Intelligent power regulation system and method for power distribution cabinet

Abstract

The invention discloses an intelligent power regulation system and method for a power distribution cabinet, and relates to the technical field of power regulation of the power distribution cabinet, and the method comprises the steps that a multi-source sensing unit carries out the early warning of harmonic overrun and local overheating risks in advance according to the collected three-phase current, three-phase voltage and harmonic components, and a temperature distribution matrix and external data generated by a thermal infrared imager; the edge prediction unit operates the LSTM-CNN-GRU hybrid neural network model to generate a future load prediction value, and stabilizes the voltage in advance; and the impedance self-adaptive execution unit enables the vibration value of the motor to be within a safety threshold value through the capacitor bank with the reconfigurable topology, the threshold value trigger type active filter and the tap switch controller according to the received optimization instruction.

Description

Technical Field

[0001] This invention relates to the field of power regulation technology for distribution cabinets, specifically to an intelligent power regulation system and method for distribution cabinets. Background Technology

[0002] Current power regulation technologies for distribution cabinets mostly employ fixed threshold protection mechanisms and static compensation strategies, which have the following significant drawbacks in industrial field applications: 1. When 200 spot welding machines on the production line are started, the total harmonic distortion (THD) of the current instantaneously exceeds 25%. Traditional solutions pose a risk of harmonic overheating in the welding line, resulting in blind spots in harmonic monitoring and a lack of overheating warnings.

[0003] 2. During die replacement on a stamping machine, the voltage load fluctuation drops sharply from 100% to 15%. Traditional static compensation systems are prone to triggering a chain reaction of voltage runaway and inaccurate regulation. Therefore, it is essential to design a smart power regulation system and method for distribution cabinets with a safer, more economical, and more efficient dynamic control strategy. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent power regulation system and method for power distribution cabinets to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: including a multi-source sensing unit, an edge prediction unit, a cloud optimization unit, an impedance adaptive execution unit, and a cooperative communication module; The multi-source sensing unit transmits the collected three-phase current, three-phase voltage, harmonic components, temperature distribution matrix generated by the infrared thermal imager, and external data collected via industrial Ethernet access to the edge prediction unit through the collaborative communication module. The edge prediction unit runs an LSTM-CNN-GRU hybrid neural network model. It extracts temperature spatial features from the received data through the CNN layer of the convolutional kernel and fuses them with the LSTM temporal features. The GRU layer generates future load prediction values, which are then transmitted to the cloud optimization unit via the collaborative communication module. The cloud-based optimization unit generates optimization instructions based on the received load forecast, combined with the electricity price signal and temperature gradient, and transmits them to the impedance adaptive execution unit via the collaborative communication module. The impedance adaptive execution unit performs capacitive reactance adjustment, harmonic compensation, and gear adjustment through a reconfigurable topology capacitor bank, threshold-triggered APF, and tap changer controller according to the received optimization instructions, and feeds back the execution status to the edge prediction unit and cloud optimization unit through the collaborative communication module. The collaborative communication module uses the MQTT protocol to establish a dual-channel redundant link between the edge prediction unit and the cloud optimization unit, and connects to the impedance adaptive execution unit via a CAN bus.

[0006] According to the above technical solution, the infrared thermal imager array is installed in a grid layout on the top plate of the distribution cabinet facing the busbar and circuit breaker contact area, generating a temperature distribution matrix including spatial coordinates and corresponding temperature values, and analyzing temperature anomalies in the matrix in real time.

[0007] According to the above technical solution, the reconfigurable topology capacitor bank includes at least three sets of capacitors with different capacitance values. The capacitors are deployed in the reactive power compensation compartment and directly connected to the busbar via copper busbars. A parallel switchable network is formed by an IGBT switch matrix, which supports the capacitive reactance value to be adjusted within a specified range in steps of 0.625%. This is achieved through two parallel combinations of the three sets of capacitors, where n is the number of IGBT switches.

[0008] According to the above technical solution, the threshold-triggered APF has a built-in FFT spectrum analysis module installed upstream of the nonlinear load. Its FFT module uses 128-point real-time spectrum analysis and automatically activates compensation when the total harmonic distortion rate exceeds the preset value.

[0009] According to the above technical solution, the cloud optimization unit generates optimization instructions through the dynamic rule engine of the PPO algorithm. The dynamic rule engine is updated in real time through reinforcement learning. Combined with the optimization instructions generated by the PPO algorithm, the compensation logic is adjusted. The policy network adopts an Actor-Critic architecture, and the state space includes: load forecast value, electricity price signal, and temperature gradient. The electricity price signals are the electricity prices for peak, flat, and valley periods.

