A collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires based on MPC

Abstract

This invention discloses a collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires based on MPC. It achieves synchronous iteration of simulation, control, and physical hardware by constructing a full-size FDS numerical model and a Python-FDS-hardware-in-the-loop three-in-one millisecond-level bidirectional coupling interface. A spatiotemporal topology of tunnel fire smoke evolution is constructed to quickly and accurately predict the spatiotemporal distribution of smoke throughout the tunnel. A two-layer deep MPC architecture combining global collaborative optimization and zoned rolling control is designed, integrating multi-objective optimization functions for evacuation safety, smoke risk, energy consumption, and adjustment smoothness, and embedding hard constraints such as equipment rating, differential pressure, and safety thresholds. Adaptive rolling time-domain optimization and adaptive reconfiguration mechanism for fan failures are introduced. Finally, distributed collaborative closed-loop intelligent control of multiple smoke exhaust fans is achieved through the coupling interface, outputting control commands and fire situation data, significantly improving response speed.

Description

Technical Field

[0001] This invention relates to the field of tunnel fire protection engineering technology, and in particular to a method for coordinated intelligent control of multiple smoke exhaust systems in tunnel fires based on MPC. Background Technology

[0002] The long, narrow, and enclosed spatial structure of long tunnels, coupled with their segmented operation with multiple smoke control zones, means that after a fire breaks out, high-temperature, toxic smoke can rapidly spread longitudinally along the tunnel and cross smoke control zones. When the ventilation volume of multiple units becomes unbalanced, the high-temperature smoke can spread in the opposite direction upstream and downstream of the fire source, creating smoke backflow and turbulent diffusion. This seriously threatens the safety of personnel evacuation throughout the tunnel and the access routes for fire and rescue personnel, making it the primary cause of casualties in long tunnel fires. Precise and coordinated control of multiple smoke exhaust fans is the core means to suppress the cross-zone spread of smoke in long tunnel fires and ensure the safety of personnel evacuation and rescue routes.

[0003] Currently, the most widely used engineering technologies for controlling smoke exhaust fans in tunnel fires are fixed air volume or fixed frequency control modes and PID-based feedback control methods. Fixed air volume mode, based on pre-set fixed fan air volume and switching logic under design conditions, cannot adapt to the highly time-varying characteristics of dynamic increases in fire source power and the spatiotemporal evolution of smoke spread during fire development. In long-distance, multi-zone scenarios, it is prone to problems such as airflow collision, cross-zone smoke transmission, and smoke backflow, resulting in extremely poor smoke control effectiveness. While PID control can achieve single-variable dynamic air volume adjustment, its essence is a single-input, single-output error feedback architecture, which cannot solve the global collaborative optimization problem of strongly coupled, multi-input, multi-output systems with multiple units and multiple smoke control zones. Furthermore, pure feedback control has inherent lag and cannot achieve global control of smoke risk throughout the entire tunnel.

[0004] Existing advanced control technologies mainly rely on traditional model predictive control (MPC) and deep learning intelligent control. While traditional MPC has a theoretical basis for multivariable collaborative control, existing solutions are mostly designed for normal tunnel operation ventilation scenarios, employing simplified linear ventilation models that cannot adapt to the strong nonlinear and time-varying characteristics of smoke evolution during fires. This results in a severe mismatch between the model and the actual fire situation. Furthermore, the centralized architecture has a large computational load and poor real-time performance, failing to meet the emergency control requirements of distributed collaborative operation of multiple units in long-distance tunnels. Existing deep learning control solutions mostly adopt a purely data-driven neural network architecture, lacking the strict physical constraints of the smoke flow conservation law. This easily leads to non-physical prediction results that do not conform to engineering realities. Moreover, they mostly focus on offline or online tuning of PID parameters, failing to break through the architectural limitations of single-loop feedback control. They also do not consider the spatial topological coupling relationships of multiple measurement points and multiple units, resulting in severely insufficient model generalization and robustness to extreme conditions.

[0005] In summary, existing technologies in the field of multi-unit smoke exhaust control for long-distance tunnel fires face four major insurmountable technical bottlenecks: First, the prediction model bottleneck, which cannot simultaneously consider the strong nonlinear spatiotemporal distribution characteristics, physical consistency, and millisecond-level real-time prediction speed of smoke evolution in tunnel fires; second, the control architecture bottleneck, which cannot adapt to the distributed, strongly coupled, multi-input multi-output system composed of multiple units and multiple zones in long-distance tunnels, and cannot achieve a balance between global collaborative optimization and precise zone control; third, the optimization objective bottleneck, which cannot simultaneously consider the multi-objective collaborative optimization of personnel evacuation safety throughout the tunnel, global smoke control, and equipment operating energy consumption and service life; and fourth, the robustness bottleneck, which lacks adaptive fault-tolerant control capabilities for fan equipment failures, making it prone to control failures under extreme conditions and leading to serious safety accidents. Summary of the Invention

[0006] This invention aims to at least partially address one of the technical problems in related technologies. To this end, the first objective of this invention is to propose a collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires based on MPC (Multi-Process Control). This method achieves high-precision, high-physical-consistency, and millisecond-level rapid prediction of smoke evolution throughout the tunnel by constructing a spatiotemporal graph neural operator prediction model embedded with hard physical constraints on smoke flow. It also constructs a two-layer deep MPC control architecture with distributed collaboration of multiple fans, overcoming the real-time bottleneck of centralized MPC and achieving a balance between global collaborative optimization and precise zone control. Furthermore, it constructs a multi-objective optimization objective function integrating evacuation safety, smoke control, equipment energy consumption, and adjustment smoothness, embedding multi-dimensional hard constraints. Finally, it designs adaptive rolling time-domain optimization and fan fault adaptive reconfiguration modules to ensure real-time control and fault tolerance under extreme conditions. This invention achieves distributed collaborative closed-loop intelligent control of multiple smoke exhaust fans in long-distance tunnels while maintaining millisecond-level real-time response speed, providing a novel core control technology support for smoke control, personnel evacuation, and emergency rescue in long-distance tunnel fires.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for coordinated intelligent control of multiple smoke exhaust systems in tunnel fires based on MPC, comprising the following steps:

[0008] S1. Construct a full-size FDS numerical model of long-distance tunnels with multiple smoke prevention zones and multiple types of smoke exhaust units, and build a millisecond-level bidirectional coupling interface integrating Python, FDS and hardware-in-the-loop to realize synchronous iteration of CFD numerical calculation, MPC control algorithm and physical fan control hardware.

[0009] S2. Construct the spatiotemporal topology of tunnel fire smoke evolution, define the spatial relationship between multi-source sensor nodes, fan control nodes and smoke prevention zone nodes, and construct node feature matrix and adjacency matrix to provide structured input for the prediction model.

[0010] S3. Construct a global prediction model for the spatiotemporal evolution of ST-GNO flue gas with embedded physical hard constraints. Using the node feature matrix of historical time series as input, output the spatiotemporal distribution of flue gas features of each node in the entire tunnel in the future prediction time domain, so as to achieve accurate and rapid prediction of fire evolution.

[0011] S4 constructs a two-layer deep MPC control architecture for distributed collaboration of multiple sets of exhaust fans, including a global collaborative optimization layer and a partitioned rolling control layer, and clarifies the control objectives, inputs and outputs and constraint boundaries of each layer;

[0012] S5 constructs a multi-objective hybrid optimization objective function that integrates evacuation path safety, global flue gas risk, equipment operating energy consumption and adjustment smoothness, and embeds hard constraints such as fan rated operating conditions, smoke prevention zone pressure difference and evacuation safety threshold.

[0013] S6, designed with an adaptive rolling time domain optimization module and a wind turbine fault adaptive reconfiguration module, to achieve dynamic adjustment of the prediction time domain and the control time domain, as well as rapid redistribution of control commands under wind turbine fault conditions;

[0014] The S7, based on a tuned dual-layer depth MPC controller, achieves distributed collaborative closed-loop intelligent control of multiple smoke exhaust fans in tunnel fires through a three-in-one bidirectional coupling interface, and outputs timing control commands for each fan group, full tunnel fire situation data, and emergency prevention and control auxiliary decision-making information.

[0015] In addition, the MPC-based collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires according to the above embodiments of the present invention may also have the following additional technical features:

[0016] According to an embodiment of the present invention, step S1 includes:

[0017] S11. Construct a full-size FDS numerical model that matches the actual long-distance tunnel project, and divide the tunnel into multiple smoke control zones along the longitudinal direction of the tunnel according to the preset length; each smoke control zone is equipped with 1 set of jet fan units, 2 sets of axial flow smoke exhaust fan units and corresponding linkage smoke exhaust outlets, and 1 set of make-up air units to form a multi-type and multi-group distributed smoke exhaust ventilation system.

