Tunnel modular prefabricated cabin power supply and distribution intelligent substation self-adaptive regulation and control system based on edge calculation

Abstract

The invention relates to the technical field of tunnel power supply and distribution, in particular to a tunnel modular prefabricated cabin power supply and distribution intelligent substation self-adaptive regulation and control system based on edge calculation. Comprising an edge calculation and AI decision-making unit which is used for realizing rapid acquisition, processing and instant decision-making of tunnel power supply and distribution multi-dimensional data, generating a power supply and distribution adaptive regulation and control strategy by deploying calculation resources and a machine learning algorithm at edge nodes close to a data source, and converting the strategy into an executable regulation and control instruction; a cloud platform collaborative management unit; and an intelligent sensing and internet-of-things unit. According to the invention, hierarchical decision control of the tunnel power supply and distribution system is realized by constructing a hybrid architecture of edge computing and cloud platform collaboration and a priority judgment mechanism; according to the invention, multi-modal data are integrated through the multi-protocol communication link module and the full-scene data fusion analysis module, and data association analysis is realized through Kalman filtering, D-S evidence theory and other algorithms.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel power supply and distribution, and in particular to an edge computing-based adaptive control system for a tunnel modular prefabricated cabin power supply and distribution intelligent substation. Background Art

[0002] As tunnel infrastructure evolves towards intelligentization, tunnel power supply and distribution systems face the dual challenges of real-time control and global energy efficiency optimization. Traditional centralized cloud platform control models suffer from high network latency and low device coordination efficiency in complex tunnel environments. This is especially true in emergency response scenarios for equipment failures, where it is difficult to strike a dynamic balance between local security protection and cloud resource scheduling. The introduction of edge computing technology provides a distributed computing architecture for tunnel power supply and distribution systems. However, achieving collaborative control between edge nodes and cloud platforms, fusion analysis of multi-source data, and decision-making across multiple systems remain technical challenges that need to be addressed.

[0003] For example, Chinese patent CN202410055014.7 discloses a tunnel comprehensive integrated cloud control system, which includes an application module, a data processing module, a communication module, a perception module and a control module; the perception module is used to collect monitoring data of various main electromechanical equipment in the tunnel; the communication module is used to upload the monitoring data; the data processing module is used to process the monitoring data and confirm the operating conditions of various main electromechanical equipment in the tunnel; the application module is used to query the operating conditions of various main electromechanical equipment in the tunnel and issue control instructions; the control module is used to transmit the control instructions to the corresponding main electromechanical equipment. Through the above method, the operating conditions of various main electromechanical equipment in the tunnel can be monitored, the monitoring difficulty is low, and it is conducive to the effect of comprehensive integrated management of the tunnel. For example, Chinese patent CN202210826831.9 discloses a tunnel management and control system based on cloud-edge-end collaboration, which includes a tunnel edge controller and a tunnel terminal system and a cloud platform server interconnected with the tunnel edge controller; the tunnel terminal system is used to detect the lighting, fire protection, environment, traffic and power distribution data in the tunnel and feed it back to the tunnel edge controller; the tunnel edge controller is used to process the data detected by the tunnel terminal system and transmit it to the cloud platform server; the cloud platform server is used to store the data in the tunnel edge controller. In the present invention, a tunnel edge controller is set up for data processing, so that it has multi-device AI connection capability, access to RSU, intelligent sensing terminal equipment, and intelligent control terminal equipment, and realizes the access, aggregation and management functions of tunnel main body and ancillary facilities monitoring data, driving vehicle operation information and roadside environment perception data.

[0004] While the aforementioned existing technical solutions all have their design advantages, they also suffer from the following technical drawbacks: First, a lack of a control decision-making hierarchy: The pure cloud-based architecture of Chinese Patent CN202410055014.7 and the "edge-only forwarding" model of Chinese Patent CN202210826831.9 fail to establish a "layered coordination mechanism between local emergency decision-making (such as fault protection) and cloud-based global optimization (such as energy efficiency scheduling)." Short circuits and arc faults in tunnels require millisecond-level responses, but the command transmission delays of pure cloud-based architectures or the edge-only forwarding logic prevent the prioritized triggering of local protection at the moment of a fault. It also struggles to reconcile conflicts between safety commands and energy efficiency scheduling commands, making it difficult to balance the timeliness and rationality of system responses in emergency conditions. Second, data fusion is insufficiently deep: Chinese patent CN202410055014.7 focuses on the electrical parameters of electromechanical equipment, while Chinese patent CN202210826831.9 extends this to data from multiple systems, including lighting and fire protection. However, both remain at the basic level of "data collection, aggregation, and storage," failing to construct a multimodal correlation model encompassing "electrical parameters + environmental parameters + equipment lifecycle status." Due to the lack of deep data fusion and fault deduction capabilities, complex operating conditions such as equipment aging and compound faults are difficult to provide early warning, hindering the effectiveness of preventative maintenance for tunnel power supply and distribution systems. Third, physical deployment is poorly adaptable: Both Chinese patents CN202410055014.7 and CN202210826831.9 focus on system functional architecture, failing to design modular equipment integration and physical protection solutions tailored to the high humidity, dusty, narrow, and dispersed environments of tunnels. Existing solutions involve dispersed equipment installation, making it impossible to enhance protection levels through cabin isolation. Rapid deployment and fault isolation of segmented tunnel power supply and distribution are also difficult. During implementation, these solutions present significant challenges, such as equipment crosstalk and low maintenance efficiency, increasing the operational complexity of the system throughout its lifecycle. To address this, we propose an adaptive control system for intelligent substations in tunnel power supply and distribution, built on edge computing. Summary of the Invention

[0005] The purpose of the present invention is to provide an adaptive control system for tunnel modular prefabricated cabin power supply and distribution intelligent substation based on edge computing to solve the problems raised in the above background technology.

[0006] To solve the above technical problems, the present invention aims to provide an edge computing-based adaptive control system for a tunnel modular prefabricated cabin power supply and distribution intelligent substation, comprising: The edge computing and AI decision-making unit is used to rapidly collect, process, and make instant decisions on multi-dimensional data on tunnel power supply and distribution. By deploying computing resources and machine learning algorithms at edge nodes close to data sources, it generates adaptive power supply and distribution control strategies and converts these strategies into executable control instructions. Based on the collaborative control instructions (including strategy type, parameter thresholds, and execution priority) issued by the cloud platform collaborative management unit, the final control instructions are generated through a priority determination mechanism (local emergency instructions > cloud optimization instructions). The cloud platform collaborative management unit builds a global data storage and analysis center to receive edge computing node data and provide a global perspective on power grid operation. Through big data analysis and resource optimization algorithms, it issues collaborative control instructions to edge computing and AI decision-making units to assist in optimizing local decision-making. The intelligent sensing and IoT unit is used to monitor the operating parameters of the tunnel power supply and distribution equipment, the cabin environment, and the fire protection status in real time. It uses a multi-protocol communication link module to transmit data to the edge computing and AI decision-making unit and the cloud platform collaborative management unit, thereby building a full-scene monitoring system through the full-scene data fusion analysis module; The modular prefabricated cabin unit integrates power supply and distribution equipment, fire protection systems, and edge computing components. It adopts a split steel structure design of high-pressure and low-pressure cabins, combined with fireproof and heat insulation technology to achieve physical isolation, rapid installation, and environmental protection. The automated adaptive control unit is used to execute the final control instructions of the edge computing and AI decision-making units. Through the multi-system linkage logic of power supply and distribution circuit switching, micro-environment control, and decision-making and control modules, it realizes real-time monitoring of power flow and emergency response to sudden incidents.

[0007] As a further improvement of this technical solution, the edge computing and AI decision-making unit includes a data processing module, a strategy generation module and an instruction conversion module, wherein: The data processing module is used to clean and normalize the raw data transmitted by the intelligent sensing and IoT units, and remove noise using a sliding window filtering algorithm with a preset window width; The strategy generation module generates a power supply and distribution adaptive control strategy based on a multi-layer neural network, the neural network input is historical load data, and the output is the power forecast value for the future period; The instruction conversion module is used to convert the control strategy into executable control instructions, adopting the industrial control standard protocol, and the instruction generation delay meets the real-time control requirements.

[0008] As a further improvement of this technical solution, the edge computing and AI decision-making unit generates the final control instruction through the priority determination mechanism, including the following steps: S100.1. Parse local real-time commands (such as emergency power-off commands triggered by equipment failures) and coordinated control commands (such as power supply optimization strategies during peak and off-peak periods) received by the edge computing and AI decision-making unit, and extract key features for priority calculation. These key features include command type, target, parameter value range, and timeliness attributes. Specifically: Instruction type (divided into emergency and optimization types); Target (targeted to specific device or system module); Parameter value range (safety boundaries of key parameters in acquisition instructions); Timeliness attribute (records the time difference between instruction generation time and trigger time).

[0009] S100.2. Calculate the local command priority value using a two-factor weight model based on device status and command attributes. Coordinated control instruction priority value , the calculation formula is: in, represents the urgency factor; Indicates the current device temperature rise; Indicates the average temperature rise under the same historical load, used to reflect the degree of equipment abnormality; ; in, represents the optimization factor; Indicates the difference between the real-time load rate and the rated load rate; Indicates the rated load rate, reflecting the urgency of system regulation; S100.3. Based on local instruction priority value Coordinated control instruction priority value The hierarchical decision logic is executed based on the calculation result of the hierarchical decision logic, which specifically includes: First priority strategy (corresponding to emergency scenarios): when When the instruction type is emergency (such as equipment short circuit, over-temperature alarm), the edge computing and AI decision-making unit generates the highest priority control instruction, and executes the equipment protection action (such as cutting off the fault circuit, starting the backup power supply) through the automated adaptive control unit. The threshold value of equipment safety urgency (value range 0-1) is used to distinguish between normal operation and emergency failure scenarios. Determination of equipment safety margin based on modular prefabricated cabin units 400; Specifically, the edge computing and AI decision-making unit immediately triggers the highest priority control process and interrupts the cloud command queue; the automated adaptive control unit performs fault isolation (such as cutting off the circuit and starting cooling) based on the equipment's safety response characteristics, and after the action is completed, it transmits the system clock synchronization status information through the intelligent perception and IoT unit; the cloud platform collaborative management unit receives the command conflict alarm, including the priority difference ( ), fault type (determined by local sensor data); Parameter fusion strategy (corresponding to collaborative optimization scenario): when Or if the instruction type is optimization (such as peak-valley scheduling and reactive power compensation), the control parameters of the local real-time instruction and the coordinated control instruction are weighted and integrated: ; in, represents the final control parameters sent to the automated adaptive control unit after fusion, and Meet the electrical safety boundaries of the modular prefabricated cabin unit (such as voltage regulation range ±α% rated value, α is provided by the equipment manufacturer) and mechanical life constraints (such as tap switching interval ≥ minimum equipment operation cycle); Indicates the local control parameters generated by edge computing and AI decision-making units; Represents the cloud control parameters generated by the cloud platform collaborative management unit; Second priority strategy (corresponding to global optimization scenario): when When the edge computing and AI decision-making units give priority to executing cloud collaborative instructions, the system operating parameters (such as load distribution and reactive power compensation) are adjusted through the automated adaptive control unit. It is a global optimization priority threshold used to balance local security and cloud energy efficiency goals, and the threshold Determination of the global energy efficiency model based on the collaborative management unit of the cloud platform.