[0010] According to the above technical solution, the LSTM-CNN-GRU hybrid neural network model is deployed on the FPGA chip of the edge computing box for load prediction, including: The LSTM layer is used to output time-series features, which refer to power data collected in chronological order. The CNN layer with convolutional kernels is used to extract spatial features from the temperature matrix, where spatial features refer to the collected temperature inside the cabinet. The feature fusion layer is used to concatenate temporal and spatial features into a feature vector through a fully connected layer. The GRU layer is used to output the load forecast for the next 5 minutes.

[0011] According to the above technical solution, it also includes a safety protection module for early warning of harmonic over-limit and local overheating risks; when the multi-source sensing unit detects that the local temperature or total harmonic distortion rate in the cabinet exceeds the preset value, the safety protection module triggers a signal that is directly transmitted to the impedance adaptive execution unit via the collaborative communication module, skipping the cloud optimization process to perform emergency actions, and at the same time pushes alarm information. The content directly triggered by the safety protection module skipping the cloud optimization process includes: complete disconnection of capacitor bank, full power compensation of APF, and reset of tap changer controller to the initial position.

[0012] According to the above technical solution, the intelligent power regulation method for power distribution cabinets, which employs an intelligent power regulation system for power distribution cabinets, further includes the following steps: A. Real-time acquisition of multi-source data, including three-phase current, three-phase voltage, harmonic components, temperature distribution matrix generated by infrared thermal imager, and external data acquired via industrial Ethernet access; B. Edge-side load prediction: Run the LSTM-CNN-GRU hybrid neural network model. Extract the temperature spatial features from the data collected in step A through the CNN layer of the convolutional kernel and fuse them with the LSTM temporal features. The GRU layer generates the future load prediction value. C. Cloud-based dynamic optimization decision-making: Based on the received load forecast, combined with electricity price signals and temperature gradients, optimization instructions are generated. D. Impedance adaptive execution: Based on the received optimization instructions, the capacitor bank with reconfigurable topology, threshold-triggered APF and tap changer controller perform capacitive reactance adjustment, harmonic compensation and tap adjustment to achieve dynamic adaptation of reactive power compensation and harmonic suppression. E. Safety protection mechanism: When the local temperature or total harmonic distortion rate inside the cabinet exceeds the preset value, the safety protection mechanism is triggered, skipping the cloud dynamic optimization decision-making steps and executing emergency actions. At the same time, alarm information is pushed. The emergency actions include: complete disconnection of capacitor bank, full power compensation of APF, and resetting of tap changer controller to the initial position.

[0013] According to the above technical solution, step B includes: B1. Data Preprocessing: Normalize the historical load series, temperature matrix, and event labels. B2. Feature Extraction: LSTM layers extract temporal features, and convolutional kernel CNN layers extract spatial features. B3. Feature Fusion: Concatenate feature vectors and input them into the GRU layer; B4. Output: Load change curve for the next 5 minutes.

