Power equipment self-repairing surge protection system based on artificial intelligence

Abstract

The invention relates to a power equipment self-repairing surge protection system based on artificial intelligence, and the system is characterized in that a surge detection module is connected between a live wire and a protection ground wire of power equipment, collects voltage and current data, and transmits the data to an artificial intelligence analysis unit; and the artificial intelligence analysis unit extracts the surge characteristic parameters, generates protection strategy parameters through a deep neural network, and controls the trigger thresholds of all levels of protection devices in the programmable protection circuit. The programmable protection circuit comprises a gas discharge tube capable of adjusting trigger voltage, a metal oxide varistor and a transient voltage suppressor diode, and is used for discharging surges with different energies step by step. The resettable fuse cuts off the circuit when the temperature exceeds a threshold value and recovers conduction after the temperature is reduced. And the controller receives and applies the protection strategy parameter output by the artificial intelligence analysis unit, and dynamically adjusts the trigger threshold value of the protection device, so that the response and protection performance of the system to the surge is improved, and the self-repairing and stable operation of the power equipment are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based self-repairing surge protection system for power equipment. Background Art

[0002] In the prior art, surge protection systems mainly rely on fixed configuration protection devices, such as gas discharge tubes, metal oxide varistors and transient voltage suppression diodes, which are usually designed with fixed trigger thresholds to provide protection for different surge energy levels. Such systems usually use simple hardware circuits for surge detection and response, cannot dynamically adjust the protection strategy, and lack adaptability to environmental changes. In addition, traditional surge protection systems often require manual maintenance or replacement of protection devices after a surge event, which increases maintenance costs and downtime.

[0003] However, there are some problems in the existing technology, mainly reflected in the inability to dynamically adjust the protection strategy according to the real-time detected surge characteristics, which makes it difficult for the system to strike a balance between protection performance and response speed. In addition, the traditional surge protection system lacks self-learning and intelligent analysis capabilities, and cannot effectively predict the occurrence of surge events or adapt to different types of surge characteristics, affecting the long-term stability and protection effect of the system. Protection components are prone to aging or damage after repeated use, and the system usually requires regular maintenance, which reduces overall reliability and operating efficiency.

[0004] Therefore, it is necessary to develop an improved surge protection system. Summary of the invention

[0005] The present application provides an artificial intelligence-based self-repairing surge protection system for power equipment to improve the response and protection performance to surges.

[0006] The present application provides an artificial intelligence-based self-repairing surge protection system for power equipment, comprising:

[0007] A surge detection module, which is connected between the live wire and the protective ground wire of the power equipment, and is used to collect voltage data and current data between the live wire and the protective ground wire according to a preset period; and transmit the collected data to the artificial intelligence analysis unit;

[0008] An artificial intelligence analysis unit is used to extract surge characteristic parameters from the data transmitted by the surge detection module, wherein the surge characteristic parameters include the voltage change rate, voltage peak value, voltage abnormality duration of adjacent sampling points, the current change rate, current peak value, and current abnormality duration of adjacent sampling points; based on the extracted surge characteristic parameters, a pre-trained deep neural network model is used to generate protection strategy parameters of the protection device in the programmable protection circuit, wherein the protection strategy parameters include the trigger thresholds of the protection devices at each level in the programmable protection circuit;

[0009] A programmable protection circuit includes a first-stage gas discharge tube, a second-stage metal oxide varistor, and a third-stage transient voltage suppression diode connected in parallel between a live wire and a protective ground wire; the trigger voltage of the first-stage gas discharge tube can be adjusted within a first voltage range to discharge high-energy surges including lightning strikes; the threshold voltage of the second-stage metal oxide varistor can be adjusted within a second voltage range to suppress medium-energy surges; the clamping voltage of the third-stage transient voltage suppression diode can be adjusted within a third voltage range to limit the residual voltage to the ground;

[0010] A resettable fuse is connected in series between the live wire and the power equipment. When the temperature exceeds a first temperature threshold, the resistance increases to cut off the circuit, and when the temperature drops below a second temperature threshold, the conduction is automatically restored;

[0011] The controller is used to receive the protection strategy parameters output by the artificial intelligence analysis unit; and adjust the trigger thresholds of various levels of protection devices in the programmable protection circuit according to the received protection strategy parameters.

[0012] This application has the following beneficial technical effects:

[0013] (1) The present invention uses an artificial intelligence analysis unit to perform deep learning and real-time analysis on surge characteristic parameters, and can generate dynamic protection strategy parameters according to different surge types and characteristics. Compared with the traditional fixed threshold protection system, the present invention significantly improves the response accuracy and effectiveness of surge protection, ensuring that the system can respond to surge events of different intensities and durations more flexibly and efficiently.

[0014] (2) Using a deep neural network model, the present invention can extract and analyze the characteristics of surge data, predict possible surge events in advance, and then adjust the trigger threshold of the protection device. This intelligent prediction capability effectively reduces the risk of equipment being damaged by surges and provides a more comprehensive and proactive protection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of an artificial intelligence-based self-repairing surge protection system for power equipment provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0016] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0017] The first embodiment of the present application provides a self-repairing surge protection system for power equipment based on artificial intelligence. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 The first embodiment of the present application provides a self-repairing surge protection system for power equipment based on artificial intelligence for detailed description.

[0018] The artificial intelligence-based self-repairing surge protection system for electric power equipment includes a surge detection module 101, an artificial intelligence analysis unit 102, a programmable protection circuit 103, a self-recovering fuse 104 and a controller 105.

[0019] The surge detection module 101 is connected between the live wire and the protective ground wire of the power equipment, and is used to collect voltage data and current data between the live wire and the protective ground wire according to a preset period; and transmit the collected data to the artificial intelligence analysis unit.

[0020] The surge detection module 101 is a key monitoring unit in this embodiment, which is designed to monitor and collect electrical parameters of power equipment in real time, especially to provide accurate data support when surge or voltage anomaly occurs. The module is connected to the live wire and protective ground wire of the power equipment through an electrical connection device, so that the voltage and current signals between the two can be obtained. This connection method ensures that the module can capture the transient change characteristics of the power system, including high-frequency voltage pulses and current anomalies.

[0021] In order to ensure the accuracy and timeliness of data collection, the surge detection module 101 includes a high-precision voltage sensor and a current sensor, which are used to detect the voltage value and current intensity between the live wire and the protective ground wire respectively. These sensors must have a wide-band response capability so that they can still accurately measure in the case of fast surges or transient overvoltages, and have low noise and high sensitivity characteristics to reduce the impact of environmental interference on signal acquisition. In addition, in order to improve the acquisition efficiency, the module is provided with a timing control unit, which controls the sensor to periodically collect voltage and current according to a preset sampling period. The sampling period can be adjusted according to different application scenarios to ensure that the acquisition frequency matches the time variation of the surge characteristics.

[0022] The surge detection module 101 also includes a signal conditioning circuit for preprocessing the collected voltage and current signals. The functions of the conditioning circuit include filtering, amplifying and noise suppression of the signal to optimize data quality and reduce errors. The filtering circuit can be designed as a bandpass filter to eliminate low-frequency interference and high-frequency noise from the outside while retaining useful signals related to the surge event. The amplification circuit is used to enhance weak voltage or current signals so that they meet the input requirements of the subsequent artificial intelligence analysis unit 102. The entire signal conditioning process can be automatically controlled by an integrated circuit to quickly adapt to different electrical environments and surge characteristics.

[0023] The signal conditioned data is transmitted to a data processing unit, which converts the voltage and current signals into a digital format that can be used for subsequent analysis. The data processing unit can use an analog-to-digital converter (ADC), and its resolution should be high enough to ensure signal accuracy. The converted digital signal is temporarily stored in a local memory and then transmitted to the artificial intelligence analysis unit 102 via a high-speed data interface. The interface can use a serial communication protocol, such as SPI or I2C, to achieve fast and reliable data transmission and ensure real-time performance.

[0024] The artificial intelligence analysis unit 102 is the core part responsible for data analysis and decision-making. It is designed to extract and analyze surge characteristic parameters from the electrical data transmitted by the surge detection module 101, and generate a dynamic protection strategy for protecting power equipment. The function of this unit relies on advanced artificial intelligence technology, specifically based on a deep neural network model to process complex surge data and optimize the protection strategy to improve the system's responsiveness and protection performance.

[0025] The artificial intelligence analysis unit 102 first receives the voltage and current data collected by the surge detection module 101 according to a preset cycle. These data include the instantaneous voltage and current values ​​between the live wire and the protective ground wire, which are usually expressed as highly dynamic signals. In order to make these signals available for intelligent analysis, the artificial intelligence analysis unit 102 has a preprocessing unit responsible for basic processing of the data, including normalization and noise suppression, to improve the accuracy of the analysis. The normalization process unifies the data in different voltage and current ranges to a standard scale for the processing of the neural network model, while noise suppression removes interference components in the environment and circuit through signal filtering technology.

