A method and system for controlling ventilation in a highway tunnel

Abstract

The application belongs to the technical field of high-speed tunnel ventilation control, and provides a highway tunnel ventilation control method and system. The method comprises the following steps: multi-source heterogeneous data acquisition, multi-source data fusion, state estimation, traffic flow prediction, dynamic pollution index calculation and multi-target weight generation, rolling optimization with constraints, control execution, model updating and feedback correction. Through multi-source data fusion, the state sensing accuracy is improved. Through heterogeneous sensor and radar and visual fusion technology, real-time and accurate estimation of the pollution concentration field and the traffic flow state is realized, the error is reduced, the sensing deviation problem caused by single data in the traditional method is solved, the future traffic flow is predicted based on the graph neural network, the fan power is optimized combined with the model predictive control, the ventilation intensity is adjusted in advance, the pollution exceeds the standard is avoided, the response speed is improved compared with the threshold control, and the energy consumption is reduced.

Description

Technical Field

[0001] This application relates to the field of ventilation control technology for high-speed tunnels, and in particular to a ventilation control method and system for highway tunnels. Background Technology

[0002] Highway tunnel ventilation control is a technology that uses equipment such as fans and dampers to control airflow, pollutant concentration and visibility in the tunnel, so as to ensure driving safety and environmental comfort. Its core is to balance pollutant emissions and ventilation energy consumption in the tunnel, improve air quality in the tunnel in real time, and cope with traffic flow fluctuations and extreme working conditions.

[0003] Existing technologies mainly employ fixed threshold control or simple PID regulation. In the fixed threshold control method, the fan is started when the carbon monoxide concentration or visibility exceeds a preset threshold and shut down when it is below the threshold. In the PID regulation method, the fan power is calculated based on feedback from a single sensor (such as carbon monoxide concentration), without considering the spatial distribution of pollutants and the dynamic impact of traffic flow.

[0004] However, these existing technologies still have some shortcomings, such as: relying on threshold triggering, making it impossible to predict changes in traffic flow, such as peak-hour congestion, resulting in ventilation only starting after pollutants exceed the standard, which leads to a response lag; not combining the spatial distribution of pollution, such as local high-concentration areas, to optimize fan power, often using full-power operation, which leads to high energy consumption; and lacking a gradient smoke exhaust strategy in case of fire or heavy vehicle congestion, relying only on a fixed smoke exhaust mode, resulting in low emergency efficiency and insufficient handling of extreme conditions. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for controlling ventilation in highway tunnels, so as to solve at least one technical problem mentioned in the background art.

[0006] To address the aforementioned technical problems, this application provides a method for controlling ventilation in highway tunnels, employing the following technical solution:

[0007] A method for controlling ventilation in a highway tunnel includes the following steps:

[0008] Environmental data, traffic data, vehicle characteristic data, fire alarm signals and vehicle data inside the tunnel are collected in real time by heterogeneous sensors to form a multi-source raw dataset; the raw dataset is cleaned, spatiotemporally aligned and fused with radar and vision to generate a tunnel grid dataset with a unified spatiotemporal reference.

[0009] Based on extended Kalman filter estimation of pollutant concentration field, traffic flow state and extreme conditions, the current gridded pollution distribution and condition judgment results are output.

[0010] A graph neural network model is used to predict future traffic flow parameters, including traffic density, average vehicle speed, and the proportion of heavy vehicles, and the accuracy of the prediction results is verified.

[0011] The dynamic pollution index of the entire tunnel is obtained by weighted fusion of node pollution indices, and the weights of safety and energy-saving objectives are dynamically allocated through Nash equilibrium.

[0012] Based on model predictive control, wind turbine power is optimized. Under the premise of satisfying state constraints and control constraints, wind turbine power is dynamically adjusted to balance safety and energy saving goals, and rolling optimization is carried out. Emergency control strategies are triggered under extreme conditions.

[0013] The system issues control commands to the fan and uses measured data to correct model parameters and error covariance online, forming a closed-loop feedback.

[0014] To address the aforementioned technical problems, this application also provides a highway tunnel ventilation control system, comprising:

[0015] Data acquisition module: Integrates a heterogeneous sensor array deployed in the tunnel to collect environmental data, traffic data, vehicle characteristic data, fire alarm signals and vehicle data in real time, forming a multi-source raw dataset;

[0016] Data fusion module: Cleans, spatiotemporally aligns, and performs radar-visual fusion processing on the original dataset to generate a tunnel grid dataset with a unified spatiotemporal reference;

[0017] State estimation module: Based on the extended Kalman filter algorithm, it estimates the pollutant concentration field, traffic flow state and extreme conditions in real time, and outputs gridded pollution distribution and condition judgment results;

[0018] Traffic flow prediction module: Employs a graph neural network model to predict traffic density, average vehicle speed, and the proportion of heavy vehicles in future time periods;

[0019] Dynamic pollution index calculation module: calculates the node pollution index and dynamic pollution index, and dynamically allocates the weights of safety and energy-saving targets through Nash equilibrium;

[0020] Optimized control module: Based on the model predictive control framework, the wind turbine power is optimized to minimize the weighted sum objective function of safety and energy saving while satisfying state constraints and control constraints;

[0021] Execution feedback module: issues fan control commands and uses measured data to correct model parameters and error covariance online, forming a closed-loop feedback control.