[0014] According to the above technical solution, step C includes: C1. State Space Construction: Input load forecast P(t), electricity price signal, and temperature gradient through multi-source sensing units. ; Solving using the C2.PPO algorithm: The action space output by the Actor network includes: target capacitance. APF trigger threshold Tap switch K, in The target capacitive reactance value optimized for the PPO algorithm is the ideal capacitive reactance of the capacitor bank dynamically calculated through reinforcement learning, which is used for precise reactive power compensation. This is the APF activation threshold. When the real-time harmonic distortion rate exceeds this preset value, the active filter APF will automatically activate compensation. K is the tap position, an integer variable, which controls the position of the transformer tap to adjust the voltage. The Critic network evaluates the value of actions and optimizes the objective function through policy updates. : ,in These are the weighting coefficients; C3. Generate an optimized instruction set.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: This invention, by setting up a multi-source sensing unit, actively mitigates the risk of harmonic overheating when the THD suddenly rises to over 25% due to the simultaneous start-up of a group of welding machines. This invention, by incorporating an edge prediction unit, suppresses voltage runaway caused by transient conditions during die changing of a stamping press where the load drops abruptly from over 80% of the rated load to 10%-15% within 1-3 seconds. This invention optimizes peak-valley scheduling of photovoltaic power plants by setting up a cloud-based optimization unit, and automatically adjusts during off-peak periods using algorithms to reduce the amount of capacitors required. This invention optimizes and improves system reliability by incorporating an impedance adaptive execution unit, which allows multiple resistance welding machines connected in parallel to the same power distribution network in the load cabinet to inject a large number of harmonics into the power grid during operation. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a diagram of the internal structure of the power distribution cabinet of the present invention; Figure 3 This is a diagram of the neural network model structure of the present invention; Figure 4 This is the logic diagram of the PPO algorithm of the present invention; Figure 5 This is a flowchart of the execution unit action of the present invention; Figure 6 This is a diagram of the safety protection module of the present invention; Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1-6 The present invention provides a technical solution: an intelligent power regulation system for a power distribution cabinet, comprising a multi-source sensing unit, an edge prediction unit, a cloud optimization unit, an impedance adaptive execution unit, a collaborative communication module, and a safety protection module; In applications where welding line distribution cabinets supply power to welding machine groups, traditional monitoring and control methods have significant technical limitations: Under normal operating conditions, while current sensors can collect electrical parameters and calculations can maintain the total harmonic distortion (THD) within a safe threshold, ensuring cable joint temperatures are ≤85℃, and capacitor banks can dynamically compensate for reactive power, while active power filters (APFs) suppress harmonic components as needed; however, relying solely on electrical parameter monitoring creates blind spots. When welding machine groups start simultaneously, causing a sudden increase in THD above 25%, traditional systems can only detect electrical anomalies, but lack non-electrical parameter monitoring methods such as infrared thermal imagers, making it difficult to obtain information on cable joint temperature changes (which may briefly exceed 85℃). The risk of exceeding the ℃ safety threshold is high, and it is impossible to accurately determine whether the harmonic exceedance is due to physical faults. In addition, the static compensation strategy is difficult to quickly adjust the capacitive reactance and switching strategy combination of the capacitor bank, resulting in reactive power compensation mismatch: insufficient compensation will aggravate the electrical aging of the equipment, while overcompensation may cause resonance phenomena. The above problems can eventually lead to serious faults such as cable joint melting and capacitor explosion, posing a potential risk of inducing electrical fires, resulting in equipment damage and production interruption. The multi-source sensing unit transmits the collected three-phase current, three-phase voltage, harmonic components, temperature distribution matrix generated by infrared thermal imager, and external data accessed via industrial Ethernet to the edge prediction unit through the collaborative communication module, realizing feedforward early warning of harmonic exceedance and local overheating risks.

[0019] During the power distribution and control process of die changing in a stamping machine, traditional systems lack load prediction capabilities and are unable to effectively cope with transient conditions where the load drops from over 80% of the rated load to 10%-15% within 1-3 seconds. They rely on real-time current and voltage feedback and fixed threshold triggering adjustment mechanisms. They need to detect that the current is continuously below the threshold and maintain it for 2-3 sampling cycles before initiating the capacitor bank disconnection operation. As a result, within 0.5-2 seconds after the load drops, the capacitor bank adapted to the high load remains in operation, causing overcompensation and causing the bus voltage to surge to 115%-125% of the rated value. This overvoltage not only damages the precision control module of the stamping machine and equipment such as the welding machine and conveying pump on the same bus, shortening the service life of the capacitor bank to less than 50% of the design value, but also, due to the lack of a fast response mechanism in the system, it relies on the voltage protector to trigger the trip with a delay of 1-3 seconds. The continuous voltage fluctuations during this period have caused irreversible damage to the equipment, ultimately leading to mold change interruption and production delays. The edge prediction unit runs an LSTM-CNN-GRU hybrid neural network model, extracts spatial features from the input data through convolutional kernels, and fuses them with the LSTM temporal features. The GRU layer generates future load prediction values, which are transmitted to the cloud optimization unit through the collaborative communication module to achieve voltage stabilization.