[0018] S12 sets multiple fire source locations in each smoke control zone to cover the center and boundary of each smoke control zone; the fire source power ranges from 3MW to 50MW, and the fire source power growth follows the t² ultra-fast fire model; the fuel types cover four typical tunnel fire fuels: gasoline, diesel, lithium battery and plastic, and corresponding heat release rate curves, combustion product generation rates and thermal radiation parameters are set respectively.

[0019] S13, a dynamic grid partitioning strategy is used to partition the tunnel space;

[0020] S14, multi-source composite sensor arrays are deployed according to smoke prevention zones. Each smoke prevention zone is equipped with one set of composite sensors every 10m along the longitudinal direction. The composite sensors integrate five monitoring functions: temperature, CO concentration, visibility, wind speed and pressure difference.

[0021] S15 establishes a three-in-one bidirectional coupling interface integrating Python, FDS, and hardware-in-the-loop, enabling 0.2s-level synchronous iteration across the three stages via asynchronous Python scripts: real-time extraction of full-tunnel sensor data and fan operation data from the FDS numerical model, or real-time data from physical sensors and fan control cabinets in the hardware-in-the-loop system; synchronously updating the airflow, air velocity, and smoke exhaust vent opening control commands output by the MPC controller to the corresponding boundaries of the FDS numerical model and the PLC control cabinet of the hardware-in-the-loop system; synchronizing fire evolution data and control process data to the tunnel fire digital twin emergency platform; and simultaneously setting hard constraints on the rated operating range of each fan and the action logic hard constraints on the fireproof roller shutters and smoke exhaust vents in the interface.

[0022] According to an embodiment of the present invention, step S2 includes:

[0023] S21, Define a graph structure G=(V,E,A), where V is the set of nodes, E is the set of edges, and A is a weighted adjacency matrix;

[0024] S22, the node set V is divided into three types of functional nodes: Sensing and monitoring nodes Vs: corresponding to all multi-source composite sensors in the entire tunnel, the feature vector of each node is the real-time monitoring value of temperature, CO concentration, visibility, wind speed and differential pressure at the measuring point of the multi-source composite sensor; Fan control nodes Vc: corresponding to all fan units and smoke exhaust outlets, the feature vector of each node is the current operating air volume, opening degree, speed and rated operating parameters of the equipment; Smoke prevention zone nodes Vz: corresponding to each independent smoke prevention zone, the feature vector of each node is the fire source power, average smoke concentration, evacuation route safety status and fireproof roller shutter opening and closing status of the zone;

[0025] S23, Construction of edge set E and adjacency matrix A: Based on the longitudinal spatial location of the tunnel, the direction of smoke flow, and the relationship between equipment control, the edge connection relationship between nodes is defined: there is a bidirectional edge between sensor nodes in adjacent spatial locations; there is a bidirectional edge between the fan control node and all sensor nodes in its smoke control zone; there is a bidirectional edge between the partition nodes of adjacent smoke control zones; there is a weighted directed edge between the node of the zone where the fire source is located and the nodes of the upstream and downstream adjacent partitions, and the weight is determined by the direction of smoke flow and wind speed; based on the edge connection relationship, a weighted adjacency matrix A is constructed to realize the structured representation of the tunnel spatial topology, smoke flow characteristics, and equipment control relationship.

[0026] S24. Construct the temporal node feature matrix, perform min-max normalization on the feature data of all nodes, and generate the temporal feature matrix [X(t),X(t-1),…,X(t-19)] of the current time and the 20 historical time steps, which serves as the input of the ST-GNO flue gas spatiotemporal evolution global prediction model.

[0027] According to an embodiment of the present invention, step S3 includes:

[0028] S31 establishes the basic architecture for constructing the ST-GNO flue gas spatiotemporal evolution global prediction model. The ST-GNO flue gas spatiotemporal evolution global prediction model consists of three cascaded parts: a spatial graph neural operator module, a time series neural operator module, and an output decoding module. The spatial graph neural operator module uses graph Fourier transform and kernel integral operator to capture the spatial topological coupling characteristics of flue gas evolution; the time series neural operator module uses time-domain integral operator to capture the temporal dynamic characteristics of flue gas evolution; and the output decoding module outputs the feature prediction values ​​of all nodes in the future prediction time domain H, which supports adaptive adjustment of 5 to 20 control time steps.

[0029] S32, a physical hard constraint embedding mechanism is constructed to transform the control equations of smoke flow in tunnel fires into hard constraint residual terms of the model and embed them into the solution process of the neural operator. The control equations include the continuity equation, momentum conservation equation, energy conservation equation and component transport equation. During model training and inference, the physical conservation laws are enforced to ensure the physical consistency of the prediction results from the root and avoid non-physical interpretations that do not conform to the laws of smoke flow.

[0030] S33, Construct the hybrid loss function of the global prediction model for the spatiotemporal evolution of ST-GNO flue gas, which consists of a weighted average of prediction accuracy loss, physical constraint loss, and generalization loss, and is expressed as follows:

[0031] ;

[0032] In the formula, L pred The mean square error loss between model predictions and FDS simulation or actual monitoring data; L phy L represents the residual loss of the flue gas flow control equations. reg L2 regularization loss is used to avoid model overfitting; λ phy λ represents the physical constraint weighting coefficient, with a value ranging from 1.0 to 1.5. reg This is the regularization weight coefficient, with a value ranging from 0.001 to 0.01;

[0033] S34 employs structured pruning, 8-bit integer quantization, and operator fusion techniques to lightweight the prediction model, compressing the number of floating-point parameters to 125,000. The Flash storage space occupied by the prediction model file after 8-bit integer quantization in the edge controller is stably controlled within 500KB, and the single-step inference time is less than 2ms. The prediction model is pre-trained based on the multi-condition FDS simulation dataset, and an online incremental update mechanism is constructed to fine-tune the prediction model every 5 seconds using the latest monitoring data to adapt to the dynamic evolution of the actual fire situation.

[0034] S35. Based on the prediction results of future time-domain smoke characteristics output by the prediction model, a spatiotemporal distribution matrix of smoke risk in the entire tunnel is constructed, and the smoke risk coefficient of each smoke prevention zone and evacuation path is calculated, providing a prediction basis for the rolling optimization of the MPC controller.

[0035] According to an embodiment of the present invention, step S4 includes:

[0036] S41, the global collaborative optimization layer, takes the future temporal smoke risk distribution of the entire tunnel output by the ST-GNO global prediction model of smoke spatiotemporal evolution as input, and takes the global optimization of the safety of the evacuation path of the entire tunnel and the minimization of the total smoke risk as the top-level objective. It optimizes and determines the control target value of each smoke control zone, the opening and closing status of the fireproof roller shutter of the smoke control zone, and the opening and closing strategy of the smoke exhaust outlet, and provides boundary conditions and target settings for the MPC controller of each zone. The optimization cycle of the global collaborative optimization layer is 1 second, which is synchronized with the incremental update cycle of the prediction model.

[0037] S42, the zoned rolling control layer, sets up an independent MPC controller for each smoke control zone. It takes the sensor monitoring data of this zone and adjacent zones, the zoned smoke prediction results of the ST-GNO smoke spatiotemporal evolution global prediction model, and the zoned control objectives issued by the global collaborative optimization layer as inputs. It takes the jet fan air volume, axial smoke exhaust fan air volume, smoke exhaust port opening and makeup air volume in this zone as control variables. It completes rolling optimization within each 0.2s control time step and outputs real-time control commands for each device in this zone.

[0038] S43, a zoned collaborative coupling mechanism, sets up an information exchange channel between the MPC controllers of adjacent smoke control zones to share real-time data on smoke status, wind speed and pressure difference at the boundary. During the rolling optimization process, it considers the impact of the control actions of adjacent zones on this zone, avoiding the problem of airflow collision between adjacent fans and cross-smoke between smoke control zones, and realizing distributed collaborative control of multiple zones.

[0039] S44 is a hierarchical control authority mechanism that sets four control levels: normal operating conditions, fire development conditions, extreme change conditions, and equipment failure conditions. The permissions, optimization cycles, and constraint boundaries of the global collaboration layer and the regional control layer are adaptively adjusted under different levels to ensure control priority and response speed under extreme conditions.

