Method for inhibiting internal misfire caused by turn-to-turn short circuit of dry-type air-core reactor

Abstract

The invention provides a method for suppressing internal misfire caused by turn-to-turn short circuit of a dry-type air-core reactor, and belongs to the technical field of dry-type air-core reactors. A turn-to-turn short circuit fault detection device combining differential protection and infrared thermal image detection is established to form a three-dimensional turn-to-turn short circuit feature vector, a four-dimensional internal misfire minimum ignition vector is obtained, and a fire development prediction model is established based on an Arrhenius equation. A double-layer game model for minimizing the upper-layer fire loss and minimizing the lower-layer fire extinguishing cost is constructed to optimize five-dimensional suppression vector parameters, and coordinated actions of suppression equipment such as power-off protection, forced ventilation, temperature suppression and inert gas release are controlled in real time through a suppression vector parameter adjustment function. The technical problem of fire suppression measure starting lag or excessive response caused by difficulty in accurately predicting the opportunity and degree of development of the turn-to-turn short circuit fault of the dry-type air-core reactor into the internal fire disaster is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of dry-type air-core reactors, and in particular relates to a method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactors. Background Art

[0002] Dry-type air-core reactor is an important reactive power compensation device in power system. Its structure is as follows: Figure 6 As shown, reactors are widely used in high-voltage transmission lines, substations, and industrial power distribution systems, and their safe operation is directly related to grid stability. Traditional reactor fault detection technology relies on single methods such as differential protection, temperature monitoring, and gas detection to identify faults. These methods trigger protection actions by setting fixed thresholds and implement unified power outages or fire extinguishing measures when an anomaly is detected.

[0003] However, traditional technologies have significant drawbacks: a single monitoring method cannot fully reflect the complex fault evolution process within the reactor, fixed threshold settings cannot adapt to the changing fault characteristics under different operating conditions, and unified suppression strategies lack specificity and have limited cost control capabilities. These shortcomings lead to low fault detection accuracy, frequent false alarms and missed alarms, and inappropriate timing of suppression measures.

[0004] Current power equipment safety protection technologies lack the ability to accurately predict the transition from turn-to-turn short-circuit faults to internal fires. This makes it difficult to implement precise preventative suppression measures before a fire occurs, nor can it dynamically adjust suppression strategy parameters based on fire development trends. Specifically, existing technologies struggle to accurately predict the timing and extent of the progression of turn-to-turn short-circuit faults in dry-type air-core reactors into internal fires, leading to delayed or overly responsive fire suppression measures. Summary of the Invention

[0005] In view of this, the present invention provides a method for suppressing internal fire caused by inter-turn short circuit of dry-type air-core reactor, which can solve the technical problem in the prior art that it is difficult to accurately predict the timing and extent of the development of inter-turn short circuit fault of dry-type air-core reactor into internal fire, resulting in delayed activation or excessive response of fire suppression measures.

[0006] The present invention is implemented as follows: The present invention provides a method for suppressing internal fire caused by inter-turn short circuit of dry-type hollow reactor, comprising: establishing an inter-turn partition monitoring device of dry-type hollow reactor, dividing each encapsulation layer of the reactor into multiple monitoring partitions according to radial and axial positions, and installing an optical fiber current sensor in each partition for real-time monitoring of inter-turn current distribution changes; constructing an inter-turn partial discharge detection device, using an ultrasonic sensor and a high-frequency current transformer to obtain partial discharge signals, and obtaining the minimum partial discharge voltage, partial discharge amplitude, partial discharge phase angle, and partial discharge frequency characteristic values ​​by measurement to form a four-dimensional inter-turn partial discharge vector; establishing an inter-turn short circuit fault detection device, using the differential protection principle to monitor the current difference of each encapsulation layer, At the same time, an infrared thermal imager is used to detect areas with abnormal surface temperature distribution to form a turn-to-turn short-circuit characteristic vector; a minimum ignition vector warning device for internal fire is established, and a distributed optical fiber temperature sensor is used to monitor the temperature gradient, an electrochemical gas sensor is used to monitor the rate of change of carbon dioxide concentration, and a laser scattering smoke detector is used to measure the visibility attenuation coefficient. Combined with the current abnormal deviation rate, the minimum ignition vector for internal fire is formed; based on a pre-trained internal fire prediction model, the fire development trend is predicted, and the probability of fire occurrence and expected fire intensity are output; a game model is used to solve and determine an optimal suppression vector; and the operating parameters of each suppression device are adjusted in real time through the optimal suppression vector parameter adjustment function.

[0007] Among them, the step of inter-turn zoning is specifically to divide the encapsulation layer into three areas: inner layer, middle layer and outer layer radially, and into three areas: upper end, middle part and lower end axially, forming nine independent monitoring zones, each zone has independent current monitoring and protection functions.

[0008] The inter-turn partial discharge vector is a four-dimensional vector including four components: minimum partial discharge voltage, partial discharge amplitude, partial discharge phase angle, and partial discharge frequency characteristic value. The minimum partial discharge voltage represents the minimum voltage value at which the insulating medium begins to generate partial discharge.

[0009] Among them, the internal fire minimum ignition vector is a four-dimensional vector, which includes four components: temperature gradient, CO2 concentration change rate, visibility attenuation coefficient, and current abnormal deviation rate. The temperature gradient represents the rate of change of temperature difference between different positions inside the reactor.

[0010] Among them, the inter-turn short-circuit characteristic vector is a three-dimensional vector, which includes three components: the current difference between adjacent encapsulation layers, the surface temperature distribution anomaly coefficient, and the short-circuit ring magnetic field strength. The current difference between adjacent encapsulation layers represents the degree of difference in current distribution between adjacent encapsulation layers.

[0011] Among them, the specific structure of the internal fire prediction model is a time series prediction neural network based on the Transformer architecture, which includes four main components: input encoding layer, multi-head attention layer, feedforward neural network layer, and output decoding layer.

[0012] The game model includes an upper-level game model with the goal of minimizing fire losses and a lower-level game model with the goal of minimizing fire extinguishing costs.

[0013] The inhibition vector is a five-dimensional vector, which includes five component parameters: power-off time window, forced ventilation power, temperature control threshold, inert gas release flow, and coolant injection pressure. The power-off time window represents the time interval from fault detection to complete power off.

[0014] Among them, the suppression vector parameter adjustment function is used to adjust the operating parameters of each suppression device in real time according to the current fire status. The input includes five parameters: the minimum ignition vector of internal fire, the inter-turn partial discharge vector, the inter-turn short circuit characteristic vector, the current device status, and the predicted fire intensity.

[0015] Among them, after the step of adjusting the operating parameters of each suppression device in real time through the suppression vector parameter adjustment function, it also includes starting the gas fire extinguishing device when the aforementioned suppression measures cannot effectively control the development of the fire, and releasing inert gas into the reactor to reduce the oxygen concentration to below 12%.

[0016] Furthermore, it also includes establishing a fire development prediction model based on the thermochemical reaction mechanism, and using the Arrhenius equation to describe the thermal decomposition process of the insulating material. The Arrhenius equation is used to describe the relationship between the thermal decomposition reaction rate constant and temperature of the insulating material under high temperature conditions. The input includes five parameters: pre-exponential factor, activation energy, gas constant, insulating material temperature, and oxygen concentration.

[0017] Among them, the inter-turn partial discharge refers to the incomplete breakdown phenomenon occurring in the insulating medium between the turns of the reactor, which manifests as a charge discharge process in a local area. It usually occurs at weak points of insulation or areas of concentrated electric fields and is an early sign of insulation degradation.

[0018] The inter-turn short-circuit fault refers to the direct electrical connection formed after the insulation between adjacent turns of the reactor completely fails, resulting in large current and strong magnetic field in the short-circuit loop, which will cause serious thermal and electrodynamic effects.

[0019] The fire loss minimization function is used to calculate the minimum overall loss caused by a fire. Its inputs include four parameters: equipment damage duration caused by the fire, fire spread, suppression cost, and equipment replacement cost. Its output is the total fire loss. The fire extinguishing cost minimization function is used to calculate the minimum overall cost during the fire extinguishing process. Its inputs include five parameters: forced ventilation power consumption, temperature suppression device operating cost, equipment damage duration caused by the fire, inert gas consumption, and coolant consumption.

[0020] Among them, before the step of constructing the inter-turn partial discharge detection device, it also includes starting the partition rapid power-off protection mechanism when the inter-turn partition monitoring device detects that the fault layer current distribution deviation exceeds 20% and the lowest partial discharge voltage in the inter-turn partial discharge vector is lower than the set threshold.

[0021] Among them, the attention weight adjustment function is used to adjust the weight of the multi-head attention mechanism of the internal fire prediction model. An attention adjustment factor is calculated based on the three data of the current temperature gradient, CO2 concentration change rate, and current abnormal deviation rate. A weight adjustment function with corresponding sensitivity is adopted according to the different ranges of the attention adjustment factor.

[0022] This invention establishes a multi-dimensional fault feature vector monitoring system, constructs a machine learning-based fire development prediction model, designs a double-layer game optimization suppression strategy, and employs multi-dimensional monitoring devices such as turn-to-turn partition monitoring, partial discharge detection, short-circuit fault detection, and internal fire warning. This method comprehensively captures key characteristic parameters during the fault evolution process, overcoming the incomplete information limitations of traditional single monitoring methods. By using a Transformer-based time series prediction neural network and a multi-head attention mechanism to capture the dependencies between different time steps, accurate predictions of fire development trends are achieved, addressing the problem that traditional fixed threshold methods are unable to adapt to dynamic fault characteristics.

[0023] The two-layer game model established by the present invention can dynamically adjust the suppression vector parameters according to the predicted degree of fire development through the coordinated optimization of minimizing the upper-layer fire losses and minimizing the lower-layer fire extinguishing costs, thereby realizing the formulation of accurate suppression strategies. The technical problem in the prior art is that it is difficult to accurately predict the timing and degree of the development of the inter-turn short-circuit fault of the dry-type air-core reactor into an internal fire, which leads to delayed initiation of fire suppression measures or excessive response. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the method of the present invention.

[0025] Figure 2 Schematic diagram of the internal fire prediction model structure.

[0026] Figure 3This is a schematic diagram of the composition of the internal fire suppression monitoring equipment caused by inter-turn short circuit in dry-type air-core reactors.

[0027] Figure 4 Schematic diagram of the manually set inter-turn short-circuit ring.

[0028] Figure 5 Circuit deployment diagram used in the process of establishing the training dataset for the interior fire prediction model.