[0026] After preprocessing, the data is passed to the surge feature extraction module. This module uses the algorithm in the artificial intelligence analysis unit 102 to extract key surge feature parameters from the preprocessed data. These parameters include the voltage change rate between adjacent sampling points, which can reflect the transient change characteristics of the voltage when a surge occurs; the voltage peak value, which is used to evaluate the maximum intensity of the surge; and the voltage anomaly duration, which indicates the duration of the surge state. The corresponding current characteristic parameters also include the current change rate, current peak value and current anomaly duration of adjacent sampling points to fully capture the impact of the surge on the power system. This feature extraction process is achieved through complex calculations and pattern recognition techniques to ensure that the multi-dimensional characteristics of the surge can be accurately captured.

[0027] After feature extraction, these surge characteristic parameters are input into a pre-trained deep neural network model. The model consists of a multi-layer neural network that can automatically learn and identify nonlinear relationships in surge data. After being trained with a large amount of historical surge data, the deep neural network model has efficient pattern recognition capabilities and can generate protection strategy parameters based on the input characteristic parameters. To achieve this function, each layer of neurons in the model will perform complex matrix operations and nonlinear activation to output accurate protection strategies.

[0028] The protection strategy parameters generated by the artificial intelligence analysis unit 102 include the trigger thresholds of the protection devices at each level. These thresholds directly determine the response timing of the protection devices in the programmable protection circuit 103, so that the appropriate protection mechanism can be activated in time when a surge occurs. For example, if a high-intensity surge is detected, the analysis unit will adjust the trigger threshold of the gas discharge tube to quickly release the surge energy; when a medium-intensity surge is detected, the system will give priority to adjusting the trigger threshold of the metal oxide varistor. All these protection strategy parameters are transmitted to the controller 105 through a high-speed communication interface to ensure that the entire system can respond in real time and adjust dynamically.

[0029] Furthermore, the deep neural network model used in the artificial intelligence analysis unit includes a surge feature encoder, a surge development predictor, and a protection strategy generator connected in sequence;

[0030] Among them, the surge feature encoder is implemented by a multi-layer perceptron structure; the surge feature encoder includes a normalization layer, a feature mapping layer and a feature fusion layer; the normalization layer is used to normalize the extracted surge feature parameters to obtain normalized feature parameters; the feature mapping layer maps the normalized feature parameters to a feature space of a first preset dimension through a multi-layer fully connected network to obtain multiple feature vectors of the first preset dimension; the feature fusion layer uses a self-attention mechanism to calculate the weight relationship between multiple feature vectors of the first preset dimension to obtain a fused feature vector of a second preset dimension;

[0031] The surge development predictor is implemented using a recurrent neural network structure; the surge development predictor includes a timing coding layer, a trend analysis layer and a prediction output layer; wherein the timing coding layer uses a long short-term memory network to extract timing features from a fused feature vector sequence of a second preset dimension of the most recent first preset number of time points to obtain a timing feature vector of a third preset dimension; the trend analysis layer uses multiple parallel attention heads to focus on the changing trends of multiple different time scales based on the timing feature vector of the third preset dimension to obtain multiple trend feature vectors of a fourth preset dimension; the prediction output layer concatenates multiple trend feature vectors of the fourth preset dimension and processes them through a fully connected network to obtain voltage and current prediction values ​​at each preset sampling moment in the future second preset time period;

[0032] The protection strategy generator is implemented by reinforcement learning, and the protection strategy generator includes a state representation network, an action generation network and a value evaluation network; wherein the state representation network is used to map the fused feature vector provided by the surge feature encoder and the prediction result provided by the surge development predictor into a state vector; the action generation network is used to generate the trigger threshold adjustment strategy of each level of protection devices based on the state vector; the value evaluation network is used to evaluate the expected effect of the generated protection strategy; wherein the state space of the reinforcement learning includes the current surge feature, the surge prediction value, and the working state of the protection device; wherein the action space of the reinforcement learning is the trigger threshold adjustment amount of each level of protection devices; the working state of the protection device includes the currently set trigger threshold, the conduction state and the accumulated number of actions; the reward function of the reinforcement learning comprehensively considers the surge suppression effect, the device service life and the energy loss.

[0033] The AI ​​analysis unit uses a deep neural network model to analyze and process surge characteristic parameters to achieve efficient and intelligent surge protection. The model is composed of a surge characteristic encoder, a surge development predictor, and a protection strategy generator connected in sequence. Each part works together through complex machine learning algorithms and structures to provide an accurate surge response strategy.

[0034] The surge feature encoder adopts a multi-layer perceptron structure, which is responsible for processing and converting the original surge feature parameters obtained from the surge detection module. This encoder consists of three main parts: normalization layer, feature mapping layer and feature fusion layer. The function of the normalization layer is to normalize the extracted surge feature parameters (such as voltage change rate, voltage peak value, voltage abnormality duration, etc.). This process ensures that all features are processed consistently and numerically stable in the model by scaling and adjusting the data range. The normalized feature parameters are sent to the feature mapping layer, which maps these parameters to a predefined feature space through a multi-layer fully connected network. This space has multiple feature vectors of the first preset dimension, which can capture the complex relationships and patterns between different features. The feature fusion layer uses the self-attention mechanism to calculate the weight relationship between these feature vectors, identify and amplify key features, and thus generate a fused feature vector with higher expressive power, which is mapped to the second preset dimension to provide rich input data for subsequent prediction and strategy generation.

[0035] Next, the surge development predictor is implemented using a recurrent neural network (RNN) structure, focusing on extracting and analyzing the dynamic change pattern of surge characteristics from time series data. The time series encoding layer uses a long short-term memory network (LSTM) to capture the long-term and short-term dependencies of surge characteristics from several recent time points. LSTM can effectively memorize and update information in the time series, thereby obtaining a time series feature vector of the third preset dimension, reflecting the evolution law of surge characteristics. The trend analysis layer uses multiple parallel attention heads, each focusing on a different time scale, to analyze the trend of surge characteristics over time. This multi-head attention mechanism allows the system to focus on both short-term and long-term change characteristics at the same time, generating multiple trend feature vectors of the fourth preset dimension. Finally, the prediction output layer splices these trend feature vectors together and further processes them through a fully connected network to generate predicted values ​​for voltage and current changes in the second preset time period in the future. This prediction result provides the system with a decision-making basis for responding to surges in advance.

[0036] The protection strategy generator is the unit responsible for generating protection decisions in the system. It is implemented using reinforcement learning technology and includes a state representation network, an action generation network, and a value assessment network. The state representation network combines the fused feature vector provided by the surge feature encoder with the prediction results of the surge development predictor to generate a state vector representing the current surge condition. This state vector combines the surge characteristics, future predictions, and the working status of the system protection devices to provide input for the action generation network. The action generation network generates a trigger threshold adjustment strategy for the protection device based on the state vector to ensure that each protection device can respond to different surge events with the optimal timing and threshold. The value assessment network is used to evaluate the expected effect of the generated strategy and provide feedback and adjustment basis for the strategy. This network is optimized through the reward function of reinforcement learning. The reward function comprehensively considers the surge suppression effect, the service life of the protection device, and the overall energy consumption to ensure that the system strikes a balance between protection performance and resource efficiency.

[0037] The entire analysis unit realizes a highly automated and intelligent surge protection process.

[0038] Furthermore, the normalization layer adopts a piecewise adaptive normalization method to cope with the dynamic change environment of voltage and current characteristic parameters, specifically including:

[0039] Based on the physical characteristics of voltage and current characteristic parameters, the value range of each characteristic parameter is divided into multiple dynamic intervals, and the initial statistical parameters in each dynamic interval are defined, including mean and variance;

[0040] Maintain a set of dynamic statistical parameters for each dynamic interval, update the mean and variance within the interval through the exponential moving average method, and adjust the interval range according to the distribution changes of the characteristic parameters;

[0041] When a surge event detects a sharp fluctuation in voltage or current, the update weight of the corresponding dynamic interval is adjusted, and the update frequency of the data in the corresponding interval is increased;

[0042] Standardization is performed based on the statistical parameters of the interval to which the current characteristic parameters belong, so as to maintain the numerical stability and electrical characteristic consistency of the normalized results and improve the response efficiency of the deep neural network to different surge characteristics.

[0043] The normalization layer adopts a piecewise adaptive normalization method to cope with the dynamic change characteristics of voltage and current characteristic parameters under different surge environments, ensuring that the deep neural network can accurately and efficiently process these data. This normalization method improves the model's adaptability to complex electrical environments by performing piecewise processing and adaptive adjustment of characteristic parameters.

[0044] First, based on the physical characteristics of the voltage and current characteristic parameters, the value range of each characteristic parameter is divided into multiple dynamic intervals. These intervals are defined based on the common value ranges of the characteristic parameters to ensure that all possible change scenarios can be covered under different operating conditions. Each dynamic interval has a set of initial statistical parameters, including mean and variance, which are used to describe the distribution characteristics of the data in the interval. The initial mean and variance can be set based on historical data or system preset values ​​to provide a basis for normalization processing.