[0020] In the scheduling of electricity during peak and off-peak periods in the stamping workshop, the daytime stamping peak usually corresponds to the peak electricity price period. Multiple stamping machines in the workshop operate continuously at high frequency, resulting in a stable load and high power demand. At this time, the traditional system uses a fixed-capacity capacitor bank configuration for reactive power compensation—by setting a fixed switching threshold, the compensation amount is basically matched with the high reactive power demand during peak hours, temporarily maintaining the power factor at the target level. Entering the off-peak electricity price period at night, the load characteristics of the stamping workshop change significantly: most stamping machines are in intermittent shutdown or low-frequency operation, with only a few devices maintaining standby or auxiliary operations. The actual reactive power demand of the system is lower than expected. The peak load drops by 60%-80%; due to the lack of a dynamic adjustment mechanism linked to the time-of-use electricity price signal in traditional systems, their compensation strategy still uses fixed parameters during the daytime peak hours and fails to adjust the compensation logic according to the low load characteristics during the off-peak hours, resulting in delayed switching of capacitor banks or insufficient compensation, forming a significant reactive power gap; the cloud optimization unit adopts a dynamic rule engine with a near-end strategy optimization (PPO) algorithm, which generates optimization instructions based on load forecast data, integrates electricity price signals and temperature gradients, and transmits them to the impedance adaptive execution unit through a collaborative communication interface to achieve dynamic coordination of reactive power compensation and harmonic suppression.

[0021] In load bank power supply scenarios, multiple resistance welding machines connected in parallel to the same power distribution network act as nonlinear loads, injecting a large number of characteristic integer harmonics such as the 3rd, 5th, and 7th harmonics into the power grid during operation. Traditional power distribution systems lack precise harmonic monitoring and dynamic suppression mechanisms, relying solely on static compensation using fixed-parameter active power filters (APFs). When the welding machine group starts simultaneously, causing the total harmonic distortion (THD) of the power grid to exceed the 15% safety threshold, the distorted current will induce periodic pulsations in the electromagnetic torque of the press drive motor. This pulsation component is transmitted to the mechanical body via the motor shaft system, inducing high-frequency vibration of the press worktable: the measured amplitude reaches 0.15 mm, significantly exceeding the 0.05 mm process allowable limit, causing inaccurate positioning between the stamping die and the blank. Inaccurate measurement leads to a decrease in workpiece qualification rate; synchronous action causes accelerated wear of motor bearings, temperature rise exceeding the limit by more than 20K, and equipment lifespan reduced to 60% of the design value; the vibration transmission process may also cause secondary faults such as loosening of motor foundation bolts and poor contact of cable joints, forming new harmonic emission sources, which in turn triggers a vicious cycle of harmonics and vibration; the impedance adaptive execution unit, based on the received optimization command signal, coordinates the capacitive reactance parameter adjustment, dynamic harmonic compensation, and impedance level adjustment through the reconfigurable topology capacitor bank, threshold-triggered APF, and tap changer controller, and feeds back the execution status parameters to the edge prediction unit and cloud optimization unit in real time through the collaborative communication module to ensure that the motor vibration amplitude is always maintained within the safe threshold range.

[0022] The collaborative communication module uses the MQTT protocol to establish a dual-channel redundant link between the edge prediction unit and the cloud optimization unit, and connects to the impedance adaptive execution unit via the CAN bus.

[0023] The MQTT protocol is a lightweight publish / subscribe messaging protocol that features low power consumption and low bandwidth usage, supports asynchronous communication, and ensures reliable message transmission. The cloud-based dual-channel redundant link establishes MQTT connections in parallel through two independent physical / logical communication channels, achieving high availability: when the primary link fails, the backup link automatically takes over; load balancing: data splitting reduces the pressure on a single channel; data integrity: dual-channel verification avoids transmission loss.

[0024] An infrared thermal imager array is installed in a grid pattern on the top panel inside the distribution cabinet, with its detection direction facing the busbars and circuit breaker contact areas. This allows for comprehensive temperature monitoring of this high-risk area without blind spots. The system generates a temperature distribution matrix that includes spatial coordinates and corresponding temperature values, and performs real-time analysis on temperature anomalies within the matrix. This effectively identifies localized temperature rises caused by poor contact or insulation aging, providing early warnings for arc faults, which are typically caused by high temperatures igniting surrounding insulation materials or air ionization. This allows for timely intervention before the fault escalates.

[0025] The reconfigurable topology capacitor bank comprises at least three sets of capacitors with different capacitance values, deployed in the reactive power compensation compartment and directly connected to the busbar via copper busbars; it forms a parallel switchable network through an IGBT switch matrix, supporting capacitive reactance adjustment within a specified range in steps of 0.625%, achieved through two parallel combinations of the three sets of capacitors, where n is the number of IGBT switches, and the capacitive reactance is calculated using the following formula: ,in This represents the total capacitance of the capacitor bank. For capacitive reactance, This is the fundamental frequency of the power grid.