[0029] Figure 6 This is a schematic diagram of the dry-type air-core reactor structure mentioned in the background technology. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0031] like Figure 1 FIG. 1 is a flow chart of a method for suppressing internal fire caused by inter-turn short circuit in a dry-type air-core reactor provided by the present invention. The method comprises the following steps:

[0032] S01. Establish a turn-to-turn partition monitoring device for dry-type hollow reactors. Divide each reactor envelope into multiple monitoring partitions according to radial and axial positions. Install a fiber optic current sensor in each partition to monitor the changes in the turn-to-turn current distribution in real time. The fiber optic current sensor uses a flexible fiber optic sensing ring around the conductor to measure the changes in the magnetic field intensity generated by the current.

[0033] S02. Construct an inter-turn partial discharge detection device, using a combination of ultrasonic sensors and high-frequency current transformers to obtain partial discharge signals. The ultrasonic sensors are installed on the surface of the reactor housing to detect the acoustic signals generated by the discharge, and the high-frequency current transformers are installed in the main circuit to detect the high-frequency current pulses caused by the discharge. By measuring the partial discharge signals of the inter-turn insulating medium, the minimum partial discharge voltage, partial discharge amplitude, partial discharge phase angle, and partial discharge frequency characteristic values ​​are obtained to form a four-dimensional inter-turn partial discharge vector for evaluating the degree of insulation degradation.

[0034] S03. Establish a turn-to-turn short-circuit fault detection device. Use the differential protection principle to monitor the current difference between each envelope layer. When the current difference between adjacent envelope layers exceeds 5% of the current during normal operation, it is determined to be a turn-to-turn short-circuit fault. Simultaneously, use an infrared thermal imager to detect areas with abnormal temperature distribution on the reactor surface, and generate a turn-to-turn short-circuit feature vector for fault location.

[0035] S04, establish an internal fire minimum ignition vector early warning device, use distributed optical fiber temperature sensors to monitor the internal temperature gradient of the reactor, use electrochemical gas sensors to monitor the carbon dioxide concentration change rate, use laser scattering smoke detectors to measure the visibility attenuation coefficient, combine the four key parameters of current abnormal deviation rate to form the internal fire minimum ignition vector for early warning of fire;

[0036] S05, based on the mechanism of thermal chemical reaction, a fire development prediction model is established, the Arrhenius equation is used to describe the thermal decomposition process of insulating materials, the Arrhenius equation set is solved to predict the trend of fire development and optimize the suppression vector, when the fault layer current distribution deviation detected by the inter-turn partition monitoring device exceeds 20% and the lowest partial discharge voltage in the inter-turn partial discharge vector is lower than the set threshold or the inter-turn short circuit characteristic vector exceeds the fault judgment threshold, the partition quick power-off protection mechanism is started;

[0037] S06, based on the pre-trained internal fire prediction model, the trend of fire development is predicted, the internal fire prediction model receives the current temperature gradient, the carbon dioxide concentration change rate, the current abnormal deviation rate and the inter-turn short circuit characteristic vector as input parameters, and outputs the fire occurrence probability and the expected fire intensity, and adjusts the suppression vector parameters according to the prediction results;

[0038] S07, build an upper game model with the minimum fire loss as the target and a lower game model with the minimum fire extinguishing cost as the target, the upper model target function is the minimum fire loss function, and the lower model target function is the minimum fire extinguishing cost function, wherein the coupling term represents the device damage time caused by the fire, and the optimal suppression vector is determined by solving the game equilibrium, the suppression vector includes five component parameters of power-off time window, forced ventilation power, temperature control threshold, inert gas release flow and coolant injection pressure;

[0039] S08, adjust the operating parameters of each suppression device in real time through the suppression vector parameter adjustment function, the suppression vector parameter adjustment function receives the internal fire minimum ignition vector, the inter-turn partial discharge vector, the inter-turn short circuit characteristic vector, the current device state and the predicted fire intensity as input parameters, and outputs the adjusted suppression vector parameters, opens the forced ventilation device to exhaust high temperature smoke, activates the temperature suppression device to control the local temperature below 200℃ by spraying fire-retardant coolant to the fault partition, and continuously monitors the change trend of the internal fire minimum ignition vector until all parameters return to the normal range;

[0040] S09, optionally, when the above suppression measures cannot effectively control the fire development and the internal fire minimum ignition vector continues to deteriorate, start the gas fire extinguishing device as the final suppression means, release inert gas into the reactor to reduce the oxygen concentration to below 12%, and suppress the combustion reaction.

[0041] Among them, inter-turn zoning refers to dividing the encapsulation layer of the dry-type hollow reactor into three areas: inner layer, middle layer and outer layer in the radial direction, and into three areas: upper end, middle part and lower end in the axial direction, forming nine independent monitoring zones. Each zone has independent current monitoring and protection functions. The zoning is based on the internal magnetic field distribution of the reactor and the characteristics of the fault propagation path.

[0042] Among them, inter-turn partial discharge refers to the incomplete breakdown phenomenon that occurs in the insulating medium between the turns of the reactor. It manifests as a charge discharge process in a local area. It usually occurs at weak points of insulation or areas of concentrated electric fields. It is an early sign of insulation degradation and is essentially different from inter-turn short-circuit faults.

[0043] Among them, inter-turn short-circuit fault refers to the direct electrical connection formed after the insulation between adjacent turns of the reactor completely fails, resulting in large current and strong magnetic field in the short-circuit loop. It is the final stage of local discharge development and will cause serious thermal and electrodynamic effects.

[0044] Among them, the inter-turn partial discharge vector is a four-dimensional vector, which contains four components: minimum partial discharge voltage, partial discharge amplitude, partial discharge phase angle, and partial discharge frequency characteristic value. The minimum partial discharge voltage represents the minimum voltage value at which the insulating medium begins to generate partial discharge. The partial discharge amplitude reflects the discharge intensity. The partial discharge phase angle indicates the location where the discharge occurs in the AC voltage waveform. The partial discharge frequency characteristic value describes the spectral distribution characteristics of the discharge signal.

[0045] Among them, the minimum ignition vector of internal fire is a four-dimensional vector, which includes four components: temperature gradient, CO2 concentration change rate, visibility attenuation coefficient, and current abnormal deviation rate. The temperature gradient represents the rate of change of temperature difference between different positions inside the reactor. The CO2 concentration change rate reflects the growth rate of gas concentration generated in the initial stage of combustion. The visibility attenuation coefficient describes the degree of obstruction of smoke particles to light propagation. The current abnormal deviation rate represents the degree of deviation of current distribution from the normal state during a fault.

[0046] Among them, the inter-turn short-circuit characteristic vector is a three-dimensional vector, which contains three components: the current difference between adjacent encapsulation layers, the surface temperature distribution anomaly coefficient, and the short-circuit ring magnetic field strength. The current difference between adjacent encapsulation layers indicates the degree of difference in current distribution between adjacent encapsulation layers. The surface temperature distribution anomaly coefficient reflects the degree of abnormality of the reactor surface temperature distribution. The short-circuit ring magnetic field strength describes the change in local magnetic field strength caused by the short-circuit fault.

[0047] Among them, the suppression vector is a five-dimensional vector, which contains five component parameters: power-off time window, forced ventilation power, temperature control threshold, inert gas release flow rate, and coolant injection pressure. The power-off time window represents the time interval from fault detection to complete power failure. The forced ventilation power represents the power of the ventilation equipment that discharges high-temperature flue gas. The temperature control threshold represents the temperature critical value that triggers the temperature suppression device. The inert gas release flow rate represents the volume of inert gas released into the reactor per unit time. The coolant injection pressure represents the pressure value of the flame-retardant coolant injected into the fault partition.

[0048] Among them, the Arrhenius equation is used to describe the relationship between the thermal decomposition reaction rate constant of insulating materials under high temperature conditions and temperature. The input includes five parameters: pre-exponential factor, activation energy, gas constant, insulating material temperature, and oxygen concentration. The output is the thermal decomposition reaction rate constant of the material. By solving the Arrhenius equation, the decomposition rate of insulating materials and the development trend of fire under different temperature conditions can be predicted.

[0049] Among them, the fire loss minimization function is used to calculate the minimum overall loss caused by the fire. The input includes four parameters: the equipment damage time caused by the fire, the fire spread range, the suppression cost, and the equipment replacement cost. The output is the overall fire loss value. The optimization algorithm is used to solve the suppression vector parameter combination that minimizes the overall fire loss value.

[0050] Among them, the fire extinguishing cost minimization function is used to calculate the minimum overall cost in the fire extinguishing process. The input includes five parameters: forced ventilation power consumption, temperature suppression device operating cost, equipment damage time caused by fire, inert gas consumption, and coolant consumption. The output is the overall fire extinguishing cost value. The optimization algorithm is used to solve the suppression vector parameter combination that minimizes the overall fire extinguishing cost value.

[0051] Among them, the suppression vector parameter adjustment function is used to adjust the operating parameters of each suppression device in real time according to the current fire status. The input includes five parameters: the minimum ignition vector of internal fire, the inter-turn partial discharge vector, the inter-turn short-circuit characteristic vector, the current equipment status, and the predicted fire intensity. The output is the adjusted suppression vector parameter. The optimal parameter configuration is achieved by balancing the fire suppression effect and cost control through a multi-objective optimization algorithm.

[0052] The attention weight adjustment function is used to adjust the multi-head attention mechanism weight of the internal fire prediction model, and the attention weight adjustment function is calculated based on the current temperature gradient, CO2 concentration change rate and current abnormal deviation rate to obtain an attention adjustment factor. When the attention adjustment factor belongs to the range of 0 to 0.3, a low sensitivity weight adjustment function is used to adjust the multi-head attention weight parameter of the internal fire prediction model; when the attention adjustment factor belongs to the range of 0.3 to 0.7, a medium sensitivity weight adjustment function is used to adjust the multi-head attention weight parameter of the internal fire prediction model; and when the attention adjustment factor belongs to the range of 0.7 to 1.0, a high sensitivity weight adjustment function is used to adjust the multi-head attention weight parameter of the internal fire prediction model.

[0053] The internal fire prediction model has a specific structure of a time series prediction neural network based on a Transformer architecture, and includes four main components of an input encoding layer, a multi-head attention layer, a feedforward neural network layer and an output decoding layer. The input encoding layer is responsible for converting the temperature gradient, CO2 concentration change rate, current abnormal deviation rate and inter-turn short circuit feature vector into a high-dimensional feature vector. The multi-head attention layer captures the dependency between different time steps through parallel calculation of multiple attention heads. The feedforward neural network layer performs nonlinear transformation and feature extraction on the features output by the attention layer. The output decoding layer maps the processed features into two output values of fire occurrence probability and expected fire intensity. The weight parameters of the multi-head attention layer are dynamically adjusted by the attention weight adjustment function according to the current temperature gradient, CO2 concentration change rate and current abnormal deviation rate.

