback data on the system's actual operational performance, including load forecasting deviations, contact resistance measurement errors, and electricity cost differences. This engine triggers a model fine-tuning process every 24 hours, employing a stochastic gradient descent algorithm with a learning rate of 0.001 to optimize 128 key parameters, including the weights of fully connected layers in the deep neural network model, the forget gate bias in the long short-term memory network, and the crossover probability in the non-dominated sorting genetic algorithm. This allows the system to gradually adapt to the long-term evolution of the power distribution network.

[0016] Compared with existing technologies, the advantages and positive effects of this invention are as follows: This invention achieves comprehensive perception and proactive control of the operating status of low-voltage distribution cabinets by constructing a system architecture integrating multi-dimensional perception, edge intelligent decision-making, and collaborative execution. The multi-time-scale adaptive control mechanism adopted by the system separates protection, optimization, and reconfiguration decisions into different cycles, effectively solving the problems of response lag and target conflict in single control loops, and significantly improving the system's rapid response and adaptive adjustment capabilities under dynamic loads and distributed energy access scenarios. The deeply integrated equipment health status assessment and prediction function transforms the operation and maintenance mode from post-maintenance to predictive maintenance, greatly reducing the risk of unplanned outages. The energy efficiency decision-making unit, based on multi-objective optimization, fully utilizes real-time electricity price signals and load information to achieve a synergistic improvement in operational economy and energy utilization efficiency. The system's internal information interaction mode based on a shared bus and a central arbitration mechanism ensures consistency and execution of decisions, breaking down information barriers between traditional subsystems. Ultimately, through its highly integrated, intelligent, and adaptive characteristics, the system provides comprehensive technical support for the safe, efficient, and reliable operation of low-voltage switchgear in the context of smart grids. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the structure of the intelligent control system for low-voltage distribution cabinets applied to smart grids proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-timescale adaptive control unit in this invention; Figure 3 This is a logical flowchart of the equipment health status assessment and prediction unit in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the energy efficiency optimization decision unit in this invention; Figure 5 This is a schematic diagram of the information interaction framework of each unit within the edge computing and decision-making hub of the present invention, based on a shared memory data bus. Detailed Implementation

[0019] Example 1 To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] like Figure 1As shown in the embodiment of the present invention, an intelligent control system for a low-voltage distribution cabinet applied to a smart grid includes a multi-dimensional sensing module, an edge computing and decision-making center, and an execution and communication interface module. The multi-dimensional sensing module is physically deployed inside the distribution cabinet and is responsible for continuously collecting various raw data during the cabinet's operation. The edge computing and decision-making center, as the intelligent core of the system, performs in-depth processing and feature extraction on the data uploaded by the sensing module and generates control decisions. The execution and communication interface module is responsible for translating decision commands into specific physical actions and realizing information interaction with external systems.

[0021] Specifically, the implementation of the multi-dimensional sensing module comprises three key sub-modules: a high-precision electrical sensor array, a vibration and temperature monitoring sub-module, and a partial discharge detection unit. The high-precision electrical sensor array is composed of a combination of three-phase voltage transformers and three-phase current transformers, with a sampling period strictly set to 1 millisecond, simultaneously measuring three-phase voltage, three-phase current, power factor, and harmonic components up to the 13th harmonic. The vibration and temperature monitoring sub-module uses magnetic attachment to attach piezoelectric accelerometers to the rotating shaft sleeves of the key circuit breaker operating mechanism and the fastening bolts of the busbar connection points. Simultaneously, distributed fiber optic temperature sensors are arranged at the same locations to capture mechanical vibration acceleration signals and temperature change data in real time. The vibration signal sampling frequency is 10 kHz, and the temperature data sampling interval is 1 second. The partial discharge detection unit uses an ultra-high frequency sensor array installed near the busbar support insulators inside the cabinet. Its detection frequency band is 300 MHz to 1500 MHz, specifically designed to monitor partial discharge electromagnetic wave signals caused by insulation material aging, surface contamination, or internal bubbles.