[0040] According to an embodiment of the present invention, step S5 includes:

[0041] S51, construct the multi-objective hybrid optimization objective function for the partitioned MPC controller, with the following expression:

[0042] ;

[0043] In the formula, J is the objective function of rolling optimization, with minimizing J as the optimization objective; ω1 is the weight coefficient of the flue gas risk sub-objective, ω2 is the weight coefficient of the evacuation safety sub-objective, ω3 is the weight coefficient of the equipment energy consumption sub-objective, and ω4 is the weight coefficient of the adjustment smoothness sub-objective. Each weight coefficient is adaptively adjusted according to the control condition level.

[0044] S52, the construction of each sub-objective item includes:

[0045] Flue gas risk sub-objective J risk : This is the sum of squares of the predicted flue gas risk values ​​of all sensor nodes within this partition in the prediction time domain, used to measure the effectiveness of flue gas control within the partition;

[0046] Evacuation safety sub-target J safe : To predict the sum of squares of the deviations between the temperature, CO concentration, visibility and safety thresholds at the safety monitoring points at the entrance of the evacuation route in this zone within the time domain, and to ensure the safety of the evacuation route.

[0047] Equipment energy consumption sub-target J energy : To predict the sum of the operating power of all wind turbines in this zone within the predicted time domain, and to measure the economic efficiency of equipment operation;

[0048] Adjusting the smoothness sub-target J smooth To predict the sum of squares of the changes of each control variable in adjacent time steps within the time domain, suppress drastic fluctuations in fan air volume and smoke exhaust port opening, and avoid frequent equipment start-ups and shutdowns and damage;

[0049] S53, Hard Constraint Setting: The following hard constraints are enforced during the rolling optimization process, without using soft penalties:

[0050] Equipment rated operating condition constraints: The air volume, speed, and operating power of each fan shall not exceed the upper and lower limits of the equipment rating;

[0051] Smoke control zone pressure difference constraint: The static pressure difference between adjacent smoke control zones shall not be less than 50Pa to prevent smoke cross-contamination;

[0052] Evacuation safety threshold constraints: The temperature at the entrance of the evacuation route shall not exceed 60℃, the CO concentration shall not exceed 200ppm, and the visibility shall not be less than 10m;

[0053] Physical constraints on flue gas flow: The ventilation volume corresponding to the control command is used to avoid flue gas backflow and turbulent diffusion. The physical constraint module of the ST-GNO flue gas spatiotemporal evolution global prediction model is used for verification.

[0054] S54, the weight coefficient adaptive adjustment mechanism, adaptively adjusts the weight coefficients of each sub-objective based on the fire level, smoke risk coefficient, and personnel evacuation status: when personnel are in the evacuation phase, the weight of ω2 is increased to prioritize evacuation safety; when the fire situation changes suddenly and smoke spreads rapidly, the weight of ω1 is increased to prioritize controlling smoke diffusion; when the fire situation is in a steady state, the weights are balanced to take into account both control effectiveness and equipment economy.

[0055] According to an embodiment of the present invention, step S6 includes:

[0056] S61, an adaptive rolling time-domain optimization module design, adaptively adjusts the prediction time domain H and control time domain U of the MPC based on the fire evolution rate, flue gas risk change rate, and degree of sudden change in operating conditions; wherein, each control time step is set to a first preset time:

[0057] When the fire is in a stable development stage and the rate of change of flue gas risk is below the threshold, the prediction time domain H is set to 20 control time steps and the control time domain U is set to 5 control time steps, taking into account both optimization accuracy and computational efficiency.

[0058] When the fire situation changes abruptly, smoke spreads rapidly, and the rate of change of smoke risk exceeds the threshold, the prediction time domain H is shortened to 10 control time steps, and the control time domain U is shortened to 2 control time steps to improve optimization speed and response sensitivity.

[0059] When an emergency scenario such as personnel evacuation or equipment failure is detected, the prediction time domain H is set to 5 control time steps and the control time domain U is set to 1 control time step to achieve a millisecond-level rapid response.

[0060] S62, the wind turbine fault detection and identification module, uses the isolated forest algorithm based on real-time data from wind turbine operating status sensors to identify fault types such as wind turbine shutdown, abnormal speed, and insufficient air volume, locates the faulty wind turbine and the fault level, and has a fault detection response time of less than 0.5s.

[0061] S63, the fan fault adaptive reconfiguration module, immediately updates the adjacency matrix of the spatiotemporal graph topology when a fan fault is detected, removes the faulty node, and the ST-GNO flue gas spatiotemporal evolution global prediction model quickly updates the prediction results; the global collaborative optimization layer readjusts the control objectives of each smoke control zone, and the MPC controller of the faulty zone and the MPC controller of the adjacent zone work together to complete rolling optimization, reallocate the control commands of normal fans, make up for the smoke exhaust capacity gap of the faulty fan, and realize non-disruptive fault-tolerant control under fault conditions;

[0062] S64, an extreme condition backup control mechanism, automatically switches to a distributed robust control mode based on fixed pressure difference of smoke control zones when the ST-GNO flue gas spatiotemporal evolution global prediction model inference is abnormal, multiple sensors fail, or multiple fans fail, ensuring the basic operating capability of the smoke exhaust system and avoiding control failure.

[0063] According to an embodiment of the present invention, step S7 includes:

[0064] S71, Pre-training and Parameter Tuning: Based on the full-condition FDS simulation dataset, complete the pre-training of the ST-GNO flue gas spatiotemporal evolution global prediction model, parameter tuning of the two-layer MPC controller, initial value determination of multi-objective optimization weight coefficients, and calibration of hard constraint boundaries.

[0065] S72, System Initialization: After a tunnel fire occurs, the system automatically starts, loads the pre-trained model and controller parameters, completes self-checks of the sensing system, fan equipment and coupling interfaces, locates the fire source and fire power based on the initial monitoring data, and initializes the spatiotemporal topology.

[0066] S73, Real-time Data Acquisition and Preprocessing: Through the three-in-one coupling interface, multi-source sensor data and wind turbine operation data of the entire tunnel are acquired according to the preset time step, and data normalization, outlier processing and spatiotemporal alignment are completed, and the time-series node feature matrix is ​​updated.

[0067] S74, Flue Gas Evolution Prediction and Risk Assessment: The ST-GNO flue gas spatiotemporal evolution global prediction model is based on the historical time series feature matrix, outputs the spatiotemporal distribution of flue gas characteristics in the entire tunnel in the future time domain, and calculates the flue gas risk matrix and evacuation path safety status of the entire tunnel.

[0068] S75, dual-layer MPC rolling optimization: The global collaborative optimization layer completes the global target optimization and issues control targets to each zone; the MPC controller of each zone synchronously completes the rolling optimization, calculates and outputs the control commands of each fan unit and smoke outlet, and outputs them after hard constraint verification.

[0069] S76, Control Command Execution and Closed-Loop Iteration: Control commands are synchronously sent to the PLC control cabinet of the FDS numerical model or hardware-in-the-loop system through the coupling interface to complete the adjustment of the fan and smoke exhaust port, and enter the next control time step iteration to realize full closed-loop intelligent control.

[0070] The present invention has the following technical effects:

[0071] This invention proposes a collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires based on MPC (Multi-Process Control). It achieves high-precision, high-physical-consistency, and millisecond-level rapid prediction of smoke evolution across the entire tunnel by constructing a spatiotemporal graph neural operator prediction model embedded with hard physical constraints on smoke flow. A two-layer deep MPC control architecture with distributed collaboration of multiple fans is constructed, overcoming the real-time bottleneck of centralized MPC and achieving a balance between global collaborative optimization and precise zone control. A multi-objective optimization objective function integrating evacuation safety, smoke control, equipment energy consumption, and adjustment smoothness is constructed, embedding multi-dimensional hard constraints. An adaptive rolling time-domain optimization and fan fault adaptive reconfiguration module are designed to ensure real-time control and fault tolerance under extreme conditions. This invention achieves distributed collaborative closed-loop intelligent control of multiple smoke exhaust fans in long-distance tunnels while maintaining millisecond-level real-time response speed, providing a novel core control technology support for smoke control, personnel evacuation, and emergency rescue in long-distance tunnel fires.

[0072] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0073] Figure 1 This is a flowchart of a collaborative intelligent control method for multiple smoke exhaust systems in a tunnel fire based on MPC, according to an embodiment of the present invention.

[0074] Figure 2 A schematic diagram of a multi-smoke-prevention zone, multi-unit FDS numerical model for a long-distance tunnel according to an embodiment of the present invention;

[0075] Figure 3 This is a flowchart of a Python-FDS-hardware-in-the-loop triad bidirectional coupling interface according to an embodiment of the present invention;

[0076] Figure 4 This is a diagram illustrating the architecture of a global prediction model for the spatiotemporal evolution of ST-GNO flue gas with embedded physical hard constraints, according to an embodiment of the present invention. Detailed Implementation

[0077] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0078] The following description, with reference to the accompanying drawings, describes the collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires based on MPC, as proposed in an embodiment of the present invention.