[0045] In order to cope with the dynamic changes of feature parameters over time, the normalization layer maintains a set of real-time updated statistical parameters for each dynamic interval. These statistical parameters are updated by the exponential moving average method, which gives higher weights to recent data and lower weights to older data, so that the mean and variance can flexibly track the changing trend of feature parameters. Specifically, when a new data point enters a dynamic interval, the system calculates the impact of the data point on the mean and variance in the interval, and updates these statistical parameters using the exponential moving average method. This method can not only ensure the real-time performance of the normalization processing, but also smooth out short-term fluctuations in data and improve the system's ability to resist noise.

[0046] When a surge event occurs, such as when a sharp fluctuation in voltage or current is detected, the normalization layer will further adjust the update mechanism of the dynamic interval. At this time, the system will immediately increase the data update frequency of the interval where the sharp fluctuation occurs, and appropriately increase the update weight of the interval to ensure that the mean and variance can quickly reflect the changes in the current surge characteristics. This adjustment mechanism enables the system to maintain the accuracy and stability of the normalized results when responding to sudden surge events, and will not cause distortion of the analysis results due to too fast parameter changes.

[0047] During each normalization process, the system uses the updated mean and variance to perform normalization operations based on the dynamic range to which the current characteristic parameters belong. The normalization process adjusts the characteristic parameters to a uniform numerical scale so that the deep neural network can perform subsequent processing more efficiently. At the same time, this piecewise adaptive normalization method retains the consistency of electrical characteristics and avoids the loss of the physical meaning of the data due to normalization, thereby improving the efficiency and accuracy of the deep neural network's response to different surge characteristics.

[0048] Furthermore, the self-attention mechanism of the feature fusion layer adopts a hierarchical weighted fusion structure to fully explore the deep correlation between voltage and current features, including:

[0049] The first layer of attention calculates the association weights between voltage features, including:

[0050] The self-attention mechanism is used to calculate the similarity between the three feature vectors of voltage change rate, voltage peak and voltage duration;

[0051] Generate an attention weight matrix of voltage features based on similarity, and adaptively weight them according to the importance of features to obtain a voltage comprehensive feature vector;

[0052] By embedding position coding, the time series information of voltage features is retained, further optimizing the feature fusion effect;

[0053] The second layer of attention calculates the association weights between current features, including:

[0054] The self-attention mechanism is used to calculate the similarity between the three feature vectors of current change rate, current peak value and current duration;

[0055] Generate an attention weight matrix of current features based on similarity, and adaptively weight them according to the mutual correlation of current features to obtain a comprehensive current feature vector;

[0056] Through the multi-scale attention mechanism, the model's sensitivity to the changing trends of different current characteristics is enhanced;

[0057] The third level of attention integrates the voltage comprehensive features and the current comprehensive features, including:

[0058] Comprehensively consider the phase difference between voltage characteristics and current characteristics, and adjust the fusion weight according to the power factor characteristics to optimize power quality analysis;

[0059] Use a bidirectional attention mechanism to model the interdependence between voltage and current comprehensive features and generate the final fused feature vector;

[0060] The final fused feature vector retains the physical correlation between voltage and current to improve the deep neural network's understanding and prediction accuracy of surge characteristics.

[0061] The feature fusion layer uses a self-attention mechanism and a hierarchical weighted fusion structure to fully explore the deep correlation between voltage and current features, thereby enhancing the surge protection system's ability to understand and respond to complex surge signals. The self-attention mechanism plays a key role in this design. By analyzing and fusing feature data in layers, a highly accurate fused feature vector is generated for subsequent deep learning model processing.

[0062] At the first level, the system focuses on the internal correlation of voltage features. Specifically, the self-attention mechanism is used to calculate the similarity between the three feature vectors of voltage change rate, voltage peak and voltage duration. Through this calculation, the system is able to quantify the correlation between feature vectors and generate an attention weight matrix for voltage features. This matrix is ​​adaptively weighted according to the importance of the features, so that key features receive higher weights in the final voltage comprehensive feature vector, improving the accuracy of feature expression. In addition, embedded position encoding is used to retain the time series information of voltage features. This encoding method ensures that the voltage feature vector does not lose its time order during the fusion process, thereby optimizing the feature fusion effect and making the system more sensitive and accurate in dealing with time-varying surges.

[0063] Next, the second layer focuses on the association weights between current features. This layer also uses the self-attention mechanism to calculate the similarity of the three feature vectors of current change rate, current peak and current duration. The generated attention weight matrix is ​​used for adaptive weighting to identify and highlight key current features to form a comprehensive current feature vector. Through the multi-scale attention mechanism, the system is able to analyze the changing trends of current features at different time scales, thereby enhancing the model's sensitivity to current feature fluctuations. This mechanism is particularly helpful in capturing the dynamic characteristics of current changes in complex surge phenomena, providing precise support for subsequent decision-making.

[0064] Finally, the third layer fuses the voltage comprehensive features with the current comprehensive features to achieve overall analysis and optimization. The system comprehensively considers the phase difference between the voltage and current features, and adjusts the fusion weight according to the power factor characteristics. In this way, the feature fusion layer can optimize the analysis of power quality and ensure that the generated feature vector fully reflects the complex physical relationship between voltage and current. In addition, the use of a bidirectional attention mechanism to model the interdependence between voltage and current features allows the model to capture the bidirectional mutual influence between them. This bidirectional modeling method significantly improves the expressiveness of the fused features, so that the final fused feature vector not only retains the physical correlation between voltage and current, but also improves the accuracy and robustness of the deep neural network in predicting surge characteristics.

[0065] Furthermore, the trend analysis layer adopts a multi-scale progressive attention mechanism to hierarchically focus on the changes of different time scales of surge characteristics, including:

[0066] Micro-scale attention heads are used to focus on high-frequency changes between adjacent sampling points and detect instantaneous and violent fluctuations in voltage and current; analyze instantaneous change characteristics, such as voltage peaks and rapid current fluctuations, to capture the sudden behavior of surges; and provide instant response surge prediction information to deal with short-term high-energy surge events;

[0067] Mesoscale attention heads are used to focus on the changes in electrical signals within a single power frequency cycle, identify the periodic modulation characteristics of surges on power frequency signals; analyze the amplitude, phase, and duration characteristics of voltage and current within a power frequency cycle; predict the mid-term impact of surges and provide data support for coordinating the actions of protection devices;

[0068] The macro-scale attention head is used to focus on the overall trend changes within multiple power frequency cycles and analyze the energy attenuation and propagation characteristics of surges. It observes the slow changes of voltage and current over a long time window to evaluate the long-term impact of surges on power equipment. It is used to optimize protection strategies to ensure the stability and protection performance of the system over a long period of time.

[0069] The trend analysis layer adopts a multi-scale progressive attention mechanism, which aims to achieve accurate prediction and effective response to complex surge phenomena by analyzing the changes of surge characteristics at different time scales in a layered manner. The design of this mechanism enables the system to capture the instantaneous and violent fluctuations, medium-term change patterns, and long-term trends of surge characteristics, thereby providing comprehensive surge analysis and protection.

[0070] The microscale attention head focuses on high-frequency changes in surge characteristics. This attention head can detect sharp fluctuations in voltage and current between adjacent sampling points and identify instantaneous electrical signal peaks and sudden changes. Such high-frequency fluctuations usually reflect the sudden characteristics of surges, such as sudden increases in voltage or sharp changes in current. By analyzing these instantaneous features, the system can capture the sudden behavior of surges and provide instant prediction information so that appropriate protective measures can be taken in a short time. This fast response is particularly critical for responding to high-energy surge events (such as instantaneous surges caused by lightning strikes or equipment failures), which can significantly reduce damage to power equipment.

[0071] The mesoscale attention head focuses on the changes in surge characteristics within a power frequency cycle. The power frequency cycle is usually the basic time scale in the power system. Analyzing the characteristics within this range helps to identify the modulation effect of surges on power frequency signals. Specifically, the mesoscale attention head can identify the amplitude changes, phase shifts, and duration characteristics of voltage and current within a power frequency cycle. These analysis results can reveal how surges affect the stability of the power system and provide medium-term predictions of surge impacts. This prediction information plays a key role in the timing and coordination of the action of protection devices, enabling the system to take balanced protection measures in the medium term to ensure the safe operation of power equipment throughout the power frequency cycle.

[0072] The macro-scale attention head focuses on surge trends over long time scales, observing the overall voltage and current changes over multiple power frequency cycles. It captures the gradual attenuation and propagation characteristics of surge energy through a long time window, thereby evaluating the long-term impact of surges on power equipment. This long-term analysis can identify slow but continuous surge phenomena and the potential risks these phenomena may bring to the power system. The macro-scale attention head provides a basis for optimizing long-term protection strategies, ensuring that the system remains stable over a long period of time and can effectively manage continuous or chronic surge events. By comprehensively analyzing the long-term variation characteristics of voltage and current, the system can dynamically adjust the protection strategy, improve the overall protection performance, and extend the service life of power equipment.

[0073] Furthermore, the prediction output layer adopts a hierarchical prediction strategy to perform hierarchical prediction in different time windows to enhance the accurate response to surge characteristics, specifically including:

[0074] The first-level prediction is used to predict the voltage and current change trends within the shortest time window in the future. The prediction is based on the output of the micro-scale attention head to capture the instantaneous surge characteristics. The prediction results are directly used to trigger the immediate response of the surge protection unit to ensure that the equipment is protected from sudden surge damage.