[0022] The edge computing and decision-making hub is implemented using an industrial-grade ruggedized computing platform, which integrates a multi-core processor and a field-programmable gate array (FPGA), possessing powerful parallel computing capabilities. Internally, the hub is logically divided into four core functional units: a data fusion and feature extraction unit, a multi-timescale adaptive control unit, an equipment health status assessment and prediction unit, and an energy efficiency optimization decision-making unit. The data fusion and feature extraction unit first performs timestamp alignment processing on the multi-source heterogeneous data from the multi-dimensional sensing modules, unifying all data onto an absolute timeline based on the Global Positioning System (GPS) clock signal. Then, data standardization preprocessing is performed, mapping the original data of different dimensions to the 0 to 1 interval through linear transformation; In some embodiments, for electrical waveform data, the unit opens an independent data buffer, uses a sliding window with a length of 256 sampling points to perform a fast Fourier transform on the instantaneous value sequence of three-phase current and voltage, calculates the amplitude and phase of the fundamental wave and the 2nd to 13th harmonics, and calculates the total harmonic distortion index according to the international standard formula. In some embodiments, for vibration signals, the data fusion and feature extraction unit calls the wavelet packet analysis algorithm library to perform 4-layer wavelet packet decomposition on the time-domain signal collected by the accelerometer, extracts the energy ratio of the 3rd and 4th sub-frequency bands and the 4th and 7th sub-frequency bands as mechanical loosening features, and extracts the energy entropy of the 4th and 2nd sub-frequency bands as early bearing wear indication features. In some embodiments, for temperature sequences, the unit calculates the absolute value of the temperature change gradient between adjacent sampling points and uses the Pearson correlation coefficient algorithm to calculate the dynamic correlation coefficient between the temperature sequence and the corresponding loop load current sequence within a 10-minute time window. For partial discharge signals, the unit statistically analyzes the amplitude distribution histogram of discharge pulses within each power frequency cycle, calculates the moving average of the pulse repetition frequency, and analyzes the difference in the distribution density of discharge pulses in the two intervals of 0 to 90 degrees and 180 to 270 degrees of power frequency voltage phase. All extracted 28-dimensional feature vectors, together with the selected 5-dimensional key original electrical parameters, constitute a 33-dimensional high-dimensional state feature set, which is updated every 20 milliseconds.

[0023] Please refer to the attached document. Figure 2 The multi-timescale adaptive control unit implements three parallel decision-making threads at the software level, corresponding to control tasks at three different timescales: millisecond, second, and minute. The millisecond-level fast protection thread runs in the field-programmable gate array hardware logic. This thread continuously reads the instantaneous three-phase current values ​​uploaded by the high-precision electrical sensor array at a period of 50 microseconds and compares them with a dynamically updated instantaneous overcurrent threshold in real time. It should be noted that this instantaneous overcurrent threshold is not a fixed value, but rather calculated using historical load current effective values ​​over the most recent 100 milliseconds provided by the data fusion and feature extraction unit, combined with an adaptive coefficient. This adaptive coefficient dynamically adjusts within the range of 1.2 to 1.8 based on the load curve type of the day. Once any phase current instantaneous value is detected to exceed this dynamic threshold, the protection thread immediately generates a highest-priority trip command signal. This signal transmission path is entirely direct hardware connection, ensuring a response latency of less than 2 milliseconds. The second-level optimization control thread runs on an independent core of the processor. This thread is activated every 500 milliseconds, and its input is the latest high-dimensional state feature set released by the data fusion and feature extraction unit. It should be noted that this thread embeds a deep neural network model pre-trained on massive amounts of historical data. The model structure consists of an input layer with 33 nodes, three hidden layers with 64, 32, and 16 nodes respectively, and an output layer with 3 nodes. The model uses the state feature sequence of the previous 5 seconds as input to predict the changing trends of total active power, total reactive power, and power factor within the next 2 seconds. Based on the prediction results, this thread dynamically calculates the number of capacitor banks that need to be switched on and off for the reactive power compensation device. The control objective is to maintain the power factor above 0.95 while avoiding equipment lifespan reduction due to frequent switching. The minute-level reconfiguration decision thread performs a comprehensive evaluation every 5 minutes. This thread first calculates the real-time load rate of each outgoing circuit and then calculates the variance of the load rate of all circuits as a balance index. In some embodiments, when the variance exceeds a preset threshold of 0.15, it is determined to be a severely unbalanced state, and the thread immediately initiates the network reconstruction algorithm. The reconstruction algorithm employs an improved ant colony optimization method, which models the distribution network topology as a weighted graph, where nodes represent load points, edges represent tie switch states, and edge weights comprehensively consider the resistance loss of the transfer path, the limit on the number of switching operations, and load priority. The algorithm initializes 50 artificial ants for parallel search, with a pheromone evaporation coefficient set to 0.1 and a heuristic factor weight of 2.0. After a maximum of 100 iterations, the algorithm outputs a set of switching operation sequences that minimize the load rate variance, generating specific load transfer instructions.