[0022] The beneficial effects of this invention are as follows:

[0023] This application provides a ventilation control method for highway tunnels. It improves the accuracy of state perception by fusing multi-source data and achieves real-time and accurate estimation of pollutant concentration field and traffic flow state by using heterogeneous sensors (environment, traffic, and vehicle characteristics) and radar-visual fusion technology. This reduces errors and solves the perception bias problem caused by single data in traditional methods.

[0024] This method achieves proactive regulation through a prediction-control closed loop. It predicts future traffic flow (such as congestion trends) based on graph neural networks, and optimizes fan power by combining model predictive control to adjust ventilation intensity in advance, thereby avoiding pollutant exceedances. Compared with threshold control, it improves response speed and reduces energy consumption. It achieves dynamic allocation of weights for safety and energy-saving objectives through dynamic pollution index, and performs rolling optimization based on model predictive control, enabling ventilation strategies to adapt to changes in traffic and pollution in advance, avoiding the lag of traditional control.

[0025] This method quantifies the threat level through a dynamic pollution index, dynamically allocates safety and energy-saving weights using Nash equilibrium, and triggers gradient negative pressure smoke exhaust and fire-fighting linkage under extreme conditions, thereby improving emergency response efficiency. It upgrades the traditional passive response into a three-in-one active defense system of "threat prediction - dynamic weighting - precise handling", overcoming the problems of delayed response and energy efficiency imbalance in tunnel fires and ensuring the safe operation of tunnels. Attached Figure Description

[0026] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a highway tunnel ventilation control method provided in an embodiment of this application;

[0028] Figure 2 This is an exemplary system architecture diagram in which this application can be applied. Detailed Implementation

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0030] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0031] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0032] like Figures 1-2 As shown, a ventilation control method for highway tunnels includes S1 to S7, wherein:

[0033] S1. Multi-source heterogeneous data acquisition: Real-time acquisition of environmental data, traffic data, vehicle characteristic data, fire alarm signals and vehicle data in the tunnel through heterogeneous sensors to form a multi-source raw dataset.

[0034] As another embodiment, the environmental data includes carbon monoxide concentration, visibility, wind speed, wind direction, temperature, and humidity; the traffic data includes traffic flow, average vehicle speed, and the proportion of heavy vehicles; and the vehicle characteristic data includes vehicle location, speed, and vehicle type, collected by millimeter-wave radar and cameras.

[0035] Heterogeneous sensors collect environmental data, traffic data, vehicle characteristic data, fire alarm signals, and vehicle data inside highway tunnels.

[0036] The content and equipment to be collected are as follows:

[0037] 1.1. Environmental data: including carbon monoxide concentration (C). CO Units: ppm, visibility C VI Wind speed (m / s), wind direction (°), temperature (°C), humidity (%).

[0038] Carbon monoxide concentration is collected by a distributed infrared carbon monoxide sensor, visibility is collected by a forward scattering visibility meter, wind speed / direction is collected by an ultrasonic anemometer, and temperature / humidity is collected by an integrated temperature and humidity sensor.

[0039] 1.2. Traffic data: including traffic flow (vehicles / hour), average vehicle speed (km / h), and percentage of heavy vehicles (unit).

[0040] Traffic flow is collected by geomagnetic coil detectors buried in the road surface, average vehicle speed is collected by millimeter-wave radar + AI camera, and the proportion of heavy vehicles is collected by AI video vehicle type recognition system. For example, the vehicle type recognition system uses the YOLOv5 model, and the vehicle type classification accuracy is 98%.

[0041] 1.3. Vehicle characteristic data: including vehicle position, speed, and model. Data is acquired using millimeter-wave radar and a 4K camera. For example, the radar used is a Continental ARS540 model with an accuracy of ±0.1 m / s, and the camera is a DS-2CD4.

[0042] 1.4. Fire alarm signal: including fire temperature signal, in °C. Acquired through distributed fiber optic temperature sensors, for example, using an APSensing sensor, with a temperature measurement accuracy of ±1 °C.

[0043] 1.5. Vehicle-mounted data: including electric vehicle battery temperature, in °C. This data is collected via the vehicle-to-everything (OBU) terminal and transmitted to the vehicle-to-everything (RSU) via the 5G V2X protocol. For example, the electric vehicle battery temperature is collected via the vehicle-to-infrastructure (V2X) unit, using dedicated short-range communication (DSRC) or a 5G network for real-time interaction. If vehicle-mounted data is unavailable, fused data from heterogeneous sensors such as radar and cameras is used as a substitute to ensure system robustness.

[0044] S2. Multi-source data fusion: The original dataset is cleaned, spatiotemporally aligned and fused with radar and vision to generate a tunnel grid dataset with a unified spatiotemporal benchmark.