[0075] The second level prediction is used to predict the voltage and current change trends in the medium time window in the future. The outputs of the micro- and meso-scale attention heads are integrated to provide a medium-term prediction of surge characteristics. The prediction results are used to coordinate the action timing of multi-level protection devices and balance the relationship between fast response and energy consumption.

[0076] The third-level prediction is used to predict the voltage and current changing trends within the longest time window in the future. It integrates the outputs of micro-, meso-, and macro-scale attention heads to provide long-term predictions of surge propagation and attenuation characteristics. The prediction results are used to dynamically optimize protection strategy parameters, improve the system's ability to respond to continuous surge events, and extend the service life of protection devices.

[0077] The prediction output layer adopts a hierarchical prediction strategy, which aims to make hierarchical predictions based on the changes in surge characteristics in different time windows, so as to achieve efficient and accurate response to surge events. This hierarchical prediction strategy optimizes the responsiveness and overall stability of the surge protection system by decomposing the changing trends of voltage and current into three levels: short-term, medium-term and long-term.

[0078] The first-level prediction focuses on voltage and current changes within the shortest time window in the future. This prediction is based on the output of the microscale attention head and quickly captures the instantaneous surge characteristics. The microscale attention head is highly sensitive to the drastic fluctuations between adjacent sampling points. Therefore, the first-level prediction can identify sudden surges that may cause damage to the equipment, such as high-frequency and short-term voltage or current peaks. The prediction results are directly used to trigger the immediate response of the surge protection unit to cut off or transfer the surge energy in the shortest time and protect the equipment from the damage of the instantaneous surge. This fast response mechanism ensures that the system can effectively protect in the first time, greatly reducing the impact of sudden surges on power equipment.

[0079] The second-level prediction focuses on the trend changes of voltage and current in the medium-term time window in the future. This prediction combines the outputs of micro- and meso-scale attention heads to generate medium-term surge feature predictions. The meso-scale attention head can analyze the characteristics of electrical signals within a single power frequency cycle. Therefore, the second-level prediction combines short-term and periodic changes to provide a more comprehensive perspective for surge protection systems. The prediction results are used to coordinate the action timing of multi-level protection devices in the system, especially when it is necessary to balance fast response and energy consumption. For example, when the system detects a continuous but not violent surge feature, the second-level prediction can delay or trigger the protection device in steps to save energy while ensuring safety, thereby optimizing the operating efficiency of the overall system.

[0080] The third-level prediction focuses on the voltage and current trends in the longest time window in the future. This level of prediction combines the outputs of micro-, meso-, and macro-scale attention heads to form a comprehensive analysis of surge propagation and attenuation characteristics. The macro-scale attention head provides observations of surge behavior over a long period of time, including energy decay and voltage fluctuation trends. Therefore, the third-level prediction can provide an accurate assessment of the long-term evolution of surges. This prediction result is used to dynamically optimize the parameters of the surge protection strategy, such as adjusting the trigger threshold and action strategy of the protection device to ensure that the system maintains efficient protection in the face of long-lasting surge events. By planning and adjusting the protection scheme in advance, the system can not only effectively manage chronic surges, but also extend the service life of the protection device and reduce the wear caused by frequent actions.

[0081] This hierarchical prediction strategy enables refined management of surge events, and each level of prediction is optimized around a specific time window to meet protection needs at different time scales.

[0082] Furthermore, the value assessment network comprehensively evaluates the protection strategy through a hierarchical assessment mechanism to optimize system performance. The hierarchical assessment mechanism includes an instantaneous effect assessment layer, a cumulative effect assessment layer, and a comprehensive assessment layer;

[0083] Among them, the instantaneous effect evaluation layer is responsible for analyzing the immediate suppression effect of the strategy, measuring the effectiveness of surge suppression by calculating the ratio of residual voltage to rated voltage, and evaluating the rationality of the distribution of discharge current among protection devices to ensure the accuracy of instantaneous protection; the cumulative effect evaluation layer analyzes the long-term impact of the strategy on the life of protection devices, calculates the load balance of each protection device, and evaluates the reasonable distribution of energy loss to balance protection performance and device durability; the comprehensive evaluation layer adaptively adjusts the weight of each indicator according to real-time surge characteristics and historical evaluation data, dynamically optimizes the strategy scoring criteria, and outputs the final strategy score to improve the stability and efficiency of the overall protection effect.

[0084] The value assessment network uses a hierarchical assessment mechanism to conduct a comprehensive evaluation of the generated protection strategy to improve the overall performance and reliability of the surge protection system. This hierarchical assessment mechanism optimizes the system's performance in dealing with instantaneous and long-term surges by analyzing the protection effects at different levels, ensuring that power equipment receives continuous and efficient protection.

[0085] First, the transient effect evaluation layer specifically analyzes the immediate suppression effect of the protection strategy when a surge event occurs. This layer quantifies the effectiveness of the protection strategy at the moment of the surge by calculating the ratio of the residual voltage to the rated voltage. The lower the residual voltage, the more effective the protection strategy is in suppressing the surge and reducing the risk of damage to the power equipment. In addition, the transient effect evaluation layer also evaluates the distribution of the discharge current among the various levels of protection devices to ensure that the current is reasonably distributed to avoid overloading a single protection device. This detailed analysis ensures that the system can respond quickly and accurately when dealing with sudden surges, minimizing the impact on the equipment.

[0086] Secondly, the cumulative effect evaluation layer focuses on the long-term impact of the protection strategy on the entire system and protection devices. Since surge protection devices may age or degrade after multiple actions, this evaluation layer analyzes the impact of the protection strategy on the life of the device over a long period of time by calculating the load balance of each protection device. Specifically, this layer tracks the frequency of use and stress level of each protection device to ensure that the protection tasks are reasonably distributed among different devices to prevent some devices from failing prematurely due to frequent use. At the same time, the cumulative effect evaluation layer evaluates the distribution of energy loss and analyzes whether the protection strategy can minimize energy consumption while ensuring the surge suppression effect. This analysis helps the system maintain efficient protection during long-term operation while extending the service life of protection devices.

[0087] Finally, the comprehensive evaluation layer integrates instantaneous and cumulative evaluation results to form a comprehensive evaluation of the overall protection strategy. This layer adaptively adjusts the weights of each evaluation indicator based on the real-time detected surge characteristics and historical evaluation data. In other words, the system dynamically optimizes the strategy scoring criteria based on the current surge environment and the historical performance of the protection effect to ensure that the score can reflect the actual protection needs of the system. The output of the comprehensive evaluation layer is the final strategy score, which is used to guide the system to adjust and optimize the protection strategy to make the overall protection effect more stable and efficient. Through this dynamic adjustment process, the system can continuously adapt to the complex and changing surge environment and always maintain the best protection performance.

[0088] Furthermore, the action generation network adopts a dual optimization mechanism to coordinate short-term protection response and long-term strategy optimization;

[0089] Among them, the dual optimization mechanism includes a short-term optimization loop and a long-term optimization loop; the short-term optimization loop quickly generates an immediate protection strategy according to the current surge characteristics, and optimizes the trigger timing of protection devices at all levels to achieve rapid response and precise coordination to sudden surges; the long-term optimization loop analyzes historical protection effects, adjusts baseline strategy parameters, balances short-term protection effects and the service life of protection devices, and achieves continuous improvement and optimization of strategies based on accumulated system performance.

[0090] The action generation network uses a dual optimization mechanism to coordinate the short-term rapid response and long-term strategy optimization of the surge protection system to improve the efficiency and reliability of the overall system. The dual optimization mechanism is designed to enable the system to respond quickly when a transient surge event occurs, while continuously improving the protection strategy based on historical data and long-term performance analysis to ensure the continued efficient operation of the system.

[0091] The short-term optimization loop is specifically designed to make quick decisions based on the current surge characteristics and generate an immediate protection strategy. This loop obtains real-time surge characteristic parameters such as voltage change rate and current peak value from the artificial intelligence analysis unit, and uses these data to quickly calculate the trigger timing of each level of protection devices. The optimization of the trigger timing is designed to enable each protection device to operate at the most appropriate time, thereby providing accurate surge protection. For example, when a sudden high-energy surge is detected, the short-term optimization loop will immediately adjust the trigger threshold of the gas discharge tube and varistor to quickly discharge excess energy and protect power equipment from damage. This fast response mechanism ensures that the system can respond within milliseconds by accurately coordinating the actions of each protection device, significantly improving the protection effect in the event of a sudden surge.

[0092] The long-term optimization loop focuses on the continuous improvement and overall optimization of the protection strategy. This loop adjusts the baseline strategy parameters by analyzing the historical protection effects of the system, including the operating frequency, service life, and energy consumption of each protection device, to optimize the performance of the system in long-term operation. Specifically, the long-term optimization loop will conduct a comprehensive evaluation based on the cumulative system performance, balancing the short-term protection effect and the durability of the protection device. For example, if a protection device frequently operates in a short period of time, the system will adjust the strategy to reduce the frequency of use of the device, thereby extending its life. This long-term adjustment can not only reduce the wear of the device, but also optimize the energy consumption of the entire system and improve the reliability and efficiency of the protection device.