[0054] Among them, the steps of establishing the training data set of the internal fire prediction model specifically include collecting historical operating data of dry-type air-core reactors under different operating conditions as a basic data set, and using an artificial short-circuit ring method to simulate the short-circuit fault between the turns of the reactor to obtain experimental data. The artificial short-circuit ring production method is to cut off the 45th and 46th turns of the conductor of the third layer coil of the reactor, and then connect them with aluminum wires of the same specification to form two turns of short-circuit rings, cut off the 78th and 79th turns of the conductor of the fifth layer coil of the reactor, and connect them with aluminum wires of the same specification to form two turns of short-circuit rings, cut off the 112th and 113th turns of the conductor of the seventh layer coil of the reactor, and connect them with aluminum wires of the same specification to form two turns of short-circuit rings, and make multiple short-circuit rings at different encapsulation layer positions for batch experiments to obtain experimental samples with different fault degrees, and each batch of experiments is repeated 1 5 times to ensure data reliability. The basic data set contains four types of samples: normal operating status data, partial discharge status data, turn-to-turn short-circuit fault status data, and internal fire occurrence status data. Each type of sample contains temperature gradient time series data, CO2 concentration change rate time series data, current abnormal deviation rate time series data, and turn-to-turn short-circuit feature vector time series data as input features, and corresponding fire occurrence labels and fire intensity labels as output labels. The original data is normalized, outlier detection and missing value filling are performed through data preprocessing steps, and then the sliding window method is used to construct time series samples. The continuous time series data is divided into fixed-length input sequences and corresponding output labels, finally forming a complete training data set containing 50,000 training samples, 10,000 validation samples, and 5,000 test samples.

[0055] Among them, the steps of training the internal fire prediction model specifically include using the stochastic gradient descent optimization algorithm to iteratively optimize the neural network parameters, setting the learning rate to 0.001, the batch size to 32, and the number of training rounds to 200 rounds, using the cross-entropy loss function to calculate the difference between the predicted output and the true label, and updating the network weight parameters through the backpropagation algorithm. After each training round, the validation data set is used to evaluate the model performance, and the early stopping mechanism is used to prevent overfitting. When the validation set loss does not decrease for 10 consecutive rounds, the training is stopped, and the model parameters with the best validation set performance are saved as the final training results. After the training is completed, the test data set is used to evaluate the generalization ability of the model to ensure that the model can still maintain good prediction accuracy on unseen data.

[0056] Among them, the upper-level game model aims to minimize fire losses, and the constraints include three constraints: fire extinguishing agent supply limit, power outage time window limit, and temperature control range limit. The fire extinguishing agent supply limit represents the maximum fire extinguishing dose that the system can provide, the power outage time window limit represents the maximum allowable power outage time range, and the temperature control range limit represents the upper and lower limits of the temperature range that the temperature suppression device can control.

[0057] The lower game model takes the lowest fire extinguishing cost as the target, and the constraint conditions include three constraint conditions of forced ventilation power limit, temperature suppression device capacity limit and coolant injection pressure limit, the forced ventilation power limit represents the maximum power capacity of the ventilation equipment, the temperature suppression device capacity limit represents the maximum heat load that the temperature suppression device can handle, and the coolant injection pressure limit represents the maximum injection pressure that the temperature suppression device can provide, and the two models realize strategy coordination optimization through the equipment damage time caused by fire as a coupling term.

[0058] The dry-type air-core reactor inter-turn short-circuit internal fire suppression monitoring device includes eight main components of inter-turn partition monitoring device, inter-turn partial discharge detection device, inter-turn short-circuit fault detection device, internal fire minimum ignition vector early warning device, gas extinguishing device, temperature suppression device, forced ventilation device and upper computer control system. Nine optical fiber current sensors in the inter-turn partition monitoring device are respectively arranged in nine monitoring partitions of the reactor, and are connected with the upper computer control system through an optical fiber communication line. The ultrasonic sensors in the inter-turn partial discharge detection device are installed at 12 monitoring points on the surface of the reactor shell, and the high-frequency current transformers are installed on the three phase lines of the main circuit, and are connected with the upper computer control system through a data acquisition module. The differential protection device in the inter-turn short-circuit fault detection device is connected with each encapsulation layer current monitoring circuit, and the infrared thermal imager is installed at 6 observation positions around the reactor, and is connected with the upper computer control system through an Ethernet interface. The distributed optical fiber temperature sensor in the internal fire minimum ignition vector early warning device is laid along the inside of the reactor to form a temperature monitoring network, and the electrochemical gas sensor and the laser scattering smoke detector are respectively installed at the key positions inside the reactor, and are connected with the upper computer control system through a field bus. The gas extinguishing device, the temperature suppression device and the forced ventilation device are all connected with the upper computer control system through a control interface. The upper computer control system is responsible for collecting all monitoring data, running an internal fire prediction model, executing a game optimization algorithm, outputting suppression vector parameters and controlling the coordinated action of each suppression device.

[0059] The specific implementation of the above steps is described in detail below.

[0060] The specific implementation of step S01 is to establish a dry-type air-core reactor turn-to-turn partition monitoring device. First, according to the internal magnetic field distribution characteristics and fault propagation path law of the reactor, each encapsulation layer is divided into three regions in the radial position, i.e., an inner layer, a middle layer, and an outer layer, and is divided into three regions in the axial position, i.e., an upper end, a middle part, and a lower end, forming nine independent monitoring partitions. In each partition, a flexible optical fiber current sensor is installed. The sensor uses the Faraday magneto-optical effect principle to measure the change in the magnetic field strength around the conductor to deduce the current size. The optical fiber sensing ring forms a closed loop around the conductor. When the magnetic field generated by the current changes, the polarized light propagating in the optical fiber rotates, and the rotation angle is proportional to the magnetic field strength, thereby realizing non-contact current measurement. The monitoring device uses a distributed data acquisition method. The current data of each partition sensor is transmitted to the upper computer control system in real time through the optical fiber communication line. The system continuously monitors the current distribution change of each partition at a sampling frequency of 1000 Hz. When the current difference between adjacent partitions exceeds 3% of the normal operation, the early warning mechanism is triggered.

[0061] The specific implementation of step S02 is to construct a turn-to-turn partial discharge detection device. An ultrasonic sensor and a high-frequency current transformer combination detection method are used to obtain the partial discharge signal. The ultrasonic sensor is based on the piezoelectric effect principle and is installed at 12 monitoring points on the surface of the reactor shell for detecting ultrasonic signals generated by partial discharge. The sensor has a working frequency range of 20 kHz to 200 kHz, which can effectively capture the sound wave pulses generated during the discharge process. The high-frequency current transformer uses a Rogowski coil structure and is installed on the three-phase conductor of the main circuit to detect high-frequency current pulses caused by discharge with a frequency response range of 1 MHz to 50 MHz. The detection device extracts partial discharge characteristic parameters through time and frequency domain signal processing algorithms, including four components, i.e., the minimum partial discharge voltage, the partial discharge amplitude, the partial discharge phase angle, and the partial discharge frequency characteristic value, forming a four-dimensional turn-to-turn partial discharge vector for evaluating the degree of insulation deterioration. The minimum partial discharge voltage threshold is set to 60% of the rated voltage, the partial discharge amplitude threshold is set to 1000 pC, the partial discharge phase angle range is 0 degrees to 360 degrees, and the partial discharge frequency characteristic value is calculated by a fast Fourier transform algorithm to obtain the main frequency component.

[0062] The specific implementation of step S03 involves establishing a turn-to-turn short-circuit fault detection device, employing a fault identification method that combines differential protection principles with infrared thermal imaging technology. Based on Kirchhoff's current law, the differential protection device determines whether a turn-to-turn short-circuit fault exists by comparing the difference between the inflow and outflow currents of each encapsulation layer. A turn-to-turn short-circuit fault is identified when the current difference between adjacent encapsulation layers exceeds 5% of the normal operating current and persists for more than 100ms. The infrared thermal imager, using an uncooled focal plane detector, is installed at six observation locations around the reactor. Its detection wavelength range is 8μm to 14μm, enabling real-time monitoring of the reactor's surface temperature distribution. The thermal imaging detection system employs an image processing algorithm to identify areas of abnormal temperature distribution and triggers an abnormality alarm when the local temperature exceeds 50°C above ambient temperature and the rate of temperature rise exceeds 10°C / min. The fault detection device combines the current difference, the temperature distribution anomaly coefficient, and the short-circuit ring magnetic field strength to form a three-dimensional turn-to-turn short-circuit feature vector for precise fault location and severity assessment.

[0063] The specific implementation method of step S04 is to establish an internal fire minimum ignition vector warning device and use multi-sensor fusion technology to achieve early fire warning. The distributed fiber optic temperature sensor is based on the Raman scattering principle and is laid along the inside of the reactor to form a temperature monitoring network. The spatial resolution is 1m and the temperature measurement accuracy is ±1°C. It can monitor the temperature gradient distribution inside the reactor in real time. The electrochemical gas sensor uses the electrochemical cell principle to detect changes in CO2 concentration. The sensor response time is less than 30s and the detection accuracy is ±50ppm. It warns of fire by monitoring the growth rate of CO2 concentration generated in the early stage of combustion. The laser scattering smoke detector is based on the Mie scattering theory and uses an infrared laser as a light source to detect changes in the scattering intensity of light by smoke particles. The visibility attenuation coefficient threshold is set to 0.1m. -1 The early warning device integrates four parameters: temperature gradient, CO2 concentration change rate, visibility attenuation coefficient, and current abnormal deviation rate to form a four-dimensional internal fire minimum ignition vector. When the vector modulus exceeds the set threshold of 0.8, the fire warning program is activated.

[0064] Step S05 is an optional step, and its specific implementation method is to establish a fire development prediction model based on the thermochemical reaction mechanism and use the Arrhenius equation to describe the thermal decomposition kinetics of the insulating material. The model input parameters include five variables: pre-exponential factor, activation energy, gas constant, insulating material temperature, and oxygen concentration. By solving the Arrhenius equations, the thermal decomposition reaction rate constant of the material under different temperature conditions is predicted. The prediction model uses the finite difference method to numerically solve the partial differential equations. The time step is set to 0.1s and the spatial grid size is set to 0.05m. The spatiotemporal distribution of the temperature field and concentration field is obtained through iterative calculation. When the inter-turn partition monitoring device detects that the fault layer current distribution deviation exceeds 20% and the lowest partial discharge voltage in the inter-turn partial discharge vector is lower than 50% of the rated voltage or the modulus of the inter-turn short-circuit characteristic vector exceeds 1.5, the model automatically activates the partition rapid power-off protection mechanism, and the power-off response time does not exceed 50ms.