[0045] As another embodiment, the data cleaning includes removing carbon monoxide concentration, visibility anomalies, and stationary vehicle data; the spatiotemporal alignment maps sensor data to the tunnel control grid based on the inverse distance weighting method; and the radar-visual fusion is achieved through spatiotemporal calibration, feature matching, and anomaly removal.

[0046] The multi-source data fusion method is as follows:

[0047] 2.1. Data cleaning.

[0048] Remove outliers from the collected data:

[0049] For carbon monoxide concentration / visibility data: C under normal traffic conditions CO ≤100ppm, if three consecutive sampling points satisfy C CO If the value is greater than 300 ppm and the traffic flow is 0, it is determined to be a sensor malfunction. This data is then discarded and the malfunction is marked.

[0050] For vehicle feature data: vehicles passing through at a speed less than the threshold of 5 km / h for 10 seconds are identified as stationary vehicles, and stationary vehicle data, such as vehicles that have stopped for emergency purposes, are excluded.

[0051] 2.2. Spatiotemporal alignment.

[0052] When collecting data inside the tunnel, due to differences in collection location and frequency, it is necessary to uniformly map discrete sensor data onto the tunnel control grid.

[0053] Taking a 3km long tunnel as an example, the tunnel is divided into 100 nodes at 30m intervals, covering the entire tunnel. The tunnel control grid refers to dividing the entire tunnel along its length into several continuous unit areas. Each grid represents a fixed-length section of the tunnel, with 100 grids corresponding to the 3km tunnel. Grid nodes are representative points / control points of each grid, usually located at the grid center, numbered from 1 to 100, corresponding to the tunnel entrance to exit.

[0054] Alignment method: Spatiotemporal calibration is performed based on the inverse distance weighted method (IDW). IDW is a classic spatial interpolation algorithm based on the principle that the influence of nearby points is greater than that of distant points. The formula is as follows:

[0055] ;

[0056] Among them, Grid j (t) represents the fused data of the j-th grid node, Raw i (t) represents the raw data from the i-th sensor, d ij Let n be the distance (m) from sensor i to grid j, and n be the number of sensors.

[0057] After spatiotemporal alignment, a tunnel grid dataset with a unified spatiotemporal benchmark is generated. Each grid node contains fused data such as carbon monoxide concentration, visibility, and traffic flow characteristics (volume, speed, and vehicle type ratio) at that location.

[0058] Spatiotemporal alignment provides standardized input for subsequent traffic flow prediction and pollutant concentration field estimation, avoiding model errors caused by spatiotemporal misalignment of data.

[0059] 2.3. Radar-Vision Fusion.

[0060] Radar excels at high-precision ranging and speed measurement, while video excels at visual feature recognition. By fusing vehicle feature data collected by millimeter-wave radar and 4K cameras, a more comprehensive and accurate perception of traffic targets (vehicles) can be achieved.

[0061] Rayvision fusion is achieved through "spatiotemporal calibration + feature matching + anomaly removal", with the specific steps as follows:

[0062] Spatiotemporal calibration: Through timestamp synchronization (ensuring time consistency between radar and video data) and spatial coordinate transformation (mapping the radar's three-dimensional coordinates and the video's two-dimensional pixel coordinates to the same tunnel coordinate system), the radar data and video data of the same vehicle are ensured to correspond.

[0063] Feature matching: The "vehicle position-speed" output by radar is matched with the "vehicle appearance-model" output by video. For example, the same vehicle is determined by the position overlap and the consistency of the motion trajectory. If the position overlap is greater than 80% and the consistency of the motion trajectory is greater than 90%, it is determined to be the same vehicle. After fusion, a complete vehicle feature containing "position, speed and model" is generated.

[0064] Outlier removal: Remove invalid data, such as stationary vehicle data mentioned in step 2.1 data cleaning; filter radar false alarms (such as interference from tunnel wall reflections) and video misidentification (such as misjudgment of vehicle type caused by light and shadow).

[0065] S3. State estimation: Based on extended Kalman filter, estimate pollutant concentration field, traffic flow state and extreme conditions, output the current gridded pollution distribution and condition judgment results. The pollutant diffusion model is used when estimating pollutant concentration field.

[0066] As another embodiment, the pollutant concentration field includes the spatial distribution of carbon monoxide concentration and visibility. The extended Kalman filter prediction includes a prediction stage and a correction stage. The prediction stage derives the state transition equation based on Fick's diffusion law and the pollutant emission model. The correction stage fuses sensor data through the observation equation and outputs an estimated value of the pollutant concentration field.

[0067] The extreme operating conditions include fires involving new energy vehicles, traffic congestion caused by heavy vehicles, and accumulation of debris on steep slopes.

[0068] The pollutant concentration field is estimated using Extended Kalman Filter (EKF), a state estimation algorithm for nonlinear systems. Through a predictive-update iterative process, EKF estimates the true state of the system from sensor data in the presence of noise. In this application, it is used to fuse multi-source data to accurately estimate the spatial distribution, i.e., the concentration field, of carbon monoxide and visibility within the tunnel. The pollutant diffusion model is the core of the EKF state transition process.