[0024] Please refer to the attached document. Figure 3 The core function of the equipment health status assessment and prediction unit is to construct a comprehensive health index model for critical circuit breakers and insulation systems. This model comprehensively considers three major degradation mechanisms: electrical stress factor, mechanical wear factor, and thermal aging factor. The electrical stress factor is quantified by statistically analyzing historical data of inrush current peak values ​​and operating overvoltage amplitudes at the moment of circuit breaker opening and closing; the mechanical wear factor is quantified by analyzing the long-term changing trend of frequency band energy characteristics representing mechanical loosening and wear in vibration signals; and the thermal aging factor is quantified by the equivalent thermal aging time when the cumulative operating temperature exceeds the benchmark value by 80 degrees Celsius. This unit uses the feature set continuously output by the data fusion and feature extraction unit to train and run a long short-term memory network prediction model. The network input layer receives time series containing 15-dimensional features such as historical contact resistance values, vibration characteristics, and temperature gradients. The network structure includes two long short-term memory layers, each with 64 memory units, followed by a fully connected output layer. The network performs a prediction every 6 hours, outputting the predicted change trajectory of the contact resistance of the main contacts of the critical circuit breaker and the predicted evolution path of the relative permittivity loss tangent of the main busbar insulation material over the next 7 days. For example, when the predicted daily increase rate of contact resistance exceeds 5% of the baseline value or the predicted remaining life is less than a 30-day safety margin, the unit generates three levels of warning signals: Attention Level, Warning Level, and Emergency Level, and recommends a specific maintenance time window based on the inflection point analysis of the prediction curve, which is accurate to a 6-hour interval.

[0025] Please refer to the attached document. Figure 4The design objective of the energy efficiency optimization decision unit is to maximize operational economy and green energy consumption while ensuring power supply reliability. This unit receives real-time electricity price curves for 96 points from the power grid dispatch center and 24-hour load forecast data from the regional load forecasting center via a communication protocol conversion gateway. Internally, a mixed-integer programming model is established with the dual objectives of minimizing total electricity cost and maximizing local distributed energy consumption. Model decision variables include the switching times for eight types of controllable loads, the 15-minute interval charging and discharging power settings for the energy storage system, and the contract execution flags for three types of interruptible loads. Model constraints cover 22 inequalities, including distribution line capacity limits, upper and lower limits of the energy storage system's state of charge, minimum operating time for controllable loads, and maximum number of interruptions for interruptible loads. This multi-objective optimization problem is solved using a non-dominated sorting genetic algorithm with an elite retention strategy. The algorithm initializes the population to 200, with a crossover probability of 0.85, a mutation probability of 0.02, and a maximum of 500 generations. Each solution outputs a Pareto optimal front containing approximately 30 non-dominated solutions. The edge computing and decision-making hub has a built-in preference decision-making module. This module sets target weights based on the energy policy orientation of the area where the distribution cabinet is located. Typically, the economic weight is set to 0.6 and the green consumption weight is set to 0.4. Based on this, the final combination of energy efficiency optimization strategies is selected from the Pareto frontier.