[0079] like Figure 1 As shown in the figure, the method for coordinated intelligent control of multiple smoke exhaust systems in tunnel fires based on MPC according to an embodiment of the present invention includes the following steps:

[0080] S1. Construct a full-size FDS numerical model of long-distance tunnels with multiple smoke prevention zones and various types of smoke exhaust units. Build a millisecond-level bidirectional coupling interface integrating Python, FDS, and hardware-in-the-loop to achieve synchronous iteration of CFD numerical calculations, MPC control algorithms, and physical fan control hardware.

[0081] According to an embodiment of the present invention, step S1 includes:

[0082] S11. Construct a full-size FDS numerical model that matches the actual long-distance tunnel project, and divide the tunnel into multiple smoke control zones along the longitudinal direction of the tunnel according to the preset length; each smoke control zone is equipped with 1 set of jet fan units, 2 sets of axial flow smoke exhaust fan units and corresponding linkage smoke exhaust outlets, and 1 set of make-up air units to form a multi-type and multi-group distributed smoke exhaust ventilation system.

[0083] For example, such as Figure 2 As shown, the full-size FDS numerical model matches the actual long-distance tunnel project. The tunnel dimensions are 1500m long, 10m wide, and 5.5m high. The tunnel is divided into 5 independent smoke control zones every 300m along its longitudinal direction. Each smoke control zone is equipped with 1 set of jet fan units, 2 sets of axial flow smoke exhaust fan units and corresponding linked smoke exhaust outlets, and 1 set of make-up air units, forming a multi-type and multi-group distributed smoke exhaust ventilation system.

[0084] S12 sets multiple (e.g., 7) fire source locations in each smoke control zone to cover the center and boundary of each zone; the fire source power ranges from 3MW to 50MW, with 10 levels in 5MW increments; the fire source power increases according to the t² ultra-fast fire model; the fuel types cover four typical tunnel fire fuels: gasoline, diesel, lithium batteries, and plastics, with corresponding heat release rate curves, combustion product generation rates, and thermal radiation parameters set for each.

[0085] S13 uses a dynamic grid partitioning strategy to divide the tunnel space.

[0086] For example, a high-precision grid of 0.2m×0.2m×0.2m is used in the 80m radius area of ​​the fire source center and the 10m area around the fan and smoke exhaust outlet. The smoke front movement area is automatically densified to a 0.3m×0.3m×0.3m grid, and the remaining areas use a standard grid of 0.5m×0.5m×0.5m. The boundaries at both ends of the tunnel are set to OPEN, the boundaries of the fireproof roller shutters between each smoke control zone are set to dynamically switchable OPEN / CLOSED, and the boundaries of the smoke exhaust outlets are set to VENT type, supporting continuous adjustment of the opening degree from 0 to 100%.

[0087] S14, multi-source composite sensor arrays are deployed according to smoke prevention zones. Each smoke prevention zone is equipped with one set of composite sensors every 10m along the longitudinal direction. The composite sensors integrate five monitoring functions: temperature, CO concentration, visibility, wind speed and pressure difference.

[0088] For example, a multi-source composite sensor array is deployed according to smoke prevention zones. Each smoke prevention zone has one set of composite sensors every 10m along the longitudinal direction, integrating five monitoring functions: temperature, CO concentration, visibility, wind speed, and pressure difference. A dedicated safety monitoring point is set at the entrance of the evacuation passage of each smoke prevention zone, and each fan unit is equipped with an operating status sensor to collect real-time data on fan air volume, speed, and operating voltage and current. The sensor sampling time interval is 0.2s, and the total number of sensors in the entire tunnel is no less than 160.

[0089] S15 establishes a three-in-one bidirectional coupling interface integrating Python, FDS, and hardware-in-the-loop. This interface uses asynchronous Python scripts to achieve 0.2-second synchronous iteration across three stages: real-time extraction of full-tunnel sensor data and fan operation data from the FDS numerical model, or real-time data from physical sensors and fan control cabinets in the hardware-in-the-loop system; synchronously updating the airflow, velocity, and smoke exhaust vent opening control commands output by the MPC controller to the corresponding boundaries of the FDS numerical model and the PLC control cabinet in the hardware-in-the-loop system; synchronizing fire evolution data and control process data to the tunnel fire digital twin emergency platform; and simultaneously setting hard constraints on the rated operating range of each fan and the action logic hard constraints on the fireproof roller shutters and smoke exhaust vents within the interface. The specific setup process is as follows: Figure 3 As shown.

[0090] S2 constructs the spatiotemporal topology of tunnel fire smoke evolution, defines the spatial relationships between multi-source sensor nodes, fan control nodes and smoke prevention zone nodes, and constructs node feature matrix and adjacency matrix to provide structured input for the prediction model.

[0091] According to an embodiment of the present invention, step S2 includes:

[0092] S21, Define a graph structure G=(V,E,A), where V is the set of nodes, E is the set of edges, and A is a weighted adjacency matrix;

[0093] S22, the node set V is divided into three types of functional nodes: Sensing and monitoring nodes Vs: corresponding to all multi-source composite sensors in the entire tunnel, the feature vector of each node is the real-time monitoring value of temperature, CO concentration, visibility, wind speed and differential pressure at the measuring point of the multi-source composite sensor; Fan control nodes Vc: corresponding to all fan units and smoke exhaust outlets, the feature vector of each node is the current operating air volume, opening degree, speed and rated operating parameters of the equipment; Smoke prevention zone nodes Vz: corresponding to each independent smoke prevention zone, the feature vector of each node is the fire source power, average smoke concentration, evacuation route safety status and fireproof roller shutter opening and closing status of the zone;

[0094] S23, Construction of edge set E and adjacency matrix A: Based on the longitudinal spatial location of the tunnel, the direction of smoke flow, and the relationship between equipment control, the edge connection relationship between nodes is defined: there is a bidirectional edge between sensor nodes in adjacent spatial locations; there is a bidirectional edge between the fan control node and all sensor nodes in its smoke control zone; there is a bidirectional edge between the partition nodes of adjacent smoke control zones; there is a weighted directed edge between the node of the zone where the fire source is located and the nodes of the upstream and downstream adjacent partitions, and the weight is determined by the direction of smoke flow and wind speed; based on the edge connection relationship, a weighted adjacency matrix A is constructed to realize the structured representation of the tunnel spatial topology, smoke flow characteristics, and equipment control relationship.

[0095] S24. Construct the temporal node feature matrix, perform min-max normalization on the feature data of all nodes, and generate the temporal feature matrix [X(t),X(t-1),…,X(t-19)] of the current time and the 20 historical time steps, which serves as the input of the ST-GNO flue gas spatiotemporal evolution global prediction model.

[0096] S3 constructs a global prediction model for the spatiotemporal evolution of ST-GNO flue gas with embedded physical hard constraints. It takes the node feature matrix of historical time series as input and outputs the spatiotemporal distribution of flue gas features of each node in the entire tunnel in the future prediction time domain, so as to achieve accurate and rapid prediction of fire evolution.

[0097] According to an embodiment of the present invention, step S3 includes:

[0098] S31 establishes the basic architecture for constructing the ST-GNO flue gas spatiotemporal evolution global prediction model. The ST-GNO flue gas spatiotemporal evolution global prediction model consists of three cascaded parts: a spatial graph neural operator module, a time series neural operator module, and an output decoding module. The spatial graph neural operator module uses graph Fourier transform and kernel integral operator to capture the spatial topological coupling characteristics of flue gas evolution; the time series neural operator module uses time-domain integral operator to capture the temporal dynamic characteristics of flue gas evolution; and the output decoding module outputs the feature prediction values ​​of all nodes in the future prediction time domain H, which supports adaptive adjustment of 5 to 20 control time steps.

[0099] S32, a physical hard constraint embedding mechanism is constructed to transform the control equations of smoke flow in tunnel fires into hard constraint residual terms of the model and embed them into the solution process of the neural operator. The control equations include the continuity equation, momentum conservation equation, energy conservation equation and component transport equation. During model training and inference, the physical conservation laws are enforced to ensure the physical consistency of the prediction results from the root and avoid non-physical interpretations that do not conform to the laws of smoke flow.

