A substation structure system compatible with dry-type and oil-immersed transformers

Abstract

The application relates to a substation structure system compatible with dry-type and oil-immersed transformers, and in particular to the field of electric power engineering construction, which automatically identifies the transformer type through a multi-modal heterogeneous data acquisition module, a space-time risk modeling module uses a graph attention network to output a fire risk probability and a sub-type distribution in real time, a dynamic strategy generation module generates a detector switching, fire extinguishing agent adjustment and the like instruction according to a state space by using a deep Q network, and a transfer learning deployment module realizes rapid strategy deployment by means of digital twinning and domain adversarial training, so that self-adaptation to transformer type change, early risk early warning and intelligent decision-making are realized, a safety vacuum period is eliminated, and the transformation cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of power engineering construction, and more specifically, to a substation structural system compatible with both dry-type and oil-immersed transformers. Background Technology

[0002] Currently, 110kV substations widely use dry-type or oil-immersed transformers as core equipment. Dry-type transformers, due to their oil-free and flame-retardant characteristics, are typically equipped with aspirating smoke detectors to capture smoke particles during the early smoldering stage. Oil-immersed transformers, however, contain a large amount of insulating oil, and the fire risk mainly manifests as oil combustion caused by high-temperature electric arcs. Therefore, cable-type linear heat detectors are suitable for use along the tank walls to monitor temperature anomalies in real time. In actual operation, due to factors such as urban load growth, tightening environmental policies, or equipment retirement cycles, the same substation often needs to switch between dry-type and oil-immersed transformers multiple times. However, existing substation fire protection systems are designed and constructed in a one-time manner based on the initially installed transformer type—detector model, installation location, type of extinguishing agent (gas, foam, or...). The fire protection system (including fine water mist) and pipeline layout are fixed configurations, making them unsuitable for subsequent equipment type changes. When a transformer needs to be switched, the fire protection system must undergo large-scale physical modifications: removing the original detectors and their wiring, reinstalling detection devices adapted to the new transformer, adjusting the extinguishing agent storage tanks and release pipelines, and even modifying the fire control logic. This process not only consumes a lot of money and time, but more importantly, during the transition period, the original fire protection system has been dismantled while the new system is not yet ready, leaving the substation in a "safety vacuum" with insufficient or severely weakened fire protection capabilities. In addition, existing fire protection systems generally follow a "post-event response" model, that is, alarms are only triggered after a fire has occurred and obvious smoke or temperature rise has been generated, lacking the ability to identify and predict fire hazards in the early stages and unable to use multi-dimensional sensor data to make forward-looking assessments of potential risks.

[0003] The aforementioned technological status quo reveals a core problem: existing fire protection systems lack the ability to adapt to changes in transformer type. Their fixed detection and extinguishing configurations cannot match dynamically changing equipment types, resulting in high modification costs and significant safety risks with each transformer switchover. Simultaneously, the system lacks intelligent analysis and early warning mechanisms based on multi-source data, failing to identify abnormal signs during the fire incubation stage, thus missing the optimal response time. Specifically, when a substation switches from oil-immersed to dry-type transformers, if the original cable-type linear heat detectors are still used, they cannot promptly alarm during the smoldering stage when smoke is generated due to localized overheating of the dry-type transformer coils. Conversely, if aspirating smoke detectors are continued after switching from dry-type to oil-immersed, ... The initial fire in an oil-immersed transformer may be characterized by open flames and minimal smoke, leading to detection delays. Furthermore, the physical replacement of fire-fighting equipment involves switching extinguishing agents (e.g., foam extinguishing for oil-immersed transformers and gas extinguishing for dry-type transformers). Different extinguishing agents require different pipeline pressures and spray devices, making the modification work extremely complex. More seriously, during the switching process, substation entrances and exits may be blocked by the accumulation of dismantled old equipment or construction materials, further weakening emergency response capabilities. Therefore, there is an urgent need for an intelligent fire management system that can automatically identify the current transformer type, dynamically configure detector types and fire-fighting strategies, and provide early risk warnings, to fundamentally address the adaptability deficiencies and safety vulnerabilities of traditional fire-fighting systems in transformer type switching scenarios. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a substation structure system compatible with both dry-type and oil-immersed transformers. It solves the problems mentioned in the background art by using a multimodal heterogeneous data acquisition module, a spatiotemporal risk modeling module, a dynamic strategy generation module, and a transfer learning deployment module.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including: Multimodal heterogeneous data acquisition module: In response to the state change signal of the magnetic induction switch on the switchable wall module, it acquires environmental sensing data from the universal adapter interface and operating parameters from the transformer body. It performs time stamp synchronization and feature alignment on the state change signal, environmental sensing data and operating parameters, determines the current transformer type identifier based on the state change signal, performs statistical analysis on the environmental sensing data to obtain the average environmental temperature and humidity, obtains the current working mode of the universal adapter interface, and generates the initial feature vector of the node in a unified semantic space. Spatiotemporal risk modeling module: Receives the initial feature vector of the node, abstracts all sensors, equipment, fire protection facilities and wall structure in the substation into a spatiotemporal graph network, uses graph attention network to perform multi-layer message passing, calculates the attention coefficient between nodes and updates the node features, and outputs the fire risk probability and fire subtype probability distribution at the current moment. Dynamic strategy generation module: The state space is composed of fire risk probability, fire subtype probability distribution, current transformer type identifier, average ambient temperature and humidity, and the current working mode of universal adapter interface. The optimal strategy is solved by deep Q network. The optimal action is selected according to the reward function. The module generates detector type switching instructions, fire extinguishing agent type adjustment instructions, fire damper control instructions and alarm level instructions, and sends them to universal adapter interface and fire protection facilities. Transfer learning deployment module: Construct a high-fidelity digital twin environment based on the structural parameters of a switchable wall module, a universal adapter interface, and the transformer body. Use a deep reinforcement learning agent to train in this digital twin environment for multiple rounds to obtain an initial policy. Use a domain adversarial training method to align the feature distribution of the digital twin environment to the feature distribution of the real substation, so that the initial policy can be deployed in the real substation after a few minor adjustments.

[0006] In a preferred embodiment, the specific operations for timestamp synchronization and feature alignment of state change signals, environmental sensor data, and operating parameters in the multimodal heterogeneous data acquisition module are as follows: First, the optimal time delay between the state change signal and the environmental sensing data, as well as the optimal time delay between the state change signal and the operating parameters of the transformer body, are calculated based on the cross-correlation function. The optimal time delay is then used to shift the environmental sensing data and the operating parameters of the transformer body to align the time bases of the three, resulting in an aligned discrete time series. Then, using the kernel canonical correlation analysis method, radial basis kernel function kernel matrices of environmental sensing data and transformer body operating parameters are constructed respectively. The generalized eigenvalue problem is solved to obtain the eigenvectors corresponding to the first few largest eigenvalues, which form a projection matrix. Then, the environmental sensing data and transformer body operating parameters in the aligned discrete time series are mapped to the common semantic space to obtain environmental alignment features and equipment alignment features with consistent dimensions.

[0007] In a preferred embodiment, the specific operation for generating the initial feature vector of a node is as follows: Calculate the arithmetic mean of the state change signal within the current time window to obtain the mean of the state change signal; use the mode of the state change signal within the current time window as the current transformer type identifier, with a value of zero or one, corresponding to dry-type transformer or oil-immersed transformer respectively. For the temperature and humidity components in the environmental sensing data, the weighted average method is used to calculate the weighted average, where the weights decrease exponentially as the sample moves away from the current time, thus obtaining the mean environmental temperature and mean environmental humidity. The current working mode of the universal adapter interface is read directly from the universal adapter interface register. The value is zero or one, which corresponds to the air intake sampling mode or the cable temperature sensing mode, respectively. Finally, the environmental alignment features, device alignment features, and mean state change signals are fused by outer product to construct a third-order tensor, which is then flattened into a one-dimensional vector. This vector is then compressed through a fully connected layer and activated by the hyperbolic tangent function to obtain the initial feature vector of the node.