[0069] The state vector X(t) is defined as:

[0070] X(t) = [C CO,1 C VI,1 C CO,2 C VI,2 ,…,C CO,N C VI,N ] T ;

[0071] Where N is the number of tunnel grids; C COi C represents the carbon monoxide concentration in the i-th grid, in ppm. VIi Let be the visibility of the i-th grid, in meters (m).

[0072] The EKF estimation process is as follows:

[0073] 3.1. Prediction Stage: Describe the pollutant variation patterns, derive the state transition equation for pollutant diffusion based on Fick's diffusion law and the Emission Model for Pollutants (EMFAC), and use process equations:

[0074] X(t+1)=A(t)·X(t)+B·U(t)+B d ·D(t)+W(t);

[0075] The system state X(t+1) inside the tunnel at the next time step (t+1) is calculated. Where:

[0076] A(t) is the state transition matrix used to describe the natural diffusion of pollutants. Based on Fick's diffusion law, the expression for A(t) is:

[0077] A ij =δ ij -k·Δt / (Δx) 2 δ ij For Kronecker delta, when i=j, δ ij If 1, then δ ij 0, d ij The grid spacing is in meters (m), k = 0.1m. 2 / s is the diffusion coefficient, Δt=1s is the time step, and Δx is the grid spacing in meters.

[0078] B is the control matrix, used to describe the dilution effect of the fan on pollutants. The expression for B is: B mj =0.8·U m (t), U m This represents the power percentage of the m-th fan group. The power percentage is only used to indicate the power level and ranges from 0 to 100%. The direction is indicated separately by the fan's rotation direction or speed. A coefficient of 0.8 indicates the fan efficiency.

[0079] D(t) is the pollutant emission rate, calculated using the EMFAC model, and its expression is D. CO (t), see below for details.

[0080] W(t) is the process noise, which follows a Gaussian distribution with a covariance Q = 0.01I, where I is the identity matrix.

[0081] 3.2. Correction Phase: Correct prediction biases using actual sensor data and fuse model and measured information. This is achieved through the observation equation:

[0082] Z(t) = H∙X(t) + V(t);

[0083] The sensor measurement value Z(t), such as carbon monoxide concentration, is calculated. Where: V(t) is the observation noise; H is the observation matrix, the sensor position mapping, expressed as:

[0084] ;

[0085] 3.3. The EKF iteration process is as follows:

[0086] 3.3.1. Prediction Step:

[0087] The pollutant diffusion model, based on the state estimate from the previous time step, uses process equations to predict the current state and covariance. State prediction formula:

[0088] ;

[0089] in: For prior state estimation, it means the predicted value of the system state at time t using data from time t-1 and earlier. In the tunnel scenario, it is the predicted result of pollutant concentration, traffic flow state, etc. in each grid of the tunnel at time t.

[0090] A(t-1) is the state transition matrix, which describes the transition law of the system state from time t-1 to time t. In tunnel ventilation, it reflects the natural evolution characteristics such as pollutant diffusion and traffic flow changes. For example, it is derived based on Fick's diffusion law and traffic flow model, reflecting the state transmission relationship over time.

[0091] The posterior state estimate is the optimal estimate of the system state after the update step correction at time t-1, which is a precise description of the tunnel state at time t-1 that combines model prediction and sensor observation.

[0092] B is the control matrix, which describes the way and intensity of the influence of control variables (such as tunnel fan power, ventilation strategy, etc.) on the system state. For example, different fan operating power has different degrees of intervention on the diffusion of pollutant concentration, and the matrix elements reflect this relationship.

[0093] U(t-1) is the control input vector, which is the control quantity applied at time t-1, such as the actual operating power of the tunnel ventilation fan, the opening degree of the damper, and other control parameters.

[0094] Covariance prediction formula:

[0095] ;

[0096] in: For prior covariance, measure The greater the covariance of this predicted state, the less reliable the prediction result is, reflecting the error range of state estimation when relying solely on model prediction.

[0097] P(t-1) is the posterior covariance, which is the state estimate after the update step at time t-1. The uncertainty measure is usually reduced after sensor observation correction;

[0098] A(t-1) T Let A(t-1) be the transpose matrix, used for the correct propagation calculation of covariance;

[0099] Q represents the process noise covariance, which indicates the statistical characteristics of errors caused by the uncertainty of the system model itself and unmodeled dynamics (such as sudden small airflow disturbances in the tunnel, random fluctuations in traffic flow, etc.). It is pre-defined based on experience or system identification.

[0100] 3.3.2. Update Steps:

[0101] By using sensor observation data at the current moment, the predicted prior state and covariance are corrected to obtain a more accurate posterior estimate.

[0102] Kalman gain formula:

[0103] ;

[0104] Where: K(t) is the Kalman gain, which determines the weight of the observation data in the state correction process. The larger the gain, the stronger the influence of the observation data on the state update. It is a key coefficient for balancing the reliability of model prediction and sensor observation.