[0100] S33, Construct the hybrid loss function of the global prediction model for the spatiotemporal evolution of ST-GNO flue gas, which consists of a weighted average of prediction accuracy loss, physical constraint loss, and generalization loss, and is expressed as follows:

[0101] ;

[0102] In the formula, L pred The mean square error loss between model predictions and FDS simulation or actual monitoring data; L phy L represents the residual loss of the flue gas flow control equations. reg L2 regularization loss is used to avoid model overfitting; λ phy λ represents the physical constraint weighting coefficient, with a value ranging from 1.0 to 1.5. reg This is the regularization weight coefficient, with a value ranging from 0.001 to 0.01;

[0103] S34 employs structured pruning, 8-bit integer quantization, and operator fusion techniques to lightweight the prediction model, compressing the number of floating-point parameters to 125,000. The Flash storage space occupied by the prediction model file after 8-bit integer quantization in the edge controller is stably controlled within 500KB, and the single-step inference time is less than 2ms. The prediction model is pre-trained based on the multi-condition FDS simulation dataset, and an online incremental update mechanism is constructed to fine-tune the prediction model every 5 seconds using the latest monitoring data to adapt to the dynamic evolution of the actual fire situation.

[0104] S35, based on the predicted future time-domain flue gas characteristics output by the prediction model, constructs a spatiotemporal distribution matrix of flue gas risk for the entire tunnel, calculates the flue gas risk coefficients for each smoke control zone and evacuation path, and provides a predictive basis for the rolling optimization of the MPC controller. The architecture of the ST-GNO flue gas spatiotemporal evolution global prediction model with embedded physical hard constraints is as follows: Figure 4 As shown.

[0105] S4 constructs a two-layer deep MPC control architecture for distributed collaboration of multiple sets of exhaust fans, including a global collaborative optimization layer and a partitioned rolling control layer, and clarifies the control objectives, inputs, outputs and constraint boundaries of each layer.

[0106] According to an embodiment of the present invention, step S4 includes:

[0107] S41, the global collaborative optimization layer, takes the future temporal smoke risk distribution of the entire tunnel output by the ST-GNO global prediction model of smoke spatiotemporal evolution as input, and takes the global optimization of the safety of the evacuation path of the entire tunnel and the minimization of the total smoke risk as the top-level objective. It optimizes and determines the control target value of each smoke control zone, the opening and closing status of the fireproof roller shutter of the smoke control zone, and the opening and closing strategy of the smoke exhaust outlet, and provides boundary conditions and target settings for the MPC controller of each zone. The optimization cycle of the global collaborative optimization layer is 1 second, which is synchronized with the incremental update cycle of the prediction model.

[0108] S42, the zoned rolling control layer, sets up an independent MPC controller for each smoke control zone. It takes the sensor monitoring data of this zone and adjacent zones, the zoned smoke prediction results of the ST-GNO smoke spatiotemporal evolution global prediction model, and the zoned control objectives issued by the global collaborative optimization layer as inputs. It takes the jet fan air volume, axial smoke exhaust fan air volume, smoke exhaust port opening and makeup air volume in this zone as control variables. It completes rolling optimization within each 0.2s control time step and outputs real-time control commands for each device in this zone.

[0109] S43, a zoned collaborative coupling mechanism, sets up an information exchange channel between the MPC controllers of adjacent smoke control zones to share real-time data on smoke status, wind speed and pressure difference at the boundary. During the rolling optimization process, it considers the impact of the control actions of adjacent zones on this zone, avoiding the problem of airflow collision between adjacent fans and cross-smoke between smoke control zones, and realizing distributed collaborative control of multiple zones.

[0110] S44 is a hierarchical control authority mechanism that sets four control levels: normal operating conditions, fire development conditions, extreme change conditions, and equipment failure conditions. The permissions, optimization cycles, and constraint boundaries of the global collaboration layer and the regional control layer are adaptively adjusted under different levels to ensure control priority and response speed under extreme conditions.

[0111] S5 constructs a multi-objective hybrid optimization objective function that integrates evacuation path safety, global flue gas risk, equipment operating energy consumption, and regulation smoothness, and embeds hard constraints such as fan rated operating conditions, smoke prevention zone pressure difference, and evacuation safety threshold.

[0112] According to an embodiment of the present invention, step S5 includes:

[0113] S51, construct the multi-objective hybrid optimization objective function for the partitioned MPC controller, with the following expression:

[0114] ;

[0115] In the formula, J is the objective function of rolling optimization, with minimizing J as the optimization objective; ω1 is the weight coefficient of the flue gas risk sub-objective, ω2 is the weight coefficient of the evacuation safety sub-objective, ω3 is the weight coefficient of the equipment energy consumption sub-objective, and ω4 is the weight coefficient of the adjustment smoothness sub-objective. Each weight coefficient is adaptively adjusted according to the control condition level.

[0116] S52, the construction of each sub-objective item includes:

[0117] Flue gas risk sub-objective J risk : This is the sum of squares of the predicted flue gas risk values ​​of all sensor nodes within this partition in the prediction time domain, used to measure the effectiveness of flue gas control within the partition;

[0118] Evacuation safety sub-target J safe : To predict the sum of squares of the deviations between the temperature, CO concentration, visibility and safety thresholds at the safety monitoring points at the entrance of the evacuation route in this zone within the time domain, and to ensure the safety of the evacuation route.

[0119] Equipment energy consumption sub-target J energy : To predict the sum of the operating power of all wind turbines in this zone within the predicted time domain, and to measure the economic efficiency of equipment operation;

[0120] Adjusting the smoothness sub-target J smooth To predict the sum of squares of the changes of each control variable in adjacent time steps within the time domain, suppress drastic fluctuations in fan air volume and smoke exhaust port opening, and avoid frequent equipment start-ups and shutdowns and damage.

[0121] S53, Hard Constraint Setting: The following hard constraints are enforced during the rolling optimization process, without using soft penalties:

[0122] Equipment rated operating condition constraints: The air volume, speed, and operating power of each fan shall not exceed the upper and lower limits of the equipment rating, and the opening of the smoke exhaust port shall be within the range of 0~100%;

[0123] Smoke control zone pressure difference constraint: The static pressure difference between adjacent smoke control zones shall not be less than 50Pa to prevent smoke cross-contamination;

[0124] Evacuation safety threshold constraints: The temperature at the entrance of the evacuation route shall not exceed 60℃, the CO concentration shall not exceed 200ppm, and the visibility shall not be less than 10m;

[0125] Physical constraints on flue gas flow: The ventilation volume corresponding to the control command is used to avoid flue gas backflow and turbulent diffusion. The physical constraint module of the ST-GNO flue gas spatiotemporal evolution global prediction model is used for verification.

[0126] S54, the weight coefficient adaptive adjustment mechanism, adaptively adjusts the weight coefficients of each sub-objective based on the fire level, smoke risk coefficient, and personnel evacuation status: when personnel are in the evacuation phase, the weight of ω2 is increased to prioritize evacuation safety; when the fire situation changes suddenly and smoke spreads rapidly, the weight of ω1 is increased to prioritize controlling smoke diffusion; when the fire situation is in a steady state, the weights are balanced to take into account both control effectiveness and equipment economy.

[0127] S6 is designed with an adaptive rolling time-domain optimization module and a wind turbine fault adaptive reconfiguration module to achieve dynamic adjustment of the prediction time domain and the control time domain, as well as rapid reconfiguration of control commands under wind turbine fault conditions.

[0128] According to an embodiment of the present invention, step S6 includes:

[0129] S61, an adaptive rolling time-domain optimization module design, adaptively adjusts the prediction time domain H and control time domain U of the MPC based on the fire evolution rate, flue gas risk change rate, and degree of sudden change in operating conditions; wherein, each control time step is set to a first preset time (for example, the first preset time can be 0.2s):

[0130] When the fire is in a stable development stage and the rate of change of smoke risk is lower than the threshold, the prediction time domain H is set to 20 control time steps (corresponding to 4s) and the control time domain U is set to 5 control time steps (corresponding to 1s), taking into account both optimization accuracy and computational efficiency.

[0131] When the fire situation changes abruptly, smoke spreads rapidly, and the rate of change in smoke risk exceeds the threshold, the prediction time domain H is shortened to 10 control time steps (corresponding to 2s), and the control time domain U is shortened to 2 control time steps (corresponding to 0.4s), thereby improving the optimization speed and response sensitivity.

[0132] When an emergency scenario such as personnel evacuation or equipment failure is detected, the prediction time domain H is set to 5 control time steps (corresponding to 1 second), and the control time domain U is set to 1 control time step (corresponding to 0.2 seconds) to achieve a millisecond-level rapid response.