[0026] Please refer to the attached document. Figure 5 Data interaction between functional units within the edge computing and decision-making hub is achieved through a high-efficiency shared memory data bus. This bus is physically composed of dual-port random access memory chips and logically employs a publish-subscribe communication model. The data fusion and feature extraction unit, as the primary data producer, encapsulates the latest generated 33-dimensional high-dimensional state feature set into a specific data structure and publishes it to a designated address area on the bus every 20 milliseconds. The multi-timescale adaptive control unit, the equipment health status assessment and prediction unit, and the energy efficiency optimization decision-making unit, as data consumers, subscribe to data subsets on the bus relevant to their respective functions. The multi-timescale adaptive control unit subscribes to all electrical and vibration / temperature characteristics; the equipment health status assessment and prediction unit subscribes to characteristics related to vibration, temperature, partial discharge, and contact resistance; and the energy efficiency optimization decision-making unit subscribes to characteristics related to load power, electricity price, and energy storage status. After each decision-making unit independently runs its core algorithm, the resulting output instructions, including protection action codes, control setpoint arrays, early warning information structures, and energy efficiency strategy matrices, are all written into a unified output queue on the bus. The system includes a central arbitrator module that scans the output queue every 10 milliseconds and prioritizes decisions based on a predefined conflict resolution rule base. The rule base defines priority as follows: protection commands have a priority of 9, control commands have a priority of 6, warning commands have a priority of 3, and energy efficiency commands have a priority of 1. When multiple commands exist simultaneously and conflict with resources, only the command with the highest priority is allowed to be output. The arbitrator also performs command security checks, such as automatically verifying the current state of the target switch and network topology constraints before issuing a load transfer command to prevent accidental operation.

[0027] In some embodiments, the specific hardware implementation of the execution and communication interface module comprises three components: an intelligent circuit breaker drive circuit, a communication protocol conversion gateway, and a local human-machine interface. The intelligent circuit breaker drive circuit employs opto-isolation and power amplification design, receiving 24V digital level commands from the central arbitrator. It generates control pulses to drive the specific circuit breaker's tripping coil or energy storage motor through an internal pulse width modulation circuit. The pulse width is precisely set within the range of 50 to 200 milliseconds depending on the circuit breaker model. The communication protocol conversion gateway uses a multi-core processor architecture and has three independent data processing channels, compatible with the IEC61850 manufacturing message specification, Modbus TCP application protocol data unit, and DNP3 transmission protocol data unit, respectively. The gateway performs protocol parsing, data mapping, communication encryption, and link redundancy switching functions, ensuring second-level data synchronization with the station control layer monitoring system and the cloud management platform. The local human-machine interface uses a 7-inch color LCD touchscreen. The main screen dynamically displays real-time waveforms of three-phase voltage and current, real-time power factor curves, key circuit breaker health status indices, energy storage system charge status, and a list of the latest warning information in different areas. The interface provides a three-level menu for querying historical event records for the past 30 days, modifying system operating parameter thresholds, and manually performing on / off remote control operations.

[0028] In some embodiments, the system possesses a continuously self-optimizing adaptive learning mechanism, implemented by an independent online learning engine within the edge computing and decision-making hub. This engine runs continuously in the background, its core task being to collect real-world feedback data on the system's actual operating performance. For the multi-timescale adaptive control unit, the engine records the sequence of deviations between each load prediction and the actual load; for the equipment health status assessment and prediction unit, the engine records the absolute error between the measured and predicted contact resistance values ​​after each actual maintenance; for the energy efficiency optimization decision-making unit, the engine compares the actual electricity expenditure after strategy execution with the predicted value. These error data are statistically processed to generate parameter adjustments for each algorithm model. The online learning engine triggers a model fine-tuning process every 24 hours, employing a stochastic gradient descent algorithm with momentum to slightly optimize 128 key parameters, including the weights of fully connected layers in the deep neural network model, the forget gate bias in the long short-term memory network, and the crossover probability in the non-dominated sorting genetic algorithm, at a learning rate of 0.001. This gradual parameter update enables the system to automatically adapt to changes in line impedance, load characteristics, and equipment aging trends that occur during long-term operation of the power distribution network, maintaining the accuracy and robustness of the decision model.