[0008] In a preferred embodiment, the specific operation of abstracting all sensors, equipment, fire protection facilities, and wall structures within the substation into a spatiotemporal graph network in the spatiotemporal risk modeling module is as follows: First, each sensor, device, fire protection facility, and wall structure is defined as a node, and the total number of nodes is a positive integer; each node carries an initial feature vector from the multimodal heterogeneous data acquisition module; Then, edges are constructed based on the physical proximity and logical control relationships between nodes: for any two nodes, if their Euclidean distance in three-dimensional space is less than a preset spatial distance threshold, an undirected edge is established. If there is a direct signal transmission or control link between two nodes, an additional directed edge is established, with the direction from the controlling party to the controlled party. Finally, each edge is assigned an initial weight, which is determined by the spatial distance and logical relationship indicator function between the two nodes: the smaller the spatial distance and the stronger the logical relationship, the greater the initial weight.

[0009] In a preferred embodiment, the specific operation of multi-layer message passing using a graph attention network is as follows: In each layer of the network, the attention coefficient of each node to all its neighboring nodes is first calculated. The attention coefficient integrates the node feature similarity, the initial weight, and the time decay factor. The time decay factor is determined by the time difference of the last state update of two nodes and the time decay constant. The time decay constant is set to 300 seconds, so that nodes that have recently interacted receive higher attention weights. Then, a multi-head attention mechanism is adopted, which sets up multiple attention heads, each of which independently calculates the attention coefficients, and the results of the calculations of multiple attention heads are concatenated and then linearly transformed to obtain the updated node features; After multiple layers of message propagation, each node obtains the final node feature containing global context information. Subsequently, a subset of key nodes representing transformers in the equipment, wall modules in the wall structure, detectors in the sensors, and fans in the equipment are selected from all nodes. The final node feature of each node in the key node subset is then globally pooled based on an attention mechanism to obtain a global pooled feature vector. The global pooling feature vector is input into two parallel fully connected classification heads: the first classification head is a single neuron and uses the sigmoid activation function to output the fire risk probability at the current time, with a value ranging from zero to one; The second classification head consists of three neurons and uses the softmax activation function to output the probability distribution of fire subtypes. The three output components correspond to the three subtypes of coil overheating, oil leakage, and electric arc fire, respectively, and the sum of each component is equal to one.

[0010] In a preferred embodiment, the specific operations for constructing the state space in the dynamic strategy generation module are as follows: First, the fire risk probability, fire subtype probability distribution, current transformer type identifier, ambient temperature and humidity mean, and universal adapter interface current working mode are taken as seven state components. The fire subtype probability distribution contains three components. The temperature and humidity components in the ambient temperature and humidity mean are linearly normalized, and the temperature component is scaled to between zero and one. The humidity component is also scaled to between zero and one. The current transformer type identifier and universal adapter interface current working mode remain as binary discrete values. Then, a mutual information-guided attention fusion mechanism is adopted to calculate the mutual information estimate between each state component and the action sampled from the historical running data. The mutual information estimate is approximated by the kernel density estimation method. Then, all mutual information estimates are normalized by the softmax function with temperature coefficient to obtain the attention weight of each state component. Finally, each state component is mapped to an embedding vector through a dedicated embedding function, multiplied by the corresponding attention weight, and all weighted embedding vectors are concatenated to obtain the fused encoded state vector.

[0011] In a preferred embodiment, the specific operation of selecting the optimal action according to the reward function is as follows: The reward function consists of four terms. The first term is the matching reward, which is positive when the selected action makes the detector type match the current transformer type, and zero otherwise. The second term is the cost penalty, which is proportional to the energy consumption and equipment loss required to perform the action. The third term is the risk change penalty, which is measured by the KL divergence between the fire risk probability before and after the action, and penalizes behaviors with a sudden increase in risk. The fourth term is the safety constraint violation penalty, which is the penalty when the fire risk probability after the action exceeds a preset safety threshold, and the excess part is multiplied by a preset weighting coefficient. A priority experience replay buffer is constructed to store historical transfer samples. Two deep Q-networks, an online network and a target network, are constructed and trained using a dual deep Q-learning algorithm. The loss function is constructed based on the difference between the Q-values ​​output by the online network and the target network. After training, an epsilon greedy strategy is used to select the optimal action during the inference phase. The selected optimal action is broken down into detector type switching command, extinguishing agent type adjustment command, fire damper control command, and alarm level command, which are then sent to the universal adapter interface and fire protection facilities via industrial Ethernet.

[0012] In a preferred embodiment, the specific operation of constructing a high-fidelity digital twin environment in the transfer learning deployment module is as follows: Based on the geometric dimensions of the switchable wall module, the physical interface parameters of the universal adapter interface, and the thermal characteristic parameters of the transformer body, a digital twin environment is constructed using computational fluid dynamics software and electromagnetic field simulation software. The digital twin environment can accurately simulate the evolution process of various fire scenarios, such as arc ignition in oil-immersed transformers and overload smoke in dry-type transformers, and generate sensor readings, equipment status, and wall structure status consistent with those of real substations. The time step of the digital twin environment is set to match the frequency of real data acquisition to ensure that the simulation data matches the real data in terms of time granularity.

[0013] In a preferred embodiment, the specific operation of aligning the feature distribution of the digital twin environment to the feature distribution of the real substation using the domain adversarial training method is as follows: The system defines three components: a feature extractor, a deep Q-network, and a domain discriminator. The feature extractor maps the input state to a common feature vector, and the domain discriminator outputs the probability that the feature belongs to the digital twin environment. State samples are collected from the digital twin environment and the real substation, respectively, and the domain discriminant loss function is calculated. A gradient reversal layer is inserted between the feature extractor and the domain discriminator. The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a preset gradient reversal tradeoff coefficient during backward propagation. The total loss function is composed of the weighted sum of the loss function in the digital twin environment and the domain discriminator loss. By alternately updating the parameters of the domain discriminator and the parameters of the feature extractor and the deep Q network, adversarial training is performed to make the feature distribution output by the feature extractor tend to be consistent in the two domains.

[0014] In a preferred embodiment, the specific operation of deploying the initial strategy, after minor adjustments, to a real substation is as follows: After the domain adversarial training is completed, the feature extractor parameters are fixed, a small number of labeled samples are collected in a real substation, and the parameters of the deep Q network are fine-tuned. The fine-tuning adopts the same dual deep Q learning algorithm as the training in the digital twin environment, but the learning rate is reduced to one-tenth of the original. The fine-tuned deep Q network is embedded in the edge computing gateway and interacts in real time with the multimodal heterogeneous data acquisition module and the spatiotemporal risk modeling module via industrial Ethernet to achieve online adaptive decision-making for substation fire protection. After deployment, real-world environmental data is continuously collected, and incremental fine-tuning is triggered periodically to address domain drift caused by equipment aging.

[0015] The beneficial effects of this invention are as follows: the multimodal heterogeneous data acquisition module automatically identifies the transformer type; the spatiotemporal risk modeling module uses a graph attention network to output the fire risk probability and subtype distribution in real time; the dynamic strategy generation module uses a deep Q-network to generate instructions such as detector switching and fire extinguishing agent adjustment based on the state space; and the transfer learning deployment module uses digital twins and domain adversarial training to achieve rapid strategy deployment, thereby realizing adaptive response to transformer type changes, early risk warning and intelligent decision-making, eliminating safety vacuum periods and reducing renovation costs. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

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

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1 This embodiment provides, for example Figure 1-2 The substation structure system shown is compatible with both dry-type and oil-immersed transformers, specifically including: Multimodal heterogeneous data acquisition module: In response to the state change signal of the magnetic induction switch on the switchable wall module, it acquires environmental sensing data from the universal adapter interface and operating parameters from the transformer body. It performs time stamp synchronization and feature alignment on the state change signal, environmental sensing data and operating parameters, determines the current transformer type identifier based on the state change signal, performs statistical analysis on the environmental sensing data to obtain the average environmental temperature and humidity, obtains the current working mode of the universal adapter interface, and generates the initial feature vector of the node in a unified semantic space. Spatiotemporal risk modeling module: Receives the initial feature vector of the node, abstracts all sensors, equipment, fire protection facilities and wall structure in the substation into a spatiotemporal graph network, uses graph attention network to perform multi-layer message passing, calculates the attention coefficient between nodes and updates the node features, and outputs the fire risk probability and fire subtype probability distribution at the current moment. Dynamic strategy generation module: The state space is composed of fire risk probability, fire subtype probability distribution, current transformer type identifier, average ambient temperature and humidity, and the current working mode of universal adapter interface. The optimal strategy is solved by deep Q network. The optimal action is selected according to the reward function. The module generates detector type switching instructions, fire extinguishing agent type adjustment instructions, fire damper control instructions and alarm level instructions, and sends them to universal adapter interface and fire protection facilities. Transfer learning deployment module: Construct a high-fidelity digital twin environment based on the structural parameters of a switchable wall module, a universal adapter interface, and the transformer body. Use a deep reinforcement learning agent to train in this digital twin environment for multiple rounds to obtain an initial policy. Use a domain adversarial training method to align the feature distribution of the digital twin environment to the feature distribution of the real substation, so that the initial policy can be deployed in the real substation after a few minor adjustments.