[0010] Specifically, fusion instructions must pass parameter legitimacy, device compatibility, and grid stability verification. Any failure triggers manual intervention: Parameter validity: Check whether the electrical / mechanical parameters exceed the rated range of the modular prefabricated cabin unit. If they exceed the range, it will be marked as "parameter abnormality"; Equipment compatibility: Verify the number of operations / intervals allowed for the remaining life of circuit breakers, capacitors, etc. If exceeded, mark "life warning"; Grid stability: Power flow calculations are used to verify whether the line / busbar operating status complies with system safety boundaries (e.g., active power flow and voltage deviation meet grid specifications). If the limits are exceeded, a "power flow abnormality" is marked.

[0011] As a further improvement of this technical solution, the edge computing and AI decision-making unit further includes a strategy verification module, a parameter optimization module, and a version management module, wherein: The strategy verification module verifies the effectiveness of the control strategy through a Monte Carlo simulation algorithm. The preset number of simulations is the industry standard value. Based on the device operating parameters collected by the intelligent sensing and IoT units, an assessment report containing the risk probability distribution is generated. The assessment report is used to correct the priority determination threshold parameters built into the edge computing and AI decision-making units. The parameter optimization module dynamically adjusts policy parameters based on a deep reinforcement learning framework, with minimizing grid losses as the optimization goal. The optimization cycle is dynamically issued by the cloud platform collaborative management unit based on load fluctuation characteristics, and directly outputs the dual-factor weight parameters used for priority calculation in the edge computing and AI decision-making units. The version management module uses blockchain hash storage technology to record historical versions of the strategy (including the priority determination strategy configuration of the edge computing and AI decision-making units), supports one-click rollback to the previous valid version in the event of an exception, and the rollback time meets the equipment fault tolerance requirements of the modular prefabricated cabin unit. The rollback trigger condition is linked to the security verification results of the edge computing and AI decision-making units.

[0012] As a further improvement to this technical solution, the global data storage and analysis center of the cloud platform collaborative management unit adopts a distributed time series database cluster architecture to perform millisecond-level sampling and storage of device operation data (voltage, current, temperature, etc.) uploaded by the edge computing and AI decision-making units, and achieve global visualization of the power grid operation status. The distributed time series database cluster architecture specifically includes: A digital twin is constructed based on a three-dimensional tunnel topology model and power flow algorithms, rendering a real-time heat map of the load distribution of the power supply and distribution network. The long short-term memory network (LSTM) is used to predict the trend of historical data and generate a warning map of power grid operation risks in the next 24 hours.

[0013] As a further improvement of this technical solution, the cloud platform collaborative management unit uses big data analysis and resource optimization algorithms to generate collaborative control instructions, including the following steps: S200.1. Use consistent hashing to manage the tunnel power distribution network in different zones and monitor the load rate of each zone in real time. When the load rate of a zone exceeds a preset overload threshold (e.g., 80%) of the rated value, a load transfer strategy is triggered for the adjacent zone, and the transfer power threshold is calculated using the following formula: : ; in, is the regional rated capacity, is a dynamic adjustment factor (obtained by training the inter-regional line impedance matrix), and the preset overload threshold is a conventional load rate critical value in the industry (such as the overload standard defined in GB / T50598-2010).

[0014] S200.2. With minimizing grid losses, maximizing new energy consumption, and maximizing power supply reliability as the objective functions, a non-dominated sorting genetic algorithm is used for iterative optimization to generate coordinated control instructions that include transformer tap adjustment gears (adjustment step size ≤ 1.25% of rated voltage), the number of capacitor switching groups (single switching capacity difference ≤ 500 kvar), and the execution timing of the load transfer strategy (action interval between adjacent areas ≥ 10 seconds); the coordinated control instructions are approximate Pareto frontier solutions for edge computing and AI decision-making units to select and execute based on local real-time status.

[0015] As a further improvement of this technical solution, the multi-protocol communication link module includes an industrial bus sub-link sub-module, a wireless sensor sub-link sub-module and a 5G core sub-link sub-module, wherein: The industrial bus sub-link submodule uses the Modbus protocol to build a dedicated communication channel for power equipment, performing millisecond-level periodic acquisition of voltage, current, and temperature parameters of transformers and switchgear in the tunnel, adapting to the strong real-time monitoring requirements of power equipment; The wireless sensor sub-link sub-module: builds an environmental monitoring network based on the LoRaWAN protocol, collects temperature, humidity, and smoke concentration parameters, optimizes node energy consumption through a sleep-wake low-power communication mechanism, and adapts to tunnel wiring-restricted scenarios; The 5G core sub-link submodule uses 5G URLLC slice channels to perform real-time priority transmission of fire alarm and gas concentration data, and the transmission delay meets the fire emergency response standard requirements; The full-scenario data fusion analysis module includes a multimodal data fusion submodule, an equipment health assessment submodule, and an adaptive sampling control submodule, wherein: The multimodal data fusion submodule is used to align the electrical parameters (current, voltage), environmental parameters (temperature and humidity), and fire protection parameters (smoke concentration) in time and space, and uses the Kalman filter algorithm to eliminate data noise. The fusion accuracy meets the power equipment status monitoring standard; The equipment health assessment submodule builds a fault diagnosis model based on DS evidence theory, extracts the circuit breaker opening and closing coil current waveform and transformer oil temperature change rate as core feature quantities, and realizes early warning of equipment failures. After receiving the abnormal working condition signal sent by the edge computing and AI decision-making unit, the adaptive sampling control submodule automatically adjusts the sampling frequency of key equipment and constructs a three-level sampling mechanism of normal periodic acquisition-abnormal trigger acquisition-high-frequency tracking acquisition to meet the dynamic monitoring accuracy requirements under different working conditions.

[0016] As a further improvement of this technical solution, the modular prefabricated cabin unit includes a high-pressure cabin module, a low-pressure cabin module and a fire protection integrated module, wherein: The high-voltage cabin module includes a switch cabinet submodule and a transformer submodule, wherein: The switch cabinet submodule has a built-in 10kV metal-clad removable switch cabinet with a Q355B steel plate frame and an arc protection sensor installed inside the cabinet; The transformer submodule adopts a dry-type transformer, the shell adopts a structure with a protection level of not less than IP54, and the heat dissipation fan is linked with the temperature controller; The low-pressure cabin module includes an intelligent distribution submodule and an edge computing server submodule, wherein: The intelligent distribution submodule integrates a molded case circuit breaker and an intelligent power meter and supports Modbus protocol communication; The edge computing server submodule adopts an industrial-grade server and is installed with a shock-absorbing bracket to adapt to a wide temperature operating environment; The fire protection integrated module includes a fire extinguishing device submodule and a fire detection submodule, wherein: The fire extinguishing device submodule integrates a heptafluoropropane gas fire extinguishing system, with nozzles arranged at preset intervals and equipped with a linkage trigger mechanism; The fire detection submodule adopts a smoke and heat composite detector, and the detection sensitivity complies with relevant national standards.

[0017] As a further improvement of this technical solution, the automated adaptive control unit includes an intelligent power distribution execution module and a microenvironment control module, wherein: The intelligent power distribution execution module uses a solid-state circuit breaker to build an intelligent power distribution network. The intelligent power distribution execution module includes a fault isolation submodule and a dynamic load distribution submodule, wherein: The fault isolation submodule is based on a high-speed current sensor and a solid-state circuit breaker to achieve rapid removal of short-circuit faults; The dynamic load distribution submodule dynamically adjusts the transformer tap position through a fuzzy PID controller based on the load prediction results of the edge computing and AI decision-making unit; The microenvironment control module realizes adaptive adjustment of the tunnel microenvironment through the IoT edge controller, including a multi-parameter collaborative control submodule and a partition temperature balancing submodule, wherein: The multi-parameter coordinated control submodule uses a model predictive control algorithm to dynamically adjust the ventilation fan speed based on CO concentration, visibility and traffic flow; The zone temperature balancing submodule adjusts the tunnel lighting system power through a PID controller to achieve cabin temperature difference control.

[0018] As a further improvement of this technical solution, the decision and control module includes a collaborative control submodule and an emergency response decision tree submodule, wherein: The collaborative control submodule builds an inter-system communication mechanism based on the event bus architecture, specifically including: Assign dynamic priorities to fire protection, ventilation, lighting and other systems (e.g., fire protection system priority is the highest level 1) to ensure that the response time of key systems meets the real-time requirements of fire emergency response; The Raft consensus algorithm is used to achieve state synchronization between control units, and the synchronization accuracy meets the collaborative control standards of industrial-grade distributed systems.

[0019] The emergency response decision tree submodule builds an emergency response model based on the Bayesian network inference engine, specifically including: Predefine a set of typical failure modes for the entire life cycle of the tunnel power supply and distribution system (covering basic faults such as single-phase grounding and three-phase short circuit, as well as derived compound faults); Bayesian reasoning is used to update the fault probability distribution in real time and generate the optimal emergency response strategy. The strategy generation timeliness meets the rapid response requirements of power system fault handling.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves hierarchical decision-making and control for the tunnel power supply and distribution system by building a hybrid architecture combining edge computing and cloud platforms and a priority determination mechanism. When an emergency fault such as a short circuit or overtemperature occurs, the edge computing unit prioritizes local protection actions (such as disconnecting the faulty circuit), while the cloud platform provides global resource optimization assistance. This mechanism effectively reconciles conflicts between local safety instructions and cloud-based energy efficiency instructions, ensuring timely and reasonable system responses in emergency situations and avoiding the command delays associated with traditional pure cloud-based architectures.

[0021] 2. This invention integrates multimodal data such as electrical parameters, environmental parameters, and fire protection parameters through a multi-protocol communication link module and a full-scenario data fusion and analysis module. It also implements data correlation analysis using algorithms such as Kalman filtering and DS evidence theory. This solution improves the accuracy of equipment fault diagnosis and can predict complex faults such as insulation aging and surface discharge. Compared with traditional single-dimensional monitoring solutions, it effectively enhances the preventive maintenance capabilities of tunnel power supply and distribution systems.