[0093] Through the collaborative work of these two optimization loops, the action generation network can dynamically adjust the protection strategy, providing fast and efficient protection in the event of a sudden surge, and continuously optimizing the strategy based on long-term data analysis to keep the system stable during long-term operation. The dual optimization mechanism can also adaptively learn and adjust the protection strategy to cope with the ever-changing power environment and different types of surge characteristics. This design makes the system highly adaptable and intelligent, and can meet various complex surge protection needs.

[0094] Furthermore, the reward function of the reinforcement learning adopts the following formula 1:

[0095]

[0096] Among them, w 1 、w 2 、w 3 、w 4 is the weight parameter; R 1 Rating for surge suppression effectiveness; R 2 To protect the service life of the device; P 1 is the energy consumption penalty term; R 3 Reward for load stability;

[0097] The surge suppression effect score is achieved using the following formula 2:

[0098]

[0099] Among them, V res is the residual voltage, which indicates the voltage amplitude that the power equipment is subjected to after the surge protection is activated; V th is the preset safety voltage threshold; 1 is the exponential decay factor;

[0100] The protection device service life score is implemented using the following formula 3:

[0101]

[0102] Among them, N act N is the cumulative number of actions, indicating the total number of times the protection device has been triggered; max is the maximum operating life of the device; β is the exponential decay factor; γ is the decay coefficient of time t;

[0103] The energy consumption penalty term is implemented using the following formula 4:

[0104]

[0105] Among them, E cons Indicates the total energy consumed by the protection device during protection action; E base is an energy reference value used for normalization; μ is the adjustment factor; ω is the energy consumption fluctuation frequency coefficient;

[0106] The load stability reward is implemented using the following formula 5:

[0107]

[0108] Where, ΔV load Indicates the fluctuation of load voltage after surge protection is activated; ΔI load It indicates the fluctuation of load current after the surge protection is activated; k is the adjustment parameter.

[0109] The reinforcement learning reward function uses a complex and sophisticated calculation method to comprehensively balance the surge suppression effect, protection device life, energy consumption, and load stability. The reward function optimizes the performance of the overall surge protection system through multiple combinations and adjustment weights to ensure that the system can provide the best protection under various operating conditions. The following is a detailed explanation of the formula and each of its components.

[0110] The overall definition of the reward function is as follows:

[0111]

[0112] Weight parameter w 1 , w 2 , w 3 and w 4 :

[0113] These weight parameters are used to adjust the contribution of each score or penalty item in the total reward. These parameters can be adjusted dynamically according to the system's priority for different performance indicators. For example, increasing w 1 This will make the surge suppression effect account for a larger proportion of the total score and reduce w 3 Will reduce the negative impact of energy consumption.

[0114] Surge suppression effectiveness rating R1 Use the following formula:

[0115]

[0116] V res :This is the residual voltage, which indicates the voltage amplitude that the power equipment is subjected to after the surge protection is activated. It is obtained through real-time monitoring of the protection device and reflects the quality of the protection effect. The smaller the residual voltage, the more effectively the surge is suppressed.

[0117] V th :The preset safe voltage threshold represents the maximum voltage that the electrical equipment can safely withstand. This value is usually determined based on the voltage resistance performance of the equipment.

[0118] λ 1 : Exponential decay factor used to adjust the effect of residual voltage on the score. Larger λ 1 Will cause R 1 It approaches 1 quickly at smaller residual voltages, emphasizing the importance of low residual voltage.

[0119] Protection device service life rating R 2 Use the following formula:

[0120]

[0121] N act : Cumulative number of actions, indicating the total number of times the protection device has been triggered. This parameter reflects the intensity of use of the protection device and is usually recorded by the controller's counter.

[0122] N max :The maximum operating life of the device, that is, the maximum number of triggers that the protection device can withstand. This is a fixed value provided by the device manufacturer for life management.

[0123] β: Exponential decay factor, controlling the effect of the number of actions on the score. A larger β will increase the adverse effect of high action frequency on the score.

[0124] γ: Time attenuation coefficient, reflecting the trend of device performance gradually decreasing over time. The time represented by t can be the total operating time of the device or the time since the last maintenance.

[0125] In the formula, R 2 The device’s operating frequency and time effects are taken into account, ensuring that protection devices that are frequently used or operate for a long time have lower scores, reflecting their potential risk of aging or loss.

[0126] Energy consumption penalty term P 1 The calculation is as follows:

[0127]

[0128] E cons : The total energy consumed by the protection device in each protection action, usually calculated by the energy monitoring module, reflects the energy cost of the action.

[0129] E base : Energy baseline value, used to normalize energy consumption for comparison at different energy levels. This value can be determined based on the average energy consumption of the system.

[0130] μ: A modifier that controls the overall impact of the energy consumption penalty. A larger μ will amplify the negative impact of energy consumption on the total reward.

[0131] ω: Energy consumption fluctuation frequency coefficient, used to introduce periodic fluctuations and simulate the nonlinear characteristics of actual energy consumption.

[0132] The formula reflects the adverse impact of energy consumption on the overall performance of the system by normalizing energy and smoothing fluctuation penalties. Higher energy consumption will significantly increase the penalty, pushing the system to optimize energy usage.

[0133] The load stability bonus is calculated as follows:

[0134]

[0135] ΔV load :The fluctuation amplitude of the load voltage after the surge protection is activated is obtained through real-time voltage monitoring and reflects the voltage stability. Large fluctuations mean that the voltage is unstable and affects the normal operation of the equipment.

[0136] ΔI load : The fluctuation amplitude of the load current after the surge protection is activated is also obtained through real-time monitoring, indicating the impact of current fluctuations on the equipment.

[0137] κ: A tuning parameter that controls the effect of voltage and current fluctuations on the load stability score. A larger κ will significantly reduce the bonus score for fluctuations.

[0138] The first part of the formula reduces the impact of fluctuation amplitude through normalization and square penalty mechanism, and the second part uses the hyperbolic tangent function tanh to adjust the difference between voltage and current fluctuations, optimize load stability analysis, and ensure that the system maintains a stable operating state after surge protection.

[0139] To train a deep neural network model and ensure that it is capable of handling surge protection tasks, you need to follow the following step-by-step process to optimize the model parameters using input data and continuously improve the protection performance through reinforcement learning strategies. The following are the detailed training steps, including data preparation, forward propagation, loss calculation, strategy optimization, and model evaluation.

[0140] First, prepare the training data. The data should include a variety of surge characteristic parameter samples, which can be collected by actual power equipment under different surge conditions. The data needs to cover characteristics such as voltage change rate, voltage peak, current change rate, current peak, etc., and provide corresponding labels or reward values ​​for reward calculation in reinforcement learning. In the data preprocessing stage, the original features need to be normalized to ensure the numerical stability of the data and facilitate effective model training.

[0141] At the beginning of training, the model receives a batch of input feature data, which is first normalized. The normalization step uses a piecewise adaptive normalization method to map the features to a uniform numerical scale to reduce the impact of changes in the dynamic range of the input data on the model. Subsequently, the features are mapped through a multi-layer perceptron structure to gradually transform low-dimensional features into high-dimensional feature representations. This process uses nonlinear activation functions (such as ReLU) to enhance the expressiveness of the model.

[0142] Next, the mapped features enter the self-attention mechanism layer. This mechanism calculates the similarity between feature vectors and generates an attention weight matrix to focus on key features, improving the model's sensitivity and understanding of surge features. Through the self-attention mechanism, the model can flexibly capture the complex dependencies between features and generate fused feature vectors. These fused features are then input into the surge development predictor.

[0143] The surge development predictor uses LSTM (Long Short-Term Memory Network) to extract time series features and analyze the trend of input features over time. This step is particularly important because surge phenomena usually have significant time correlation. The time series feature vector output by the LSTM layer is further processed by the multi-scale attention mechanism. The model can analyze the short-term violent fluctuations and long-term trend changes of surge characteristics according to different time scales, providing more comprehensive prediction information for protection strategy generation.

[0144] The protection strategy generator generates protection strategies using fused features and prediction results. First, the state representation network maps the input features into a state vector, which represents the current protection state of the power equipment. Then, the action generation network generates the trigger threshold adjustment strategy for each level of protection devices based on the state vector, ensuring that the surge protection device can quickly respond appropriately in different surge scenarios. The value evaluation network evaluates the generated strategy, calculates the possible benefits or losses of the strategy in the future, and helps the model optimize decisions in reinforcement learning.

[0145] The reinforcement learning training process includes policy optimization and model update. By interacting with the environment, the model collects the current state, takes actions, obtains rewards, and observes the next state in each surge event. Based on these interaction data, the model calculates the value of each policy and evaluates the policy according to the reward function. The reward function comprehensively considers the surge suppression effect, protection device life, and energy consumption, balancing short-term and long-term goals.