[0026] Multiple resistance welding machines on the same power distribution network, acting as nonlinear loads, will inject significant characteristic harmonics. A threshold-triggered APF (Automatic Power Filter) with a built-in FFT spectrum analysis module, installed upstream of the nonlinear load, uses a 128-point real-time spectrum analysis. When the total harmonic distortion (THD) exceeds a preset value, compensation is automatically activated. The formula for calculating the THD is: This formula can accurately calculate the degree of harmonic distortion, ensuring that the APF (Advanced Harmonic Filter) activates compensation at the appropriate time to effectively suppress harmonics. This is the effective value of the fundamental current. Let be the effective value of the i-th harmonic current, n be the highest harmonic order analyzed, and THD be the total harmonic distortion rate. A. When the multi-source sensing unit detects that 200 spot welding machines on the production line start simultaneously, the real-time current data is as follows: A1. Effective value of fundamental current: 50Hz power frequency; A2. Effective value of harmonic current: 5th harmonic 250Hz: , 7th harmonic 350Hz: , 11th harmonic at 550Hz: , A3. Preset security threshold: ; B. THD Formula Calculation Flow: B1. Calculate the square of each harmonic current: B2. Find the sum of squares of harmonic currents: B3. Calculate the square of the fundamental current: B4. Substitute into the THD formula: C. Actual effect C1. Threshold-triggered APF action: Calculated The APF's built-in FFT module is activated immediately; An APF injects a compensating current into the power grid that is out of phase with the harmonics, achieving a suppression rate >85%. D. Performance advantages: Response speed: It only takes 15ms from detection to action, while traditional solutions take more than 3 seconds, making it 200 times faster.

[0027] Suppression effect: Harmonic current is reduced to a safe level, THD is reduced from 67.82% to <10%, eliminating the risk of bus overheating.

[0028] The dynamic rule engine updates in real time through reinforcement learning and combines the PPO algorithm to generate optimization instructions to adjust compensation logic, thereby improving economic efficiency and operational efficiency. The dynamic rule engine runs on the photovoltaic power plant cloud platform, and the policy network adopts an Actor-Critic architecture. The state space includes: Load forecast, Peak, flat, and valley electricity price signals Temperature gradient The calculation formula is: , where ΔT is the temperature difference between adjacent sampling points, and Δd is the distance between sampling points; When changing the stamping die, the load power drops suddenly from 100% to 15%, causing a sudden change in the temperature field. Multi-source sensing units collect real-time data: A1. Load forecast value drops by 85% within 5 minutes: A2. Electricity Price Signal: 1.2 yuan / kWh during peak electricity periods A3. Temperature Matrix: Infrared thermal imager detects two adjacent points in the circuit breaker contact area: Temperature at point A: Temperature at point B: Distance between two points: A4. Temperature gradient formula calculation process A5. Calculate the temperature difference: A6. Substitute into the gradient formula: PPO algorithm dynamic optimization decision process B1. State Space Input: The load dropped sharply. B2. Electricity price = peak electricity period. B3 Risk of localized overheating; PPO action space output, Actor network generation: C1. Target capacitive reactance: , Due to the decrease in load, the reactive power compensation needs to be reduced. Output: Original value rise, Economic logic: Increase capacitive reactance during peak electricity periods to reduce the amount of compensation capacitors required and save on electricity costs; C2.APF trigger threshold : A sudden load drop may trigger harmonic oscillations. Output: A 15% decrease from the original value is more sensitive. C3. Tap switch position k: Risk of a sudden voltage surge due to a decrease in load. Output: k=-2 downshift gear; Critic Network Evaluation Value: D1. Optimization Objective: , D2. Weighting of this decision: Electricity cost takes priority; Actual effect E1. Impedance adaptive execution: E2. Capacitor bank reconfiguration: Target Capacitive → Switch capacitor combinations via IGBT switching matrix; Actual capacitive reactance Stepping accuracy: 0.625%; Result: 50% capacity was withdrawn to avoid voltage spikes caused by overcompensation; E3. Tap switch adjustment: gear → Output voltage decreases by 1.25%, offsetting the reduction in line voltage drop. Result: Voltage fluctuations were controlled within... Internally, traditional solutions >5% E4. Improved safety and economy: Overheating risk mitigation: , Triggering the PPO active cooling strategy Reducing capacitor heat generation increases capacitive reactance, which in turn decreases current. Tap changer reduces voltage and alleviates contact resistance and overheating of contacts; E5. Peak electricity period cost optimization: Actual measurement data: Reduced capacitor input → 38% savings in ineffective compensation electricity costs; APF threshold reduced to 10% → Harmonic suppression in advance.