[0105] H is the observation matrix, which establishes the mapping relationship between the system state and sensor observations. It projects the high-dimensional system state (such as the pollutant concentration of multiple grids in a tunnel) to the sensor observation dimension (such as the measured values ​​of a finite number of sensors). For example, if the sensors are deployed in a specific grid, H reflects the correspondence between the grid state and the sensor observations.

[0106] H T This is the transpose of H, used for covariance and observation calculation adaptation;

[0107] R is the observation noise covariance, which reflects the statistical characteristics of the sensor's own measurement error, such as the observation uncertainty caused by sensor accuracy limitations and environmental interference (electromagnetic interference affecting sensor readings, etc.). It is also set in advance according to the sensor characteristics.

[0108] To comprehensively calculate the observation covariance and reflect the combined effect of model prediction uncertainty and sensor observation uncertainty, its inverse operation is used to reasonably allocate the weights of the Kalman gain.

[0109] State correction formula:

[0110] ;

[0111] in, Posterior state estimation is the optimal estimate of the system state at time t, combining prediction results with sensor observations. Compared to... It incorporates observational information, resulting in higher precision and a more accurate reflection of the actual distribution of pollutants and traffic flow conditions in the tunnel;

[0112] Z(t) is the sensor observation vector, which represents the actual measurement value of the sensor at time t, such as the measured data of carbon monoxide concentration sensor, visibility sensor, traffic flow sensor, etc. in the tunnel.

[0113] To observe the residual, which reflects the difference between the sensor observations and the theoretical observations derived from the predicted state, the residual is weighted by the Kalman gain K(t) and used to correct the predicted state and reduce the estimation error.

[0114] Covariance update formula:

[0115] ;

[0116] Where: P(t) is the empirical covariance, and the corrected state estimate. The uncertainty measure, because it incorporates sensor observation information, is generally more uncertain than the prior covariance. A low value reflects the reliability of the updated state estimate;

[0117] I is the identity matrix, with dimensions I... Consistency is used to ensure dimensional compliance in matrix operations and to achieve reasonable updates of covariance.

[0118] The prediction results of EKF are mainly prior state estimates of the prediction step output. and prior covariance :

[0119] It is a prediction of the system's state at a future time (time t). In the tunnel ventilation scenario, it is based on the state at time t-1 and the system model to infer the state quantities such as pollutant concentration and traffic flow parameters (flow rate, vehicle speed, etc.) of each grid in the tunnel at time t. However, this prediction result is not integrated with the sensor observations at the current time (time t), and there is a certain degree of uncertainty.

[0120] This quantifies the degree of uncertainty in the prediction results, such as the size of the covariance matrix and eigenvalues, in order to understand the credibility range of the predicted state in each dimension and provide a basis for the balance model prediction and sensor observation in subsequent update steps.

[0121] Estimation results: Output the estimated carbon monoxide concentration C for each grid cell. est_ coi(t) and visibility estimate C est_VIi (t), forming a complete pollutant concentration field distribution. Application: As a core input for "Dynamic Pollution Index (DPI) calculation" and "fan control optimization", it directly determines the direction of ventilation strategy adjustment (e.g., enhanced ventilation is required in high-concentration areas).

[0122] Extreme operating condition judgment result: Based on the above state estimation and rule judgment, the judgment result of "normal operating condition", "extreme operating condition" or "sensor failure" is output, which provides a basis for subsequent control processes, such as routine optimization, emergency response and fail-safe mode.

[0123] 3.4. Calculation of pollutant emission rate.

[0124] The EMFAC model calculates D(t):

[0125] The EMFAC model is a mobile source emissions model used by the California Air Resources Board (CARB) to estimate pollutant emission rates under different vehicle models and operating conditions. The formula is:

[0126] ;

[0127] Where, ρ pred To predict traffic density, the unit is vehicles / km; v pred Predicted vehicle speed, unit: km / h; EF CO Emission factor, in g / km, is determined with reference to the "Specifications for Ventilation Design of Highway Tunnels" (JTG / T D70 / 2-02) and in conjunction with on-site measured data. For example, EF=2.5 for small vehicles and EF=15 for heavy-duty diesel vehicles.

[0128] 3.5. Determination of extreme working conditions.

[0129] Operating condition type: fire in a new energy vehicle; triggering condition: activation of fire alarm signal, or temperature rise rate > 50℃ / min and visibility < 100m; judgment logic: battery thermal runaway characteristics: sudden temperature rise + rapid smoke accumulation, resulting in a sudden drop in visibility;

[0130] Operating Condition Type: Heavy Vehicle Congestion; Triggering Condition: Radar vision data detects 20 heavy vehicles staying within 1km for more than 5 minutes; Judgment Logic: Congestion leads to concentrated emissions, and carbon monoxide concentration may surge by 30%;

[0131] Operating condition type: longitudinal slope accumulation; triggering condition: road section with longitudinal slope > 2%, wind speed less than preset wind speed threshold; judgment logic: insufficient natural wind leads to the accumulation of pollutants along the slope.

[0132] S4. Traffic Flow Prediction: A graph neural network model is used to predict future traffic flow parameters, including traffic density, average vehicle speed, and the proportion of heavy vehicles, and the accuracy of the prediction results is verified.