[0021] In this embodiment, it is specifically necessary to explain the specific operations of timestamp synchronization and feature alignment of state change signals, environmental sensor data, and operating parameters in the multimodal heterogeneous data acquisition module as follows: First, the optimal time delay between the state change signal and the environmental sensing data, and the optimal time delay between the state change signal and the transformer's operating parameters are calculated based on the cross-correlation function. The optimal time delays are then used to shift the environmental sensing data and the transformer's operating parameters, aligning their time bases to obtain an aligned discrete-time series. The calculation process of the cross-correlation function is as follows: the state change signal is considered as the reference signal, and the environmental sensing data is considered as the signal to be aligned. Within a preset time delay range (e.g., from -500 milliseconds to +500 milliseconds, with a step size of one sampling interval), the product integral of the two signals under different delays is calculated sequentially to obtain the cross-correlation function value sequence. The delay corresponding to the maximum value in the sequence is taken as the optimal time delay. Similarly, the above process is repeated for the state change signal and the operating parameters of the transformer body to obtain the second optimal time delay. The translation operation refers to shifting the environmental sensing data backward by the number of sampling points corresponding to the first optimal time delay, and shifting the operating parameters of the transformer body backward by the number of sampling points corresponding to the second optimal time delay, so that the three signals reach peak alignment on the time axis. After alignment, the overlapping time interval of the three signals is extracted based on the sampling time of the state change signal to obtain an aligned discrete time sequence of consistent length. The sequence length contains at least one thousand sampling points to ensure statistical stability. Then, using kernel canonical correlation analysis, radial basis function kernel matrices for environmental sensing data and transformer operating parameters are constructed respectively. The generalized eigenvalue problem is solved to obtain the eigenvectors corresponding to the first few largest eigenvalues, forming projection matrices. This maps the environmental sensing data and transformer operating parameters in the aligned discrete time series to a common semantic space, resulting in consistent environmental alignment features and equipment alignment features. The specific implementation process of kernel canonical correlation analysis is as follows: First, the environmental sensing data matrix and the transformer operating parameter matrix are extracted from the aligned discrete time series. The number of rows in both matrices is the number of sampling points, and the number of columns is their respective feature dimensions. Then, a radial basis function is selected, with its kernel width parameter set to the reciprocal of the median of the Euclidean distance between all sample pairs multiplied by a constant (e.g., 0.5). The kernel matrix for the environmental sensing data and the kernel matrix for the transformer operating parameters are calculated respectively. The size of each kernel matrix is ​​the number of sampling points multiplied by the number of sampling points, where the element in the i-th row and j-th column is equal to the result of the radial basis function applied to the i-th sample and the j-th sample. Next, the generalized eigenvalue problem is solved. Mathematically, the left-hand matrix consists of zero blocks and the product of the environment kernel matrix and the device kernel matrix, while the right-hand matrix consists of square blocks of the environment kernel matrix and the device kernel matrix. The generalized eigenvalue problem is solved using a power iteration method or the Jacobian method, yielding a set of eigenvalues ​​and corresponding eigenvector pairs. The eigenvector pairs corresponding to the ten largest eigenvalues ​​are selected, with the length of each pair equal to the number of sampling points. These eigenvectors are then used as column vectors to construct the environment projection matrix and the device projection matrix, respectively. Finally, the environment kernel matrix is ​​multiplied by the environment projection matrix to obtain the environment-aligned feature matrix, whose size is ten times the number of sampling points. Similarly, the device kernel matrix is ​​multiplied by the device projection matrix to obtain the device-aligned feature matrix, also whose size is ten times the number of sampling points. Thus, the environmental sensing data and the transformer's operating parameters are mapped into a ten-dimensional common semantic space, achieving feature alignment between modes. The specific operation for generating the initial feature vector of a node is as follows: The arithmetic mean of the state change signals is calculated within the current time window to obtain the mean value of the state change signals. The length of the current time window is set to two hundred sampling points, corresponding to a duration of two seconds (assuming a sampling frequency of one hundred Hz). The calculation process of the arithmetic mean is as follows: add the values ​​of all state change signals within the current time window, divide by two hundred, and obtain the mean value of the state change signals. This mean value reflects the average transparency of the wall structure within the time window, and its value ranges from zero to one. The mode value of the state change signal within the current time window is used as the current transformer type identifier, which can be either zero or one, corresponding to a dry-type transformer or an oil-immersed transformer, respectively. The mode value is determined by counting the number of times the state change signal is zero and the number of times it is one within the current time window, and taking the value with the higher frequency as the mode value. If the two frequencies are equal, the default value is zero. This identifier is used in the subsequent dynamic strategy generation module to distinguish the currently running transformer type. For the temperature and humidity components in the environmental sensor data, an exponentially weighted moving average method is used to calculate the weighted average. The weights decay exponentially as the sample moves further away from the current time point, resulting in the average environmental temperature and humidity. The specific calculation process for the exponentially weighted moving average is as follows: a decay coefficient of 0.9 is set, indicating that the sample closest to the current time point has the highest weight; the weight of each preceding sample is multiplied by 0.9. For the temperature component, starting from the last sample in the current time window and traversing backwards to the first sample, the product of each temperature value and its corresponding weight is accumulated, along with the weight values. Finally, the average environmental temperature is obtained by dividing the weighted sum by the total weights. The calculation process for the humidity component is exactly the same. The unit for the average environmental temperature is degrees Celsius, and the unit for the average environmental humidity is percentage. These two averages reflect the current thermal and humidity state of the environment and are used for subsequent risk assessment. The current operating mode of the universal adapter interface is read directly from the universal adapter interface register. The value is either zero or one, corresponding to either the air-feed sampling mode or the cable-sensing mode, respectively. The reading method is as follows: a command to read the register is sent to the universal adapter interface via industrial Ethernet. The command includes the register address (e.g., address 0x01), and the interface is waited for a byte of data to be returned. If the return value is 0x00, it indicates that the current mode is air-feed sampling mode; if the return value is 0x01, it indicates that the current mode is cable-sensing mode. This operating mode determines the execution direction of subsequent detector type switching commands. Finally, the environment alignment features, device alignment features, and the mean of the state change signal are fused by outer product to construct a third-order tensor, which is then flattened into a one-dimensional vector. This vector is then compressed through a fully connected layer and activated with a hyperbolic tangent function to obtain the initial feature vector of the node. The specific process of outer product fusion is as follows: the environment alignment features are matrices of size 10 times the number of sampling points, the device alignment features are matrices of the same size, and the mean of the state change signal is a scalar. First, the environment alignment features and device alignment features are outer producted to obtain a matrix of size 10 times 10. Then, this matrix is ​​multiplied by the mean of the state change signal (i.e., each element of the matrix is ​​multiplied by the scalar) to obtain a two-dimensional matrix of size 10 times 10. This two-dimensional matrix is ​​then expanded row-wise into a one-dimensional vector with a length of 100. Finally, this one-dimensional vector is input into a fully connected layer with 64 neurons, a weight matrix of size 64 times 100, and a bias vector of length 64. The calculation process of the fully connected layer is as follows: multiply the input vector by the weight matrix, and add the bias vector to obtain an intermediate vector of length 64; finally, apply the hyperbolic tangent activation function to each element of the intermediate vector to compress all output values ​​to between negative one and positive one, and obtain the final node initial feature vector; this vector simultaneously encodes high-order interaction information of environmental state, device state and structural state, providing rich node initial attributes for the subsequent spatiotemporal risk modeling module.