[0022] 3. This invention utilizes a separate steel structure for the high-pressure and low-pressure compartments, combined with fireproofing and thermal insulation technology and an IP54 protection rating, making it suitable for the complex, humid, and dusty environments of tunnels. This design achieves physical isolation and rapid installation, supporting the deployment requirements of segmented power supply and distribution in tunnels. Furthermore, the modular design of the compartments reduces the risk of equipment crosstalk and improves maintenance efficiency, effectively addressing the adaptability issues of traditional substation cabinet designs in tunnel scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the system framework of the present invention; The meaning of each number in the figure is: 100, edge computing and AI decision-making unit; 110, data processing module; 120, strategy generation module; 130, instruction conversion module; 140, strategy verification module; 150, parameter optimization module; 160, version management module; 200. Cloud platform collaborative management unit; 300, intelligent sensing and IoT unit; 310, multi-protocol communication link module; 311, industrial bus sub-link sub-module; 312, wireless sensor sub-link sub-module; 313, 5G core sub-link sub-module; 320, full-scenario data fusion analysis module; 321, multimodal data fusion sub-module; 322, equipment health assessment sub-module; 323, adaptive sampling control sub-module; 400, modular prefabricated cabin unit; 410, high-voltage cabin module; 411, switchgear submodule; 412, transformer submodule; 420, low-voltage cabin module; 421, intelligent distribution submodule; 422, edge computing server submodule; 430, fire protection integration module; 431, fire extinguishing device submodule; 432, fire detection submodule; 500, automated adaptive control unit; 510, intelligent power distribution execution module; 511, fault isolation submodule; 512, dynamic load distribution submodule; 520, micro-environment control module; 521, multi-parameter collaborative control submodule; 522, zone temperature balancing submodule; 530, decision and control module; 531, collaborative control submodule; 532, emergency response decision tree submodule. DETAILED DESCRIPTION

[0024] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this embodiment provides an edge computing-based tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system, including: The edge computing and AI decision-making unit 100 is used to rapidly collect, process, and make real-time decisions on multi-dimensional data on tunnel power supply and distribution. By deploying computing resources and machine learning algorithms at edge nodes close to data sources, it generates adaptive power supply and distribution control strategies and converts these strategies into executable control instructions. Based on the collaborative control instructions (including strategy type, parameter thresholds, and execution priority) issued by the cloud platform collaborative management unit 200, the final control instructions are generated through a priority determination mechanism (local emergency instructions > cloud optimization instructions). In this embodiment, the edge computing and AI decision-making unit 100 includes a data processing module 110, a strategy generation module 120, and an instruction conversion module 130, wherein: The data processing module 110 is used to clean and normalize the raw data transmitted by the intelligent sensing and IoT unit 300, and remove noise using a sliding window filtering algorithm with a preset window width; As a further illustration of this embodiment, the data processing module 110 performs real-time filtering and preprocessing on the raw data transmitted by the intelligent sensing and IoT unit 300, employing a dynamic sliding window filtering algorithm to enhance data reliability. Specifically, the algorithm adaptively adjusts the window width based on the device's operating status. Under normal operating conditions, a 100-point window (corresponding to a 100ms sampling period and a sampling frequency of 1kHz) is used to eliminate random noise through mean filtering. When the device's temperature rise rate exceeds a preset threshold (e.g., 0.5°C / s), the algorithm automatically switches to a narrower 50-point window to improve filtering accuracy for high-frequency noise. This sliding window filtering algorithm is implemented through the following mechanism: a circular buffer is used to store historical sampled data, with new data overwriting the oldest data. Filtering is performed through cumulative summation and mean calculation. When the data missing rate within the window exceeds 10%, a forward interpolation strategy is triggered to ensure data continuity. At the hardware level, a Xilinx Artix-7 FPGA can be used to build a parallel computing architecture. A pipelined design enables zero-delay processing of 100Hz sampled data, meeting the DL / T860 standard for real-time data processing in power systems. Furthermore, data processing module 110 incorporates built-in data validation logic to identify out-of-limit conditions and flag anomalies in filtered data, providing a reliable data foundation for subsequent strategy generation. This design, through dynamic window adjustment and hardware acceleration, effectively suppresses data noise in tunnels with high electromagnetic interference, ensuring the accuracy of edge computing unit decisions.

[0026] The strategy generation module 120 generates a power supply and distribution adaptive control strategy based on a multi-layer neural network. The neural network input is historical load data and the output is the power forecast value for the future period. As a further illustration of this embodiment, the strategy generation module 120 constructs a load forecasting model based on a multi-layer neural network to generate an adaptive control strategy for power supply and distribution. First, the strategy generation module 120 adopts a three-layer fully connected neural network structure: the input layer contains 12 neurons, corresponding to the historical 12 hours of load data; the hidden layer has 24 neurons and uses the ReLU activation function (the mathematical expression is ( , ), by retaining the linear response characteristics of non-negative input to avoid the gradient vanishing problem; the number of neurons is designed to correspond to the double expansion of the input dimension (12-dimensional historical load data), strengthening the dimensional support capability of nonlinear feature mapping; the output layer has 1 neuron, which outputs the power forecast value for the next hour; Secondly, the neural network was implemented using the TensorFlow framework and trained using the Adam optimizer to minimize grid losses. The initial learning rate was set to 0.001 and decayed by 10% every 50 iterations. The mean squared error (MSE) loss function was used to ensure prediction stability. Before model training, the historical load data was normalized and mapped to a distribution with a mean of 0.8 and a standard deviation of 0.2 using the Z-Score method to improve network convergence. The input sequence was constructed using a sliding window mechanism, sampling 12 consecutive hours of data with a step size of 1 hour. The batch size was set to 32, and training was repeated for 200 rounds, with a validation set accounting for 20% to avoid overfitting. To cope with sudden load fluctuations unique to tunnel scenarios, the module integrates a residual correction mechanism: when the deviation between the predicted value and the actual value exceeds 15% of the rated power, the LSTM-based residual network is automatically triggered. This network uses environmental parameters such as humidity and traffic volume in the tunnel as input, as well as the initial prediction error, and captures time series features through a 1-layer 16-unit LSTM layer. The LSTM layer uses the tanh activation function (mathematical expression: , used for cell state update, where The candidate value of the cell state corresponds to the time series feature input of the residual correction network: 8-dimensional data such as tunnel humidity, traffic flow, initial prediction error, etc., and the calculation result after weighting by the LSTM input gate; Represents state update amplitude adjustment) and Sigmoid activation function (mathematical expression is: ,in, represents the weighted sum of the inputs to the gating mechanism; The system uses a combination of 16 units (gated switches), gating logic for input gates, forget gates, and output gates) to adapt to the 8-dimensional input features composed of environmental parameters (humidity, traffic flow) and initial prediction errors. This system uses a double-redundant dimension to balance the accuracy and computational efficiency of low-dimensional time series modeling. A residual correction value is then output and added to the initial prediction result to form the final power prediction value, providing a more accurate decision-making basis for subsequent control strategies.

[0027] In addition, at the hardware level, the strategy generation module 120 can be deployed on the ARM Cortex-A72 architecture processor (such as NXP i.MX8MPlus) of the edge computing node, using its built-in NPU coprocessor to accelerate neural network calculations, and the single prediction time is controlled within 1 second, meeting the real-time requirements of the tunnel power supply and distribution system.

[0028] The instruction conversion module 130 is used to convert the control strategy into executable control instructions, using industrial control standard protocols, and the instruction generation delay meets real-time control requirements.

[0029] In this embodiment, generating the final control instruction through the priority determination mechanism includes the following steps: S100.1. Parse local real-time commands (e.g., emergency power-off commands triggered by equipment failures) and coordinated control commands (e.g., power supply optimization strategies during peak and off-peak periods) received by the edge computing and AI decision-making unit 100, and extract key features for priority calculation, including command type, target, parameter range, and timeliness attributes. Instruction type (divided into emergency and optimization types); Target (targeted to specific device or system module); Parameter value range (safety boundaries of key parameters in acquisition instructions); Timeliness attribute (records the time difference between instruction generation time and trigger time).

[0030] S100.2. Calculate the local command priority value using a two-factor weight model based on device status and command attributes. Coordinated control instruction priority value , the calculation formula is: in, represents the urgency factor; Indicates the current device temperature rise; Indicates the average temperature rise under the same historical load, used to reflect the degree of equipment abnormality; ; in, represents the optimization factor; Indicates the difference between the real-time load rate and the rated load rate; Indicates the rated load rate, reflecting the urgency of system regulation; S100.3. Based on local instruction priority value Coordinated control instruction priority value The calculation results are used to execute the hierarchical decision logic, which specifically includes: First priority strategy (corresponding to emergency scenarios): when When the instruction type is emergency (such as equipment short circuit, over-temperature alarm), the edge computing and AI decision-making unit 100 generates the highest priority control instruction, and executes the equipment protection action (such as cutting off the fault circuit, starting the backup power supply) through the automated adaptive control unit 500. The threshold value of equipment safety urgency (value range 0-1) is used to distinguish between normal operation and emergency failure scenarios. Determination of equipment safety margin based on modular prefabricated cabin units 400; Specifically, the edge computing and AI decision-making unit 100 immediately triggers the highest priority control process and interrupts the cloud command queue; the automated adaptive control unit 500 performs fault isolation based on the device safety response characteristics, such as cutting off the circuit and starting cooling. After the action is completed, the system clock synchronization status information is returned through the intelligent sensing and IoT unit 300; the cloud platform collaborative management unit 200 receives the command conflict alarm, including the priority difference ,The fault type is determined by local sensor data; Parameter fusion strategy (corresponding to collaborative optimization scenario): when Or if the instruction type is optimization (such as peak-valley scheduling and reactive power compensation), the control parameters of the local real-time instruction and the coordinated control instruction are weighted and integrated: ; in, represents the final control parameters sent to the automated adaptive control unit 500 after fusion, and The electrical safety limits (e.g., voltage regulation range ±α% of rated value, where α is provided by the equipment manufacturer) and mechanical life constraints (e.g., tap change interval ≥ minimum equipment operation cycle) of the modular prefabricated cabin unit 400 are met; represents the local control parameters generated by the edge computing and AI decision-making unit 100; represents the cloud control parameters generated by the cloud platform collaborative management unit 200; Second priority strategy (corresponding to global optimization scenario): when When the edge computing and AI decision-making unit 100 executes the cloud collaborative instructions first, the system operating parameters (such as load distribution and reactive power compensation) are adjusted through the automatic adaptive control unit 500. It is a global optimization priority threshold used to balance local security and cloud energy efficiency goals, and the threshold Determined based on the global energy efficiency model of the cloud platform collaborative management unit 200.