[0133] As another embodiment, the graph neural network model input includes radar-visual fusion data, current traffic status, tunnel topology information and historical disturbance data. In the graph structure, the nodes are tunnel control grids, and the node features include vehicle type, average speed and position. The edges are directed connections between nodes, and the weights are determined by the node spacing. A spatiotemporal graph convolutional network is adopted, which aggregates the features of neighboring nodes through spatial convolution and extracts temporal features through temporal convolution.

[0134] Traffic flow prediction is based on real-time traffic data fused from multiple sources and tunnel topology information. It uses a graph neural network (GNN) spatiotemporal propagation model for prediction. The core idea is to treat tunnel traffic flow as a spatiotemporal dynamic graph and capture the spatial correlation and temporal evolution of vehicles through the graph structure.

[0135] GNN is a deep learning model that processes graph-structured data, propagating information through nodes (sensor locations) and edges (spatial associations); in this application, it is used to fuse tunnel topology and vehicle dynamic features to predict future traffic flow changes.

[0136] The inputs to the GNN model include:

[0137] Rayvision-Fusion Data: Real-time vehicle location (coordinates inside the tunnel), speed, and vehicle type are mapped to the tunnel control grid after spatiotemporal calibration.

[0138] Current traffic status data: traffic flow, average vehicle speed, and proportion of heavy vehicles, generated by the fusion of traffic detector and radar-visual data.

[0139] Tunnel topology information: static structural features such as tunnel length, number of lanes (e.g., two-way four lanes), and longitudinal slope (1.5%) are used to construct spatial constraints for traffic flow propagation.

[0140] Historical disturbance data: Inject 20% of abnormal scenario data, such as sudden drop in traffic and accident simulation, to improve the model's ability to predict emergencies.

[0141] 4.1. The steps for building a GNN model are as follows:

[0142] 4.1.1. Definition of Graph Structure:

[0143] Nodes: Tunnel control grid, with a total of 100 nodes, corresponding to different locations in the tunnel;

[0144] The characteristics of each node are:

[0145] NodeFeature v =[Type,v,pos];

[0146] Where: Type represents vehicle type: 0 = small car, 1 = heavy car;

[0147] v represents the average speed of vehicles within the grid, derived from radar speed measurement data, in km / h.

[0148] pos represents the location: grid number 1 to 100, representing the longitudinal position of the tunnel.

[0149] Edge: A directed connection between nodes, with the direction consistent with the direction of vehicle travel. The weight of an edge is determined by the distance between nodes; the closer the nodes are, the greater the weight, reflecting the spatial correlation of vehicle flow.

[0150] Edge features: vehicle movement direction, based on the tunnel's one-way traffic rule, only retaining forward edges.

[0151] 4.1.2. Model Architecture: Spatiotemporal Graph Convolutional Network (ST-GCN) is adopted.

[0152] Spatial convolution, aggregating features from neighboring nodes, formula:

[0153] ;

[0154] Where u and v are nodes in the graph; l is the number of layers in the graph neural network; The output feature of node v in the (l+1)th layer of the network; In the l-th layer network, N(v) represents the output feature of node u; N(v) represents the set of neighboring nodes of node v; N(u) represents the set of neighboring nodes of node u; W (l) σ represents the convolution weights of the l-th layer, and σ is the activation function.

[0155] This formula is the core of the spatial convolutional layer in ST-GCN, used to aggregate spatiotemporal features from neighboring nodes and ultimately output high-level features of node v, such as predicted future traffic density.

[0156] Temporal convolution: 1D-CNN extracts temporal features.

[0157] 4.2. Model training and prediction.

[0158] Training data: Traffic flow data from the past year (including morning and evening rush hours, off-peak hours, severe weather, etc.), with 20% of disturbance data injected (such as simulated traffic drops caused by accidents) to ensure the model's adaptability to abnormal situations.

[0159] Model training: Parameters are optimized through backpropagation, with the goal of minimizing the mean squared error (MSE) between the predicted and actual values, and iterative training continues until convergence.

[0160] Prediction process:

[0161] Based on the constructed spatiotemporal graph, the GNN model extracts spatial features (such as the mutual influence of upstream and downstream vehicles) through "graph convolutional layers" and captures the temporal evolution pattern (such as the inertial propagation of traffic flow) through "temporal cyclic layers", and finally outputs the traffic flow prediction results for the next 5 minutes.

[0162] Prediction formula:

[0163] ;

[0164] in, This represents the traffic flow prediction result, which is an output over a time interval. It covers traffic flow-related data from the next time interval (t+1) after the current time t to the end of the next 5 minutes (i.e., time t+5min). Specific indicators such as traffic density and average vehicle speed will be output later.

[0165] Graph 拓扑 This represents graph data constructed based on the tunnel topology, where different locations in the tunnel (such as nodes divided by control grids) are treated as nodes in the graph, and the vehicle travel relationships between nodes (such as the influence relationship of vehicle flow between upstream and downstream locations) are treated as edges, thus depicting the spatial constraints and relationships of tunnel traffic flow propagation.