[0065] The specific implementation of step S06 is to predict the development trend of the fire based on a pre-trained internal fire prediction model. The model receives the current temperature gradient, CO2 concentration change rate, current abnormal deviation rate, and inter-turn short circuit feature vector as input parameters, and outputs two numerical values: the probability of fire occurrence and the expected fire intensity. The prediction model uses a sliding time window method to process input data. The time window length is set to 60s, and the prediction results are updated every 10s. The fire probability output by the model ranges from 0 to 1. When the probability exceeds 0.7, a high-level warning is triggered. The expected fire intensity is represented by the heat release rate in kW. When the expected fire intensity exceeds 100kW, the automatic suppression program is activated. The prediction model dynamically adjusts internal parameters based on real-time monitoring data and continuously optimizes the prediction accuracy through online learning algorithms to ensure the reliability and timeliness of the prediction results.

[0066] The specific implementation of step S07 involves constructing a two-layer game optimization model. The upper-layer model minimizes fire damage as its objective function, while the lower-layer model minimizes fire extinguishing costs. Constraints in the upper-layer model include a fire extinguishing agent supply limit of no more than 500L, a power outage window of no more than 5 minutes, and a temperature control range of 50°C to 300°C. Constraints in the lower-layer model include a forced ventilation power limit of no more than 50kW, a temperature suppression device capacity limit of no more than 200kW, and a coolant injection pressure limit of 0.5MPa to 3.0MPa. The game model employs Nash equilibrium theory to solve for the optimal strategy combination. An iterative algorithm searches for an equilibrium solution in the strategy space, with convergence criterion being that the Euclidean distance between two consecutive iterations is less than 0.01. The model uses the equipment damage time caused by the fire as a coupling term to coordinate the upper and lower-layer strategies. The optimization solution yields an optimal suppression vector consisting of five components: the power outage window, forced ventilation power, temperature control threshold, inert gas release flow rate, and coolant injection pressure.

[0067] Step S08 is implemented by adjusting the operating parameters of each suppression device in real time using a suppression vector parameter adjustment function. The function inputs include five parameters: the minimum ignition vector for internal fires, the inter-turn partial discharge vector, the inter-turn short-circuit characteristic vector, the current device status, and the predicted fire intensity. The output of the adjusted suppression vector parameters is a multi-objective optimization algorithm based on particle swarm optimization principles, with a population size of 50 particles and a maximum number of iterations of 100. The forced ventilation device activates smoke exhaust based on the power parameter output by the adjustment function. The ventilation power range is 10 kW to 50 kW, effectively exhausting high-temperature smoke and toxic gases. The temperature suppression device injects a water-based flame retardant coolant into the faulty zone based on the temperature control threshold parameter. The coolant is a water-based flame retardant with an injection pressure range of 0.5 MPa to 3.0 MPa, quickly controlling the local temperature below 200°C. The system continuously monitors the trend of the minimum ignition vector for internal fires at a frequency of 10 times per second until all parameters return to normal ranges.

[0068] Step S09 is an optional step. Its specific implementation method is to start the gas fire extinguishing device as the final suppression method when the above suppression measures cannot effectively control the development of the fire and the minimum ignition vector of the internal fire continues to deteriorate. The gas fire extinguishing device adopts the inert gas suppression principle to release inert gases such as nitrogen or carbon dioxide into the reactor to suppress the combustion reaction by diluting the oxygen concentration. The fire extinguishing device calculates the required inert gas release amount based on the internal space volume of the reactor, and the release flow range is 50m 3 / h to 200m 3 / h, with the goal of reducing the oxygen concentration to below 12%. The gas release process uses a graded control method, initially releasing 30% of the designed amount for tentative suppression. If the fire continues to develop, the remaining 70% is released for full suppression. The fire extinguishing system is equipped with an oxygen concentration monitoring system, which monitors changes in oxygen concentration within the reactor in real time. Fire extinguishing is confirmed as successful when the concentration drops below the set threshold and remains below it for more than 10 minutes.

[0069] The internal fire prediction model adopts a time series prediction neural network structure based on the Transformer architecture. The model includes four main components: input encoding layer, multi-head attention layer, feedforward neural network layer, and output decoding layer. The input encoding layer uses linear transformation to convert the temperature gradient, CO2 concentration rate of change, current abnormal deviation rate, and inter-turn short circuit feature vector into a 512-dimensional high-dimensional feature vector, and adds position encoding information to maintain the time sequence characteristics. The multi-head attention layer contains 8 parallel attention heads, each with a dimension of 64. The self-attention mechanism captures the dependencies between different time steps, and the attention weights are dynamically adjusted according to the current temperature gradient, CO2 concentration rate of change, and current abnormal deviation rate through the attention weight adjustment function. The feedforward neural network layer contains two fully connected layers, with 2048 neurons in the first layer and 512 neurons in the second layer. The activation function uses the ReLU function to perform nonlinear transformation and feature extraction on the output features of the attention layer. The output decoding layer contains two parallel fully connected layers, which output the fire occurrence probability and the expected fire intensity, respectively. The fire occurrence probability is limited to between 0 and 1 by the Sigmoid activation function, and the expected fire intensity is directly output by the linear activation function.

[0070] The training dataset was established by collecting historical operating data from dry-type air-core reactors under various operating conditions as a baseline dataset. Experimental data was also obtained by artificially creating short-circuit rings to simulate inter-turn short-circuit faults in the reactor. The artificial short-circuit rings were created by cutting the 45th and 46th turns of wire on the third coil layer of the reactor and connecting them with aluminum wire of the same specification to create two short-circuit rings. In the fifth coil layer, two short-circuit rings were created by cutting the 78th and 79th turns of wire to create two short-circuit rings. In the seventh coil layer, two short-circuit rings were created by cutting the 112th and 113th turns of wire to create two short-circuit rings. Multiple short-circuit rings were created at different locations on the encapsulation layer for batch experiments to obtain experimental samples with varying fault severity. Each batch of experiments was repeated 15 times to ensure data reliability. The basic dataset contains four types of samples: normal operation, partial discharge, turn-to-turn short-circuit fault, and internal fire. Each type of sample includes time series data for temperature gradient, CO2 concentration change rate, current anomaly deviation rate, and turn-to-turn short-circuit feature vectors as input features, along with corresponding fire occurrence and fire intensity labels as output labels. Data preprocessing steps include zero-mean normalization of the raw data, outlier detection based on the 3σ criterion, and missing value filling using linear interpolation. Time series samples are then constructed using a sliding window method with a window length of 120 time steps and a sliding step size of 10 time steps. The continuous time series data is segmented into fixed-length input sequences and corresponding output labels, resulting in a complete training dataset consisting of 50,000 training samples, 10,000 validation samples, and 5,000 test samples. The model training adopts the stochastic gradient descent optimization algorithm, the learning rate is set to 0.001, the batch size is set to 32, and the number of training rounds is set to 200 rounds. The cross entropy loss function is used to calculate the difference between the fire probability prediction output and the true label, and the mean square error loss function is used to calculate the difference between the fire intensity prediction output and the true label. The network weight parameters are updated by the back propagation algorithm. The validation dataset is used to evaluate the model performance after each training round. The early stopping mechanism is used to prevent overfitting. When the validation set loss does not decrease for 10 consecutive rounds, the training is stopped and the model parameters with the best validation set performance are saved as the final training results.

[0071] Furthermore, in order to further improve the model training during the process of artificially making the short-circuit ring, the physical mechanism of thermal distribution analysis can be combined with the statistical method of orthogonal experimental design to build a systematic short-circuit ring position optimization framework. First, based on the thermal field distribution characteristics of the reactor under normal operating conditions, a three-dimensional heat conduction equation is established:

[0072]

[0073] to solve the steady-state temperature field distribution. This equation takes into account the differences in the thermal physical properties of the materials inside the reactor, including the high thermal conductivity of aluminum conductors, the low thermal conductivity of epoxy insulating materials, and the convective heat transfer characteristics of air gaps. Through finite element numerical solving, the internal space of the reactor is discretized into about 5 million calculation units using tetrahedral meshing technology. The temperature distribution of each unit is obtained by iterative solving, with a convergence accuracy of

[0074] On this basis, the temperature sensitivity coefficient is defined: characterizing the influence of power loss changes at each location inside the reactor on the overall temperature distribution. Through numerical differentiation calculation, a unit power disturbance is applied at each spatial location: Δpower consumption = 1 W, and the response change of the entire temperature field is calculated to obtain the thermal sensitivity coefficient of that location. At the same time, a detailed thermal resistance network model is established: Analyzing the transmission path and heat transfer efficiency of the heat generated by the short-circuit ring to each temperature monitoring point.

[0075] Through thermal sensitivity analysis and heat transfer path analysis, the internal part of the reactor is divided into three regions according to the sensitivity coefficient: high sensitivity region (sensitivity > 0.8), medium sensitivity region (0.3 < sensitivity ≤ 0.8), and low sensitivity region (sensitivity ≤ 0.3). The high sensitivity region is mainly distributed inside the coil and the axial center region. Changes in power loss at these locations can quickly affect the temperature monitoring system. This theoretical framework ensures that the short-circuit ring location selection is consistent with the heat transfer physical laws and can be accurately optimized through systematic testing methods, providing strong physically interpretable basic data for subsequent machine learning model training.

[0076] In the high sensitivity region determined by the thermal distribution analysis, orthogonal experimental design method is used for fine position optimization. Considering the geometric structure characteristics of the reactor, the key factors affecting the effect of the short-circuit ring are determined as the axial layer number (1st-9th layer), the radial position (inner, middle, and outer relative positions), the circumferential angle (0°, 72°, 144°, 216°, 288° five equal division positions), the short-circuit turn number (1 turn, 2 turns, 3 turns, 4 turns, 5 turns), and the installation depth (conductor surface, conductor center, and conductor bottom relative positions). Five factors are set to 5 levels in the thermal sensitivity region. L 25 (5 5 ) orthogonal table is established for experimental design. Compared with the 5 5 =3125 combinations of full-factor experiments, orthogonal experiments only need 25 experiments to obtain the main factor effect information, reducing the experimental workload by about 99%. However, considering the need for interaction and error analysis, the actual reduction in effective experimental workload is about 12%.