[0031] Specifically, fusion instructions must pass parameter legitimacy, device compatibility, and grid stability verification. Any failure triggers manual intervention: Parameter validity: Check whether the electrical / mechanical parameters exceed the rated range of the modular prefabricated cabin unit 400. If they exceed the range, it will be marked as "parameter abnormal"; Equipment compatibility: Verify the number of operations / intervals allowed for the remaining life of circuit breakers, capacitors, etc. If exceeded, mark "life warning"; Grid stability: Power flow calculations are used to verify whether the line / busbar operating status complies with system safety boundaries (e.g., active power flow and voltage deviation meet grid specifications). If the limits are exceeded, a "power flow abnormality" is marked.

[0032] As a further illustration of this embodiment, the urgency factor Dynamic correction based on historical fault handling results: When a short circuit occurs in the equipment, if the protection action successfully eliminates the fault (confirmed by the status signal sent back by the intelligent sensing unit), Keep the original value 1.5; if protection failure occurs (such as over-tripping), then The correction value is increased by 0.1 (upper limit 2.0) and stored in the EEPROM of the edge computing and AI decision-making unit 100. This mechanism can improve the success rate of handling similar faults after learning multiple fault samples; at the same time, the optimization factor Associated with the number of iterations of global optimization of the cloud platform: the initial value is 0.8 (load rate < 30%), and each time the cloud platform completes global energy efficiency optimization, it is reassigned according to the current load rate range. For example, when the load rate increases from 25% to 40%, Adjusted from 0.8 to 1.0 and synchronized to register 40023 of the edge computing and AI decision-making unit 100 through the Modbus protocol.

[0033] Furthermore, in this embodiment, for a 10kV dry-type transformer (rated temperature 105°C), when the current operating temperature is 80°C: Safety margin factor = (105-80) / 105 = 0.238, =0.238×0.3≈0.07; for a circuit breaker (rated current 1250A), when the real-time current is 1000A: Safety margin factor = (1250-1000) / 1250 = 0.2, at this time Take a fixed value of 0.2 (because the safety margin factor is less than 0.6); also, The initial value is 0.5 (the number of iterations on the cloud platform is n=0), and n+1 is added after each optimization. When n=10, =0.5+10×0.05=1.0, and will not increase after reaching the upper limit. This threshold is verified by the global energy efficiency model of the cloud platform collaborative management unit 200 to ensure a balance between local security and cloud optimization.

[0034] It is understood that the priority determination algorithm in this embodiment has a built-in abnormal operating condition fault tolerance mechanism. When the system detects the following operating anomalies, the fault tolerance process is automatically started to ensure the continuity and reliability of the control strategy, including: (1) Fault-tolerant processing of missing historical temperature rise data If the intelligent sensing and IoT unit 300 cannot obtain the historical average temperature rise of the device (Scenarios such as the first commissioning of equipment, storage module failure, etc.), based on the safety redundancy principle of power equipment temperature rise design, 80% of the equipment rated temperature rise is taken as a temporary replacement value. Taking a 10kV dry-type transformer as an example, when its rated temperature rise is 65℃, the temporary The assigned value is 52°C. This value refers to the "temperature rise margin under non-full load conditions" requirement in the thermal design specifications for power equipment to avoid misjudgment or failure of the protection strategy due to missing data. (2) Fault-tolerant processing of unsynchronized cloud target load rates: When the cloud platform collaborative management unit 200 fails to issue the peak and valley target load rates in a timely manner, the edge computing unit automatically switches to the preset economic operation mode, using a 60% load rate as the control benchmark by default. This load rate is derived from the engineering practice of economic operation of the power supply and distribution system. When most power equipment operates within the 60% load range, the comprehensive performance of energy efficiency and equipment life is optimal, ensuring that the system maintains efficient operation even when there are no global optimization instructions. (3) Fault-Tolerant Processing for Hardware Computation Failures If a hardware anomaly (such as register overflow or arithmetic unit error) occurs during the priority calculation process of the ARM Cortex-A72 processor, the independent watchdog module triggers a processor reset and simultaneously calls the version management module 160 to roll back to the previous valid policy version. The total time required for the reset and rollback process is controlled within a few hundred milliseconds, ensuring that the system can quickly restart the control logic after the fault is recovered, avoiding prolonged out-of-control.

[0035] In this embodiment, the edge computing and AI decision-making unit 100 further includes a policy verification module 140, a parameter optimization module 150, and a version management module 160, wherein: The strategy verification module 140 verifies the effectiveness of the control strategy through the Monte Carlo simulation algorithm, and the preset number of simulations is the industry standard value. Based on the equipment operating parameters collected by the intelligent sensing and Internet of Things unit 300, an assessment report containing the risk probability distribution is generated. The assessment report is used to correct the priority judgment threshold parameters built into the edge computing and AI decision-making unit 100; the intelligent sensing and Internet of Things unit 300 encapsulates parameters such as the transformer load rate and the circuit breaker temperature into a Modbus frame through the industrial bus sub-link sub-module 311 of the multi-protocol communication link module 310, and transmits registers 40011-40020 to the strategy verification module 140.

[0036] As a further explanation of this embodiment, the strategy verification module 140 in this embodiment verifies the effectiveness of the control strategy through a Monte Carlo simulation algorithm, sets 10,000 iterations of sampling (this design can cover a 99.7% confidence interval and meet the requirements of industrial-grade simulation accuracy), the voltage fluctuation range is set to ±15% of the rated value (according to the definition of extreme working conditions in "GB / T12325-2008 Power Quality Supply Voltage Deviation"), and the equipment failure probability follows a Poisson distribution (parameters =0.02 times / hour, based on the statistical data fitting of power equipment failures in the past five years), the convergence condition is that the deviation of the results of two adjacent iterations is ≤0.5% (balancing computing efficiency and accuracy, single-threaded simulation time ≤120 seconds); based on the equipment operating parameters transmitted by the intelligent sensing and Internet of Things unit 300 via the industrial bus sub-link sub-module 311, random disturbances are injected into the voltage and load, and with the help of the parallel computing capability of the Xilinx Artix-7 FPGA, a risk probability distribution assessment report containing a 95% confidence interval is generated; if the simulation finds that the fault trigger probability exceeds 5%, the priority judgment threshold is automatically lowered by 10% (for example, the equipment safety urgency threshold is revised from 0.3 to 0.27), and the correction result is fed back to the threshold register of the strategy generation module 120.

[0037] The parameter optimization module 150 dynamically adjusts the policy parameters based on the deep reinforcement learning framework, with minimizing grid losses as the optimization goal. The optimization cycle is dynamically issued by the cloud platform collaborative management unit 200 based on the load fluctuation characteristics, and directly outputs the dual-factor weight parameters used for priority calculation in the edge computing and AI decision-making unit 100; As a further explanation of this embodiment, the parameter optimization module 150 in this embodiment dynamically adjusts the strategy parameters based on the deep reinforcement learning framework, extracts the 3D state characteristics (current load rate , Equipment temperature rise , grid loss ), defining the 6-dimensional action space of transformer tap ±2 gear adjustment and capacitor 1-3 group switching, using the reward function ,in is the loss change rate, is the temperature rise change, To control the number of actions; trained with Q-Learning algorithm (learning rate , discount factor , completing 1000 iterations per cycle); the cloud platform collaborative management unit 200 dynamically issues an optimization cycle based on the load fluctuation standard deviation (set to 10 minutes when the load fluctuation standard deviation is greater than 5%, and extended to 30 minutes when it is less than 5%) through the 5G core sub-link sub-module 313, and directly outputs the dual-factor weight parameter ( 、 Correction value) to the parameter register of the strategy generation module 120 to achieve dynamic optimization of the strategy.

[0038] The version management module 160 uses blockchain hash storage technology to record historical versions of the strategy (including the priority determination strategy configuration of the edge computing and AI decision-making unit 100), and supports one-click rollback to the previous valid version in the event of an exception. The rollback time meets the equipment fault tolerance requirements of the modular prefabricated cabin unit 400, and the rollback trigger condition is linked to the security verification results of the edge computing and AI decision-making unit 100.

[0039] As a further explanation of this embodiment, the version management module 160 in this embodiment integrates an STM32L4R9 encryption chip, uses the SHA-256 algorithm to perform hash operations on the policy version (priority determination strategy, two-factor parameters), the block header encapsulates the version number, timestamp, and previous block hash, and the block body stores parameter binary data; when the edge computing and AI decision unit 100 fails to execute the policy for three consecutive times, or the security verification parameter exceeds the limit (such as ) or hardware calculation error (Watchdog trigger), the previous valid version of the blockchain is called, the historical strategy loading is completed within 100ms, and the intelligent power distribution execution module 510 is synchronized through the industrial bus.

[0040] The cloud platform collaborative management unit 200 builds a global data storage and analysis center to receive edge computing node data and provide a global perspective on power grid operation. It also issues collaborative control instructions to the edge computing and AI decision-making unit 100 through big data analysis and resource optimization algorithms to assist in optimizing local decision-making. In this embodiment, the global data storage and analysis center of the cloud platform collaborative management unit 200 adopts a distributed time series database cluster architecture to perform millisecond-level sampling and storage of device operation data (voltage, current, temperature, etc.) uploaded by the edge computing and AI decision-making unit 100 and achieve global visualization of the power grid operation status. The distributed time series database cluster architecture specifically includes: A digital twin is constructed based on a three-dimensional tunnel topology model and power flow algorithms, rendering a real-time heat map of the load distribution of the power supply and distribution network. As a further illustration of this embodiment, this embodiment can use Blender + GIS data to build a 3D tunnel model with an accuracy of 0.5 meters and integrate the following elements: Spatial location of power supply and distribution equipment (transformers, switchgear, cable trenches); Tunnel environmental parameters (humidity, temperature field distribution); In addition, a query interface for geometric coordinates and device parameters can be provided to the outside world through OpenAPI.

[0041] As a further explanation of this embodiment, the Newton-Raphson method can be used to solve the power flow equation. The specific parameters include: Iteration convergence threshold: voltage amplitude error < 0.001 pu, phase angle error < 0.005 rad; Node type: PV node (high-voltage side of main transformer), PQ node (load point), balancing node (grid-connected side of tunnel group); In addition, this embodiment can also use WebGL technology to render the load heat map on the browser side, and its color mapping relationship is: Load rate < 30%: green (RGB0,255,0); 30%-70%: yellow (RGB255,255,0); >70%: red (RGB255,0,0).

[0042] The long short-term memory network (LSTM) is used to predict the trend of historical data and generate a warning map of power grid operation risks in the next 24 hours.