[0166] NodeFeature t It is the feature vector of the node (corresponding to the tunnel control grid) in the graph at time t, which contains information such as vehicle type (distinguishing between small cars, heavy vehicles, etc.), speed, and location number of the vehicle in the grid at that time. It is used as the input of the GNN model to capture the traffic flow state at the current time.

[0167] Uses of prediction results:

[0168] Support for pollutant emission rate calculation: Combining the predicted traffic flow status (especially the proportion of heavy vehicles), the future emission rate (D(t)) of pollutants (carbon monoxide, particulate matter) is calculated using the EMFAC model, providing input for the pollutant diffusion model.

[0169] Optimize fan control strategy: Based on predicted traffic flow changes (such as upcoming peak traffic), adjust fan power in advance to avoid the "response lag" of traditional control (such as increasing air volume only after pollutants exceed the standard).

[0170] Extreme Condition Warning: If a sudden drop in traffic flow is predicted (such as a 20% decrease in traffic volume), the accident warning mechanism can be triggered in advance to enhance the response speed to extreme conditions such as congestion and accidents.

[0171] 4.3. Accuracy Verification:

[0172] Verification method: Compare the predicted results with the measured traffic data and calculate the relative error;

[0173] Accuracy requirements: Prediction error ≤ 10%. For example, if the actual traffic flow is 1000 vehicles / h, the predicted value should be within the range of 900 to 1100 vehicles / h.

[0174] Validation results: Based on historical data testing, the model achieves an accuracy of ≥90% in normal scenarios and ≥85% in disturbed scenarios.

[0175] S5. Calculation of dynamic pollution index and generation of multi-objective weights: The dynamic pollution index of the entire tunnel is obtained by weighted fusion of the node pollution indices, and the weights of safety and energy-saving objectives are dynamically allocated through Nash equilibrium.

[0176] As another embodiment, the node pollution index integrates the amount of carbon monoxide exceeding the standard and the degree of visibility deterioration. The dynamic pollution index is obtained by weighted summation of the node pollution indices, with the weights dynamically adjusted according to the degree of pollution. The safety objective is to minimize the maximum amount of carbon monoxide exceeding the standard, and the energy-saving objective is to minimize the total energy consumption of the wind turbine. The weights are dynamically allocated through the dynamic pollution index.

[0177] The distribution of pollutants (carbon monoxide, particulate matter) inside the tunnel is complex. If local congestion causes a sudden increase in concentration, the dynamic pollution index (DPI) is weighted and fused through the node pollution index (NPI) to output a quantitative value of the pollution threat to the entire tunnel.

[0178] 5.1. NPI Calculation: Calculate the pollution level of a single grid cell, taking carbon monoxide and visibility VI as examples. i The formula comprehensively reflects the dual threats of excessive pollutants and deteriorating visibility:

[0179] ;

[0180] in, Let be the node pollution index of the i-th grid at time t, which combines the pollution contributions of carbon monoxide concentration and visibility; C represents the estimated carbon monoxide concentration in the i-th grid at time t, in ppm, derived from the EKF-corrected result; threshold_CO =100ppm, which is the safe threshold for carbon monoxide, that is, the highest allowable concentration of carbon monoxide in the tunnel, used for pollution index calculation; C represents the visibility estimate for the i-th grid at time t, in meters, derived from the EKF-corrected result; threshold_VI =500m, which is the visibility threshold, i.e., the minimum allowable visibility inside the tunnel.

[0181] 5.2. DPI Calculation:

[0182] ;

[0183] Among them, w i(t) represents the weight of the i-th grid at time t.

[0184] ;

[0185] Where j is the grid index, ranging from j=1,2,…,100; 0.1 is the smoothing coefficient; γ is the power exponent hyperparameter, used to adjust the distribution of weights. In the normal mode, γ=1.5, and in the emergency mode, γ=2.0, which enhances the weight of highly polluted areas.

[0186] DPI is used to drive the switching of control targets:

[0187] Low DPI (e.g., <0.5): indicates low pollution threat, prioritize energy conservation, and reduce fan power;

[0188] High DPI (e.g., >1.5): indicates a high risk of pollution, requiring safety to be prioritized, such as full-power ventilation, or even triggering emergency mode.

[0189] 5.3. Nash Equilibrium Multi-Objective Weights:

[0190] 5.3.1. Safety Objectives:

[0191] ;

[0192] Where, max i The maximum value operator (min) calculates the maximum value for all grid i. Its function is to iterate through all grid i and find the area with the highest carbon monoxide concentration exceeding the standard, i.e., the most dangerous polluted area. U The minimization operator indicates the goal of the optimization problem: to find a set of wind turbine power values ​​U. m (t) ensures that the maximum carbon monoxide exceedance is minimized, thus preventing severe exceedances without a grid.