[0029] The software portion of the system described in this embodiment adopts a modular design, with the core algorithm encapsulated as a dynamic link library, and the main program scheduling framework separated from the hardware driver layer. After power-on, the system executes a complete self-test process, including sensor communication testing, memory data bus verification, model loading status checks for each decision unit, and actuator loop impedance measurement. After passing the self-test, the system enters normal operation, with each module operating in parallel according to a preset timing sequence. The system provides comprehensive fault log recording functionality; any abnormality in any module triggers detailed event logging, including an abnormality timestamp, abnormal code, relevant data snapshots, and automatically implemented recovery measures. All logs are cyclically stored in non-volatile memory, with a maximum retention period of one year. Through a highly integrated hardware platform and deeply optimized software algorithms, this system achieves accurate perception, intelligent decision-making, and rapid execution of the all-around status of low-voltage distribution cabinets, significantly improving the safety, economy, and reliability of the power distribution system in complex operating environments.

[0030] Example 2 This embodiment provides an alternative implementation scheme for a low-voltage distribution cabinet intelligent control system applied to smart grids. Its core feature is that the technical path of the multi-dimensional perception module and the specific algorithm unit in the edge computing and decision-making center has been optimized and upgraded to adapt to application scenarios with higher precision and more stringent real-time requirements.

[0031] In one example, within the multi-dimensional sensing module, the high-precision electrical sensor array is upgraded to a wideband measurement scheme employing a combination of Rogowski coils and capacitive voltage dividers. Its effective measurement bandwidth extends to 0.1 Hz to 2 kHz, enabling more accurate capture of voltage dips and swells caused by impact loads. The vibration and temperature monitoring submodule, while retaining the original piezoelectric accelerometer, adds a non-contact vibration measurement probe based on the microwave Doppler effect. This probe is installed inside the cabinet door and inverts the internal mechanical state by detecting frequency shifts caused by micro-vibrations on the cabinet surface, effectively avoiding insulation risks associated with sensor installation. The partial discharge detection unit connects a high-frequency current transformer in parallel with the ultra-high frequency sensor array. This transformer is connected in series with the grounding wire to detect the conducted current signal induced in the grounding wire by partial discharge pulses, achieving joint detection using electromagnetic wave and pulse current methods, thus improving the confidence level of insulation defect identification.

[0032] Specifically, in the data fusion and feature extraction unit of the edge computing and decision-making center, an improved harmonic analysis algorithm is introduced for processing electrical waveform data. Based on the traditional Fast Fourier Transform, this algorithm combines adaptive noise complete set empirical mode decomposition technology to preprocess non-stationary load current signals, effectively suppressing spectral leakage and improving the accuracy of harmonic amplitude measurement by approximately 15%. For example, for feature extraction of vibration signals, this embodiment uses an automatic feature extraction method based on a deep convolutional neural network to replace the original wavelet packet decomposition scheme. This network takes the original vibration acceleration time-domain signal as input and automatically learns and outputs a 128-dimensional deep feature vector through a stack of three convolutional layers and two pooling layers. This feature vector demonstrates better discriminative power than manually generated features in mechanical fault classification tasks.