[0029] During the power distribution control process of die changing in a stamping press, the system extracts the temporal features of historical load data through an LSTM layer, and simultaneously uses the convolutional kernel structure of a CNN layer to extract spatial features from the temperature matrix collected by an infrared thermal imager. The obtained temporal and spatial features are fused and spliced ​​through a fully connected layer of the feature fusion layer to form a joint feature vector, which is then input to the GRU layer. The GRU layer outputs a load prediction value for the next 5 minutes to address the transient condition where the load drops from over 80% of the rated load to 10%-15% within 1-3 seconds during die changing. This prediction mechanism aims to solve the overcompensation problem caused by traditional systems lacking load prediction capabilities, which results in the continued operation of capacitor banks adapted to high load conditions for 0.5-2 seconds after a sudden load drop, thus preventing malfunctions of the reactive power compensation system. The LSTM-CNN-GRU hybrid neural network model is deployed in the FPGA chip of the edge computing box for load prediction, and its structure includes: The LSTM layer is used to output time-series features, which refer to power data collected in chronological order. The CNN layer with convolutional kernels is used to extract spatial features from the temperature matrix, where spatial features refer to the collected temperature inside the cabinet. The feature fusion layer is used to concatenate temporal and spatial features into a feature vector through a fully connected layer. The GRU layer is used to output the load forecast for the next 5 minutes.

[0030] Under the condition of a sudden emergency failure in a high-load welding line, the instantaneous impact load of the welding machine group will cause the harmonic distortion rate of the power grid to rise sharply to above the safety threshold. At the same time, the large current passing through the busbar will cause its temperature to rapidly exceed the safe operating limit of 85°C, forming a critical state of simultaneous harmonic exceedance and busbar overheating. Because traditional systems lack the ability to monitor multiple parameters in a coordinated manner, they cannot achieve rapid linkage control. The continuous excessive harmonics cause the capacitor bank to be subjected to overvoltage and overcurrent impacts, while the high temperature environment generated by the busbar overheating further accelerates the aging and breakdown process of the internal dielectric of the capacitor, which can ultimately easily lead to capacitor explosion. An explosion accident caused equipment damage and production interruption. The system includes a safety protection module for early warning of harmonic over-limit and localized overheating risks. When the multi-source sensing unit detects that the local temperature or total harmonic distortion rate inside the cabinet exceeds a preset value, the safety protection module triggers a signal that is directly transmitted to the impedance adaptive execution unit via the collaborative communication module. This bypasses the cloud optimization process and executes emergency actions, while simultaneously pushing alarm information. The actions directly triggered by the safety protection module bypassing the cloud optimization process include: complete disconnection of the capacitor bank, full power compensation of the APF, and resetting the tap changer controller to its initial position.

[0031] The method of the intelligent power regulation system for the power distribution cabinet also includes the following operating procedures: Real-time acquisition of multi-source data: A1. Electrical parameter acquisition: Acquire three-phase current and voltage, and calculate harmonic components and THD through FFT; A2. Temperature Acquisition: Infrared thermal imager generates temperature matrix. Calculate the temperature gradient , where i,j are the infrared thermal imager grid coordinate indices, and T_{i,j} is the temperature distribution matrix; A3. External data access: Obtain real-time electricity price signals.

[0032] Edge-side load forecasting: B1. Data Preprocessing: Normalize the historical load sequence, temperature matrix, and event labels; B2. Feature Extraction: LSTM layers extract temporal features, and convolutional kernel CNN layers extract spatial features; B3. Feature Fusion: Concatenate feature vectors and input them into the GRU layer; B4. Output: Load change curve for the next 5 minutes And high-temperature area warning, where t is a time variable and P(t) is the future load forecast curve.

[0033] Cloud-based dynamic optimization decision-making: C1. State Space Construction: Input load forecast P(t), electricity price signal, and temperature gradient through multi-source sensing units. ; Solving using the C2.PPO algorithm: Actor network output action space: target capacitance APF trigger threshold Tap switch K, in The target capacitive reactance value optimized for the PPO algorithm is the ideal capacitive reactance of the capacitor bank dynamically calculated through reinforcement learning, which is used for precise reactive power compensation. This is the APF activation threshold. When the real-time harmonic distortion rate exceeds this preset value, the active filter APF will automatically activate compensation. k is the tap position of the transformer, an integer variable, used to control the tap position to adjust the voltage.