[0043] As a further explanation of this embodiment, in the trend prediction model constructed by the long short-term memory network LSTM in this embodiment, the input layer contains 24 neurons (corresponding to 8 time points each for the voltage, current, and load rate in the historical 24 hours), the hidden layer uses 2 layers of 64-unit LSTM and 1 layer of 32-unit fully connected layer, and the output layer is 3 neurons (predicting the operating parameters for the next 1 hour, 3 hours, and 6 hours); the training process uses the Adam optimizer (learning rate 0.001, decay of 10% every 50 rounds), the loss function is a combination of mean square error and mean absolute error, the batch size is 64, and the iteration is 200 rounds (the validation set accounts for 20%). When the prediction error exceeds 15%, a secondary correction is triggered, and the correction value = initial prediction value + 0.3×(actual value - initial prediction value).

[0044] In this embodiment, the cloud platform collaborative management unit 200 uses big data analysis and resource optimization algorithms to generate collaborative control instructions, including the following steps: S200.1. Use consistent hashing to manage the tunnel power distribution network in different zones and monitor the load rate of each zone in real time. When the load rate of a zone exceeds a preset overload threshold (e.g., 80%) of the rated value, a load transfer strategy is triggered for the adjacent zone, and the transfer power threshold is calculated using the following formula: : ; in, is the regional rated capacity, It is a dynamic adjustment factor (obtained by training the inter-regional line impedance matrix), and the preset overload threshold is the industry's conventional load rate critical value (such as the overload standard defined in GB / T50598-2010).

[0045] As a further illustration of this embodiment, is a dynamic adjustment factor, and ,in, is the inter-area line impedance, is the average impedance of the entire network. The regional interconnection coefficient is determined by the topological connectivity of the distribution network. It is a parameter used to quantify the degree of electrical connection between different areas in the tunnel group power supply and distribution network.

[0046] The regional interconnectivity coefficient reflects the physical connection strength between two power supply and distribution areas (e.g., power supply areas in adjacent tunnel segments) based on network topology (e.g., line connection method, distance, impedance, etc.). It is used to adjust the load transfer power threshold to ensure that when one area is overloaded, the load transfer strategy for adjacent areas remains within electrical safety and feasibility.

[0047] The value of the regional interconnection coefficient is determined by the following topological characteristics: Line connection status: whether the areas are directly connected by cables / busbars, and the number of connecting lines (e.g., single circuit / double circuit); Line impedance: The impedance of the lines connecting regions (the smaller the impedance, the tighter the electrical connection and the greater the interconnection coefficient); Topological distance: The physical distance between regions in the power supply and distribution network (the closer the distance, the greater the interconnection coefficient).

[0048] In the above formula for calculating the load transfer power threshold, the role of the regional interconnection coefficient is: When inter-regional connections are dense (high interconnection coefficient), the actual transferable power threshold is lowered to prevent overloading of connecting lines due to excessive transfer power. When inter-regional connections are loose (low interconnection coefficient), the transfer power threshold can be appropriately increased to improve resource utilization efficiency. The regional interconnection coefficient is a "quantitative indicator of the closeness of connectivity" based on the topology of the power supply and distribution network, used to dynamically balance the feasibility and electrical safety of load transfer.

[0049] S200.2. Taking minimizing grid losses, maximizing new energy consumption, and maximizing power supply reliability as the objective functions, a non-dominated sorting genetic algorithm is used for iterative optimization to generate coordinated control instructions including transformer tap adjustment gears (adjustment step ≤ 1.25% of rated voltage), number of capacitor switching groups (single switching capacity difference ≤ 500kvar), and load transfer strategy execution timing (action interval between adjacent areas ≥ 10 seconds); the coordinated control instructions are approximate Pareto frontier solution sets for edge computing and AI decision-making unit 100 to select and execute based on local real-time status.

[0050] As a further explanation of this embodiment, the objective function in this embodiment specifically includes minimizing grid losses, maximizing new energy consumption, and maximizing power supply reliability, where: Minimize grid losses : Based on Kirchhoff's law and Ohm's law, a network loss calculation model is established: ; in, Indicates the Real-time current of the lines (uploaded by the edge computing and AI decision-making unit 100 with a 100ms sampling period, obtained through the MQTT protocol, and the data format complies with the IEC61850 standard); Indicates the The resistance value of the line (based on the cable model and length preset, such as YJV22-8.7 / 10kV-3×240mm 2 The resistance of the cable is 0.072Ω / km and its length is determined by the topology of the tunnel power distribution network); Maximizing the consumption of new energy : Quantified Photovoltaics and wind power Grid-connected output: ; in, Indicates the real-time output of the distributed photovoltaic power station in the tunnel group (monitored by the photovoltaic inverter, with a sampling period of 100ms and uploaded via the Modbus protocol); Indicates the real-time output of wind farms around the tunnel cluster (provided by the wind farm SCADA system and synchronized to the cloud platform via the 5G network); Maximize power supply reliability : Reliability indicators based on component failure rate and repair time: ; in, Indicates the Failure rate of each component (refer to DL / T837-2012 "Regulations for Reliability Evaluation of Power Transmission and Transformation Facilities", e.g., the failure rate of a 10kV dry-type transformer is 0.05 times / year, and the failure rate of a 10kV circuit breaker is 0.08 times / year); Indicates the The average repair time of each component (referring to DL / T837-2012, the repair time for circuit breaker faults is 4 hours, and the repair time for transformer faults is 8 hours).

[0051] Furthermore, the non-dominated sorting genetic algorithm in this embodiment performs iterative optimization on the objective function, specifically including the following process: First, an initial population of 100 was constructed, with each individual mapping the power supply and distribution control parameters using real-number encoding (e.g., transformer tap position is encoded as [-5, 5], corresponding to a ±2.5% rated voltage adjustment range with a step size of 1.25% rated voltage). To prevent the initial solution from being concentrated in a local area, Latin hypercube sampling was used to generate the population. Stratified random sampling was used to evenly cover the parameter space (e.g., combinations of tap positions and capacitor switching groups), ensuring the algorithm's global exploration capabilities for the multiple objectives of "grid losses, renewable energy consumption, and power supply reliability"; Then, non-dominated relationship discrimination and crowding distance calculation are performed, where: Domination relationship determination: traverse the individuals in the population and select non-dominated solutions that are "not inferior to other individuals in all objective functions and have at least one objective that is better" and mark them as Pareto level 1 (inferior solutions are marked with a higher level), achieving preliminary screening under multi-objective conflicts; Crowding distance calculation: For each objective function (such as grid loss and renewable energy consumption), the difference between adjacent individuals in the function value is calculated to quantify the "sparseness" of individuals in the target space. This method retains solutions in sparse areas, avoids excessive clustering of high-quality solutions, and ensures balanced coverage of multiple objectives by the solution set. Next, to simulate the recombination and mutation of biological genes, the algorithm drives population evolution through crossover and mutation operations: Simulate binary crossover (probability 0.8, distribution index 15): For the selected parent individuals, the "discrete degree" of gene exchange is controlled by a distribution index of 15 (the larger the index, the closer the new solution is to the parent), while retaining the high-quality characteristics of the parent. Explore the neighborhood parameter space; Polynomial mutation (probability 0.1, distribution index 20): A distribution index of 20 is used to constrain the mutation amplitude (the larger the index, the smoother the mutation), avoiding destroying the local structure of high-quality solutions. At the same time, it introduces global exploration capabilities to the algorithm to prevent it from falling into local optimality. Finally, set up the double termination rule: Maximum number of iterations (200): matches the computational efficiency requirements of engineering scenarios (a single iteration takes ≤ 3 seconds with GPU acceleration, and a total of 200 iterations takes ≤ 10 minutes, meeting the hourly optimization cycle of the cloud platform); Convergence determination (the change rate of the optimal solution for 20 consecutive iterations is less than 5%): When the change rate of the optimal solution on the Pareto front in the objective function value is less than 5% in 20 consecutive iterations, the solution set is determined to have converged and the iteration is terminated early to avoid invalid calculations; Through the above process, the non-dominated sorting genetic algorithm continuously generates a set of control strategies covering multiple objective demands in the dynamic balance of "global exploration-local optimization-convergence judgment", providing sufficient optimization solution support for the decision output of the cloud platform collaborative management unit 200.

[0052] It should be added that after the algorithm iteration is completed, the top 20 optimal solutions are selected by crowding distance sorting to form an approximate Pareto frontier solution set. Each solution contains: Transformer tap adjustment parameters: such as adjustment gear +2 (corresponding to a voltage increase of 2.5% of the rated value); Capacitor switching parameters: if 3 groups are put into operation (compensation capacity 300kvar); Load transfer strategy: If area 1 transfers 2 MVA to area 2 at t=0s, area 2 will receive the load at t=10s.

[0053] In summary, the edge computing and AI decision-making unit 100 selects an execution strategy from the solution set based on the local real-time status (such as equipment temperature rise and fault warning level): when the local safety margin coefficient is less than 0.6, the strategy with the least impact on the equipment load is prioritized; when the system is in normal operating conditions, the strategy with the greatest reduction in grid losses is selected to achieve coordinated regulation of global optimization and local safety.

[0054] The intelligent sensing and IoT unit 300 is used to monitor the operating parameters of the tunnel power supply and distribution equipment, the cabin environment, and the fire protection status in real time. It uses the multi-protocol communication link module 310 to transmit the data to the edge computing and AI decision-making unit 100 and the cloud platform collaborative management unit 200, thereby building a full-scene monitoring system through the full-scene data fusion analysis module 320; In this embodiment, the multi-protocol communication link module 310 includes an industrial bus sub-link sub-module 311, a wireless sensor sub-link sub-module 312, and a 5G core sub-link sub-module 313, wherein: The industrial bus sub-link sub-module 311 uses the Modbus protocol to build a dedicated communication channel for power equipment. It collects the voltage, current, and temperature parameters of transformers and switchgear in the tunnel in millisecond cycles, meeting the strong real-time monitoring requirements of power equipment. As a further illustration of this embodiment, the industrial bus sub-link sub-module 311 in this embodiment uses a Siemens S7-1200 PLC as the main controller, and connects to the transformers and switch cabinets in the tunnel via the Profibus DP bus; it is also equipped with a MOXAMGate MB3170 protocol converter to implement Modbus RTU and TCP protocol conversion, ensuring a 10ms response delay.

[0055] Wireless sensor sub-link submodule 312: Builds an environmental monitoring network based on the LoRaWAN protocol to collect temperature, humidity, and smoke concentration parameters. It optimizes node energy consumption through a sleep-wake low-power communication mechanism and adapts to tunnel wiring-restricted scenarios. As a further illustration of this embodiment, the wireless sensor sub-link submodule 312 in this embodiment is integrated with a LoRaWAN wireless mesh network, the gateway adopts DraginoLG01-N, the terminal node is based on the STM32L071 microcontroller, and integrates an SHT31-D temperature and humidity sensor and an MQ-2 smoke sensor.