[0077] For each orthogonal test combination, a complete physical simulation model needs to be established for numerical calculation. The simulation process includes four coupled modules: electromagnetic field solution, temperature field solution, fluid mechanics solution, and chemical reaction solution. The electromagnetic field module is based on Maxwell's equations:

[0078]

[0079] Calculate the current distribution changes caused by the short-circuit ring; the temperature field module solves the heat conduction equation considering the Joule heat source; the fluid module calculates the air convection and CO2 concentration diffusion process:

[0080]

[0081] Used to simulate the process of thermal decomposition of insulation materials to produce CO2.

[0082] Establish a comprehensive evaluation index system, including fault characteristic signal strength:

[0083]

[0084] Detection response time: response time = fault occurrence time - abnormal signal detection time;

[0085] Signal stability:

[0086] Fault simulation authenticity:

[0087] The weight coefficients of each evaluation index are determined by the hierarchical analysis method: Weight 信号 =0.4, weight 时间 =0.3, weight 稳定 =0.2, weight 真实 =0.1, establish the comprehensive index of the comprehensive evaluation function:

[0088]

[0089] Using the range analysis method:

[0090] Extremely poor j =max{mean j1 , mean j2 , mean j3 , mean j4, mean j5}-min{mean j1 , mean j2 , mean j3 , mean j4 , mean j5};

[0091] Determine the degree of influence of each factor on the comprehensive index through variance analysis:

[0092]

[0093] The significance of the verification factor influence is verified. At the same time, a quadratic response surface regression model is established:

[0094]

[0095] The interaction between factors and the non-linear relationship are analyzed, and the regression coefficients are solved by the least square method to determine the optimal parameter combination.

[0096] The comprehensive optimization scheme shows excellent convergence performance in the machine learning model training process, mainly reflected in the smoothness of the loss function and the stability of the gradient. From the perspective of physical principles, the short-circuit ring position determined based on thermal distribution analysis has a clear physical meaning, making the input features (temperature gradient, CO2 concentration change rate, current abnormal deviation rate) and output labels (fire occurrence probability, fire intensity) in the training data have a clear causal relationship. This physical interpretability ensures that the loss function:

[0097] Loss = a x cross-entropy loss (fire probability prediction, fire probability true value) + b x mean square error loss (fire intensity prediction, fire intensity true value);

[0098] In the parameter space, it presents a relatively smooth distribution, avoiding the problem of local extreme points and gradient explosion caused by data inconsistency.

[0099] From a mathematical point of view, the orthogonal experimental design ensures the uniform distribution of training data in the feature space. The balanced nature of the orthogonal table makes the number of occurrences of each level of each factor in the experiment completely equal. This balance translates into statistical uniformity of the training data, effectively reducing the adverse effects of data distribution skew on model training. Specifically, the condition number of the feature space:

[0100]

[0101] Significantly reduced, making the convergence speed of the gradient descent algorithm more stable. Theoretical analysis and numerical experiments show that compared with randomly sampled data sets, the convergence speed of the model training using orthogonal design data sets is improved by about 8%, and the final convergence accuracy is improved by about 12%.

[0102] In addition, the scheme pre-screens high-quality short-circuit ring positions through thermal sensitivity analysis, avoiding the generation of "noise data" in physically unreasonable positions. These noise data often have weak fault feature signals or abnormal patterns that do not conform to the real fault mechanism, which can cause the model to learn incorrect mapping relationships during training, affecting convergence and generalization performance. By introducing physical constraints, the signal-to-noise ratio of the training data is significantly improved, allowing the neural network to focus more on learning truly meaningful fault feature patterns.

[0103] Actual application verification shows that the short-circuit ring position configuration determined by this scheme improves the fault feature signal strength by an average of 28%, shortens the temperature anomaly detection time by 22%, improves the CO2 concentration anomaly detection sensitivity by 18%, and improves the current anomaly recognition accuracy by 25%. More importantly, through thermal distribution analysis, the thermal characteristics of the short-circuit ring fault are ensured to be quickly and effectively transmitted to the temperature monitoring points, shortening the fire warning time by an average of 5.2 minutes, which saves valuable disposal time for fire prevention and control. During model training, the convergence stability of the validation set loss is improved by 14%, overfitting is significantly reduced, and the generalization ability and actual application reliability of the final model are significantly improved.

[0104] Considering the possibility of limited simulation calculation resources or time constraints in actual engineering applications, this scheme can also quickly determine the basic position configuration of the short-circuit ring through simplified empirical formulas and engineering judgment. Based on statistical analysis of a large amount of reactor thermal distribution simulation data, it is found that the optimal position of the short-circuit ring has obvious regularity and can be directly obtained through the following simplified method:

[0105] The axial position selection follows the "1 / 3 division principle", that is, the short-circuit ring is set near the 1 / 3 and 2 / 3 positions of the total length of the reactor in the axial direction, with the specific calculation formula being axial position 1 = 0.33 × total height of the reactor and axial position 2 = 0.67 × total height of the reactor. These two positions usually correspond to the thermal stress concentration regions of the reactor, which can produce significant temperature anomaly signals. The radial position selection follows the "middle layer priority strategy", that is, the short-circuit ring is preferentially set in the middle layer of the reactor in the radial direction, with the calculation formula being: This position can ensure sufficient electromagnetic coupling strength and facilitate heat transfer to the outside.

[0106] For specific layer and turn selection, the "odd layer even turn" empirical rule can be used. The 3rd layer, 5th layer, and 7th layer are selected as the short-circuit ring setting layers, and these odd layers usually have strong electromagnetic field strength; in each layer, the even turn positions are selected, such as the 44-46th turn, 76-78th turn, and 110-112th turn, avoiding the mechanical stress concentration area at the end of the coil. The number of short-circuit turns is uniformly set to 2 turns, which is the best balance point between ensuring fault feature obviousness and avoiding excessive damage to the equipment.

[0107] The circumferential angle distribution adopts a "120-degree three-point layout," with short-circuit rings placed at 0°, 120°, and 240°. This ensures even distribution of fault signals around the circumference and avoids monitoring blind spots caused by single-point failures. The specific angle calculation is based on the reactor's main axis and is measured in a clockwise direction.

[0108] It should be noted that the first key technical idea of ​​the present invention is a fire early warning system based on multi-dimensional vector fusion. By constructing a four-dimensional internal fire minimum ignition vector, a four-dimensional inter-turn partial discharge vector and a three-dimensional inter-turn short circuit characteristic vector, the system has achieved a technological leap from single parameter monitoring to multi-parameter collaborative monitoring. Traditional fire warning methods mainly rely on a single sensor or a single parameter for judgment, which is prone to false alarms and missed alarms. The present invention combines multiple physical quantities such as temperature gradient, carbon dioxide concentration change rate, visibility attenuation coefficient, and current abnormal deviation rate into high-dimensional feature vectors, which can capture early signs of fire development from multiple dimensions at the same time. This multi-dimensional vector fusion technology is based on the information entropy principle in information theory. By increasing the dimension and density of monitoring information, it significantly improves the accuracy and reliability of fire warning. At the same time, through distance calculation and pattern recognition technology in vector space, it can effectively distinguish between normal operating status, fault status and fire precursor status, thereby achieving more accurate early warning.

[0109] The second key technical idea of ​​the present invention is an intelligent suppression vector optimization mechanism based on a two-layer game model. This mechanism realizes the intelligent optimization of fire suppression strategies by constructing an upper-layer game model with the goal of minimizing fire losses and a lower-layer game model with the goal of minimizing fire extinguishing costs. Traditional fire suppression methods usually adopt fixed suppression strategies and preset parameter configurations, which cannot be dynamically adjusted according to the actual situation of fire development, resulting in poor suppression effects or waste of resources. The present invention introduces the Nash equilibrium theory in game theory to model the fire suppression process as a multi-objective optimization problem. The upper-layer model focuses on minimizing fire losses, and the lower-layer model focuses on minimizing fire extinguishing costs. The two models achieve strategy coordination by using the equipment damage time caused by the fire as a coupling term. This two-layer game optimization mechanism can find the optimal balance between fire loss control and resource consumption, and automatically generate the optimal suppression vector containing parameters such as power outage time window, forced ventilation power, temperature control threshold, inert gas release flow rate, coolant injection pressure, etc., realizing the intelligence and refinement of fire suppression strategies.

[0110] The synergy of these two key technical ideas has produced a significant amplification effect of technical effects. The multi-dimensional vector fusion warning system provides high-quality input data and accurate fire status assessment for the two-layer game optimization model, while the two-layer game optimization model provides a precise suppression execution strategy for the multi-dimensional vector monitoring system. In traditional methods, warning and suppression are often two independent links, lacking organic information interaction and strategy coordination. The present invention uses the internal fire prediction model based on the Transformer architecture as a connecting bridge to achieve closed-loop control from fire warning to suppression execution. This collaborative mechanism enables the entire system to adjust the monitoring strategy and suppression parameters in real time according to the dynamic changes of the fire development, forming an adaptive intelligent fire protection system. Compared with traditional passive fire handling methods, the present invention realizes the organic combination of active fire prevention and intelligent fire suppression, fundamentally improving the technical level of dry-type air-core reactor fire protection.

[0111] Specifically, the principle of the present invention is as follows: The technical solution of the present invention is able to solve the above-mentioned core technical problems by establishing a complete closed-loop control system for fault evolution monitoring, prediction, and suppression. First, by dividing the reactor into nine independent monitoring zones and deploying a variety of sensing devices such as fiber optic current sensors, ultrasonic sensors, and high-frequency current transformers, a fault characteristic monitoring network covering multiple dimensions such as current distribution, partial discharge, temperature gradient, and gas concentration is constructed. This multi-dimensional monitoring system can comprehensively capture the characteristic changes of inter-turn short-circuit faults from initiation to development, compensating for the incomplete information provided by a single monitoring method.

[0112] Secondly, the Transformer-based time series prediction neural network constructed in this paper uses a multi-head attention mechanism to effectively learn complex dependencies between different time steps and accurately identify fault evolution patterns. The network inputs include key parameters such as temperature gradient, carbon dioxide concentration change rate, current anomaly deviation rate, and inter-turn short-circuit feature vectors. The attention weight adjustment function dynamically adjusts network parameters based on the current fault state, improving prediction accuracy and adaptability. The model can output the probability of fire occurrence and expected fire intensity, providing a quantitative decision-making basis for suppression strategies.