[0022] In this embodiment, it is specifically necessary to explain the specific operation of abstracting all sensors, equipment, fire protection facilities, and wall structures within the substation into a spatiotemporal graph network in the spatiotemporal risk modeling module as follows: First, each sensor, device, fire protection facility, and wall structure is defined as a node, with the total number of nodes being a positive integer. Each node carries an initial feature vector from the multimodal heterogeneous data acquisition module. The total number of nodes is determined as follows: based on the actual deployment of the substation, the number of all installed entities such as temperature sensors, humidity sensors, gas sensors, current transformers, voltage transformers, circuit breakers, disconnect switches, transformer bodies, cooling fans, fire extinguishing devices, fire valves, wall modules, metal grilles, and louvers is counted, with each entity serving as an independent node. The total number of nodes is typically between fifty and two hundred. The initial feature vector of each node has a dimension of sixty-four. This vector is generated by the multimodal heterogeneous data acquisition module and contains high-order interactive information about environmental status, equipment status, and structural status. Then, edges are constructed based on the physical proximity and logical control relationships between nodes: for any two nodes, if their Euclidean distance in three-dimensional space is less than a preset spatial distance threshold, an undirected edge is established; the specific value of the spatial distance threshold is set to 2.5 meters; the Euclidean distance is calculated by substituting the three-dimensional coordinates (in meters) of the two nodes into the distance formula, that is, summing the squares of the differences in the horizontal coordinates, vertical coordinates, and heights, and then taking the square root; if the calculation result is less than 2.5 meters, the two nodes are considered to be physically adjacent, and an undirected edge is established; for example, the distance between a transformer node and an adjacent cooling fan node is usually between one and two meters, which meets the condition; while the distance between a transformer node and a gas sensor node at a distant corner may exceed ten meters, so no edge is established; If a direct signal transmission or control link exists between two nodes, an additional directed edge is established, with the direction from the controller to the controlled node. The determination of a signal transmission or control link is based on: referring to the substation's secondary wiring diagram and communication topology diagram. If two nodes are directly connected via hardwiring, fieldbus, or industrial Ethernet, and there is a unidirectional or bidirectional control command flow or data acquisition flow, then a logical control relationship is considered to exist. For example, there is a control link between a fire controller and a fire extinguishing device valve, with the direction from the controller to the valve; there is a data link between a temperature sensor and a data acquisition unit, with the direction from the sensor to the acquisition unit. For such relationships, an additional directed edge is established, with the direction strictly determined according to the signal flow direction. Finally, each edge is assigned an initial weight, which is determined by the spatial distance between the two nodes and the logical relationship indicator function: the smaller the spatial distance and the stronger the logical relationship, the larger the initial weight. The initial weight serves as a prior bias for subsequent attention calculations. The specific calculation process for the initial weight is as follows: First, calculate the square of the Euclidean distance between the two nodes, divide it by the square of twice the spatial scale parameter (the spatial scale parameter is set to 1.0 meters), and obtain the distance decay exponent. Then, take the negative value of this exponent and calculate the natural exponent function to obtain the distance decay factor. The distance decay factor ranges from zero to one, with the closer the distance, the closer it is to one, and the farther the distance, the closer it is to zero. Next, determine whether there is a logical control relationship between the two nodes: if there is, set the value of the logical relationship indicator function to 1.2; if there is no logical control relationship, set it to 0.8. Finally, multiply the distance decay factor by the logical relationship indicator function to obtain the initial weight. This weight is a positive real number, ranging from 0.8 to 1.2, and serves as a prior bias for subsequent attention calculations, so that physically adjacent and logically related node pairs receive higher initial attention. The specific operations for multi-layer message passing using graph attention networks are as follows: In each layer of the network, the attention coefficients of each node with respect to all its neighbors are first calculated. These attention coefficients integrate node feature similarity, initial weights, and a time decay factor. The calculation process is as follows: For any node in the current layer and one of its neighbors, the current feature vectors of both nodes are multiplied by the trainable weight matrix of that layer to obtain transformed feature vectors. Then, the transformed feature vectors are concatenated with the product of the initial weights and the time decay factor to form a combined vector. The length of this combined vector is equal to twice the length of the transformed feature vector plus one. Next, this combined vector is multiplied by a trainable attention vector to obtain a scalar value. A leaky linear rectified activation function is applied to the scalar value to obtain unnormalized attention coefficients. The leaky linear rectified activation function outputs the input when the input is positive, and outputs the input multiplied by a very small negative slope (e.g., 0.2) when the input is negative. The time decay factor is determined by the time difference between the last state updates of two nodes and the time decay constant, which is set to 300 seconds. This allows nodes that have recently interacted to receive higher attention weights. The calculation process for the time decay factor is as follows: obtain the timestamps of the last state updates of the two nodes, calculate the time difference between them in seconds; divide the time difference by the time decay constant (300 seconds) to obtain the time ratio; take the negative value of the time ratio and calculate the natural exponential function to obtain the time decay factor. The value of this factor ranges from zero to one, with a smaller time difference being closer to one and a larger time difference being closer to zero. For example, if two nodes have both updated their states within the last minute and the time difference is 60 seconds, the time decay factor is approximately 0.8187; if the time difference is one hour (3600 seconds), the time decay factor is approximately 0.00006, which is almost zero. This factor allows nodes that have recently interacted to receive higher weights in the attention calculation. Then, a multi-head attention mechanism is adopted, setting up multiple attention heads, each independently calculating attention coefficients, and concatenating the calculation results of multiple attention heads before obtaining the updated node features through linear transformation. The specific implementation of the multi-head attention mechanism is as follows: four attention heads are set; each attention head has an independent trainable weight matrix and attention vector; for each head, the normalized attention coefficients are independently calculated according to the above process (the unnormalized coefficients are normalized using the softmax function so that the sum of the attention coefficients of the node to all its neighbors equals one), and then the transformed features of the neighboring nodes are weighted and summed to obtain the output feature vector of that head; the output feature vectors of the four heads are all 128 in length; these four vectors of length 128 are concatenated end-to-end to obtain a vector of length 5... The combined vector is 112. Finally, this combined vector is transformed linearly (multiplied by a trainable weight matrix and a bias vector is added) to obtain the updated node feature vector, which still has a length of 128. Simultaneously, after each message aggregation step, a gating mechanism is introduced: first, a gating vector is calculated, with each component between zero and one, obtained using the sigmoid function; the gating vector is determined by the current node's own features and the aggregated neighbor information; then, the gating vector is used to perform a weighted fusion of the current node's own features and the aggregated neighbor information, i.e., the gating vector is multiplied element-wise by the current node's own features, then one is added or subtracted, and finally multiplied element-wise by the aggregated information to obtain the final updated node features; this gating mechanism allows the network to adaptively balance its own memory with external information. After multiple layers of message propagation, each node obtains its final node features containing global context information. The total number of message propagation layers is set to three. The first layer takes the initial feature vector of the node as input and outputs the updated node features. The second layer takes the output of the first layer as input and outputs the updated node features. The third layer takes the output of the second layer as input and outputs the final node features. Each layer uses the same multi-head attention mechanism and gating mechanism, but the trainable parameters of each layer are independent. After three layers of propagation, the feature vector of each node incorporates the information of its neighbors within a maximum of three hops, thus possessing global context awareness. Subsequently, a subset of key nodes representing transformers in the equipment, wall modules in the wall structure, detectors in the sensors, and fans in the equipment is selected from all nodes. The final node features of each node in the key node subset are then globally pooled based on the attention mechanism to obtain globally pooled features. The key node subset is selected as follows: based on the node number list, all nodes with type labels "transformer," "wall module," "detector," and "fan" are selected; these nodes are the entities most directly related to fire risk. The specific process of global pooling is as follows: first, an importance score is calculated for each node in the key node subset, which is obtained through a trainable small neural network; this neural network consists of a fully connected layer and a hyperbolic tangent activation function, mapping the final node feature of each node to a scalar value; then, a softmax function is applied to all scalar values ​​for normalization, so that the sum of the importance scores of all key nodes equals one; finally, the final node feature of each node is multiplied by its corresponding importance score, and then summed to obtain the global pooled feature vector; this vector has a length of 512, integrates the information of all key nodes, and assigns different weights to nodes of different importance. The global pooling feature vector is input into two parallel fully connected classification heads: the first classification head is a single neuron using the sigmoid activation function, outputting the fire risk probability at the current time step, with a value ranging from zero to one; the calculation process of the first classification head is as follows: the global pooling feature vector is multiplied by a trainable weight vector (length 512), and then a trainable bias scalar is added to obtain a scalar value; then the sigmoid activation function is applied to this scalar value, compressing it to between zero and one, which is taken as the fire risk probability; the closer the probability value is to one, the higher the fire risk, and the closer it is to zero, the lower the risk; for example, when the probability is greater than 0.7, the system considers there to be a high fire risk and triggers the early warning mechanism; The second classification head consists of three neurons and uses the softmax activation function to output a probability distribution of fire subtypes. The three output components correspond to the three subtypes of coil overheating, oil leakage, and arc ignition, respectively, and the sum of each component is equal to one. The calculation process of the second classification head is as follows: the global pooling feature vector is multiplied by a trainable weight matrix (size 3 x 512), and a trainable bias vector (length 3) is added to obtain a score vector of length 3. Then, the softmax activation function is applied to each component of the score vector, that is, the exponent of each component is divided by the sum of the exponents of all components, so that the three output components are all positive and the sum is equal to one. These three components correspond to the conditional probabilities of the three fire subtypes of coil overheating, oil leakage, and arc ignition, respectively. For example, if the output is (0.1, 0.7, 0.2), it means that under the current conditions, the probability of oil leakage is the highest, which is 70%. This probability distribution is used in the subsequent dynamic strategy generation module to formulate targeted fire extinguishing strategies.