[0133] S62, the wind turbine fault detection and identification module, uses the isolated forest algorithm based on real-time data from wind turbine operating status sensors to identify fault types such as wind turbine shutdown, abnormal speed, and insufficient air volume, locates the faulty wind turbine and the fault level, and has a fault detection response time of less than 0.5s.

[0134] S63, the fan fault adaptive reconfiguration module, immediately updates the adjacency matrix of the spatiotemporal graph topology when a fan fault is detected, removes the faulty node, and the ST-GNO flue gas spatiotemporal evolution global prediction model quickly updates the prediction results; the global collaborative optimization layer readjusts the control objectives of each smoke control zone, and the MPC controller of the faulty zone and the MPC controller of the adjacent zone work together to complete rolling optimization, reallocate the control commands of normal fans, make up for the smoke exhaust capacity gap of the faulty fan, and realize non-disruptive fault-tolerant control under fault conditions;

[0135] S64, an extreme condition backup control mechanism, automatically switches to a distributed robust control mode based on fixed pressure difference of smoke control zones when the ST-GNO flue gas spatiotemporal evolution global prediction model inference is abnormal, multiple sensors fail, or multiple fans fail, ensuring the basic operating capability of the smoke exhaust system and avoiding control failure.

[0136] The S7, based on a tuned dual-layer depth MPC controller, achieves distributed collaborative closed-loop intelligent control of multiple smoke exhaust fans in tunnel fires through a three-in-one bidirectional coupling interface, and outputs timing control commands for each fan group, full tunnel fire situation data, and emergency prevention and control auxiliary decision-making information.

[0137] According to an embodiment of the present invention, step S7 includes:

[0138] S71, Pre-training and Parameter Tuning: Based on the full-condition FDS simulation dataset, complete the pre-training of the ST-GNO flue gas spatiotemporal evolution global prediction model, parameter tuning of the two-layer MPC controller, initial value determination of multi-objective optimization weight coefficients, and calibration of hard constraint boundaries.

[0139] S72, System Initialization: After a tunnel fire occurs, the system automatically starts, loads the pre-trained model and controller parameters, completes self-checks of the sensing system, fan equipment and coupling interfaces, locates the fire source and fire power based on the initial monitoring data, and initializes the spatiotemporal topology.

[0140] S73, Real-time Data Acquisition and Preprocessing: Through the three-in-one coupling interface, multi-source sensor data and wind turbine operation data of the entire tunnel are acquired at a preset time step (for example, the preset time step can be 0.2s), and data normalization, outlier processing and spatiotemporal alignment are completed, and the time-series node feature matrix is ​​updated.

[0141] S74, Flue Gas Evolution Prediction and Risk Assessment: The ST-GNO flue gas spatiotemporal evolution global prediction model is based on the historical time series feature matrix, outputs the spatiotemporal distribution of flue gas characteristics in the entire tunnel in the future time domain, and calculates the flue gas risk matrix and evacuation path safety status of the entire tunnel.

[0142] S75, dual-layer MPC rolling optimization: The global collaborative optimization layer completes the global target optimization and issues control targets to each zone; the MPC controller of each zone synchronously completes the rolling optimization, calculates and outputs the control commands of each fan unit and smoke outlet, and outputs them after hard constraint verification.

[0143] S76, Control Command Execution and Closed-Loop Iteration: Control commands are synchronously sent to the PLC control cabinet of the FDS numerical model or hardware-in-the-loop system through the coupling interface to complete the adjustment of the fan and smoke exhaust port, and enter the next control time step iteration to realize full closed-loop intelligent control.

[0144] In summary, the MPC-based collaborative intelligent control method for multiple smoke exhaust systems in tunnel fires, as described in this invention, firstly, constructs a full-size FDS numerical model matching the actual tunnel and establishes a Python-FDS-hardware-in-the-loop three-in-one millisecond-level bidirectional coupling interface to achieve 0.2-second-level synchronous data interaction and command closed-loop iteration between CFD numerical calculation, MPC control algorithm, and physical fan control hardware. Based on this, it abstracts multi-source sensors, fan control equipment, and smoke control zones within the tunnel as nodes, and constructs a spatiotemporal diagram topology of tunnel fire smoke evolution based on spatial location and smoke flow relationship, generating a node feature matrix and adjacency matrix containing historical temporal characteristics, which serves as the basis for subsequent... The system provides structured input for the prediction. Next, it employs a spatiotemporal graph neural operator (ST-GNO) prediction model, embedding continuity equations, momentum conservation equations, energy conservation equations, and component transport equations as physical hard constraints. Using historical time-series feature matrices as input, it rapidly infers and outputs the spatiotemporal distribution of smoke characteristics such as temperature, CO concentration, visibility, wind speed, and pressure difference at each node of the entire tunnel within the future prediction time domain, thereby generating a full-tunnel smoke risk matrix. Subsequently, a two-layer deep MPC control architecture consisting of a global collaborative optimization layer and a zoned rolling control layer is constructed. The global layer, based on the future risk distribution of the entire tunnel, optimizes and determines the control objectives of each smoke control zone and the strategies for fireproof roller shutters and smoke exhaust outlets with a 1-second cycle. The zoned layer... An independent MPC controller is set up for each smoke control zone, with a control step size of 0.2 seconds. Based on real-time sensor data of the zone and adjacent zones, ST-GNO zone prediction results, and control objectives issued by the upper level, the controller performs rolling optimization to solve for control commands such as jet fan airflow, axial smoke exhaust fan airflow, smoke exhaust outlet opening, and makeup airflow. During the optimization process, a multi-objective hybrid optimization function is adopted, integrating smoke risk, evacuation safety, equipment energy consumption, and adjustment smoothness, and forcibly satisfying hard constraints such as fan rated operating conditions, smoke control zone pressure difference, and evacuation safety thresholds. Simultaneously, an adaptive rolling time-domain optimization module is introduced to dynamically adjust the prediction time domain and control time domain according to the fire evolution rate, and the isolated forest algorithm is used for implementation. The system detects fan malfunctions in real time. When a malfunction occurs, it quickly updates the spatiotemporal graph adjacency matrix and triggers collaborative reconfiguration between the global layer and adjacent zone controllers, migrating control tasks to normal fans to achieve non-disruptive fault tolerance. Finally, the pre-trained and parameter-tuned dual-layer deep MPC controller sends timing control commands to the FDS model or physical PLC control cabinet for execution through a three-in-one bidirectional coupling interface. At the same time, it transmits fire situation data and emergency auxiliary decision-making information back in real time, forming a full-link distributed collaborative closed-loop intelligent control from perception, prediction, optimization, decision-making to execution. This significantly shortens the collaborative response time of multiple units, reduces the smoke risk coverage of the entire tunnel, and ensures the safety of personnel evacuation and the economic efficiency of equipment operation.

[0145] By adopting the above technical solution, the present invention has the following beneficial effects:

[0146] (1) Dual improvement in prediction accuracy and speed: By embedding the ST-GNO model with physical hard constraints, high accuracy (error less than 3%), high physical consistency and millisecond-level (less than 2ms) fast prediction of flue gas evolution in the whole tunnel are achieved.

[0147] (2) Distributed collaboration of control architecture: Through the dual-layer deep MPC control architecture, the real-time bottleneck of centralized MPC is broken through, and the balance between global collaborative optimization and precise partition control is achieved, which shortens the response time of multi-unit collaborative control by more than 65%.

[0148] (3) Multi-objective collaborative optimization: By integrating multi-objective functions of evacuation safety, smoke control, equipment energy consumption and adjustment smoothness, the system ensures personnel evacuation safety (reduces the smoke risk coverage of the entire tunnel by more than 80%) while taking into account the economy and service life of equipment operation.

[0149] (4) High robustness under extreme conditions: Through adaptive rolling time-domain optimization and fan failure adaptive reconfiguration module, the real-time control and fault tolerance under extreme conditions such as sudden fire and equipment failure are guaranteed, and the seamless switching under fault conditions is realized.