[0113] Finally, the two-layer game optimization model designed by the present invention can dynamically calculate the optimal suppression vector parameters based on the prediction results through the coordinated optimization of minimizing the upper-layer fire losses and minimizing the lower-layer fire extinguishing costs. The suppression vector contains five components: power outage time window, forced ventilation power, temperature control threshold, inert gas release flow rate, and coolant injection pressure. The operating parameters of each suppression device are adjusted in real time through the suppression vector parameter adjustment function. This multi-objective optimization mechanism not only ensures the fire suppression effect, but also achieves cost control, and solves the problem of inaccurate suppression measures of traditional methods. The entire technical solution forms a complete technical chain from fault monitoring, trend prediction to precise suppression, which can accurately predict and take targeted suppression measures before a fire occurs.

[0114] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0115] The specific implementation of step S01 is the same as above and will not be described in detail here. In addition, to better understand how the fiber optic current sensor can accurately measure current under high temperature and partial discharge conditions, its principle is explained below: The fiber optic current sensor measures current based on the Faraday magneto-optical effect, where the relationship between the polarization rotation angle and the magnetic field strength is expressed as:

[0116] θ = V·B·L;

[0117] Where θ is the polarization rotation angle in radians; V is the Verde constant, which is 4.6×10 -6 Radians per Tesla per meter; B is the magnetic field strength in Tesla; L is the effective length of the optical fiber in the magnetic field in meters. The relationship between current and magnetic field strength is expressed according to Ampere's circuit law as follows:

[0118]

[0119] Where I is the conductor current in amperes; r is the radius of the optical fiber sensing ring in meters; μ0 is the vacuum magnetic permeability, which is 4π×10 -7 Henry per meter. Combining the above two formulas, we can get the current measurement formula:

[0120]

[0121] The sensor measures the rotation angle of polarized light in a magnetic field to infer the current magnitude. This principle enables non-contact current measurement. Compared with traditional contact current transformers, it avoids the safety hazards and measurement errors caused by electrical connections. At the same time, it can work stably in high temperature and high pressure environments, providing a reliable technical means for accurate current monitoring in complex environments inside reactors, and effectively solving the problems of reduced measurement accuracy and equipment damage in harsh environments caused by traditional monitoring methods.

[0122] The specific implementation of step S02 is to construct an inter-turn partial discharge detection device, and the four-dimensional inter-turn partial discharge vector is expressed as:

[0123]

[0124] Where, is the four-dimensional inter-turn partial discharge vector; U pd is the minimum partial discharge voltage, in kilovolts; Q pd is the partial discharge amplitude, in pico-coulombs; φ pd is the partial discharge phase angle, in degrees; f pd is the characteristic value of partial discharge frequency, in Hertz. The calculation formula of partial discharge phase angle is:

[0125]

[0126] Where U sin is the sinusoidal component of the partial discharge signal; U cos is the cosine component of the partial discharge signal. The partial discharge frequency characteristic value is calculated by the fast Fourier transform algorithm:

[0127] f pd =argmax f |FFT(s(t))|;

[0128] Where s(t) is the time domain signal detected by the ultrasonic sensor; FFT is the fast Fourier transform function. The parameter acquisition method is: U pd It is obtained by gradually increasing the voltage to the voltage value when the partial discharge signal appears; Q pd The charge of partial discharge pulse is measured by high-frequency current transformer; U sin and U cos Obtained by performing orthogonal demodulation on the partial discharge signal.

[0129] The specific implementation of step S03 is to establish a turn-to-turn short circuit fault detection device, and the three-dimensional turn-to-turn short circuit feature vector is expressed as:

[0130]

[0131] Where, is the three-dimensional turn-to-turn short-circuit characteristic vector; I diff is the current difference between adjacent encapsulation layers, in amperes;

[0132] T abn is the surface temperature distribution anomaly coefficient, dimensionless; B sc is the magnetic field strength of the short-circuit ring, in Tesla. The current difference calculation formula is:

[0133] I diff=|I in -I out |;

[0134] Where, I in is the current flowing into the encapsulation layer; I out is the current flowing out of the encapsulation layer. The temperature distribution anomaly coefficient calculation formula is:

[0135]

[0136] Where, σ T is the standard deviation of temperature distribution; is the average temperature; T i is the temperature of the i-th measurement point; n is the total number of measurement points. The formula for calculating the magnetic field strength of the short-circuit ring is:

[0137]

[0138] Where, I sc is the short-circuit current in amperes; r sc is the radius of the short-circuit loop, in meters. The parameter acquisition method is: in and I out Obtained by measuring the current transformer installed at the inlet and outlet of the encapsulation layer; T i The data were obtained by measuring with an infrared thermal imager at 6 observation positions on the surface of the reactor; sc The induced current in the short-circuit ring is measured by the Hall sensor; r sc is the physical geometrical dimension of the short-circuit ring.

[0139] The specific implementation of step S04 is to establish an internal fire minimum ignition vector warning device. The four-dimensional internal fire minimum ignition vector is expressed as:

[0140]

[0141] Where, is the four-dimensional internal fire minimum ignition vector; is the temperature gradient in degrees Celsius per meter; is the rate of change of CO2 concentration, in parts per million per second; α vis is the visibility attenuation coefficient, in meters; δ I is the current abnormal deviation rate, dimensionless. The temperature gradient calculation formula is:

[0142]

[0143] Where, are the partial derivatives of temperature in the x, y, and z directions respectively. The current abnormal deviation rate calculation formula is:

[0144]

[0145] Where, I current is the current value; I normal is the current value during normal operation; || || represents the second norm of the vector. The formula for calculating the visibility attenuation coefficient is:

[0146]

[0147] Where I0 is the incident laser intensity in watts; I t is the transmitted laser intensity, in watts; L path is the length of the laser propagation path, in meters. The parameter acquisition method is: The temperature is measured at different locations by distributed optical fiber temperature sensors and the spatial derivative is calculated; The CO2 concentration is continuously measured by an electrochemical gas sensor and the time derivative is calculated; α vis Obtained by measuring the light intensity attenuation through a laser scattering smoke detector; current and I normal Obtained by comparing real-time measurement of optical fiber current sensor with historical data; I0 and I t The incident and transmitted light intensities are measured by laser and photodetector respectively; L path is the fixed distance between the laser and the detector.

[0148] The specific implementation of the optional step S05 is to establish a fire development prediction model based on the thermochemical reaction mechanism. The Arrhenius equation is expressed as:

[0149]

[0150] Where, k is the thermal decomposition reaction rate constant, in units of per second; A is the pre-exponential factor, in units of per second; E a is the activation energy in joules per mole; R is the gas constant, which is 8.314 joules per mole per Kelvin; T is the temperature of the insulating material in Kelvin; is the oxygen concentration in moles per cubic meter; m is the reaction order, dimensionless. The fire development trend is predicted by solving the partial differential equations:

[0151]

[0152] Where α is the thermal diffusivity, in square meters per second; is the Laplace operator; Q gen is the heat generation rate per unit volume, expressed in watts per cubic meter; C fuel is the fuel concentration in moles per cubic meter. The parameter acquisition method is: A and Ea Thermal decomposition curves of insulating materials were measured at different heating rates by thermogravimetric analysis and calculated using the Kissinger method. T was obtained by real-time measurement using a distributed optical fiber temperature sensor. Obtained by measuring with an oxygen sensor; m is determined by controlling the variable experiment and measuring the reaction rate at different oxygen concentrations.

[0153] The specific implementation of step S06 is the same as above and will not be described in detail here.

[0154] The specific implementation of step S07 is to construct a two-layer game optimization model, where the objective function of the upper-layer game model is the fire loss minimization function:

[0155] minF upper =w damage ·T damage +w spread ·A spread +w replace ·C replace ;

[0156] Where, F upper is the upper layer objective function value; w damage 、w spread 、w replace is the weight coefficient; T damage The equipment damage time caused by the fire, in hours; A spread is the fire spread range, in square meters; C replace is the equipment replacement cost, in ten thousand yuan. The objective function of the lower-level game model is the fire extinguishing cost minimization function:

[0157] minF lower =c1·P fan t1+c2·P cool t2+c3·V gas +c4·Va gent ;

[0158] Where, F lower is the lower layer objective function value; c1, c2, c3, c4 are cost coefficients; P fan is the forced ventilation power in kilowatts; t1 is the ventilation operation time in hours; P cool is the power of the temperature suppression device, in kilowatts; t2 is the temperature suppression operating time, in hours; V gas is the inert gas consumption, in cubic meters; V agent is the coolant consumption in liters. The optimal suppression vector is solved by Nash equilibrium:

[0159]

[0160] Where, is the optimal suppression vector; t off is the power-off time window in seconds; is the optimal forced ventilation power in kilowatts; is the optimal temperature control threshold, in degrees Celsius; is the optimal inert gas release flow rate, in cubic meters per hour; is the optimal coolant injection pressure, in MPa. The parameter acquisition method is: weight coefficient w damage 、w spread 、w replace The values ​​are determined based on the statistical analysis of historical fire loss data, ranging from 0.4 to 0.6, 0.2 to 0.4, and 0.2 to 0.4 respectively; the cost coefficients c1, c2, c3, and c4 are determined based on the equipment operating costs and consumable prices, respectively, at 0.5 yuan per kilowatt-hour, 0.8 yuan per kilowatt-hour, 2.5 yuan per cubic meter, and 1.2 yuan per liter; T damage 、A spread Obtained through calculation of fire development prediction model; C replace Calculated based on the equipment purchase price and depreciation period.

[0161] The specific implementation of step S08 is to adjust the operating parameters of each suppression device in real time through the suppression vector parameter adjustment function, and the adjustment function is expressed as:

[0162]

[0163] Where, is the adjusted inhibition vector parameter; f adj is the parameter adjustment function; is the current device state vector; I pred To predict the fire intensity, the unit is kilowatt. The multi-objective optimization function is expressed as:

[0164]

[0165] Where J is the comprehensive objective function; λ1 and λ2 are the trade-off coefficients; is the target internal fire ignition vector, and each component is set to The particle position update formula in the particle swarm optimization algorithm is:

[0166]

[0167] Where, is the velocity of the i-th particle in the d-th dimension; is the position of the ith particle in the dth dimension; w is the inertia weight; a1 and a2 are learning factors; rand1 and rand2 are random numbers between 0 and 1; is the individual optimal position of the particle; is the global optimal position. The parameter acquisition method is as follows: obtained by real-time monitoring of the equipment operating state through a sensor; pred obtained by output of an internal fire prediction model; λ1 and λ2 are determined to be optimal values of 0.6 and 0.4 through experiments; the inertia weight w is decreased from 0.9 to 0.4 using a linear decreasing strategy; the learning factors a1 and a2 are both taken as 2.0; The target values of each component are set according to the safety operation requirements.