[0023] In this embodiment, it is necessary to specifically explain the operations that constitute the state space in the dynamic policy generation module: First, the fire risk probability, fire subtype probability distribution, current transformer type identifier, average ambient temperature and humidity, and current operating mode of the universal adapter interface are considered as seven state components. The fire subtype probability distribution contains three components. The temperature and humidity components of the average ambient temperature and humidity are linearly normalized, scaling both the temperature and humidity components to between zero and one. The current transformer type identifier and current operating mode of the universal adapter interface remain as binary discrete values. The specific calculation process for linear normalization is as follows: For the temperature component, a lower limit is preset to -10 degrees Celsius, and an upper limit is set to... For 50 degrees Celsius, subtract the lower limit of the current temperature value, then divide by the difference between the upper and lower limits to obtain the normalized temperature value. For example, if the current temperature is 25 degrees Celsius, the normalized result is (25 minus -10) divided by (50 minus -10) equals 35 divided by 60, which is approximately 0.5833. For the humidity component, directly divide the current humidity percentage by 100 to obtain the normalized humidity value. For example, if the current humidity is 65%, the normalized result is 0.65. The current transformer type identifier and the current working mode of the universal adapter interface are both binary discrete values, taking values ​​of zero or one, and do not require normalization. Then, a mutual information-guided attention fusion mechanism is employed to calculate the mutual information estimate between each state component and the actions sampled from historical data. This mutual information estimate is approximated using kernel density estimation. All mutual information estimates are then normalized using a temperature-coefficient softmax function to obtain the attention weight for each state component. The specific calculation process for the mutual information estimate is as follows: First, a batch of historical actions is randomly sampled from the historical data, with each action corresponding to a historical state. For each state component at the current moment, it is paired with the sampled historical action. The joint probability is then approximated using kernel density estimation. Mutual information density and marginal probability density; the kernel density estimation uses a Gaussian kernel function, with the kernel width set to one-third of the median Euclidean distance between all sample pairs; for each pair of state components and actions, the joint probability density, the marginal probability density of the state component, and the marginal probability density of the action are calculated separately; then, the joint probability density is divided by the product of the two marginal probability densities, and the natural logarithm is taken to obtain the mutual information estimate of the sample pair; the average of the mutual information estimates calculated for all sampled historical actions is taken as the final mutual information estimate between the state component and the historical action; the temperature coefficient is set to 0.1 to control the concentration of attention weights; The softmax normalization process is as follows: divide the mutual information estimate of each state component by the temperature coefficient, take the exponent, and then divide it by the sum of the exponents of all state components to obtain the attention weight of each state component. The sum of all weights equals one. Finally, each state component is mapped to an embedding vector through a dedicated embedding function, multiplied by the corresponding attention weight, and all weighted embedding vectors are concatenated to obtain the fused encoded state vector. The specific implementation of the embedding function is as follows: For continuous state components (fire risk probability, three components in the fire subtype probability distribution, normalized temperature and humidity values), a fully connected linear layer is used as the embedding function. This linear layer has an input dimension of one and an output dimension of sixteen, mapping scalar values ​​to vectors of length sixteen. For discrete state components (current transformer type identifier and current working mode of the universal adapter interface), a... Using a lookup table embedding method, a trainable vector of length sixteen is pre-assigned for each possible value (zero and one), and the corresponding vector is directly looked up based on the current value. After each of the seven state components is processed by an embedding function, seven embedding vectors of length sixteen are obtained. Each embedding vector is multiplied by the corresponding attention weight (scalar) to obtain a weighted embedding vector. Finally, these seven weighted embedding vectors of length sixteen are concatenated to obtain a fused encoded state vector of length one hundred and twelve. This vector simultaneously contains risk information, configuration information, and environmental information, and the importance of different components is differentiated through attention weights. The specific operation of selecting the optimal action based on the reward function is as follows: The reward function consists of four terms. The first term is the matching reward, which is positive when the selected action makes the detector type match the current transformer type, and zero otherwise. The specific calculation process of the matching reward is as follows: Determine the value of the current transformer type identifier. If it is zero (dry-type transformer), the detector type should be in air-absorbing sampling mode; if it is one (oil-immersed transformer), the detector type should be in cable-type temperature sensing mode. If the detector type switching instruction in the selected action makes the universal adapter interface working mode consistent with the above requirements, the matching reward is ten, otherwise it is zero. The second item is the cost penalty, which is proportional to the energy consumption and equipment wear required to perform the action. The specific calculation process of the cost penalty is as follows: a normalized cost value is pre-set for each action combination, with the cost value ranging from zero to one; for example, the cost of keeping the current state unchanged is zero, the cost of switching the detector type is 0.2, the cost of adjusting the fire extinguishing agent type is 0.3, and the cost of switching the detector and adjusting the fire extinguishing agent at the same time is 0.5; the normalized cost value corresponding to the selected action is multiplied by the weight coefficient one to obtain the cost penalty value; The third item is the risk change penalty, which is measured by the KL divergence between the fire risk probabilities before and after the action, penalizing actions that suddenly increase risk. The specific calculation process of KL divergence is as follows: the fire risk probability before the action is denoted as p, and the fire risk probability after the action is denoted as p'; the formula for calculating KL divergence is p multiplied by the natural logarithm of p divided by p', plus one minus the natural logarithm of p multiplied by one minus the natural logarithm of p divided by one minus p'; this divergence value measures the degree of difference between two probability distributions, and the larger the value, the more drastic the risk change; multiplying the KL divergence by a weighting coefficient of five, we obtain the risk change penalty value; The fourth item is the penalty for violating safety constraints. When the fire risk probability after the action exceeds a preset safety threshold, the excess is multiplied by a preset weighting coefficient as a penalty. The specific value of the safety threshold is set to 0.8. If the fire risk probability after the action is greater than 0.8, the excess (i.e., fire risk probability minus 0.8) is multiplied by the weighting coefficient 20 to obtain the penalty value for violating safety constraints. If the fire risk probability is less than or equal to 0.8, the penalty value is zero. For example, if the fire risk probability after the action is 0.9, the penalty value is (0.9 minus 0.8) multiplied by 20, which equals 2. A priority experience replay buffer is constructed to store historical transition samples. Two deep Q-networks, an online network and a target network, are built and trained using a dual deep Q-learning algorithm. The loss function is constructed based on the difference between the Q-values ​​output by the online network and the target network. The specific construction process of the priority experience replay buffer is as follows: the buffer capacity is set to 100,000 transition samples; each transition sample includes the current state, the selected action, the immediate reward, and the next state; the priority of each sample is set according to the absolute value of its TD error, which is the immediate reward plus a discount factor multiplied by the maximum Q-value of the target network for the next state minus the Q-value of the online network for the selected action in the current state; the discount factor is set to 0.95; newly added samples are assigned the highest priority, and a batch (batch size set to 64) of samples is sampled according to the priority ratio during each training. Both the online and target networks employ the same dual-channel deep Q-network architecture: the risk channel consists of a three-layer residual network with 128 neurons per layer and ReLU activation function; the configuration channel consists of a three-layer gated recurrent unit network with 128 neurons per layer; the fusion layer uses bilinear interaction to output 128-dimensional features; finally, a fully connected layer outputs the Q-values ​​of 36 actions (action space size is 2 x 3 x 2 x 3 = 36); the parameters of the target network are soft-updated from the online network every 100 training steps, with a soft-update coefficient of 0.001, meaning the target network parameters are equal to 0.999 times the target network parameters plus 0.001 times the online network parameters; the total number of training steps is set to 500,000; after training, an epsilon greedy strategy is used to select the optimal value during the inference phase. Actions; where epsilon is linearly annealed from an initial value of 1.0 to a final value of 0.05; the specific execution process of the epsilon greedy strategy is as follows: at the beginning of the inference phase, epsilon is set to 1.0, and actions are selected completely randomly; after each action is executed, epsilon decreases according to the linear decay formula, with the decay step set to 100,000 steps, that is, each step reduces (1.0 minus 0.05) divided by 100,000 equals 0.000095; when epsilon decays to 0.05, this value remains unchanged; at each decision step, a random number between 0 and 1 is generated. If the random number is less than the current epsilon, an action is randomly selected uniformly from the 36 actions; if the random number is greater than or equal to epsilon, the action with the largest Q value output by the online network is selected; The selected optimal action is decomposed into detector type switching command, extinguishing agent type adjustment command, fire damper control command, and alarm level command, which are sent to the universal adapter interface and fire protection facilities via industrial Ethernet. Specifically, the detector type switching command switches the universal adapter interface to the detection mode matching the current transformer type; the extinguishing agent type adjustment command activates the corresponding fire extinguishing system in the six-square-meter area in the southwest corner; the fire damper control command controls the opening and closing status of the fire dampers in the corresponding area; and the alarm level command sends the corresponding alarm signal to the fire control center. The specific process of command issuance and execution is as follows: the detector type switching command is a one-bit binary value. If it is one, a switching command is sent to the universal adapter interface via industrial Ethernet. The command format is to write the register address (e.g., address 0x02) and the target mode value. (Zero indicates aspirating type, one indicates cable-type temperature sensing); the extinguishing agent type adjustment command is a two-bit binary value (00 for gas, 01 for foam, 10 for fine water mist), which is sent via industrial Ethernet to the fire extinguishing system controller in a six-square-meter area in the southwest corner. The command includes the extinguishing agent type code and activation signal; the fire damper control command is a one-bit binary value. If it is one, an open command is sent; if it is zero, a close command is sent. The command is sent via industrial Ethernet to the fire damper actuator in the corresponding area; the alarm level command is a two-bit binary value (00 for no alarm, 01 for warning, 10 for emergency), which is sent via industrial Ethernet to the fire control center. The message includes the alarm level code and substation identification; all commands use cyclic redundancy check to ensure transmission correctness. If the check fails, it is resent three times.