[0033] In this embodiment, the second-level optimization control thread in the multi-timescale adaptive control unit has undergone algorithm enhancement. Its core prediction model has been upgraded from a deep neural network to a hybrid architecture combining a temporal convolutional network and a self-attention mechanism. The temporal convolutional network is responsible for capturing local dependency patterns in the load data, while the self-attention mechanism models long-term periodic characteristics. This hybrid model extends the load prediction time span from 2 seconds to 5 seconds and reduces the root mean square error of prediction by approximately 20%. Furthermore, the improved ant colony algorithm used in the minute-level reconstruction decision thread introduces a dynamic adjustment mechanism in the pheromone update rule. This mechanism adaptively adjusts the pheromone evaporation coefficient based on the real-time connectivity status of the network topology. A low evaporation coefficient of 0.05 is used to enhance convergence when running on radial networks, and a high evaporation coefficient of 0.15 is used to promote exploration when network rings are detected, enabling the algorithm to quickly find high-quality solutions even in complex network structures.

[0034] For example, the comprehensive health index model of the equipment health status assessment and prediction unit in this embodiment incorporates the fusion of a physical model and a data-driven model. In addition to the long short-term memory network predicting contact resistance and insulation life, the unit runs a physical degradation model based on electrothermal coupling differential equations in parallel. This model uses real-time current and ambient temperature as inputs to solve the electro-erosion equation for circuit breaker contact materials and the thermal aging equation for insulation materials. The prediction results from the data-driven model and the simulation results from the physical model are assimilated every 24 hours, and a Kalman filter algorithm is used to fuse them, generating the final health status trajectory prediction. The mean square error of this fused prediction is reduced by approximately 30% compared to the prediction from a single model.

[0035] In some alternative embodiments, the algorithm for solving the energy efficiency optimization decision unit is replaced by a multi-objective particle swarm optimization algorithm. This algorithm initializes 100 particles, and the position vector of each particle encodes the values ​​of all decision variables. The particle swarm optimization algorithm updates particle velocities by tracking the individual's historical best position and the swarm's global best position. It should be understood that, compared with the genetic algorithm, the particle swarm optimization algorithm exhibits a faster convergence speed when solving the mixed integer programming problem in this embodiment, reducing the average solution time from 35 seconds to 22 seconds, while improving the uniformity of the Pareto solution set distribution by about 18%.

[0036] In this embodiment, the communication protocol conversion gateway for the execution and communication interface module enhances support for time-sensitive networks. The gateway integrates a hardware timestamp unit, assigning microsecond-accurate timestamps to all cross-protocol transmissions of manufacturing specification messages and distributed network protocol messages, ensuring data synchronization accuracy with the control center within ±2 milliseconds. The local human-machine interface has been upgraded to a 10-inch high-brightness display, ensuring clear visibility even in strong light conditions. Augmented reality-assisted maintenance functionality has also been added, automatically overlaying real-time operating parameters and historical maintenance records of the device after the camera identifies the device's nameplate.