[0146] During the back-propagation phase, the loss function is used to calculate the difference between the policy value and the target value. The target value is a discounted sum of the current reward and future value, calculated using the Bellman equation. The model uses an optimizer (such as Adam) to minimize the loss and update the parameters to improve the effectiveness and reliability of the protection strategy. The entire training process is repeated until the model demonstrates stable and excellent protection capabilities in all surge scenarios.

[0147] After training, the model is tested using an independent validation dataset to evaluate its response and protection effect under different surge conditions. The model performance can be further optimized by continuously adjusting hyperparameters such as learning rate, number of hidden units of LSTM, and number of attention heads. Those skilled in the art can adjust the model architecture and training strategy according to actual application requirements and data characteristics to adapt to specific power equipment protection tasks.

[0148] The programmable protection circuit 103 includes a first-level gas discharge tube, a second-level metal oxide varistor, and a third-level transient voltage suppression diode connected in parallel between the live wire and the protective ground wire; the trigger voltage of the first-level gas discharge tube can be adjusted within a first voltage range to discharge high-energy surges including lightning strikes; the threshold voltage of the second-level metal oxide varistor can be adjusted within a second voltage range to suppress medium-energy surges; the clamping voltage of the third-level transient voltage suppression diode can be adjusted within a third voltage range to limit the residual voltage to the ground.

[0149] The programmable protection circuit 103 is a core module for responding to surge events and protecting power equipment. It adopts a multi-level protection mechanism, including gas discharge tubes, metal oxide varistors and transient voltage suppression diodes, which are connected in parallel between the live wire and the protective ground wire. The module is designed to release and suppress surge energy in stages, and dynamically adjust the protection strategy according to the different intensities and durations of the surge to minimize damage to the power equipment.

[0150] In the first level of protection, the gas discharge tube is responsible for handling high-energy surges, especially strong current shocks such as lightning strikes. When the surge voltage is detected to reach or exceed the triggering threshold of the gas discharge tube, the device quickly turns on and discharges the high-energy surge directly to the protective ground wire. The triggering voltage of the gas discharge tube is adjustable and can be dynamically changed within a preset voltage range to adapt to different power environments and protection requirements. This adjustment mechanism is usually set through the protection strategy parameters provided by the controller 105, so as to optimize the release path of the surge energy and reduce the impact on the system.

[0151] The second level of protection is implemented by metal oxide varistors, which are specifically designed to deal with medium-intensity surges. Metal oxide varistors have voltage-dependent nonlinear impedance characteristics. When the voltage exceeds its set threshold, the impedance is significantly reduced, thereby suppressing the surge voltage and absorbing the surge energy to protect subsequent equipment from damage. Its threshold voltage can also be adjusted within the second voltage range to ensure that it can effectively play a protective role under different surge conditions. By adjusting the threshold, unnecessary triggering can be avoided, the life of the protection device can be extended, and the overall stability of the system can be improved.

[0152] In the third level of protection, the transient voltage suppressor diode is mainly responsible for limiting the residual voltage after the surge and protecting the precision circuits of the power equipment from the impact of transient voltage. The transient voltage suppressor diode has a fast response characteristic. When the surge voltage exceeds its clamping voltage, it immediately turns on and clamps the excess voltage within a safe level. The clamping voltage can be adjusted within the third voltage range to accurately control the level of the residual voltage and ensure the best protection for sensitive equipment. The adjustment mechanism of the transient voltage suppressor diode relies on the surge feature analysis results obtained from the artificial intelligence analysis unit 102, so as to achieve real-time response and optimized protection.

[0153] The entire programmable protection circuit 103 realizes multi-level division of labor through a parallel structural design. Each protection device is triggered in turn when a surge of different strength is detected, forming an efficient surge energy hierarchical release strategy. The mutual cooperation of gas discharge tubes, metal oxide varistors and transient voltage suppression diodes can not only effectively disperse surge energy, but also intelligently adjust protection parameters according to surge characteristics, providing a flexible and reliable protection mechanism. In addition, this hierarchical protection strategy can effectively reduce the cumulative loss of protection devices and improve the long-term stability and reliability of the entire surge protection system.

[0154] The resettable fuse 104 is connected in series between the live wire and the power device. When the temperature exceeds a first temperature threshold, the resistance increases to cut off the circuit. When the temperature drops below a second temperature threshold, the fuse automatically resumes conduction.

[0155] The resettable fuse 104 plays a key protective role in the present invention. It is designed to prevent power equipment from being damaged due to overheating or overcurrent, and automatically restores circuit function after environmental conditions return to normal, thereby improving the reliability and sustainability of the entire surge protection system. The module adopts an overcurrent protection mechanism that integrates the characteristics of automatic recovery, significantly reducing the need for manual maintenance.

[0156] Specifically, the resettable fuse 104 is connected in series between the live wire and the electrical equipment to provide protection when the current or equipment temperature reaches a dangerous level. When the current in the circuit exceeds the preset safety limit, or the temperature around the equipment rises to a first temperature threshold, the resistance of the resettable fuse will increase significantly, thereby quickly cutting off the circuit. This resistance change characteristic is achieved through special material design, usually composed of positive temperature coefficient (PTC) materials, which rapidly increase resistance when the temperature rises to protect the electrical equipment from overcurrent or overheating.

[0157] In actual operation, the triggering mechanism of the resettable fuse 104 is based on the thermal effect. When the current increases abnormally, or the temperature of the electrical component rises above the set first temperature threshold, the fuse material will sense this heat change and produce a rapid physical response, causing the resistance to increase sharply to a preset value. This increase in resistance will limit or completely prevent the flow of current, cut off the connection between the electrical equipment and the power supply, and effectively prevent the occurrence of dangerous situations such as equipment damage or fire.

[0158] Once the abnormal situation is under control, for example, the circuit temperature drops below the second temperature threshold, the resettable fuse 104 will gradually cool down and restore its low resistance state. At this time, the fuse automatically turns on again, allowing the circuit to return to normal working state. This recovery process does not require manual resetting or replacement of the fuse, greatly improving the automation and convenience of the surge protection system. In addition, the resettable fuse still maintains its protection function after repeated triggering, has good durability and long service life, and is suitable for protection of multiple surge events.

[0159] In order to ensure the reliability and stability of the resettable fuse 104 under various conditions, its design takes into account a variety of factors, including the thermal stability, durability and reaction speed of the material under different current and temperature conditions. The material selection and processing technology need to ensure that it will not degrade or fail in a high temperature environment, while maintaining a fast response time to provide effective overcurrent protection.

[0160] This automatic recovery function significantly reduces equipment downtime after a surge event, while also lowering the need for manual maintenance and replacement, making the power equipment protection system highly reliable and long-lasting.

[0161] The controller 105 is used to receive the protection strategy parameters output by the artificial intelligence analysis unit; and adjust the trigger thresholds of various levels of protection devices in the programmable protection circuit according to the received protection strategy parameters.

[0162] The controller 105 is the core decision-making and execution module, which is responsible for dynamically adjusting the trigger thresholds of various levels of protection devices in the programmable protection circuit 103 according to the protection strategy parameters output by the artificial intelligence analysis unit 102, so as to realize intelligent surge protection for power equipment. This module ensures that the system can respond quickly to surge events, optimize protection strategies, and adjust the working status of various protection devices in real time through the precise collaboration of a series of hardware and software components.

[0163] The controller 105 is connected to the artificial intelligence analysis unit 102 via a high-speed data interface to receive and process the protection strategy parameters generated by the deep neural network model. These parameters include specific settings for the trigger thresholds of different protection devices (such as gas discharge tubes, metal oxide varistors, and transient voltage suppression diodes) in the programmable protection circuit 103. The hardware architecture of the controller usually includes a high-performance processor, such as a microcontroller unit (MCU) or a digital signal processor (DSP), for performing complex protection strategy adjustment operations. The high-speed computing unit built into the processor can parse and execute instructions from the artificial intelligence analysis unit in real time to ensure the response speed and operating efficiency of the system.

[0164] The function realization of the controller 105 also relies on a set of precise output control circuits. This set of circuits is responsible for converting the digital signals generated by the processor into control signals to adjust the various protection devices in the programmable protection circuit 103. For example, when it is necessary to adjust the trigger voltage of the gas discharge tube, the controller will output the corresponding adjustment signal to change the electrical parameters of the gas discharge tube so that it can quickly respond to high-energy surges. Similarly, the controller can adjust the thresholds of the metal oxide varistor and the transient voltage suppression diode according to the current surge characteristic parameters, so that the system can provide optimal protection under various surge conditions.

[0165] The controller 105 is equipped with a sophisticated feedback monitoring mechanism to ensure that all instructions can be accurately executed and monitor the status of the protection devices in real time during the execution process. This feedback system continuously collects the working status and circuit parameters of each protection device and feeds this information back to the controller for it to adjust the strategy or issue a warning signal. If it is detected that a protection device is not operating as expected or an abnormal situation occurs, the controller can quickly adjust the protection plan or switch to an alternative protection strategy when necessary to maintain the overall stability and safety of the system.