[0024] In this embodiment, the specific operation of building a high-fidelity digital twin environment in the transfer learning deployment module is as follows: Based on the geometric dimensions of the switchable wall modules, the physical interface parameters of the universal adapter interface, and the thermal characteristic parameters of the transformer body, a digital twin environment is constructed using computational fluid dynamics software and electromagnetic field simulation software. This digital twin environment can accurately simulate the evolution of various fire scenarios, including arcing in oil-immersed transformers and overload smoke in dry-type transformers, and generates sensor readings, equipment status, and wall structure status consistent with a real substation. The geometric dimensions of the switchable wall modules include: strip foundations with a length of 6,720 mm, a width of 650 mm, and a height of 400 mm; beam cross-sections with a height range of 600-800 mm and a width range of 400-500 mm; and column cross-sections with dimensions ranging from 500-600 mm by 500-600 mm. The physical interface parameters of the universal adapter interface include: the total area of ​​the reserved fan opening in the side wall is more than 19.5 square meters, the number of reserved fan bases in the roof is six, and the spacing between the detector brackets is less than five meters. The thermal characteristic parameters of the transformer body include: rated loss value of dry-type transformer (unit kilowatt) and oil circulation path parameters of oil-immersed transformer (such as oil flow velocity and oil channel cross-sectional area); Fluent software is selected for computational fluid dynamics, and Maxwell software is selected for electromagnetic field simulation. The construction process of the digital twin environment is as follows: First, import the three-dimensional geometric model of the substation into Fluent, set boundary conditions (including wall thermal conductivity, fan inlet and outlet pressure, and transformer heating power), and divide the grid (approximately five million grids, using tetrahedral grids, with localized refinement areas including the transformer surface and the vicinity of ventilation openings); then, establish the transformer electromagnetic field model in Maxwell, set the excitation current and material properties, calculate eddy current loss and hysteresis loss, and import the loss results as a heat source into Fluent; finally, couple Fluent and Maxwell for simulation, setting the time step to 0.01 seconds and the total simulation duration to 300 seconds, outputting the temperature field, flow field, and electromagnetic field data for each time step, and extracting sensor readings (temperature, humidity, gas concentration, partial discharge amplitude), equipment status (fan speed, valve opening), and wall structure status (magnetic induction switch signal) from them. The time step of the digital twin environment is set to be consistent with the real data acquisition frequency to ensure that the simulation data and the real data match in terms of time granularity. The real data acquisition frequency is 100 Hz, that is, data is acquired once every 0.01 seconds. Therefore, the time step of the digital twin environment is also set to 0.01 seconds to ensure that the simulation output and the real sensor data can correspond one-to-one in time, which is convenient for aligning feature distribution in subsequent domain adversarial training. The specific operation of aligning the feature distribution of the digital twin environment to the feature distribution of the real substation using the domain adversarial training method is as follows: The system defines three components: a feature extractor, a deep Q-network, and a domain discriminator. The feature extractor maps the input state to a common feature vector, and the domain discriminator outputs the probability that the feature belongs to the digital twin environment. The feature extractor uses the first few layers of the deep Q-network, specifically: starting from the input layer, it sequentially passes through a three-layer residual network (128 neurons per layer, ReLU activation function) for the risk channel and a three-layer gated recurrent unit (128 neurons per layer) for configuring the channel. Then, the outputs of the two channels are fused through a bilinear interaction layer to obtain a 128-dimensional common feature vector. The domain discriminator consists of three fully connected layers: the first layer has 128 neurons with ReLU activation function; the second layer has 64 neurons with ReLU activation function; and the third layer has one neuron with sigmoid activation function, outputting a probability value between zero and one, representing the likelihood that the input feature comes from the digital twin environment. The closer the probability value is to one, the more likely the feature is to come from the digital twin environment; the closer it is to zero, the more likely it is to come from the real substation. State samples were collected from both the digital twin environment and the real substation, and the domain discriminant loss function was calculated. The state samples were collected as follows: ten thousand state samples were randomly selected from the digital twin environment; each state sample contained seven components: fire risk probability, fire subtype probability distribution, current transformer type identifier, average ambient temperature and humidity, and the current operating mode of the universal adapter interface. Similarly, ten thousand unlabeled state samples (excluding action and reward information) were randomly selected from the historical operation database of the real substation. The calculation process of the domain discriminant loss function was as follows: for each digital twin environment... For each real substation sample, the common feature vector output by the feature extractor is input into the domain discriminator to obtain the probability of belonging to the digital twin environment. The natural logarithm of this probability is taken, and then this value is summed over all digital twin environment samples and the negative average is taken. For each real substation sample, the common feature vector is input into the domain discriminator to obtain the probability of belonging to the digital twin environment. This probability is subtracted by 1, and the natural logarithm is taken again. Then, this value is summed over all real substation samples and the negative average is taken. The two averages are added together to obtain the domain discriminant loss function value. The smaller this value, the stronger the domain discriminator's ability to distinguish between the two domains. A gradient inversion layer is inserted between the feature extractor and the domain discriminator. The gradient inversion layer performs an identity mapping during forward propagation and multiplies the gradient by a preset gradient inversion tradeoff coefficient during backward propagation. The specific value of the gradient inversion tradeoff coefficient is set to 0.1. The working principle of the gradient inversion layer is as follows: during forward propagation, the output is exactly what is input, without any changes; during backward propagation, the gradient from the back propagation is multiplied by negative 0.1 and then passed to the feature extractor. The effect of this is that when the feature extractor updates its parameters, it will optimize in the direction of increasing the domain discriminative loss, that is, it attempts to make the domain discriminator unable to distinguish which domain the feature comes from, thereby achieving feature distribution alignment. The total loss function is a weighted sum of the loss function in the digital twin environment and the domain discriminant loss. Adversarial training is performed by alternately updating the parameters of the domain discriminator and the feature extractor against the deep Q-network, causing the feature distribution output by the feature extractor to converge across the two domains. The loss function in the digital twin