[0150] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0151] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0152] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0153] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent coordination and regulation of a multi-group smoke exhaust system for tunnel fire based on MPC, characterized in that, The method comprises the following steps: S1, a full-size FDS numerical model of long-distance tunnel multi-smoke partition and multi-type smoke exhaust unit is constructed, a Python-FDS-hardware-in-the-loop three-in-one millisecond-level bidirectional coupling interface is built, and synchronous iteration of CFD numerical calculation, MPC control algorithm and entity fan control hardware is realized; S2, a tunnel fire smoke evolution space-time graph topology is constructed, a spatial correlation relationship of multi-source sensing nodes, fan control nodes and smoke partition nodes is defined, a node feature matrix and an adjacency matrix are constructed, and a structured input is provided for a prediction model; S3, an ST-GNO smoke space-time evolution global prediction model embedded with physical hard constraints is constructed, the historical time series node feature matrix is taken as input, and the smoke feature space-time distribution of each node in the whole tunnel in the future prediction time domain is output, so that accurate and rapid prediction of fire evolution is realized; S4, a double-layer deep MPC control architecture of distributed coordination of multiple groups of smoke exhaust fans is constructed, including a global collaborative optimization layer and a partition rolling control layer, and the control target, input and output and constraint boundary of each layer are defined; S5, a multi-objective hybrid optimization objective function is constructed, which integrates evacuation path safety, global smoke risk, equipment operation energy consumption and adjustment smoothness, and hard constraint conditions of fan rated working condition, smoke partition pressure difference and evacuation safety threshold are embedded; S6, an adaptive rolling time domain optimization module and a fan fault adaptive reconfiguration module are designed, and dynamic adjustment of the prediction time domain and the control time domain and rapid redistribution of control instructions under fan failure are realized; S7, based on the double-layer deep MPC controller, through the three-in-one bidirectional coupling interface, distributed collaborative closed-loop intelligent regulation and control of the tunnel fire multiple groups of smoke exhaust fans is realized, and time sequence control instructions of each fan group, whole tunnel fire situation data and emergency prevention and control auxiliary decision information are output.

2. The MPC-based intelligent coordinated regulation and control method for a multi-group smoke exhaust system in a tunnel fire according to claim 1, characterized in that, Step S1 comprises: S11, a full-size FDS numerical model matched with the actual long-distance tunnel engineering is constructed, and the tunnel is divided into multiple smoke partitions according to the preset length along the longitudinal direction of the tunnel; each smoke partition is provided with 1 group of jet fan group, 2 groups of axial flow smoke exhaust fan group and corresponding linkage smoke exhaust port, 1 group of air supply fan group, forming a distributed smoke exhaust ventilation system of multiple types and multiple groups; S12, multiple fire source positions are set in each smoke partition to cover the center and the boundary of each smoke partition; the fire power covers 3MW-50MW, and the fire power growth follows the t² super-fast fire model; the fuel types cover four typical tunnel fire fuels of gasoline, diesel, lithium battery and plastic, and corresponding heat release rate curves, combustion product generation rate and heat radiation parameters are set respectively; S13, dynamic grid division strategy is adopted for tunnel space division; S14, a multi-source composite sensor array is arranged according to the smoke partition, and 1 group of composite sensor is arranged every 10m along the longitudinal direction of each smoke partition, and the composite sensor integrates 5 monitoring functions of temperature, CO concentration, visibility, wind speed and pressure difference; S15 establishes a three-in-one bidirectional coupling interface integrating Python, FDS, and hardware-in-the-loop, enabling 0.2s-level synchronous iteration across the three stages via asynchronous Python scripts: real-time extraction of full-tunnel sensor data and fan operation data from the FDS numerical model, or real-time data from physical sensors and fan control cabinets in the hardware-in-the-loop system; synchronously updating the airflow, air velocity, and smoke exhaust vent opening control commands output by the MPC controller to the corresponding boundaries of the FDS numerical model and the PLC control cabinet of the hardware-in-the-loop system; synchronizing fire evolution data and control process data to the tunnel fire digital twin emergency platform; and simultaneously setting hard constraints on the rated operating range of each fan and the action logic hard constraints on the fireproof roller shutters and smoke exhaust vents in the interface.

3. The MPC-based intelligent coordinated regulation and control method for a multi-group smoke exhaust system in a tunnel fire according to claim 1, characterized in that, Step S2 includes: S21, Define a graph structure G=(V,E,A), where V is the set of nodes, E is the set of edges, and A is a weighted adjacency matrix; S22, the node set V is divided into three types of functional nodes: Sensing and monitoring nodes Vs: corresponding to all multi-source composite sensors in the entire tunnel, the feature vector of each node is the real-time monitoring value of temperature, CO concentration, visibility, wind speed and differential pressure at the measuring point of the multi-source composite sensor; Fan control nodes Vc: corresponding to all fan units and smoke exhaust outlets, the feature vector of each node is the current operating air volume, opening degree, speed and rated operating parameters of the equipment; Smoke prevention zone nodes Vz: corresponding to each independent smoke prevention zone, the feature vector of each node is the fire source power, average smoke concentration, evacuation route safety status and fireproof roller shutter opening and closing status of the zone; S23, Construction of edge set E and adjacency matrix A: Based on the longitudinal spatial location of the tunnel, the direction of smoke flow, and the relationship between equipment control, the edge connection relationship between nodes is defined: there is a bidirectional edge between sensor nodes in adjacent spatial locations; there is a bidirectional edge between the fan control node and all sensor nodes in its smoke control zone; there is a bidirectional edge between the partition nodes of adjacent smoke control zones; there is a weighted directed edge between the node of the zone where the fire source is located and the nodes of the upstream and downstream adjacent partitions, and the weight is determined by the direction of smoke flow and wind speed; based on the edge connection relationship, a weighted adjacency matrix A is constructed to realize the structured representation of the tunnel spatial topology, smoke flow characteristics, and equipment control relationship. S24. Construct the temporal node feature matrix, perform min-max normalization on the feature data of all nodes, and generate the temporal feature matrix [X(t),X(t-1),…,X(t-19)] of the current time and the 20 historical time steps, which serves as the input of the ST-GNO flue gas spatiotemporal evolution global prediction model.

4. The MPC-based intelligent coordinated regulation and control method for a multi-group smoke exhaust system in a tunnel fire according to claim 1, characterized in that, Step S3 includes: S31 establishes the basic architecture for constructing the ST-GNO flue gas spatiotemporal evolution global prediction model. The ST-GNO flue gas spatiotemporal evolution global prediction model consists of three cascaded parts: a spatial graph neural operator module, a time series neural operator module, and an output decoding module. The spatial graph neural operator module uses graph Fourier transform and kernel integral operator to capture the spatial topological coupling characteristics of flue gas evolution; the time series neural operator module uses time-domain integral operator to capture the temporal dynamic characteristics of flue gas evolution; and the output decoding module outputs the feature prediction values ​​of all nodes in the future prediction time domain H, which supports adaptive adjustment of 5 to 20 control time steps. S32, a physical hard constraint embedding mechanism is constructed to transform the control equations of smoke flow in tunnel fires into hard constraint residual terms of the model and embed them into the solution process of the neural operator. The control equations include the continuity equation, momentum conservation equation, energy conservation equation and component transport equation. During model training and inference, the physical conservation laws are enforced to ensure the physical consistency of the prediction results from the root and avoid non-physical interpretations that do not conform to the laws of smoke flow. S33, Construct the hybrid loss function of the global prediction model for the spatiotemporal evolution of ST-GNO flue gas, which consists of a weighted average of prediction accuracy loss, physical constraint loss, and generalization loss, and is expressed as follows: ； In the formula, L pred is the mean square error loss of the model prediction value and the FDS simulation or actual monitoring data; L phy is the residual loss of the flue gas flow control equation; L reg is the L2 regularization loss to avoid model overfitting; λ phy is the physical constraint weight coefficient, and the value is 1.0-1.5; λ reg is the regularization weight coefficient, and the value is 0.001-0.01; S34 employs structured pruning, 8-bit integer quantization, and operator fusion techniques to lightweight the prediction model, compressing the number of floating-point parameters to 125,000. The Flash storage space occupied by the prediction model file after 8-bit integer quantization in the edge controller is stably controlled within 500KB, and the single-step inference time is less than 2ms. The prediction model is pre-trained based on the multi-condition FDS simulation dataset, and an online incremental update mechanism is constructed to fine-tune the prediction model every 5 seconds using the latest monitoring data to adapt to the dynamic evolution of the actual fire situation. S35. Based on the prediction results of future time-domain smoke characteristics output by the prediction model, a spatiotemporal distribution matrix of smoke risk in the entire tunnel is constructed, and the smoke risk coefficient of each smoke prevention zone and evacuation path is calculated, providing a prediction basis for the rolling optimization of the MPC controller.