[0056] Furthermore, the terminal node in this embodiment adopts a sleep-wake-up scheduling mechanism (wake-up cycle of 5 minutes), and only collects temperature and humidity (5-minute sampling cycle) and smoke concentration (2-minute sampling cycle) during the wake-up period; the gateway performs data aggregation every 30 minutes and transmits it to the RAK2245 receiving module of the edge computing unit via the LoRa link, reducing network energy consumption to the μW level.

[0057] The 5G core sub-link submodule 313 uses 5G URLLC slice channels to perform real-time priority transmission of fire alarm and gas concentration data, and the transmission delay meets the fire emergency response standard requirements; As a further explanation of this embodiment, this embodiment can establish a URLLC slice channel based on Huawei 5GCPEPro2, and the edge computing gateway can use IntelNUC11 to locally cache fire alarm and gas concentration data.

[0058] Furthermore, the transmission priority rules for fire alarm and gas concentration data in this embodiment are as follows: Fire alarm data: Marked with QoS level 5 (highest priority), transmitted directly to the cloud platform collaborative management unit 200 via the UDP protocol, ensuring an end-to-end latency of ≤50ms from sensor trigger to cloud reception (in compliance with the emergency response requirements of GB50116-2013, "Design Specifications for Automatic Fire Alarm Systems"). Gas concentration data: Marked as QoS level 4, transmitted via TCP protocol to ensure data integrity.

[0059] The full-scenario data fusion analysis module 320 includes a multimodal data fusion submodule 321, an equipment health assessment submodule 322, and an adaptive sampling control submodule 323, wherein: The multimodal data fusion submodule 321 is used to align the electrical parameters (current, voltage), environmental parameters (temperature and humidity), and fire protection parameters (smoke concentration) in time and space, and uses the Kalman filter algorithm to eliminate data noise. The fusion accuracy meets the power equipment status monitoring standards. As a further illustration of this embodiment, this embodiment uses the NTP network time protocol to synchronize the timestamps of each sensor during spatiotemporal alignment. Data with different sampling rates (e.g., 10ms power parameters and 50ms oil temperature data) are fitted into a unified 10ms time series using a cubic spline interpolation algorithm, achieving spatiotemporal alignment of electrical, environmental, and fire protection parameters. At the same time, the calculation formula of the Kalman filter algorithm in this embodiment is: Status prediction: ;in, is the state transition matrix, and , is the process noise, covariance ; for Prior state estimate at time t; for The posterior state estimate at time t; Observation Update: ;in, The observation matrix, , is the Kalman gain, is the observed value, and the measurement noise covariance R=0.1; for System status at the moment; During specific execution, the current state is first predicted through the state equation, and then the optimal estimate is obtained by combining the observation value correction, suppressing the current data noise from ±5% to ±1%, and the oil temperature data noise from ±0.5℃ to ±0.1℃. The fused data is transmitted to the edge computing and AI decision-making unit 100 via the SPI bus.

[0060] The equipment health assessment submodule 322 builds a fault diagnosis model based on the DS evidence theory, extracts the circuit breaker opening and closing coil current waveform and transformer oil temperature change rate as core features, and realizes early warning of equipment failures; As a further illustration of this embodiment, the equipment health assessment submodule 322 builds a multi-source feature fusion fault diagnosis model based on the DS evidence theory. It interacts with the multimodal data fusion submodule 321 in real time with the spatiotemporally aligned electrical and temperature data to implement early fault warnings for circuit breakers and transformers. The specific process is as follows: First, establish data interaction with the multimodal data fusion submodule 321: receive the time-space aligned monitoring data through the SPI bus, including the circuit breaker opening and closing coil current waveform (transmitted through the industrial bus sub-link submodule 311) and the transformer oil temperature time series data; At the same time, feature parameters are extracted in real time: For circuit breakers, analyze the peak value, rise time, and fall time of the coil current waveform and calculate the waveform distortion rate (distortion rate = (actual waveform - mean square error of standard waveform) / standard waveform energy); For transformers, the oil temperature change rate is calculated with a 30-minute sliding window. The calculation formula is: ; in, Indicates the rate of change of oil temperature; Indicates time Oil temperature; Indicates time The oil temperature 30 minutes ago; represents; and is associated with the load rate during the same period (obtained from the multimodal data fusion submodule 321).

[0061] Then, the characteristic parameters are mapped to the basic probability of failure: Circuit breaker side: When the waveform distortion rate is greater than 15%, the relationship between the distortion degree and the fault probability is fitted by the Gaussian mixture model, and the basic fault probability is output. (Fault) (The higher the distortion rate, Linear increase, under normal conditions approaches 0); Transformer side: Based on the oil temperature change rate and load rate, the basic probability of failure is mapped through the Sigmoid function (Fault) (The higher the load rate, the faster the temperature rises. Exponential increase, under normal conditions approaches 0); Then, the Dempster combination rule is used to fuse multi-source evidence and calculate the comprehensive probability of the target state (normal, warning, fault): ; Molecule: traverse all state sets 、 , and their intersection is The product of evidence (e.g. =Fault, =fault), the intersection is a fault and is counted calculate; Denominator: Conflict coefficient (quantifies the degree of contradiction between two pieces of evidence). If the conflict coefficient is greater than 0.5, the Yager rule softening calculation is automatically switched (reducing the weight of the contradictory evidence) to avoid decision-making distortion caused by extreme conflict. Finally, the failure decision is executed: When the fusion When , a high level is output through the GPIO port to trigger an interrupt response of the edge computing and AI decision-making unit 100; Synchronously mark the fault type (coil abnormality / oil temperature mutation), through I 2 The C bus transmits the diagnosis results back to the multimodal data fusion submodule 321 for dynamic correction of the subsequent data denoising model; like , only triggers early warning log records for operation and maintenance personnel to conduct retrospective analysis.

[0062] After receiving the abnormal working condition signal sent by the edge computing and AI decision-making unit 100, the adaptive sampling control submodule 323 automatically adjusts the sampling frequency of key equipment and constructs a three-level sampling mechanism of normal periodic collection-abnormal trigger collection-high-frequency tracking collection to meet the dynamic monitoring accuracy requirements under different working conditions.

[0063] As a further explanation of this embodiment, the three-stage sampling parameters and trigger conditions in this embodiment are as follows: Normal acquisition (default mode): Power parameters (transformer voltage / current, switchgear bus current): 100ms sampling period (configured through Modbus register 40001, transmitted by industrial bus sub-link submodule 311); Environmental parameters (temperature, humidity, smoke concentration): 5-minute sampling period (configured through register 0x0010 of the LoRa terminal node, transmitted by the wireless sensor sub-link submodule 312); Trigger condition: The equipment health assessment submodule 322 outputs a normal signal ( ).

[0064] Abnormal trigger collection (early warning response): Power parameters: sampling period compressed to 50ms (notified to the S7-1200 main controller via UART instructions (baud rate 9600)); Oil temperature / contact temperature: sampling period compressed to 100ms (via I 2 C bus to rewrite the sensor configuration register); Trigger condition: The equipment health assessment submodule 322 outputs an early warning signal ( ), or the edge computing and AI decision-making unit 100 sends an abnormal flag with a load rate greater than 80%.

[0065] High-frequency tracking and acquisition (fault response): Full device parameters: 10ms sampling of power parameters, 100ms sampling of environmental parameters (receiving temporary configuration instructions from the cloud platform through the 5G core sub-link submodule 313); Continuous logic: After triggering, it lasts for 1 minute and automatically falls back to normal collection through a hardware timer; Trigger condition: The equipment health assessment submodule 322 outputs a fault signal ( ), or the fire alarm submodule detects that the smoke concentration is greater than 500ppm.

[0066] The modular prefabricated cabin unit 400 integrates power supply and distribution equipment, fire protection systems, and edge computing components. It adopts a split steel structure design of high-pressure and low-pressure cabins, combined with fireproof and heat-insulating technology, to achieve physical isolation, rapid installation, and environmental protection. In this embodiment, the modular prefabricated cabin unit 400 includes a hyperbaric cabin module 410, a hypobaric cabin module 420, and a fire protection integrated module 430, wherein: The high-voltage compartment module 410 includes a switch cabinet submodule 411 and a transformer submodule 412, wherein: Switchgear Module 411 houses a 10kV metal-clad removable switchgear with a Q355B steel frame. Arc protection sensors are installed within the switchgear. When arc energy reaches a preset threshold, the circuit breaker trips within 10ms and transmits a fault signal via the Modbus protocol. Switchgear Module 411 is horizontally positioned on the left side of the high-voltage compartment (800mm from the front panel). Its bottom is welded to the compartment floor using L-shaped angle steel (75mm × 75mm × 8mm). A 5mm rubber shock-absorbing pad is installed between the angle steel and the cabinet. The cabinet top is secured to the cabin ceiling with M10 bolts to ensure mechanical stability during operation.

[0067] Transformer submodule 412 utilizes a dry-type transformer with an enclosure rated at or above IP54. The cooling fan is linked to a thermostat for control. Internally, the thermostat and cooling fan automatically activate when the winding temperature reaches a preset threshold, and the temperature status is uploaded via a communication interface. Transformer submodule 412 is vertically positioned to the right of the high-voltage compartment (1.5 meters from the switchgear). Its base is welded from 10# channel steel (1.2 meters long), and the transformer is secured to the base with M12 bolts. A 300mm cooling channel is provided between the enclosure and the compartment side panels, and louvers are installed on the top to ensure efficient heat dissipation.

[0068] The low-pressure cabin module 420 includes an intelligent distribution submodule 421 and an edge computing server submodule 422, wherein: The intelligent distribution submodule 421 integrates a molded case circuit breaker and an intelligent power meter, supporting Modbus protocol communication. It enables power parameter acquisition and circuit breaker status monitoring, responding to opening and closing control commands from the edge computing unit. Located at the front of the low-voltage compartment (500 mm from the front panel), the intelligent distribution submodule 421 is supported by a 200 mm × 100 mm × 5 mm rectangular steel pipe welded bracket. The terminal box is secured to the bracket with M8 bolts at the four corners, and the bracket is connected to the bottom of the compartment with M10 expansion bolts (spaced 300 mm apart) to ensure a secure installation.

[0069] The edge computing server submodule 422 uses an industrial-grade server and is installed with a shock-absorbing bracket to adapt to a wide temperature operating environment. The edge computing server submodule 422 is vertically installed on the 4U standard guide rail at the rear of the low-pressure cabin (1.2m from the bottom plate) and is adapted to industrial standard installation through the guide rail slider. A shock-absorbing bracket is installed at the bottom, and the bracket is fixed to the guide rail with M6 bolts. A pressure plate and cabin roof limiter are set on the top of the server to adapt to the vibration environment.