environment is the TD error loss function of the dual deep Q-learning, calculated as follows: a batch (64 samples) is sampled from the priority experience replay buffer. For each sample, the Q-value of the online network for the selected action in the current state and the maximum Q-value of the target network for the next state are calculated. The immediate reward is added with a discount factor (0.95), multiplied by the target network Q-value minus the online network Q-value, to obtain the TD error. The squared TD errors of all samples are then averaged. The total loss function equals the loss function in the digital twin environment plus 0.1 times the domain discriminant loss. The training process employs an alternating update strategy: First, the parameters of the feature extractor and deep Q-network are fixed, and only the parameters of the domain discriminator are updated to minimize the domain discriminant loss, performing one step of gradient descent; then, the parameters of the domain discriminator are fixed, and the parameters of the feature extractor and deep Q-network are updated to minimize the total loss function, performing one step of gradient descent; this alternation is repeated for a total of 10,000 steps of adversarial training; after training, the classification accuracy of the domain discriminator for samples from two domains should be close to 50%, indicating that the feature distributions have been basically aligned; The specific operation of deploying the initial strategy, after minor adjustments, in a real substation is as follows: After domain adversarial training is completed, the feature extractor parameters are fixed, and a small number of labeled samples are collected in a real substation to fine-tune the parameters of the deep Q-network. The fine-tuning uses the same dual deep Q-learning algorithm as the training in the digital twin environment, but the learning rate is reduced to one-tenth of the original. The labeled samples are collected as follows: during normal operation of the real substation, maintenance personnel manually trigger several fire-fighting actions (such as switching detector modes or adjusting the type of extinguishing agent), and record the state, action, reward, and next state before and after each action, collecting a total of one hundred transition samples. Each sample contains: current state (seven components), selected action (one of thirty-six actions), immediate reward (calculated by the reward function), and next state (seven components). The learning rate for fine-tuning is set to 0.0001 (the original learning rate was 0.001), and the number of training steps is set to one thousand steps. During the fine-tuning process, only the fully connected layer parameters of the deep Q-network (i.e., the last layer that outputs thirty-six Q values) are updated, while the feature extractor parameters remain unchanged. After fine-tuning, the performance of the deep Q-network in the real environment should reach a level similar to that in the digital twin environment. The fine-tuned deep Q network is embedded in the edge computing gateway and interacts in real time with the multimodal heterogeneous data acquisition module and the spatiotemporal risk modeling module via industrial Ethernet to achieve online adaptive decision-making for substation fire protection. The edge computing gateway is an embedded industrial computer that supports industrial Ethernet communication and has at least four gigabit Ethernet ports for connecting the multimodal heterogeneous data acquisition module, the spatiotemporal risk modeling module, the universal adapter interface, and the fire protection facility controller. The solidification process is as follows: the fine-tuned deep Q-network parameters are written into the gateway's non-volatile memory in floating-point format. When the gateway starts, it automatically loads the parameters and initializes the inference engine. The online decision-making cycle is once per second, that is, the gateway obtains the latest state data from the multimodal heterogeneous data acquisition module and the spatiotemporal risk modeling module every second, inputs it into the deep Q-network to calculate the Q value of all actions, selects the action with the largest Q value, decomposes it into four sub-instructions, and sends them out via industrial Ethernet. The entire inference process takes no more than fifty milliseconds, meeting the real-time requirements. After deployment, real-world environmental data is continuously collected, and incremental fine-tuning is triggered periodically to address domain drift caused by equipment aging or environmental changes. The trigger cycle for incremental fine-tuning is set to once every seven days. During each incremental fine-tuning, 500 transfer samples (covering different transformer types and weather conditions) are randomly selected from the real-world operational data collected over the past seven days, and the fully connected layers of the deep Q-network are fine-tuned in 100 steps with a learning rate of 0.0001. During the fine-tuning process, the feature extractor parameters remain fixed. If the average Q-value of the deep Q-network output does not change significantly after three consecutive incremental fine-tunings, the incremental fine-tuning is paused, and the domain drift is considered to have stabilized. If the probability of fire risk is consistently high or the false alarm rate increases, an additional incremental fine-tuning is immediately triggered.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A substation structural system compatible with both dry-type and oil-immersed transformers, characterized in that, Specifically, it includes: Multimodal heterogeneous data acquisition module: In response to the state change signal of the magnetic induction switch on the switchable wall module, it acquires environmental sensing data from the universal adapter interface and operating parameters from the transformer body. It performs time stamp synchronization and feature alignment on the state change signal, environmental sensing data and operating parameters, determines the current transformer type identifier based on the state change signal, performs statistical analysis on the environmental sensing data to obtain the average environmental temperature and humidity, obtains the current working mode of the universal adapter interface, and generates the initial feature vector of the node in a unified semantic space. Spatiotemporal risk modeling module: Receives the initial feature vector of the node, abstracts all sensors, equipment, fire protection facilities and wall structure in the substation into a spatiotemporal graph network, uses graph attention network to perform multi-layer message passing, calculates the attention coefficient between nodes and updates the node features, and outputs the fire risk probability and fire subtype probability distribution at the current moment. Dynamic strategy generation module: The state space is composed of fire risk probability, fire subtype probability distribution, current transformer type identifier, average ambient temperature and humidity, and the current working mode of universal adapter interface. The optimal strategy is solved by deep Q network. The optimal action is selected according to the reward function. The module generates detector type switching instructions, fire extinguishing agent type adjustment instructions, fire damper control instructions and alarm level instructions, and sends them to universal adapter interface and fire protection facilities. Transfer learning deployment module: Construct a high-fidelity digital twin environment based on the structural parameters of a switchable wall module, a universal adapter interface, and the transformer body. Use a deep reinforcement learning agent to train in this digital twin environment for multiple rounds to obtain an initial policy. Use a domain adversarial training method to align the feature distribution of the digital twin environment to the feature distribution of the real substation, so that the initial policy can be deployed in the real substation after a few minor adjustments.