5. The MPC-based intelligent coordinated regulation and control method for a multi-group smoke exhaust system in a tunnel fire according to claim 1, characterized in that, Step S4 includes: S41, the global collaborative optimization layer, takes the future temporal smoke risk distribution of the entire tunnel output by the ST-GNO global prediction model of smoke spatiotemporal evolution as input, and takes the global optimization of the safety of the evacuation path of the entire tunnel and the minimization of the total smoke risk as the top-level objective. It optimizes and determines the control target value of each smoke control zone, the opening and closing status of the fireproof roller shutter of the smoke control zone, and the opening and closing strategy of the smoke exhaust outlet, and provides boundary conditions and target settings for the MPC controller of each zone. The optimization cycle of the global collaborative optimization layer is 1 second, which is synchronized with the incremental update cycle of the prediction model. S42, the zoned rolling control layer, sets up an independent MPC controller for each smoke control zone. It takes the sensor monitoring data of this zone and adjacent zones, the zoned smoke prediction results of the ST-GNO smoke spatiotemporal evolution global prediction model, and the zoned control objectives issued by the global collaborative optimization layer as inputs. It takes the jet fan air volume, axial smoke exhaust fan air volume, smoke exhaust port opening and makeup air volume in this zone as control variables. It completes rolling optimization within each 0.2s control time step and outputs real-time control commands for each device in this zone. S43, a zoned collaborative coupling mechanism, sets up an information exchange channel between the MPC controllers of adjacent smoke control zones to share real-time data on smoke status, wind speed and pressure difference at the boundary. During the rolling optimization process, it considers the impact of the control actions of adjacent zones on this zone, avoiding the problem of airflow collision between adjacent fans and cross-smoke between smoke control zones, and realizing distributed collaborative control of multiple zones. S44 is a hierarchical control authority mechanism that sets four control levels: normal operating conditions, fire development conditions, extreme change conditions, and equipment failure conditions. The permissions, optimization cycles, and constraint boundaries of the global collaboration layer and the regional control layer are adaptively adjusted under different levels to ensure control priority and response speed under extreme conditions.

6. The MPC-based intelligent coordinated regulation and control method for a multi-group smoke exhaust system in a tunnel fire according to claim 1, characterized in that, Step S5 includes: S51, construct the multi-objective hybrid optimization objective function for the partitioned MPC controller, with the following expression: ； In the formula, J is the objective function of rolling optimization, with minimizing J as the optimization objective; ω1 is the weight coefficient of the flue gas risk sub-objective, ω2 is the weight coefficient of the evacuation safety sub-objective, ω3 is the weight coefficient of the equipment energy consumption sub-objective, and ω4 is the weight coefficient of the adjustment smoothness sub-objective. Each weight coefficient is adaptively adjusted according to the control condition level. S52, the construction of each sub-objective item includes: smoke risk sub-target J risk : to predict the sum of squares of smoke risk prediction values of all sensor nodes in the subzone in the prediction time domain, measure the smoke control effect in the subzone Evacuation safety sub-objective J safe : For the prediction time domain, the sum of squares of deviations of temperature, CO concentration, and visibility of the safety monitoring points at the entrance of the evacuation path in this partition from the safety threshold is forced to guarantee the safety of the evacuation path; Device energy consumption sub-target J energy : To predict the sum of the operating power of all fan groups in this partition within the time domain, measure the economy of device operation; Adjusting smoothness sub-target J smooth : For predicting the square sum of the change amount of each control variable at adjacent time steps in the time domain, suppressing the sharp fluctuations of the fan air volume and the exhaust port opening, and avoiding frequent start-stop and damage of equipment; S53, Hard Constraint Setting: The following hard constraints are enforced during the rolling optimization process, without using soft penalties: Equipment rated operating condition constraints: The air volume, speed, and operating power of each fan shall not exceed the upper and lower limits of the equipment rating; Smoke control zone pressure difference constraint: The static pressure difference between adjacent smoke control zones shall not be less than 50Pa to prevent smoke cross-contamination; Evacuation safety threshold constraints: The temperature at the entrance of the evacuation route shall not exceed 60℃, the CO concentration shall not exceed 200ppm, and the visibility shall not be less than 10m; Physical constraints on flue gas flow: The ventilation volume corresponding to the control command is used to avoid flue gas backflow and turbulent diffusion. The physical constraint module of the ST-GNO flue gas spatiotemporal evolution global prediction model is used for verification. S54, the weight coefficient adaptive adjustment mechanism, adaptively adjusts the weight coefficients of each sub-objective based on the fire level, smoke risk coefficient, and personnel evacuation status: when personnel are in the evacuation phase, the weight of ω2 is increased to prioritize evacuation safety; when the fire situation changes suddenly and smoke spreads rapidly, the weight of ω1 is increased to prioritize controlling smoke diffusion; when the fire situation is in a steady state, the weights are balanced to take into account both control effectiveness and equipment economy.

7. The MPC-based intelligent regulation and control method for multi-group smoke exhaust system in tunnel fire according to claim 1, characterized in that, Step S6 includes: S61, an adaptive rolling time-domain optimization module design, adaptively adjusts the prediction time domain H and control time domain U of the MPC based on the fire evolution rate, flue gas risk change rate, and degree of sudden change in operating conditions; wherein, each control time step is set to a first preset time: When the fire is in a stable development stage and the rate of change of flue gas risk is below the threshold, the prediction time domain H is set to 20 control time steps and the control time domain U is set to 5 control time steps, taking into account both optimization accuracy and computational efficiency. When the fire situation changes abruptly, smoke spreads rapidly, and the rate of change of smoke risk exceeds the threshold, the prediction time domain H is shortened to 10 control time steps, and the control time domain U is shortened to 2 control time steps to improve optimization speed and response sensitivity. When an emergency scenario such as personnel evacuation or equipment failure is detected, the prediction time domain H is set to 5 control time steps and the control time domain U is set to 1 control time step to achieve a millisecond-level rapid response. S62, the wind turbine fault detection and identification module, uses the isolated forest algorithm based on real-time data from wind turbine operating status sensors to identify fault types such as wind turbine shutdown, abnormal speed, and insufficient air volume, locates the faulty wind turbine and the fault level, and has a fault detection response time of less than 0.5s. S63, the fan fault adaptive reconfiguration module, immediately updates the adjacency matrix of the spatiotemporal graph topology when a fan fault is detected, removes the faulty node, and the ST-GNO flue gas spatiotemporal evolution global prediction model quickly updates the prediction results; the global collaborative optimization layer readjusts the control objectives of each smoke control zone, and the MPC controller of the faulty zone and the MPC controller of the adjacent zone work together to complete rolling optimization, reallocate the control commands of normal fans, make up for the smoke exhaust capacity gap of the faulty fan, and realize non-disruptive fault-tolerant control under fault conditions; S64, an extreme condition backup control mechanism, automatically switches to a distributed robust control mode based on fixed pressure difference of smoke control zones when the ST-GNO flue gas spatiotemporal evolution global prediction model inference is abnormal, multiple sensors fail, or multiple fans fail, ensuring the basic operating capability of the smoke exhaust system and avoiding control failure.

8. The MPC-based intelligent regulation and control method for a multi-group smoke exhaust system for tunnel fires according to claim 1, characterized in that, Step S7 includes: S71, Pre-training and Parameter Tuning: Based on the full-condition FDS simulation dataset, complete the pre-training of the ST-GNO flue gas spatiotemporal evolution global prediction model, parameter tuning of the two-layer MPC controller, initial value determination of multi-objective optimization weight coefficients, and calibration of hard constraint boundaries. S72, System Initialization: After a tunnel fire occurs, the system automatically starts, loads the pre-trained model and controller parameters, completes self-checks of the sensing system, fan equipment and coupling interfaces, locates the fire source and fire power based on the initial monitoring data, and initializes the spatiotemporal topology. S73, Real-time Data Acquisition and Preprocessing: Through the three-in-one coupling interface, multi-source sensor data and wind turbine operation data of the entire tunnel are acquired according to the preset time step, and data normalization, outlier processing and spatiotemporal alignment are completed, and the time-series node feature matrix is ​​updated. S74, Flue Gas Evolution Prediction and Risk Assessment: The ST-GNO flue gas spatiotemporal evolution global prediction model is based on the historical time series feature matrix, outputs the spatiotemporal distribution of flue gas characteristics in the entire tunnel in the future time domain, and calculates the flue gas risk matrix and evacuation path safety status of the entire tunnel. S75, dual-layer MPC rolling optimization: The global collaborative optimization layer completes the global target optimization and issues control targets to each zone; the MPC controller of each zone synchronously completes the rolling optimization, calculates and outputs the control commands of each fan unit and smoke outlet, and outputs them after hard constraint verification. S76, Control Command Execution and Closed-Loop Iteration: Control commands are synchronously sent to the PLC control cabinet of the FDS numerical model or hardware-in-the-loop system through the coupling interface to complete the adjustment of the fan and smoke exhaust port, and enter the next control time step iteration to realize full closed-loop intelligent control.

Images