[0037] In this embodiment, the system's adaptive learning mechanism is extended to a federated learning architecture. The edge computing and decision-making center no longer relies solely on local data for model updates; instead, with data anonymization, it periodically exchanges model parameter updates with other similar systems within the region. A coordination server is responsible for aggregating parameter gradients from multiple sites, calculating global model updates, and then distributing them to each edge node. This distributed learning model allows the system to draw on broader operational experience, accelerating model convergence while strictly protecting the data privacy of each site. Through the upgrades of the aforementioned key technologies, the entire system, while maintaining all the functions of Embodiment 1, further enhances the breadth and depth of state awareness, the accuracy and efficiency of the decision-making algorithm, and the overall learning and adaptability of the system.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A low-voltage switchgear intelligent control system for smart grids, characterized in that, include: Multi-dimensional sensing module: Deployed in the main body of the power distribution cabinet, it is used to continuously collect electrical parameters, mechanical status signals of equipment and micro-environment data inside the power distribution cabinet; It includes a high-precision electrical sensor array, a vibration and temperature monitoring submodule, and a partial discharge detection unit. The high-precision electrical sensor array synchronously measures three-phase voltage, three-phase current, power factor, and 2nd to 13th harmonic components. The vibration and temperature monitoring submodule captures mechanical vibration characteristics and temperature rise gradient data in real time through accelerometers attached to key circuit breaker operating mechanisms and busbar connection points, and distributed fiber optic temperature measurement points. The partial discharge detection unit uses an ultra-high frequency sensor array to monitor electromagnetic discharge signals caused by insulation defects inside the cabinet. Edge computing and decision-making hub: integrates a data fusion and feature extraction unit, a multi-timescale adaptive control unit, an equipment health status assessment and prediction unit, and an energy efficiency optimization decision-making unit; used to process, analyze, and generate decisions from the raw data uploaded by the multi-dimensional sensing module; the data fusion and feature extraction unit performs timestamp alignment and normalization preprocessing on multi-source heterogeneous sensing data, uses fast Fourier transform to analyze current and voltage waveforms, extracts specific harmonic amplitude and phase and total harmonic distortion rate indices, performs multi-layer wavelet packet decomposition on vibration acceleration signals, extracts the energy ratio of the 3rd and 4th sub-bands and the 4th and 7th sub-bands as mechanical loosening features, extracts the energy entropy of the 4th and 2nd sub-bands as early bearing wear indication features, calculates the absolute value of the gradient change within the time window of the temperature sequence and the Pearson correlation coefficient with the load current, and statistically analyzes the amplitude distribution histogram, pulse repetition frequency moving average, and phase resolution features of the partial discharge signal within each power frequency cycle to form a high-dimensional state feature set; Execution and Communication Interface Module: Used to drive the switching elements in the cabinet to perform corresponding actions according to the instructions issued by the edge computing and decision-making center, and to realize data interaction with the upper-level power grid dispatching system and adjacent distribution units.

2. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 1, characterized in that, The multi-timescale adaptive control unit operates in multiple parallel decision cycles; The first decision cycle is a millisecond-level fast protection. During this cycle, the unit continuously reads the instantaneous value of the three-phase current at a period of 50 microseconds and compares it in real time with the instantaneous overcurrent threshold dynamically calculated based on the historical data of the effective value of the load current. This threshold is adaptively adjusted within the range of 1.2 to 1.8 times the rated current according to the load curve type of the day. Once an over-limit is detected, the highest priority trip command is immediately triggered. The second decision cycle is a second-level optimization control. In this cycle, the unit is activated once every 500 milliseconds. Using the state feature sequence of the previous 5 seconds as input, the pre-trained deep neural network model predicts the trend of total active power, total reactive power and power factor changes in the next 2 seconds. Based on the prediction results, the switching strategy of the reactive power compensation device is dynamically adjusted. The third decision cycle is a minute-level reconfiguration decision. During this cycle, the unit performs a comprehensive evaluation every 5 minutes and calculates the load rate variance of each outgoing circuit as an equilibrium index. When the variance exceeds the threshold of 0.15, the improved ant colony optimization algorithm is launched. This algorithm models the distribution network topology as a weighted graph, initializes 50 artificial ants for parallel search, sets the pheromone evaporation coefficient to 0.1, and sets the heuristic factor weight to 2.

0. After a maximum of 100 iterations, it outputs the switching operation sequence that minimizes the load rate variance.

3. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 2, characterized in that, The specific implementation process of the millisecond-level fast protection is as follows: The instantaneous values ​​of the three-phase current uploaded by the high-precision electrical sensor array are continuously read at a period of 50 microseconds, and compared in real time with the dynamically updated instantaneous overcurrent threshold. The instantaneous overcurrent threshold is calculated by combining the historical data of the effective load current in the most recent 100 milliseconds provided by the data fusion and feature extraction unit with an adaptive coefficient. The adaptive coefficient is dynamically adjusted in the range of 1.2 to 1.8 according to the load curve type of the day. Once the instantaneous value of any phase current is detected to exceed the dynamic threshold, the highest priority trip command signal is immediately generated.

4. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 2, characterized in that, The specific implementation process of the second-level optimization control is as follows: it is activated once every 500 milliseconds. The input is the latest high-dimensional state feature set released by the data fusion and feature extraction unit. The deep neural network model takes the state feature sequence of the previous 5 seconds as input to predict the change trend of total active power, total reactive power and power factor in the next 2 seconds. Based on the prediction results, the number of capacitor banks to be switched by the reactive power compensation device is dynamically calculated.

5. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 2, characterized in that, The specific implementation process of the minute-level reconfiguration decision is as follows: A comprehensive evaluation is performed every 5 minutes to calculate the real-time load rate of each outgoing circuit and the variance of the load rate of all circuits as a balance index. When the variance exceeds the preset threshold of 0.15, the improved ant colony optimization method is started to reconfigure the network. The distribution network topology is modeled as a weighted graph, where nodes represent load points and edges represent tie switch states. The edge weights comprehensively consider the resistance loss of the transfer path, the limit of the number of switching operations, and the load priority. By initializing 50 artificial ants for parallel search, the pheromone evaporation coefficient is set to 0.1, the heuristic factor weight is 2.0, and after a maximum of 100 iterations, the switching operation sequence that minimizes the load rate variance is output.

6. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 1, characterized in that, The equipment health status assessment and prediction unit constructs a comprehensive health index model that includes electrical stress factors, mechanical wear factors, and thermal aging factors. Utilizing the feature set output by the data fusion and feature extraction unit, it recursively predicts the change trajectory of the contact resistance of critical circuit breakers and the remaining life of the insulation material through a long short-term memory network. When the predicted daily growth rate of the contact resistance exceeds the benchmark value or the predicted remaining life is lower than the safety margin, the unit generates warning signals at three levels: attention level, early warning level, and emergency level. Based on the inflection point analysis of the prediction curve, it recommends maintenance time windows accurate to a 6-hour interval.

7. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 6, characterized in that, The equipment health status assessment and prediction unit performs a prediction every 6 hours, outputting the predicted change trajectory of the contact resistance of the main contacts of the critical circuit breaker and the predicted evolution path of the loss tangent of the relative permittivity of the main bus insulation material within the next 7 days. The warning signals include three levels: attention level, warning level and emergency level, and the maintenance time window is accurate to a 6-hour interval.

8. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 1, characterized in that, The energy efficiency optimization decision unit accesses the real-time electricity price signal released by the grid side and the load forecast data of the region. With the dual objectives of minimizing total electricity cost and maximizing local distributed energy consumption, it establishes a mixed integer programming model containing several inequality constraints. The model decision variables include the switching time points of multiple types of controllable loads, the set value of the interval charging and discharging power of the energy storage system, and the interruptible load contract execution flag. The unit uses a non-dominated sorting genetic algorithm with an elite retention strategy to solve the multi-objective optimization problem, outputs the Pareto recommended operation scheme, and the edge computing and decision center selects the execution strategy according to the preset preference weights.

9. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 1, characterized in that, The various units within the edge computing and decision-making hub interact with each other via a shared memory data bus. The data fusion and feature extraction unit publishes high-dimensional state feature sets to a designated address area on the bus. The multi-timescale adaptive control unit subscribes to all electrical and vibration-temperature features. The equipment health status assessment and prediction unit subscribes to vibration-temperature, partial discharge, and contact resistance-related features. The energy efficiency optimization decision-making unit subscribes to load power, electricity price, and energy storage status-related features. The output instructions from each decision-making unit are aggregated to the central arbitrator. This arbitrator scans the output queue and performs conflict detection and priority sorting based on a predefined rule base to ensure that only the highest priority instruction is output when there is a resource conflict.

10. The intelligent control system for low-voltage distribution cabinets applied to smart grids according to claim 1, characterized in that, The system has a continuous self-optimizing adaptive learning mechanism. The edge computing and decision center has an built-in online learning engine that continuously collects feedback data on the actual operation of the system, including load prediction deviation, contact resistance measurement error and electricity cost difference. The engine triggers a model fine-tuning process every 24 hours, using a stochastic gradient descent algorithm with momentum to optimize the parameters of the deep neural network model.

Images