[0034] The Critic network evaluates the value of actions and optimizes the objective function through policy updates. ,in These are the weighting coefficients; C3. Generate an optimized instruction set.

[0035] Impedance adaptive execution: D1. Capacitor bank reconfiguration: Calculate the target reactive power compensation Q* according to the formula, where Q* is the target reactive power compensation, which is the theoretical reactive power value calculated based on the load demand. Search for IGBT switch combinations to make capacitive reactance Approaching ,in This represents the actual capacitive reactance value of the capacitor bank after real-time adjustment via the IGBT switching matrix. The target capacitive reactance value; D2. Harmonic suppression: Real-time FFT analysis of THD, when THD > At this time, APF injects reverse current compensation; D3. Tap changer adjustment: Based on the predicted voltage value, adjust the tap K in 0.625% increments to offset the line voltage drop.

[0036] Security protection mechanism: E1. Real-time monitoring: Continuously monitors the temperature matrix. and THD; E2. Exception Trigger: If If the temperature is >85℃ or the THD is >15%, proceed as follows: Disconnect all capacitor banks: ,in This represents the total capacitance of the capacitor bank. APF at full power: compensation depth 100%; Tap switch reset: K=0, where K is the tap switch reset position; E3. Alarms and Logging: Alarm information is pushed via MQTT, and fault data such as temperature, harmonics, and timestamps are stored.

[0037] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0038] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent power regulation system for a power distribution cabinet, characterized in that, It includes a multi-source sensing unit, an edge prediction unit, a cloud optimization unit, an impedance adaptive execution unit, and a cooperative communication module; The multi-source sensing unit transmits the collected three-phase current, three-phase voltage, harmonic components, temperature distribution matrix generated by the infrared thermal imager, and external data collected via industrial Ethernet access to the edge prediction unit through the collaborative communication module. The edge prediction unit runs an LSTM-CNN-GRU hybrid neural network model. It extracts temperature spatial features from the received data through the CNN layer of the convolutional kernel and fuses them with the LSTM temporal features. The GRU layer generates future load prediction values, which are then transmitted to the cloud optimization unit via the collaborative communication module. The cloud-based optimization unit generates optimization instructions based on the received load forecast, combined with the electricity price signal and temperature gradient, and transmits them to the impedance adaptive execution unit via the collaborative communication module. The impedance adaptive execution unit performs capacitive reactance adjustment, harmonic compensation, and gear adjustment through a reconfigurable topology capacitor bank, threshold-triggered APF, and tap changer controller according to the received optimization instructions, and feeds back the execution status to the edge prediction unit and cloud optimization unit through the collaborative communication module. The collaborative communication module uses the MQTT protocol to establish a dual-channel redundant link between the edge prediction unit and the cloud optimization unit, and connects to the impedance adaptive execution unit via a CAN bus.

2. The intelligent power regulation system for a power distribution cabinet according to claim 1, characterized in that, The infrared thermal imager array is installed in a grid layout on the top plate of the distribution cabinet facing the busbar and circuit breaker contact area, generating a temperature distribution matrix that includes spatial coordinates and corresponding temperature values, and analyzing temperature anomalies in the matrix in real time.

3. The intelligent power regulation system and method for a power distribution cabinet according to claim 1, characterized in that, The reconfigurable topology capacitor bank includes at least three sets of capacitors with different capacitance values. The capacitors are deployed in the reactive power compensation compartment and directly connected to the busbar via copper busbars. A parallel switchable network composed of IGBT switch matrices supports adjustment of capacitive reactance in steps of 0.625% within a specified range. This is achieved through two parallel combinations of three sets of capacitors, where n is the number of IGBT switches.

4. The intelligent power regulation system for a power distribution cabinet according to claim 3, characterized in that, The threshold-triggered APF has a built-in FFT spectrum analysis module installed upstream of the nonlinear load. Its FFT module uses 128-point real-time spectrum analysis and automatically activates compensation when the total harmonic distortion rate exceeds the preset value.