[0166] In order to enhance the intelligence of the system, the controller 105 also supports a parameter adaptive optimization function. This function continuously learns and optimizes the trigger threshold settings of each level of protection devices based on historical surge data and current surge characteristics, thereby improving the long-term protection effect of the system. This process is implemented by the controller's software algorithm, combined with the deep learning results of the artificial intelligence analysis unit 102, to improve the overall performance through continuous iteration of protection strategies.

[0167] The controller 105 also has a built-in fault diagnosis function for detecting and recording any abnormal conditions in the system, including sensor failure, protection device failure, or data transmission error. The diagnosis results can be stored in a local storage unit and, when necessary, notified to maintenance personnel through a communication interface, or trigger a system self-check process to ensure system reliability and maintainability.

[0168] Furthermore, the programmable protection circuit also includes a plurality of bidirectional thyristor switches and a self-recovery isolation circuit; wherein a bidirectional thyristor switch is connected in parallel at both ends of each protection device to control the access timing of the corresponding protection device;

[0169] The self-recovery isolation circuit is connected in series between the common connection point of each protection device and the bidirectional thyristor switch and the protection ground line, and the self-recovery isolation circuit includes a fast fuse, which is quickly disconnected when the discharge current exceeds the preset current threshold; a parallel gas discharge tube, which provides a discharge channel after the fast fuse is disconnected; a metal oxide varistor is connected in series with the fast fuse to limit the discharge current;

[0170] The control end of the bidirectional thyristor is connected to the controller, and is specifically used to detect the initial stage of the surge, first turn on the bidirectional thyristor of the third-level transient voltage suppression diode; when the discharge current exceeds the first preset value, turn on the bidirectional thyristor of the second-level metal oxide varistor; when the discharge current exceeds the second preset value, turn on the bidirectional thyristor of the first-level gas discharge tube; when the self-recovery isolation circuit of the designated protection device is activated, automatically switch to other levels of protection channels;

[0171] The self-recovery process of the self-recovery isolation circuit includes: when the fast fuse is disconnected, the current automatically switches to the gas discharge tube path; after the surge ends, the metal oxide varistor connected in series limits the recovery current; in the current limiting state, the fast fuse gradually cools down and restores the conduction characteristics; the system automatically returns to the normal protection state.

[0172] The programmable protection circuit is uniquely designed, combining multiple bidirectional thyristors and self-recovery isolation circuits to achieve precise and dynamic surge protection. Through intelligent control and self-recovery functions, this design not only effectively responds to surge events, but also reduces system downtime, significantly improving the overall protection effect and long-term reliability of the system.

[0173] The programmable protection circuit contains multiple bidirectional thyristors, with a bidirectional thyristor connected in parallel at both ends of each protection device to control the access timing of the protection device. A bidirectional thyristor is a semiconductor device that can conduct in both directions and accurately manage the opening and closing of current flow through control signals. Its control end is connected to the controller, which determines which protection devices need to be turned on and when according to the surge characteristic parameters, thereby achieving a graded response to surge events.

[0174] The specific working principle is: when the initial stage of the surge is detected, the controller first turns on the bidirectional thyristor switch of the third-stage transient voltage suppression diode (TVS). This step is designed to immediately provide primary surge suppression to prevent power equipment from being affected by high transient voltages. If the surge continues to intensify and the discharge current exceeds the preset first current threshold, the controller will immediately turn on the bidirectional thyristor switch of the second-stage metal oxide varistor (MOV) to further absorb and suppress medium-intensity surges. When the discharge current exceeds the higher second preset value, the controller will turn on the bidirectional thyristor switch of the first-stage gas discharge tube to quickly release the high-energy surge to the ground, thereby protecting the power equipment from serious damage.

[0175] At the same time, in order to deal with overcurrent in extreme cases, the self-recovery isolation circuit provides an additional protection mechanism. Each self-recovery isolation circuit is connected in series between the common connection point between the protection device and the bidirectional thyristor switch and the protective ground wire, and the structure includes a fast fuse, a parallel gas discharge tube and a series metal oxide varistor. The fast fuse is designed to quickly disconnect the circuit when the discharge current exceeds the preset current threshold to prevent the protection device from overloading and damage. Once the fast fuse is disconnected, the current automatically switches to the gas discharge tube path, and the gas discharge tube provides a safe surge discharge channel to prevent system failure.

[0176] When the surge event ends, the circuit enters the self-recovery process. The metal oxide varistor in series limits the recovery current at this stage, ensuring that the current gradually returns to a safe level while protecting the equipment from the impact of the recovery current. The fast fuse gradually cools down and restores its conduction characteristics under the current limiting state, which enables the system to automatically return to the normal protection state without manual intervention. The entire self-recovery process is precisely designed to ensure that the surge protection system can recover quickly while providing efficient protection, reducing system downtime and improving equipment availability.

[0177] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. An artificial intelligence-based self-repairing surge protection system for power equipment, characterized in that: include: A surge detection module, which is connected between the live wire and the protective ground wire of the power equipment and is used to collect voltage data and current data between the live wire and the protective ground wire according to a preset period; and transmit the collected data to the artificial intelligence analysis unit; An artificial intelligence analysis unit, used to extract surge characteristic parameters from the data transmitted by the surge detection module, wherein the surge characteristic parameters include a voltage change rate, a voltage peak value, and a voltage anomaly duration of adjacent sampling points, and a current change rate, a current peak value, and a current anomaly duration of adjacent sampling points; Based on the extracted surge characteristic parameters, a pre-trained deep neural network model is used to generate protection strategy parameters of the protection device in the programmable protection circuit, wherein the protection strategy parameters include trigger thresholds of various levels of protection devices in the programmable protection circuit; A programmable protection circuit includes a first-stage gas discharge tube, a second-stage metal oxide varistor, and a third-stage transient voltage suppression diode connected in parallel between a live wire and a protective ground wire; the trigger voltage of the first-stage gas discharge tube can be adjusted within a first voltage range to discharge high-energy surges including lightning strikes; the threshold voltage of the second-stage metal oxide varistor can be adjusted within a second voltage range to suppress medium-energy surges; The clamping voltage of the third-stage transient voltage suppression diode can be adjusted within a third voltage range to limit the residual voltage to ground; A resettable fuse is connected in series between the live wire and the power equipment. When the temperature exceeds a first temperature threshold, the resistance increases to cut off the circuit, and when the temperature drops below a second temperature threshold, the conduction is automatically restored; The controller is used to receive the protection strategy parameters output by the artificial intelligence analysis unit; and adjust the trigger thresholds of various levels of protection devices in the programmable protection circuit according to the received protection strategy parameters.

2. The artificial intelligence-based self-repairing surge protection system for power equipment according to claim 1 is characterized in that: The deep neural network model used in the artificial intelligence analysis unit includes a surge feature encoder, a surge development predictor, and a protection strategy generator connected in sequence; Among them, the surge feature encoder is implemented by a multi-layer perceptron structure; the surge feature encoder includes a normalization layer, a feature mapping layer and a feature fusion layer; the normalization layer is used to normalize the extracted surge feature parameters to obtain normalized feature parameters; the feature mapping layer maps the normalized feature parameters to a feature space of a first preset dimension through a multi-layer fully connected network to obtain multiple feature vectors of the first preset dimension; the feature fusion layer uses a self-attention mechanism to calculate the weight relationship between multiple feature vectors of the first preset dimension to obtain a fused feature vector of a second preset dimension; The surge development predictor is implemented using a recurrent neural network structure; the surge development predictor includes a timing coding layer, a trend analysis layer and a prediction output layer; wherein the timing coding layer uses a long short-term memory network to extract timing features from a fused feature vector sequence of a second preset dimension of the most recent first preset number of time points to obtain a timing feature vector of a third preset dimension; the trend analysis layer uses multiple parallel attention heads to focus on the changing trends of multiple different time scales based on the timing feature vector of the third preset dimension to obtain multiple trend feature vectors of a fourth preset dimension; the prediction output layer concatenates multiple trend feature vectors of the fourth preset dimension and processes them through a fully connected network to obtain voltage and current prediction values ​​at each preset sampling moment in the future second preset time period; The protection strategy generator is implemented by reinforcement learning, and the protection strategy generator includes a state representation network, an action generation network and a value evaluation network; wherein the state representation network is used to map the fused feature vector provided by the surge feature encoder and the prediction result provided by the surge development predictor into a state vector; the action generation network is used to generate the trigger threshold adjustment strategy of each level of protection devices based on the state vector; the value evaluation network is used to evaluate the expected effect of the generated protection strategy; wherein the state space of the reinforcement learning includes the current surge feature, the surge prediction value, and the working state of the protection device; wherein the action space of the reinforcement learning is the trigger threshold adjustment amount of each level of protection devices; the working state of the protection device includes the currently set trigger threshold, the conduction state and the accumulated number of actions; the reward function of the reinforcement learning comprehensively considers the surge suppression effect, the device service life and the energy loss.