[0168] The specific implementation of the optional step S09 is to start the gas fire extinguishing device, and the inert gas release amount calculation formula is as follows:

[0169]

[0170] In the formula, V release is the required inert gas release amount, in cubic meters; V reactor is the internal space volume of the reactor, in cubic meters; is the initial oxygen concentration, usually 21%; is the target oxygen concentration, set to 12%; η is the release efficiency coefficient, taken as 1.2. The oxygen concentration-time change relationship is expressed as:

[0171]

[0172] In the formula, is the oxygen concentration at time t; Q gas is the inert gas release flow, in cubic meters per second; t is time, in seconds. The parameter acquisition method is as follows: V reactor calculated according to the reactor design drawings; and obtained by real-time measurement through an oxygen concentration sensor; Q gas determined according to the gas release valve opening degree and the pipeline flow characteristics; η is obtained by measuring the ratio of the actual release efficiency to the theoretical value through a gas release test.

[0173] It should be noted that, in the present embodiment, the four-dimensional inter-turn partial discharge vector The voltage, amplitude, phase and frequency characteristics of partial discharge are uniformly modeled as feature points in a high-dimensional vector space based on the principle of multi-dimensional feature fusion. Compared with the traditional single-parameter monitoring method, this mathematical model can comprehensively capture the multi-dimensional feature information of partial discharge. Through the pattern recognition technology in the vector space, it significantly improves the recognition accuracy of insulation degradation status, effectively avoids the misjudgment and missed judgment phenomena that are prone to occur in single-parameter monitoring, and provides a scientific mathematical basis for the accurate evaluation of the insulation status of the reactor.

[0174] Three-dimensional turn-to-turn short-circuit eigenvector And the calculation formula of temperature distribution anomaly coefficient Based on the principle of multi-physical field coupling analysis, the three key physical quantities of current difference, temperature anomaly and magnetic field change are combined into a comprehensive characteristic vector. Compared with the traditional single fault criterion method, this model can simultaneously characterize the development status of inter-turn short-circuit faults from multiple dimensions of electromagnetic heat. By quantifying the degree of abnormality of temperature distribution through statistical methods, it can achieve precise fault location and severity assessment, and significantly improve the reliability and accuracy of fault detection.

[0175] Four-dimensional internal fire minimum ignition vector And its temperature gradient calculation formula Based on multi-sensor information fusion theory, by constructing a comprehensive fire sign model that includes heat transfer, gas generation, smoke production and current anomalies, this mathematical framework can capture early signs in multiple dimensions at the incipient stage of fire compared to traditional single fire detection methods. It accurately describes the uneven distribution characteristics of the temperature field through the spatial gradient operator, realizes ultra-early warning and precise positioning of fire, effectively shortens fire response time and improves warning accuracy.

[0176] Arrhenius equation The fire development partial differential equations are based on the theory of chemical reaction kinetics and predict the fire development trend by describing the thermal decomposition reaction mechanism of insulating materials under high temperature conditions. Compared with the traditional empirical prediction method, this theoretical model establishes a physical and chemical mechanism model of fire development. It can quantitatively predict the development speed and spread of fire according to key parameters such as material properties, temperature distribution and oxygen concentration, providing a scientific theoretical basis for the formulation of accurate fire suppression strategies, and significantly improving the accuracy of fire prediction and the targeted nature of suppression measures.

[0177] The objective function minF of the two-level game model upper and minF lowerBased on the Nash equilibrium theory in game theory, the optimal suppression strategy is sought by constructing a multi-objective optimization problem of minimizing fire losses and fire extinguishing costs. Compared with the traditional fixed suppression strategy method, this mathematical framework can achieve a dynamic balance between fire loss control and resource consumption. By solving the game equilibrium, the optimal suppression vector parameter combination is automatically generated, avoiding the problems of over-suppression or under-suppression in traditional methods, and realizing the intelligent optimization of fire suppression strategies and efficient utilization of resources.

[0178] Suppression vector parameter adjustment function and multi-objective optimization function Based on adaptive control theory, the operating parameters of the suppression equipment are dynamically adjusted by real-time monitoring of the system status and fire development trends. Compared with the traditional open-loop control method, this control model realizes closed-loop feedback control from fire warning to suppression execution. It can automatically optimize the suppression strategy according to the actual situation of fire development, ensuring that the suppression effect is accurately matched with the fire development status, significantly improving the success rate of fire suppression and the system's adaptive ability.

[0179] Calculation formula for the release amount of gas fire extinguishing and the oxygen concentration change equation Based on gas kinetic theory and combustion suppression mechanisms, efficient fire extinguishing is achieved by accurately calculating the required inert gas release amount and the time evolution law of oxygen concentration. Compared with traditional empirical gas release methods, this mathematical model can accurately calculate the optimal gas release strategy based on the specific spatial structure of the reactor and the fire development state, avoiding the problems of fire extinguishing failure caused by insufficient gas usage or resource waste caused by excessive gas usage, and realizing precise control and efficient suppression of the gas fire extinguishing process.

[0180] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 1-1 and 1-2.

[0181] Table 1-1 Variable Explanation Table

[0182]

[0183] Table 1-2 Variable Explanation Table

[0184]

[0185] To better understand and implement the present invention, Example 2 of a specific application scenario is provided below: A research team implemented the present invention's method for suppressing internal fires caused by turn-to-turn short circuits on a 35kV dry-type air-core reactor at a 220kV substation, successfully preventing a potentially serious fire. The reactor had a rated current of 800A, a rated voltage of 35kV, and an operating environment temperature of 25°C and a relative humidity of 60%.

[0186] The research team installed a complete monitoring and suppression system on the reactor in accordance with the technical solution of the present invention. The inter-turn partition monitoring device divides the reactor envelope into inner, middle and outer layers radially, and into upper, middle and lower ends axially, forming 9 independent monitoring partitions. Each partition is equipped with a fiber optic current sensor, and the sensor sampling frequency is set to 1000Hz. The inter-turn partial discharge detection device installs 12 ultrasonic sensors on the surface of the reactor casing with an operating frequency range of 20kHz to 200kHz. At the same time, a high-frequency current transformer is installed on the three-phase conductors of the main circuit with a frequency response range of 1MHz to 50MHz. The internal fire minimum ignition vector warning device lays a distributed fiber optic temperature sensor with a total length of 50m along the inside of the reactor with a spatial resolution of 1m. At the same time, 6 electrochemical CO2 sensors and 8 laser scattering smoke detectors are installed.

[0187] The fault occurred at 14:32 on March 15, 2024, when the operating current of the reactor was 720A, close to full load. The inter-turn partition monitoring device first detected an abnormal current distribution between the 78th and 79th turns of the 5th layer coil. The current difference between adjacent encapsulation layers suddenly increased from 1.2A in normal operation to 42.3A, exceeding 5.8% of the current in normal operation, triggering an inter-turn short-circuit fault warning. The infrared thermal imager simultaneously detected abnormal surface temperature in the area. The local temperature rose rapidly from the ambient temperature of 25°C to 78°C, and the temperature rise rate reached 15°C / min. The system immediately calculated the three-dimensional inter-turn short-circuit characteristic vector as The current difference is 42.3A, the temperature distribution anomaly coefficient is 2.13, and the short-circuit ring magnetic field strength is 0.087T.

[0188] As shown in Table 2, the inter-turn partial discharge detection device began to detect abnormal signals 30 minutes before the fault occurred:

[0189] Table 2 Inter-turn partial discharge monitoring data

[0190]

[0191] The modulus of the four-dimensional inter-turn partial discharge vector gradually increased from 0.67 at the beginning to 0.94 when the fault occurred, exceeding the set threshold of 0.8. The system judged that the degree of insulation degradation had reached a dangerous level.

[0192] Within 15 seconds of the fault, the distributed fiber optic temperature sensor detected a significant gradient change in the temperature distribution inside the reactor. As shown in Table 3, the temperature data at each monitoring point showed a sharp rise in the temperature of the fault area:

[0193] Table 3 Temperature monitoring data after the fault occurs

[0194]

[0195] The system calculates the temperature gradient At the same time, the electrochemical CO2 sensor detected that the CO2 concentration began to rise from the normal 420ppm, and the rate of change was 85℃ / m. Reach 12ppm / s. Visibility attenuation coefficient α measured by laser scattering smoke detector vis From the initial 0.01m -1 Increased to 0.08m -1 Abnormal current deviation rate δ I The calculated result is 0.058. The minimum ignition vector of the four-dimensional internal fire is The vector modulus reaches 85.9, far exceeding the set threshold of 0.8.

[0196] Based on the pre-trained internal fire prediction model, the short-circuit loop artificially created during the training data establishment process is as follows: Figure 4 As shown, the experimental circuit used is as follows Figure 5 As shown. The system inputs parameters such as the current temperature gradient, CO2 concentration change rate, current abnormal deviation rate, and inter-turn short circuit characteristic vector, and predicts that the probability of fire occurrence is 0.82 and the expected fire intensity is 150kW. The two-layer game model starts to calculate the optimal suppression vector. The upper model aims to minimize fire losses, and the weight coefficient w damage =0.5, w spread =0.3, w replace =0.2, the lower model takes the lowest fire extinguishing cost as the goal, cost coefficient c1 = 0.5 yuan / kWh, c2 = 0.8 yuan / kWh, c3 = 2.5 yuan / m 3 , c4 = 1.2 yuan / L.

[0197] As shown in Table 4, the system obtains the optimal inhibition vector parameters through game equilibrium solution:

[0198] Table 4 Optimal suppression vector parameters

[0199] Parameter name Numerical unit Power outage time window 45 s Forced ventilation power 35 kW Temperature control threshold 180 ℃ Inert gas release flow 85 <![CDATA[m 3 / h <!-- 21 -->]]> Coolant injection pressure 2.2 MPa

[0200] 48 seconds after the fault occurs, when the temperature at the fault point reaches 180°C, the system automatically activates the zone-specific rapid power-off protection mechanism, shutting off power to the faulty zone. Simultaneously, the forced ventilation system activates at 35kW to exhaust high-temperature fumes. A temperature suppression device sprays flame-retardant coolant at a pressure of 2.2MPa into the faulty zone, effectively keeping the local temperature below 200°C.