[0070] The fire protection integrated module 430 includes a fire extinguishing device submodule 431 and a fire detection submodule 432, wherein: Fire extinguishing submodule 431 integrates a heptafluoropropane (HFC-227ea) gas fire extinguishing system. The nozzles are arranged at preset intervals and feature a linkage trigger mechanism. Upon receiving a fire detection signal, they delay the release of the agent and trigger an audible and visual alarm, providing status feedback via dry contacts. The nozzles are evenly spaced 2.5 meters apart on the cabin roof (50 mm from the cabin ceiling) and are hoisted using DN25 galvanized steel pipes. The pipe supports are spaced 1.5 meters apart and secured with expansion bolts. The steel pipes are threaded and sealed to the nozzles to ensure uniform coverage.

[0071] The fire detection submodule 432 uses a combined smoke and heat detector with a sensitivity that meets relevant national standards. It collects smoke concentration and temperature data in real time. When the alarm threshold is reached, an audible and visual alarm is triggered, activating the fire extinguishing system. The alarm signal is then uploaded to the cloud platform collaborative management unit 200 via the 5G core sub-link submodule 313, triggering a remote early warning. The smoke and heat detector is ceiling-mounted in the center of the cabin ceiling (100 mm from the cabin roof), with a spacing of ≤ 10 m. The base is secured with M6 expansion bolts, and the detector body is connected to the base with a clip. The wiring port is protected by a waterproof connector.

[0072] As a further illustration of this embodiment, three sets of Φ100mm epoxy resin wall bushings are provided at the bottom of the high-pressure and low-pressure compartment bulkheads (200mm from the bottom plate) in this embodiment. The bushings are filled with fireproof putty and sealed with sealant after the cables pass through, meeting fire prevention and protection requirements. In addition, heat dissipation holes (600mm×400mm, 1.5m from the ground) are provided on the rear side panels of the transformer-corresponding compartments. A metal protective net and an axial flow fan are installed. The fan is welded and fixed with a 50mm×50mm×5mm angle steel bracket to ensure reliable operation of the heat dissipation system.

[0073] The automated adaptive control unit 500 is used to execute the final control instructions of the edge computing and AI decision-making unit 100, and realize real-time monitoring of power flow and emergency response to sudden incidents through power supply and distribution circuit switching, micro-environment control, and multi-system linkage logic of the decision and control module 530.

[0074] In this embodiment, the automated adaptive control unit 500 includes an intelligent power distribution execution module 510 and a micro-environment control module 520, wherein: The intelligent power distribution execution module 510 uses solid-state circuit breakers to build an intelligent power distribution network. The intelligent power distribution execution module 510 includes a fault isolation submodule 511 and a dynamic load distribution submodule 512, wherein: The fault isolation submodule 511 is based on a high-speed current sensor and a solid-state circuit breaker to achieve rapid removal of short-circuit faults; As a further illustration of this embodiment, the high-speed current sensor in this embodiment collects three-phase current in real time and calculates the waveform distortion rate through the FPGA. When the current exceeds 5 times the rated value or the distortion rate exceeds the threshold, the FPGA sends a disconnect instruction to the circuit breaker and uploads the fault type to the edge computing and AI decision-making unit 100 through the Modbus / TCP protocol. At the same time, the fault signal triggers the circuit breaker to disconnect through hard wiring, and the action status is fed back to the intelligent distribution terminal through the RS485 interface to ensure that the fault is isolated within 2ms.

[0075] The dynamic load distribution submodule 512 dynamically adjusts the transformer tap position through a fuzzy PID controller based on the load prediction results of the edge computing and AI decision-making unit 100; As a further explanation of this embodiment, the specific process of dynamically adjusting the transformer tap position by using the fuzzy PID controller is as follows: First, the load forecast data (update period is 1 minute) output by the edge computing and AI decision-making unit 100 is collected as a decision basis for load control; Then, based on the collected load data, the following two core inputs are calculated: Load forecast deviation: the difference between the actual load and the forecast load (characterizing “how big the forecast error is”); Deviation change rate: The speed at which the load forecast deviation changes (depicting whether the error is getting bigger or smaller); Then, based on preset fuzzy rules (for example, if the deviation is "positive" (actual load is much higher than the predicted value) and the rate of change is "positive" (the deviation growth slows down), the "+2 gear" adjustment instruction is output), the number of tap adjustment steps is derived (the minimum adjustment unit is ±1 gear, which directly determines the tap gear change). At the same time, the basic PID parameters (proportional coefficient , integral coefficient , differential coefficient It is not fixed, but is modified in real time through fuzzy reasoning, allowing the controller to adapt to the dynamic changes of the load; Finally, the final adjustment amount is calculated every 10 seconds to drive the tap motor: the motor step angle is 1.8 to ensure adjustment accuracy; the single adjustment step size is ≤ 1.25% of the rated voltage, achieving "low-impact, fine-grained" voltage optimization.

[0076] The microenvironment control module 520 implements adaptive adjustment of the tunnel microenvironment through the IoT edge controller, including a multi-parameter collaborative control submodule 521 and a zone temperature balancing submodule 522, wherein: The multi-parameter coordinated control submodule 521 dynamically adjusts the ventilation fan speed based on CO concentration, visibility and traffic flow using a model predictive control algorithm.

[0077] The zone temperature balancing submodule 522 uses a PID controller to adjust the tunnel lighting system power, achieving cabin temperature differential control. This design uses tunnel zone temperature differentials as the control target. Distributed temperature sensors collect real-time data. The PID controller dynamically adjusts the LED lighting power (with an adjustment range of 20%-100%) based on the temperature differential deviation and its rate of change. This utilizes the thermal effects of the lighting system to balance cabin temperature, forming a closed-loop control system of "temperature acquisition-deviation calculation-power adjustment" to ensure that zone temperature differentials remain within a reasonable range.

[0078] In this embodiment, the decision and control module 530 includes a collaborative control submodule 531 and an emergency response decision tree submodule 532, wherein: The collaborative control submodule 531 builds an inter-system communication mechanism based on the event bus architecture, specifically including: Assign dynamic priorities to fire protection, ventilation, lighting and other systems (e.g., fire protection system priority is the highest level 1) to ensure that the response time of key systems meets the real-time requirements of fire emergency response; The Raft consensus algorithm is used to synchronize states between control units, achieving synchronization accuracy that meets industrial-grade distributed system collaborative control standards. The Raft consensus algorithm divides control nodes into leader, follower, and candidate states, ensuring the consistency and reliability of multi-node control commands through log replication and an election mechanism. In this embodiment, command synchronization for systems such as fire protection, ventilation, and lighting is combined with event bus priority management (the fire protection system has the highest priority), ensuring that critical systems meet emergency response requirements and enabling multi-system collaborative control.

[0079] The emergency response decision tree submodule 532 builds an emergency response model based on the Bayesian network inference engine, specifically including: Predefine a set of typical failure modes for the entire life cycle of the tunnel power supply and distribution system (covering basic faults such as single-phase grounding and three-phase short circuit, as well as derived compound faults); Bayesian reasoning is used to update the fault probability distribution in real time and generate the optimal emergency response strategy. The strategy generation timeliness meets the rapid response requirements of power system fault handling.

[0080] As a further explanation of this embodiment, this embodiment uses a Bayesian network inference engine to construct an emergency response model for sudden events: the model is based on typical failure modes of the tunnel power supply and distribution system throughout its life cycle (covering single-phase grounding, three-phase short circuit, etc.), describes the correlation between fault characteristics and fault types through a preset conditional probability table, collects protection device signals and electrical characteristic quantities in real time, uses Bayesian reasoning to update the fault probability distribution, and quickly generates the optimal emergency response strategy (such as the circuit breaker disconnection sequence). The strategy generation timeliness meets the power system fault handling requirements to ensure the safe operation of the system.

[0081] It should be noted that the intelligent power distribution execution module 510 in this embodiment uses Modbus / TCP and RS485 interfaces, the microenvironment control module 520 supports CANopen and DALI protocols, and the decision and control module 530 implements event-driven communication based on message queues. The hardware uses an industrial-grade controller and fieldbus architecture to ensure the real-time and reliable control logic. The overall design complies with relevant standards for power systems and industrial automation, supporting rapid on-site deployment and commissioning.

[0082] Those skilled in the art will appreciate that the process of implementing all or part of the steps of the above embodiments may be accomplished by hardware, or by instructing related hardware through a program.

[0083] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing is characterized by: include: The edge computing and AI decision-making unit (100) is used to realize the rapid collection, processing and instant decision-making of multi-dimensional data of tunnel power supply and distribution, generate a power supply and distribution adaptive control strategy by deploying computing resources and machine learning algorithms at edge nodes close to data sources, and convert the strategy into executable control instructions; based on the collaborative control instructions issued by the cloud platform collaborative management unit (200), generate the final control instructions through a priority determination mechanism; The edge computing and AI decision-making unit (100) generates the final control instruction through the priority determination mechanism, including the following steps: S100.

1. Parsing local real-time instructions and collaborative control instructions received by the edge computing and AI decision-making unit (100), extracting key features for priority calculation, wherein the key features include instruction type, action object, parameter value range, and timeliness attribute; S100.

2. Calculate the local command priority value using a two-factor weight model based on device status and command attributes. Coordinated control instruction priority value , the calculation formula is: in, represents the urgency factor; Indicates the current device temperature rise; Indicates the average temperature rise under the same historical load, used to reflect the degree of equipment abnormality; ； in, represents the optimization factor; Indicates the difference between the real-time load rate and the rated load rate; Indicates the rated load rate, reflecting the urgency of system regulation; S100.

3. Based on local instruction priority value Coordinated control instruction priority value The calculation results are used to execute hierarchical decision logic; The cloud platform collaborative management unit (200) builds a global data storage and analysis center for receiving edge computing node data and providing a global perspective of power grid operation, and issues collaborative control instructions to the edge computing and AI decision-making unit (100) through big data analysis and resource optimization algorithms to assist in optimizing local decision-making; An intelligent sensing and IoT unit (300) is used to monitor the operating parameters of the tunnel power supply and distribution equipment, the cabin environment, and the fire protection status in real time, and transmit the data to the edge computing and AI decision-making unit (100) and the cloud platform collaborative management unit (200) using a multi-protocol communication link module (310), so as to build a full-scene monitoring system through a full-scene data fusion analysis module (320); The modular prefabricated cabin unit (400) integrates power supply and distribution equipment, fire protection system and edge computing components, adopts a split steel structure design of high-pressure cabin and low-pressure cabin, and combines fireproof and heat insulation technology to achieve physical isolation, rapid installation and environmental protection; The modular prefabricated cabin unit (400) comprises a high-pressure cabin module (410), a low-pressure cabin module (420) and a fire protection integrated module (430), wherein: The high-voltage cabin module (410) includes a switch cabinet submodule (411) and a transformer submodule (412); The low-pressure cabin module (420) includes an intelligent distribution submodule (421) and an edge computing server submodule (422); The fire protection integrated module (430) includes a fire extinguishing device submodule (431) and a fire detection submodule (432); The automated adaptive control unit (500) is used to execute the final control instructions of the edge computing and AI decision-making unit (100), and realizes real-time monitoring of power flow and emergency response to sudden events through multi-system linkage logic of power supply and distribution circuit switching, micro-environment control, and decision-making and control module (530).

2. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 1 is characterized in that: The edge computing and AI decision-making unit (100) includes a data processing module (110), a strategy generation module (120) and an instruction conversion module (130), wherein: The data processing module (110) is used to clean and normalize the original data transmitted by the intelligent sensing and Internet of Things unit (300), and to remove noise using a sliding window filtering algorithm, with the window width using a preset value; The strategy generation module (120) generates a power supply and distribution adaptive control strategy based on a multi-layer neural network, wherein the neural network input is historical load data and the output is a power prediction value for a future period; The instruction conversion module (130) is used to convert the control strategy into an executable control instruction, using an industrial control standard protocol, and the instruction generation delay meets the real-time control requirements.

3. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 1 is characterized in that: In step S100.3, the hierarchical decision logic specifically includes: First priority strategy: when When the instruction type is emergency, the edge computing and AI decision-making unit (100) generates the highest priority control instruction, and executes the equipment protection action through the automated adaptive control unit (500). The threshold The threshold for equipment safety urgency is used to distinguish between normal operation and emergency failure scenarios. Determination of equipment safety margin based on modular prefabricated cabin units (400); Parameter fusion strategy: when Or if the instruction type is optimization type, the control parameters of local real-time instructions and collaborative control instructions are weighted and integrated: ； in, represents the final control parameters sent to the automated adaptive control unit (500) after fusion, and Meeting electrical safety boundaries and mechanical life constraints of the modular prefabricated cabin unit (400); represents the local control parameters generated by the edge computing and AI decision-making unit (100); represents the cloud-side control parameters generated by the cloud platform collaborative management unit (200); Second priority strategy: when When the edge computing and AI decision-making unit (100) executes the cloud collaborative instructions first, the system operating parameters are adjusted by the automatic adaptive control unit (500), and the threshold It is a global optimization priority threshold used to balance local security and cloud energy efficiency goals, and the threshold A global energy efficiency model is determined based on the cloud platform collaborative management unit (200).

4. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 2 is characterized in that: The edge computing and AI decision-making unit (100) further includes a strategy verification module (140), a parameter optimization module (150) and a version management module (160), wherein: The strategy verification module (140) verifies the effectiveness of the control strategy through a Monte Carlo simulation algorithm, with the number of simulations preset to a conventional value in the industry, and generates an assessment report including a risk probability distribution based on the equipment operating parameters collected by the intelligent sensing and IoT unit (300), wherein the assessment report is used to correct the priority determination threshold parameters built into the edge computing and AI decision-making unit (100); The parameter optimization module (150) dynamically adjusts the strategy parameters based on the deep reinforcement learning framework, with minimization of power grid loss as the optimization goal. The optimization cycle is dynamically issued by the cloud platform collaborative management unit (200) according to the load fluctuation characteristics, and directly outputs the dual-factor weight parameters used for priority calculation in the edge computing and AI decision-making unit (100); The version management module (160) uses blockchain hash storage technology to record historical versions of the strategy, supports one-click rollback to the previous valid version in the event of an exception, and the rollback time meets the equipment fault tolerance requirements of the modular prefabricated cabin unit (400). The rollback trigger condition is linked to the security verification result of the edge computing and AI decision-making unit (100).

5. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 1 is characterized in that: The global data storage and analysis center of the cloud platform collaborative management unit (200) adopts a distributed time series database cluster architecture to perform millisecond-level sampling and storage of the device operation data uploaded by the edge computing and AI decision-making unit (100) and realize global visualization of the power grid operation status. The distributed time series database cluster architecture specifically includes: A digital twin is constructed based on a three-dimensional tunnel topology model and power flow algorithms, rendering a real-time heat map of the load distribution of the power supply and distribution network. The long short-term memory network (LSTM) is used to predict the trend of historical data and generate a warning map of power grid operation risks in the next 24 hours.

6. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 5 is characterized in that: The cloud platform collaborative management unit (200) uses big data analysis and resource optimization algorithms to generate collaborative control instructions, including the following steps: S200.

1. Use consistent hashing to manage the tunnel power distribution network in different zones and monitor the load rate of each zone in real time. When the load rate of a zone exceeds the preset overload threshold of the rated value, trigger the load transfer strategy for the adjacent zone and calculate the transfer power threshold using the following formula: : ； in, is the regional rated capacity, is a dynamic adjustment factor. The preset overload threshold is the industry's conventional load rate critical value. The regional interconnection coefficient is determined by the topological connectivity of the distribution network and is a parameter used to quantify the degree of electrical connection between different areas in the tunnel group power supply and distribution network. S200.

2. With minimization of grid losses, maximization of new energy consumption, and maximization of power supply reliability as objective functions, a non-dominated sorting genetic algorithm is used for iterative optimization to generate collaborative control instructions including transformer tap adjustment gears, number of capacitor switching groups, and execution sequence of load transfer strategy; the collaborative control instructions are approximate Pareto frontier solution sets for edge computing and AI decision-making units (100) to select and execute based on local real-time status.

7. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 1 is characterized in that: The multi-protocol communication link module (310) includes an industrial bus sub-link sub-module (311), a wireless sensor sub-link sub-module (312) and a 5G core sub-link sub-module (313), wherein: The industrial bus sub-link sub-module (311) uses the Modbus protocol to build a dedicated communication channel for power equipment, performs millisecond-level periodic acquisition of voltage, current, and temperature parameters of transformers and switch cabinets in the tunnel, and adapts to the strong real-time monitoring requirements of power equipment; The wireless sensor sub-link sub-module (312) builds an environmental monitoring network based on the LoRaWAN protocol, collects temperature, humidity and smoke concentration parameters, optimizes node energy consumption through a sleep-wake low-power communication mechanism, and adapts to tunnel wiring-restricted scenarios; The 5G core sub-link sub-module (313) uses a 5G URLLC slice channel to perform real-time priority transmission of fire alarm and gas concentration data; The full-scenario data fusion analysis module (320) includes a multimodal data fusion submodule (321), an equipment health assessment submodule (322), and an adaptive sampling control submodule (323), wherein: The multimodal data fusion submodule (321) is used to align the electrical parameters, environmental parameters, and fire protection parameters in time and space dimensions, and use a Kalman filter algorithm to eliminate data noise; The equipment health assessment submodule (322) constructs a fault diagnosis model based on DS evidence theory, extracts the circuit breaker opening and closing coil current waveform and transformer oil temperature change rate as core feature quantities, and realizes early warning of equipment failure; After receiving the abnormal working condition signal sent by the edge computing and AI decision-making unit (100), the adaptive sampling control submodule (323) automatically adjusts the sampling frequency of key equipment to construct a three-level sampling mechanism of normal periodic acquisition-abnormal trigger acquisition-high-frequency tracking acquisition.

8. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 1 is characterized in that: In the hyperbaric chamber module (410): The switch cabinet submodule (411) has a built-in 10kV metal-clad removable switch cabinet, which adopts a Q355B steel plate frame and an arc protection sensor is provided in the cabinet; The transformer submodule (412) adopts a dry-type transformer, the housing adopts a structure with a protection level not lower than IP54, and the heat dissipation fan is linked with the temperature controller for control; In the low-pressure cabin module (420): The intelligent distribution submodule (421) integrates a molded case circuit breaker and an intelligent power meter, and supports Modbus protocol communication; The edge computing server submodule (422) adopts an industrial-grade server, is installed using a shock-absorbing bracket, and is adaptable to a wide temperature operating environment; In the fire protection integrated module (430): The fire extinguishing device submodule (431) integrates a heptafluoropropane gas fire extinguishing system, the nozzles are arranged at a preset interval, and are equipped with a linkage trigger mechanism; The fire detection submodule (432) adopts a smoke and temperature composite detector.

9. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 2 is characterized by: The automated adaptive control unit (500) comprises an intelligent power distribution execution module (510) and a micro-environment control module (520), wherein: The intelligent power distribution execution module (510) uses a solid-state circuit breaker to construct an intelligent power distribution network. The intelligent power distribution execution module (510) includes a fault isolation submodule (511) and a dynamic load distribution submodule (512), wherein: The fault isolation submodule (511) is based on a high-speed current sensor and a solid-state circuit breaker to achieve rapid removal of short-circuit faults; The dynamic load distribution submodule (512) dynamically adjusts the transformer tap position through a fuzzy PID controller according to the load prediction result of the edge computing and AI decision-making unit (100); The microenvironment control module (520) realizes adaptive adjustment of the tunnel microenvironment through the Internet of Things edge controller, and includes a multi-parameter collaborative control submodule (521) and a partition temperature balancing submodule (522), wherein: The multi-parameter coordinated control submodule (521) dynamically adjusts the ventilation fan speed using a model predictive control algorithm based on CO concentration, visibility and traffic flow; The partition temperature balancing submodule (522) adjusts the tunnel lighting system power through a PID controller to achieve cabin temperature difference control.

10. The tunnel modular prefabricated cabin power supply and distribution intelligent substation adaptive control system based on edge computing according to claim 1 is characterized by: The decision and control module (530) includes a collaborative control submodule (531) and an emergency response decision tree submodule (532), wherein: The collaborative control submodule (531) constructs an inter-system communication mechanism based on the event bus architecture, specifically including: Assign dynamic priorities to fire protection, ventilation, lighting and other systems to ensure that the response time of key systems meets the real-time requirements of fire emergency response; The Raft consensus algorithm is used to achieve state synchronization between control units, and the synchronization accuracy meets the collaborative control standards of industrial-grade distributed systems; The emergency response decision tree submodule (532) constructs an emergency response model for emergencies based on a Bayesian network inference engine, specifically including: Predefine a set of typical failure modes for the entire life cycle of the tunnel power supply and distribution system; Bayesian reasoning is used to update the fault probability distribution in real time and generate the optimal emergency response strategy. The strategy generation timeliness meets the rapid response requirements of power system fault handling.

Images