3. The artificial intelligence-based self-repairing surge protection system for power equipment according to claim 2 is characterized in that: The normalization layer adopts a segmented adaptive normalization method to cope with the dynamic change environment of voltage and current characteristic parameters, specifically including: Based on the physical characteristics of voltage and current characteristic parameters, the value range of each characteristic parameter is divided into multiple dynamic intervals, and the initial statistical parameters in each dynamic interval are defined, including mean and variance; Maintain a set of dynamic statistical parameters for each dynamic interval, update the mean and variance within the interval through the exponential moving average method, and adjust the interval range according to the distribution changes of the characteristic parameters; When a surge event detects a sharp fluctuation in voltage or current, the update weight of the corresponding dynamic interval is adjusted, and the update frequency of the data in the corresponding interval is increased; Standardization is performed based on the statistical parameters of the interval to which the current characteristic parameters belong, so as to maintain the numerical stability and electrical characteristic consistency of the normalized results and improve the response efficiency of the deep neural network to different surge characteristics.

4. The artificial intelligence-based self-repairing surge protection system for power equipment according to claim 3 is characterized in that: The self-attention mechanism of the feature fusion layer adopts a hierarchical weighted fusion structure to fully explore the deep correlation between voltage and current features, including: The first layer of attention calculates the association weights between voltage features, including: The self-attention mechanism is used to calculate the similarity between the three feature vectors of voltage change rate, voltage peak and voltage duration; Generate an attention weight matrix of voltage features based on similarity, and adaptively weight them according to the importance of features to obtain a voltage comprehensive feature vector; By embedding position coding, the time series information of voltage features is retained, further optimizing the feature fusion effect; The second layer of attention calculates the association weights between current features, including: The self-attention mechanism is used to calculate the similarity between the three feature vectors of current change rate, current peak value and current duration; Generate an attention weight matrix of current features based on similarity, and adaptively weight them according to the mutual correlation of current features to obtain a comprehensive current feature vector; Through the multi-scale attention mechanism, the model's sensitivity to the changing trends of different current characteristics is enhanced; The third level of attention integrates the voltage comprehensive features and the current comprehensive features, including: Comprehensively consider the phase difference between voltage characteristics and current characteristics, and adjust the fusion weight according to the power factor characteristics to optimize power quality analysis; Use a bidirectional attention mechanism to model the interdependence between voltage and current comprehensive features and generate the final fused feature vector; The final fused feature vector retains the physical correlation between voltage and current to improve the deep neural network's understanding and prediction accuracy of surge characteristics.

5. The artificial intelligence-based self-repairing surge protection system for electric power equipment according to claim 4 is characterized in that: The trend analysis layer adopts a multi-scale progressive attention mechanism to hierarchically focus on the changes of different time scales of surge characteristics, including: Micro-scale attention heads are used to focus on high-frequency changes between adjacent sampling points and detect instantaneous and violent fluctuations in voltage and current; analyze instantaneous change characteristics, such as voltage peaks and rapid current fluctuations, to capture the sudden behavior of surges; and provide instant response surge prediction information to deal with short-term high-energy surge events; Mesoscale attention heads are used to focus on the changes in electrical signals within a single power frequency cycle, identify the periodic modulation characteristics of surges on power frequency signals; analyze the amplitude, phase, and duration characteristics of voltage and current within a power frequency cycle; predict the mid-term impact of surges and provide data support for coordinating the actions of protection devices; The macro-scale attention head is used to focus on the overall trend changes within multiple power frequency cycles and analyze the energy attenuation and propagation characteristics of surges. It observes the slow changes of voltage and current over a long time window to evaluate the long-term impact of surges on power equipment. It is used to optimize protection strategies to ensure the stability and protection performance of the system over a long period of time.

6. The artificial intelligence-based self-repairing surge protection system for electric power equipment according to claim 5 is characterized in that: The prediction output layer adopts a hierarchical prediction strategy to perform hierarchical prediction in different time windows to enhance the accurate response to surge characteristics, specifically including: The first-level prediction is used to predict the voltage and current change trends within the shortest time window in the future. The prediction is based on the output of the micro-scale attention head to capture the instantaneous surge characteristics. The prediction results are directly used to trigger the immediate response of the surge protection unit to ensure that the equipment is protected from sudden surge damage. The second level prediction is used to predict the voltage and current change trends in the medium time window in the future. The outputs of the micro- and meso-scale attention heads are integrated to provide a medium-term prediction of surge characteristics. The prediction results are used to coordinate the action timing of multi-level protection devices and balance the relationship between fast response and energy consumption. The third-level prediction is used to predict the voltage and current changing trends within the longest time window in the future. It integrates the outputs of micro-, meso-, and macro-scale attention heads to provide long-term predictions of surge propagation and attenuation characteristics. The prediction results are used to dynamically optimize protection strategy parameters, improve the system's ability to respond to continuous surge events, and extend the service life of protection devices.

7. The artificial intelligence-based self-repairing surge protection system for electric power equipment according to claim 2, characterized in that: The value assessment network comprehensively evaluates the protection strategy through a hierarchical assessment mechanism to optimize system performance. The hierarchical assessment mechanism includes an instantaneous effect assessment layer, a cumulative effect assessment layer, and a comprehensive assessment layer; Among them, the instantaneous effect evaluation layer is responsible for analyzing the immediate suppression effect of the strategy, measuring the effectiveness of surge suppression by calculating the ratio of residual voltage to rated voltage, and evaluating the rationality of the distribution of discharge current among protection devices to ensure the accuracy of instantaneous protection; the cumulative effect evaluation layer analyzes the long-term impact of the strategy on the life of protection devices, calculates the load balance of each protection device, and evaluates the reasonable distribution of energy loss to balance protection performance and device durability; the comprehensive evaluation layer adaptively adjusts the weight of each indicator according to real-time surge characteristics and historical evaluation data, dynamically optimizes the strategy scoring criteria, and outputs the final strategy score to improve the stability and efficiency of the overall protection effect.

8. The artificial intelligence-based self-repairing surge protection system for electric power equipment according to claim 2, characterized in that: The action generation network adopts a dual optimization mechanism to coordinate short-term protection response and long-term strategy optimization; Among them, the dual optimization mechanism includes a short-term optimization loop and a long-term optimization loop; the short-term optimization loop quickly generates an immediate protection strategy according to the current surge characteristics, and optimizes the trigger timing of protection devices at all levels to achieve rapid response and precise coordination to sudden surges; the long-term optimization loop analyzes historical protection effects, adjusts baseline strategy parameters, balances short-term protection effects and the service life of protection devices, and achieves continuous improvement and optimization of strategies based on accumulated system performance.

9. The artificial intelligence-based self-repairing surge protection system for electric power equipment according to claim 2, characterized in that: The reward function of the reinforcement learning adopts the following formula 1: Among them, w1, w2, w3, and w4 are weight parameters; R1 is the surge suppression effect score; R2 is the protection device service life score; P1 is the energy consumption penalty item; R3 is the load stability reward; The surge suppression effect score is achieved using the following formula 2: Among them, V res is the residual voltage, which indicates the voltage amplitude that the power equipment is subjected to after the surge protection is activated; V th is the preset safety voltage threshold; λ1 is the exponential decay factor; The protection device service life score is implemented using the following formula 3: Among them, N act N is the cumulative number of actions, indicating the total number of times the protection device has been triggered; max is the maximum operating life of the device; β is the exponential decay factor; γ is the decay coefficient of time t; The energy consumption penalty term is implemented using the following formula 4: Among them, E cons Indicates the total energy consumed by the protection device during protection action; E base is an energy reference value used for normalization; μ is the adjustment factor; ω is the energy consumption fluctuation frequency coefficient; The load stability reward is implemented using the following formula 5: Where, ΔV load Indicates the fluctuation of load voltage after surge protection is activated; ΔI load It indicates the fluctuation of load current after the surge protection is activated; k is the adjustment parameter.

10. The artificial intelligence-based self-repairing surge protection system for electric power equipment according to claim 1, characterized in that: The programmable protection circuit also includes a plurality of bidirectional thyristor switches and a self-recovering isolation circuit; wherein a bidirectional thyristor switch is connected in parallel at both ends of each protection device to control the access timing of the corresponding protection device; The self-recovery isolation circuit is connected in series between the common connection point of each protection device and the bidirectional thyristor switch and the protection ground line, and the self-recovery isolation circuit includes a fast fuse, which is quickly disconnected when the discharge current exceeds the preset current threshold; a parallel gas discharge tube, which provides a discharge channel after the fast fuse is disconnected; a metal oxide varistor is connected in series with the fast fuse to limit the discharge current; The control end of the bidirectional thyristor is connected to the controller, and is specifically used to detect the initial stage of the surge, first turn on the bidirectional thyristor of the third-level transient voltage suppression diode; when the discharge current exceeds the first preset value, turn on the bidirectional thyristor of the second-level metal oxide varistor; when the discharge current exceeds the second preset value, turn on the bidirectional thyristor of the first-level gas discharge tube; when the self-recovery isolation circuit of the designated protection device is activated, automatically switch to other levels of protection channels; The self-recovery process of the self-recovery isolation circuit includes: when the fast fuse is disconnected, the current automatically switches to the gas discharge tube path; after the surge ends, the metal oxide varistor connected in series limits the recovery current; in the current limiting state, the fast fuse gradually cools down and restores the conduction characteristics; the system automatically returns to the normal protection state.

Images