[0201] As shown in Table 5, the changes in various monitoring parameters after the implementation of the suppression measures:

[0202] Table 5 Changes in monitoring parameters after implementation of suppression measures

[0203]

[0204] The suppression vector parameter adjustment function continuously optimizes the operating parameters of each suppression device based on real-time monitoring data, balancing fire suppression effectiveness and cost control through a particle swarm optimization algorithm. The inertia weight w is linearly reduced from an initial value of 0.9 to 0.4, and the learning factors a1 = a2 = 2.0. The trade-off coefficients λ1 = 0.6 and λ2 = 0.4 in the multi-objective optimization function ensure that the system maintains suppression effectiveness while controlling operating costs.

[0205] After 15 minutes of continuous monitoring and suppression, all components of the internal fire minimum ignition vector returned to normal ranges, and the four-dimensional vector modulus dropped to 0.15, far below the warning threshold. The temperature at the fault point stabilized at 65°C, the CO2 concentration returned to the normal level of 420ppm, and the visibility attenuation coefficient dropped to 0.01m. -1 , the current abnormal deviation rate dropped to 0.001. The system judged that the fire risk had been completely eliminated, automatically shut down the suppression equipment, and resumed normal monitoring status.

[0206] Subsequent inspections revealed that the fault was caused by an inter-turn short circuit between the 78th and 79th turns of the fifth-layer coil due to aging insulation, forming a two-turn short-circuit loop. Without prompt containment measures, the Arrhenius equation predicts that the thermal decomposition reaction rate constant of the insulation material would reach k = 2.3 × 10 -3 s -1 , the fire will break out in 3 minutes, and the final heat release rate may exceed 1500kW.

[0207] Traditional approaches to addressing the core technical issues of this invention primarily rely on passive monitoring of a single parameter and a fixed-strategy fire extinguishing system. Traditional methods typically rely on temperature sensors or smoke detectors for fire monitoring, which use a single monitoring parameter and are prone to false alarms and missed alarms. Traditional fire extinguishing systems employ preset fixed parameters and are unable to dynamically adjust to the actual development of a fire, resulting in poor fire extinguishing effectiveness or waste of resources. Traditional fire prediction methods primarily rely on empirical judgment, lacking the support of scientific theoretical models and resulting in low prediction accuracy.

[0208] The present invention has brought significant progress compared to traditional means. In terms of monitoring accuracy, the multi-dimensional vector fusion monitoring method improves the fire identification accuracy by 15% and reduces the false alarm rate by 18% compared to traditional single-parameter monitoring. In terms of warning timeliness, the four-dimensional internal fire minimum ignition vector warning system based on multi-sensor fusion can issue a warning at the budding stage of the fire, 12 minutes earlier than the traditional method. In terms of suppression effect, the intelligent suppression strategy optimized by double-layer game improves the fire suppression success rate by 16% compared to the traditional fixed strategy, while reducing resource consumption in the suppression process by 14%. In terms of system response, the response time from fault detection to the initiation of suppression measures is shortened from the traditional 3 to 5 minutes to 48 seconds, and the response speed is increased by about 85%. Through the fire prediction model based on the Transformer architecture, the accuracy of fire development trend prediction is improved by 19% compared to the traditional empirical prediction method, providing a reliable scientific basis for the formulation of accurate suppression strategies.

[0209] The following is a simple example of manually creating a short-circuit loop in the present invention: The reactor used in the training set has a 9-layer coil structure, with 126 turns per layer, a total height of 2.4m, an inner diameter of 1.2m, and an outer diameter of 2.8m. First, a 3D finite element thermal analysis model was established, with normal operating power consumption set to Power_Normal = 180kW and an ambient temperature of 40°C. The heat conduction equation was solved: The steady-state temperature field distribution was obtained. The calculation results show that the axial temperature distribution of the reactor is high in the middle and low at both ends. The highest temperature is located at 0.8m and 1.6m axially, corresponding to the third and seventh coil layers.

[0210] By temperature sensitivity analysis: The sensitivity coefficients of the 3rd, 5th, and 7th coils were found to be 0.92, 0.87, and 0.89, respectively, all belonging to the highly sensitive area. The radial position analysis shows that the thermal resistance value of the middle coil position (radial range of 1.8-2.2m) is:

[0211]

[0212] Based on the results of thermal distribution analysis, the candidate area was determined to be the middle coil position of the 1st to 9th coil layers, providing a physical constraint range for subsequent orthogonal test design.

[0213] Based on the candidate areas determined by the thermal distribution analysis, an orthogonal experimental design with five factors and five levels was established. Factor A is the number of axial layers, and the levels are set to the 1st, 3rd, 5th, 7th, and 9th layers; Factor B is the radial position, and the levels are set to the inner ring, middle inner ring, middle ring, middle outer ring, and outer ring; Factor C is the circumferential angle, and the levels are set to 0°, 72°, 144°, 216°, and 288°; Factor D is the number of short-circuit turns, and the levels are set to 1, 2, 3, 4, and 5 turns; Factor E is the inter-turn position, and the levels are set to five equally divided positions of each layer of coil. L25(5 5 ) orthogonal table was used for experimental design, and the effect of each experimental combination was analyzed through electromagnetic-thermal-fluid multi-physics field coupling simulation.

[0214] For this reactor's specific structure, the third coil layer has 126 turns, divided into five equally spaced locations: turns 25, 50, 75, 100, and around 125. Considering mechanical strength and ease of installation, turns 45-46 were ultimately chosen as the optimal solution. Similarly, turns 78-79 were chosen for the fifth coil layer, and turns 112-113 were chosen for the seventh coil layer. These locations are located in the middle and rear sections of each coil layer, avoiding mechanical stress concentration areas at the coil ends while ensuring good electromagnetic coupling.

[0215] The influence of each factor is ranked as follows through range analysis: axial layer number (range A =0.24)> turn position (extreme difference E =0.18)>Short-circuit turns (extreme difference D =0.15)> radial position (range B =0.12)>Circumferential Angle (Range C =0.08). The results of variance analysis show that the influence of axial layer number and turn position on the comprehensive evaluation index reaches a significant level.

[0216] The optimal parameter combination was determined as follows: axial layers selected as the 3rd, 5th, and 7th, radial position selected as the middle coil, circumferential angle selected as 0°, number of short-circuited turns selected as 2, and inter-turn position selected as the middle and rear end of each layer. In practice, the wires were cut at turns 45-46 of the 3rd layer coil, turns 78-79 of the 5th layer coil, and turns 112-113 of the 7th layer coil, and reconnected with aluminum wire of the same specifications to form a two-turn short-circuit loop.

[0217] Simulation verification shows that

[0218] The detection response time was 145 seconds, the signal stability was 0.91, and the comprehensive evaluation index reached 0.86, meeting the quality requirements for fire prediction model training data. Compared to the initial empirical selection scheme, the fault detection accuracy increased by 11%, providing a high-quality standardized dataset for subsequent neural network model training.

[0219] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor, characterized in that: include: Establish a dry-type hollow reactor inter-turn partition monitoring device, divide the reactor's envelope into multiple monitoring partitions according to radial and axial positions, and install a fiber optic current sensor in each partition to monitor the inter-turn current distribution changes in real time; construct an inter-turn partial discharge detection device, use an ultrasonic sensor and a high-frequency current transformer to obtain partial discharge signals, and obtain the minimum partial discharge voltage, partial discharge amplitude, partial discharge phase angle, and partial discharge frequency characteristic values ​​through measurement to form a four-dimensional inter-turn partial discharge vector; establish an inter-turn short-circuit fault detection device, use the differential protection principle to monitor the current difference of each envelope layer, and use an infrared thermal imager to detect abnormal surface temperature distribution areas to form an inter-turn short-circuit characteristic vector; establish The internal fire minimum ignition vector warning device uses a distributed fiber optic temperature sensor to monitor the temperature gradient, an electrochemical gas sensor to monitor the rate of change of carbon dioxide concentration, and a laser scattering smoke detector to measure the visibility attenuation coefficient. Combined with the current abnormal deviation rate, it forms the internal fire minimum ignition vector. Based on a pre-trained internal fire prediction model, it predicts the fire development trend and outputs the fire probability and expected fire intensity. A game model is used to solve and determine an optimal suppression vector. The operating parameters of each suppression device are adjusted in real time through the optimal suppression vector parameter adjustment function.

2. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 1, characterized in that: The step of inter-turn zoning is specifically to divide the encapsulation layer into three areas: inner layer, middle layer and outer layer in the radial direction, and into three areas: upper end, middle part and lower end in the axial direction, forming nine independent monitoring zones, each zone has independent current monitoring and protection functions.

3. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 2, characterized in that: The inter-turn partial discharge vector is a four-dimensional vector, including four components: minimum partial discharge voltage, partial discharge amplitude, partial discharge phase angle, and partial discharge frequency characteristic value. The minimum partial discharge voltage represents the minimum voltage value at which the insulating medium begins to generate partial discharge.

4. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 3, characterized in that: The internal fire minimum ignition vector is a four-dimensional vector, including four components: temperature gradient, CO2 concentration change rate, visibility attenuation coefficient, and current abnormal deviation rate. The temperature gradient represents the rate of change of temperature difference between different positions inside the reactor.

5. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 4, characterized in that: The inter-turn short circuit characteristic vector is a three-dimensional vector, which includes three components: the current difference between adjacent encapsulation layers, the surface temperature distribution anomaly coefficient, and the short-circuit ring magnetic field strength. The current difference between adjacent encapsulation layers represents the degree of difference in current distribution between adjacent encapsulation layers.

6. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 5, characterized in that: The specific structure of the internal fire prediction model is a time series prediction neural network based on the Transformer architecture, which includes four main components: input encoding layer, multi-head attention layer, feedforward neural network layer, and output decoding layer.

7. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 6, characterized in that: The game model includes an upper-level game model with the goal of minimizing fire losses and a lower-level game model with the goal of minimizing fire extinguishing costs.

8. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 7, characterized in that: The inhibition vector is a five-dimensional vector including five component parameters: power-off time window, forced ventilation power, temperature control threshold, inert gas release flow rate, and coolant injection pressure. The power-off time window represents the time interval from fault detection to complete power off.

9. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 8, characterized in that: The suppression vector parameter adjustment function is used to adjust the operating parameters of each suppression device in real time according to the current fire status. The input includes five parameters: the minimum ignition vector of internal fire, the inter-turn partial discharge vector, the inter-turn short circuit characteristic vector, the current device status, and the predicted fire intensity.

10. The method for suppressing internal fire caused by inter-turn short circuit in dry-type air-core reactor according to claim 9, characterized in that: After adjusting the operating parameters of each suppression device in real time through the suppression vector parameter adjustment function, it also includes starting the gas fire extinguishing device when the aforementioned suppression measures cannot effectively control the development of the fire, releasing inert gas into the reactor to reduce the oxygen concentration to below 12%.

Images