5. The intelligent power regulation system for a power distribution cabinet according to claim 1, characterized in that, The cloud-based optimization unit generates optimization instructions through a dynamic rule engine based on the PPO algorithm. This dynamic rule engine is updated in real time using reinforcement learning. Combined with the optimization instructions generated by the PPO algorithm, the compensation logic is adjusted. The policy network adopts an Actor-Critic architecture, and the state space includes: load forecast, electricity price signal, and temperature gradient. The electricity price signals are the electricity prices for peak, flat, and valley periods.

6. The intelligent power regulation system and method for a power distribution cabinet according to claim 1, characterized in that, The LSTM-CNN-GRU hybrid neural network model is deployed on the FPGA chip of the edge computing box for load prediction, including: The LSTM layer is used to output time-series features, where time-series features refer to power data collected in chronological order. The CNN layer with convolutional kernels is used to extract spatial features from the temperature matrix, where spatial features refer to features that reflect the spatial patterns of the temperature field inside the cabinet. The feature fusion layer is used to concatenate temporal and spatial features into a feature vector through a fully connected layer. The GRU layer is used to output the load forecast for the next 5 minutes.

7. The intelligent power regulation system for a power distribution cabinet according to claim 4, characterized in that, It also includes a safety protection module for early warning of harmonic over-limit and local overheating risks; when the multi-source sensing unit detects that the local temperature or total harmonic distortion rate in the cabinet exceeds the preset value, the safety protection module triggers a signal that is directly transmitted to the impedance adaptive execution unit via the collaborative communication module, skipping the cloud optimization process to perform emergency actions, and at the same time pushes alarm information. The content directly triggered by the safety protection module skipping the cloud optimization process includes: complete disconnection of capacitor bank, full power compensation of APF, and reset of tap changer controller to the initial position.

8. A method for intelligent power regulation of a power distribution cabinet, employing the intelligent power regulation system for a power distribution cabinet as described in claim 1, characterized in that, Includes the following steps: A. Real-time acquisition of multi-source data, including three-phase current, three-phase voltage, harmonic components, temperature distribution matrix generated by infrared thermal imager, and external data acquired via industrial Ethernet access; B. Edge-side load prediction: Run the LSTM-CNN-GRU hybrid neural network model. Extract the temperature spatial features from the data collected in step A through the CNN layer of the convolutional kernel and fuse them with the LSTM temporal features. The GRU layer generates the future load prediction value. C. Cloud-based dynamic optimization decision-making: Based on the received load forecast, combined with electricity price signals and temperature gradients, optimization instructions are generated. D. Impedance adaptive execution: Based on the received optimization instructions, the capacitor bank with reconfigurable topology, threshold-triggered APF and tap changer controller perform capacitive reactance adjustment, harmonic compensation and tap adjustment to achieve dynamic adaptation of reactive power compensation and harmonic suppression. E. Safety protection mechanism: When the local temperature or total harmonic distortion rate inside the cabinet exceeds the preset value, the safety protection mechanism is triggered, skipping the cloud dynamic optimization decision-making steps and executing emergency actions. At the same time, alarm information is pushed. The emergency actions include: complete disconnection of capacitor bank, full power compensation of APF, and resetting of tap changer controller to the initial position.

9. The intelligent power regulation method for a power distribution cabinet according to claim 8, characterized in that, Step B includes: B1. Data Preprocessing: Normalize the historical load sequence, temperature matrix, and event labels; B2. Feature Extraction: LSTM layers extract temporal features, and convolutional kernel CNN layers extract spatial features; B3. Feature Fusion: Concatenate feature vectors and input them into the GRU layer; B4. Output: Load change curve for the next 5 minutes.

10. The intelligent power regulation system and method for a power distribution cabinet according to claim 8, characterized in that, Step C includes: C1. State Space Construction: Input load forecast P(t), electricity price signal, and temperature gradient through multi-source sensing units. ; Solving using the C2.PPO algorithm: The action space output by the Actor network includes: target capacitance. APF trigger threshold Tap switch K, in The target capacitive reactance value optimized for the PPO algorithm is the ideal capacitive reactance of the capacitor bank dynamically calculated through reinforcement learning, which is used for precise reactive power compensation. This is the APF activation threshold. When the real-time harmonic distortion rate exceeds this preset value, the active filter APF will automatically activate compensation. K is the tap position, an integer variable, which controls the position of the transformer tap to adjust the voltage. The Critic network evaluates the value of actions and optimizes the objective function through policy updates. : ,in These are the weighting coefficients; C3. Generate an optimized instruction set.

Images