2. The substation structure system compatible with both dry-type and oil-immersed transformers according to claim 1, characterized in that: In the multimodal heterogeneous data acquisition module, the specific operations for timestamp synchronization and feature alignment of state change signals, environmental sensor data, and operating parameters are as follows: First, the optimal time delay between the state change signal and the environmental sensing data, as well as the optimal time delay between the state change signal and the operating parameters of the transformer body, are calculated based on the cross-correlation function. The optimal time delay is then used to shift the environmental sensing data and the operating parameters of the transformer body to align the time bases of the three, resulting in an aligned discrete time series. Then, using the kernel canonical correlation analysis method, radial basis kernel function kernel matrices of environmental sensing data and transformer body operating parameters are constructed respectively. The generalized eigenvalue problem is solved to obtain the eigenvectors corresponding to the first few largest eigenvalues, which form a projection matrix. Then, the environmental sensing data and transformer body operating parameters in the aligned discrete time series are mapped to the common semantic space to obtain environmental alignment features and equipment alignment features with consistent dimensions.

3. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 2, characterized in that: The specific operation for generating the initial feature vector of a node is as follows: Calculate the arithmetic mean of the state change signal within the current time window to obtain the mean of the state change signal; use the mode of the state change signal within the current time window as the current transformer type identifier, with a value of zero or one, corresponding to dry-type transformer or oil-immersed transformer respectively. For the temperature and humidity components in the environmental sensing data, the weighted average method is used to calculate the weighted average, where the weights decrease exponentially as the sample moves away from the current time, thus obtaining the mean environmental temperature and mean environmental humidity. The current working mode of the universal adapter interface is read directly from the universal adapter interface register. The value is zero or one, which corresponds to the air intake sampling mode or the cable temperature sensing mode, respectively. Finally, the environmental alignment features, device alignment features, and mean state change signals are fused by outer product to construct a third-order tensor, which is then flattened into a one-dimensional vector. This vector is then compressed through a fully connected layer and activated by the hyperbolic tangent function to obtain the initial feature vector of the node.

4. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 3, characterized in that: In the spatiotemporal risk modeling module, the specific operation of abstracting all sensors, equipment, fire protection facilities, and wall structures within the substation into a spatiotemporal graph network is as follows: First, each sensor, device, fire protection facility, and wall structure is defined as a node, and the total number of nodes is a positive integer; each node carries an initial feature vector from the multimodal heterogeneous data acquisition module; Then, edges are constructed based on the physical proximity and logical control relationships between nodes: for any two nodes, if their Euclidean distance in three-dimensional space is less than a preset spatial distance threshold, an undirected edge is established. If there is a direct signal transmission or control link between two nodes, an additional directed edge is established, with the direction from the controlling party to the controlled party. Finally, each edge is assigned an initial weight, which is determined by the spatial distance and logical relationship indicator function between the two nodes: the smaller the spatial distance and the stronger the logical relationship, the greater the initial weight.

5. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 4, characterized in that: The specific operations for multi-layer message passing using graph attention networks are as follows: In each layer of the network, the attention coefficient of each node to all its neighboring nodes is first calculated. The attention coefficient integrates the node feature similarity, the initial weight, and the time decay factor. The time decay factor is determined by the time difference of the last state update of two nodes and the time decay constant. The time decay constant is set to 300 seconds, so that nodes that have recently interacted receive higher attention weights. Then, a multi-head attention mechanism is adopted, which sets up multiple attention heads, each of which independently calculates the attention coefficients, and the results of the calculations of multiple attention heads are concatenated and then linearly transformed to obtain the updated node features; After multiple layers of message propagation, each node obtains the final node characteristics containing global context information; Subsequently, a subset of key nodes representing transformers in the equipment, wall modules in the wall structure, detectors in the sensors, and fans in the equipment are selected from all nodes. The final node features of each node in the subset of key nodes are then global pooled based on an attention mechanism to obtain a global pooled feature vector. The global pooling feature vector is input into two parallel fully connected classification heads: the first classification head is a single neuron and uses the sigmoid activation function to output the fire risk probability at the current time, with a value ranging from zero to one; The second classification head consists of three neurons and uses the softmax activation function to output the probability distribution of fire subtypes. The three output components correspond to the three subtypes of coil overheating, oil leakage, and electric arc fire, respectively, and the sum of each component is equal to one.

6. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 5, characterized in that: In the dynamic policy generation module, the specific operations that constitute the state space are as follows: First, the fire risk probability, fire subtype probability distribution, current transformer type identifier, ambient temperature and humidity mean, and universal adapter interface current working mode are taken as seven state components. The fire subtype probability distribution contains three components. The temperature and humidity components in the ambient temperature and humidity mean are linearly normalized, and the temperature component is scaled to between zero and one. The humidity component is also scaled to between zero and one. The current transformer type identifier and universal adapter interface current working mode remain as binary discrete values. Then, a mutual information-guided attention fusion mechanism is adopted to calculate the mutual information estimate between each state component and the action sampled from the historical running data. The mutual information estimate is approximated by the kernel density estimation method. Then, all mutual information estimates are normalized by the softmax function with temperature coefficient to obtain the attention weight of each state component. Finally, each state component is mapped to an embedding vector through a dedicated embedding function, multiplied by the corresponding attention weight, and all weighted embedding vectors are concatenated to obtain the fused encoded state vector.

7. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 6, characterized in that: The specific operation of selecting the optimal action based on the reward function is as follows: The reward function consists of four terms. The first term is the matching reward, which is positive when the selected action makes the detector type match the current transformer type, and zero otherwise. The second term is the cost penalty, which is proportional to the energy consumption and equipment wear required to execute the action. The third term is the risk change penalty, which is measured by the KL divergence between the fire risk probability before and after the action, and penalizes behaviors with a sudden increase in risk. The fourth term is the safety constraint violation penalty, which is applied by multiplying the excess by a preset weighting coefficient when the fire risk probability after the action exceeds a preset safety threshold. A priority experience replay buffer is constructed to store historical transfer samples. Two deep Q-networks, an online network and a target network, are constructed and trained using a dual deep Q-learning algorithm. The loss function is constructed based on the difference between the Q-values ​​output by the online network and the target network. After training, an epsilon greedy strategy is used to select the optimal action during the inference phase. The selected optimal action is broken down into detector type switching command, extinguishing agent type adjustment command, fire damper control command, and alarm level command, which are then sent to the universal adapter interface and fire protection facilities via industrial Ethernet.

8. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 7, characterized in that: In the transfer learning deployment module, the specific operations for building a high-fidelity digital twin environment are as follows: Based on the geometric dimensions of the switchable wall module, the physical interface parameters of the universal adapter interface, and the thermal characteristic parameters of the transformer body, a digital twin environment is constructed using computational fluid dynamics software and electromagnetic field simulation software. The digital twin environment can accurately simulate the evolution process of various fire scenarios, such as arc ignition in oil-immersed transformers and overload smoke in dry-type transformers, and generate sensor readings, equipment status, and wall structure status consistent with those of real substations. The time step of the digital twin environment is set to match the frequency of real data acquisition to ensure that the simulation data matches the real data in terms of time granularity.

9. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 8, characterized in that: The specific operation of aligning the feature distribution of the digital twin environment to the feature distribution of the real substation using the domain adversarial training method is as follows: The system defines three components: a feature extractor, a deep Q-network, and a domain discriminator. The feature extractor maps the input state to a common feature vector, and the domain discriminator outputs the probability that the feature belongs to the digital twin environment. State samples are collected from the digital twin environment and the real substation, respectively, and the domain discriminant loss function is calculated. A gradient reversal layer is inserted between the feature extractor and the domain discriminator. The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a preset gradient reversal tradeoff coefficient during backward propagation. The total loss function is composed of the weighted sum of the loss function in the digital twin environment and the domain discriminator loss. By alternately updating the parameters of the domain discriminator and the parameters of the feature extractor and the deep Q network, adversarial training is performed to make the feature distribution output by the feature extractor tend to be consistent in the two domains.

10. A substation structure system compatible with both dry-type and oil-immersed transformers according to claim 9, characterized in that: The specific operation of deploying the initial strategy, after minor adjustments, in a real substation is as follows: After the domain adversarial training is completed, the feature extractor parameters are fixed, a small number of labeled samples are collected in a real substation, and the parameters of the deep Q network are fine-tuned. The fine-tuning adopts the same dual deep Q learning algorithm as the training in the digital twin environment, but the learning rate is reduced to one-tenth of the original. The fine-tuned deep Q network is embedded in the edge computing gateway and interacts in real time with the multimodal heterogeneous data acquisition module and the spatiotemporal risk modeling module via industrial Ethernet to achieve online adaptive decision-making for substation fire protection. After deployment, real-world environmental data is continuously collected, and incremental fine-tuning is triggered periodically to address domain drift caused by equipment aging.

Images