[0193] The safety target formula, through the logic of "finding the maximum excess amount + minimizing the maximum excess amount", ensures that the carbon monoxide concentration in all grids within the tunnel does not exceed the safety threshold. It is a mathematical embodiment of the first principle of safety in tunnel ventilation systems. Its design goal is to control the carbon monoxide concentration to within 100 ppm even in the most polluted grids through fan control.

[0194] 5.3.2. Energy Saving Target:

[0195] ;

[0196] Among them, J energy P represents the total energy consumption of all wind turbines, in kW. m Rated power of a single fan, in kW; m , For the number of wind turbines; U m(t) represents the power percentage of the m-th wind turbine at time t; min is the minimization operator, which indicates the objective of the optimization problem: to find a set of wind turbine power U. m (t), making the total energy consumption J energy Minimum.

[0197] The core logic of the energy-saving target formula is: Total energy consumption = Sum of the actual power of all fans, by minimizing J energy The system can minimize fan energy consumption while meeting pollution control requirements (such as carbon monoxide concentration and visibility).

[0198] 5.3.3. Weight Allocation: Multi-objective weights (λ) safe ,λ energy Priorities are dynamically allocated through Nash equilibrium.

[0199] Weighting formula:

[0200] λ safe (t) = 0.3 + 0.5·DPI(t);

[0201] λ energy (t) = 0.7 - 0.5·DPI(t);

[0202] The higher the DPI, the higher the λ safe The closer DPI is to 1, the higher the priority for safety; the lower the DPI, the higher the λ. energy The closer to 1, the more energy-efficient it becomes. For example, when DPI > 1, λ safe ≥0.8, prioritize safety; when DPI≤1, λ energy ≥0.2, taking energy saving into consideration.

[0203] The weights of safety and energy conservation targets are dynamically allocated through Nash equilibrium. Nash equilibrium is based on the changes in the dynamic pollution index DPI and is adaptively adjusted through a pre-defined linear relationship (see the weight formula in 5.3.3), without the need for complex mathematical proof.

[0204] S6. Constrained Rolling Optimization: Based on model predictive control, wind turbine power is optimized. Under the premise of satisfying state constraints and control constraints, the wind turbine power is dynamically adjusted to balance safety and energy saving objectives, and rolling optimization is performed. Emergency control strategies are triggered under extreme conditions.

[0205] As another embodiment, the objective function of the rolling optimization is the weighted sum of the safety objective and the energy-saving objective. The state constraint is that the predicted carbon monoxide concentration is ≤ the concentration threshold, and the control constraint is that the fan power range is 0 to 100% and the power change rate is ≤ the change threshold. Under extreme conditions, gradient negative pressure smoke exhaust is triggered, the upstream fan at the fire point rotates at full power and the downstream fan rotates in reverse, and the sprinkler system is linked to form a water curtain.

[0206] The conventional mode optimizes objectives and constraints, and the control objective function integrates safety and energy saving:

[0207] Model predictive control (MPC) is a model-based control algorithm that uses a dynamic model of the system to predict the future behavior of the system and determines the current control input by optimizing an objective function.

[0208] MPC (Multi-Purpose Control) is based on predictive models that describe the dynamic behavior of a system. In tunnel ventilation control, these predictive models can be mathematical models built upon pollutant diffusion equations, traffic flow models, and the relationship between fan control and ventilation effectiveness. Using these predictive models, the system's state changes over a future period are predicted based on the current system state (such as pollutant concentration and traffic flow within the tunnel) and potential future control inputs (such as fan power settings).

[0209] 6.1. Rolling Optimization: In each control cycle, MPC optimizes the control input within the prediction time domain based on the current system state and the prediction model. The objective function is set as follows:

[0210] J(t) = λ safe (t)·J safe +λ energy (t)·J energy；

[0211] Where, λ safe and λ energy λ represents the weight that changes with time t; J(t) is the weighted sum of safety and energy saving. By dynamically adjusting the weight λ, a "condition-adaptive" control strategy is achieved.

[0212] J safe For safety objectives, such as concentration penalties for exceeding grid limits;

[0213] ;

[0214] in, The carbon monoxide concentration at future time k in the i-th grid of the tunnel is predicted based on GNN traffic prediction and EKF state estimation.

[0215] This indicates coverage of the prediction time domain (the next 5 minutes), assessing "the accumulation of security risks over a period of time," which better aligns with the needs of MPC rolling optimization.

[0216] J energy For energy-saving targets, such as total fan power penalty:

[0217] ;

[0218] Among them, U m(k) represents the percentage of wind turbine power at future time k, predicting the control input sequence in the time domain;

[0219] This indicates that the forecast time domain is covered, and the "total energy consumption over a period of time" is assessed because MPC needs to optimize the control sequence (the wind turbine power plan for the next 5 minutes), rather than the power at a single moment.

[0220] It should be noted that in steps 5.3.1 and 5.3.2, ·J safe and J energy The formula corresponds to a simplified description of the control logic, where ·J safe and J energy The formula is the complete form of MPC rolling optimization.

[0221] Optimization logic: Um(t) represents the percentage of fan power, with a value range of 0 to 100%; the fan direction is controlled by independent parameters, with forward rotation indicating air supply and reverse rotation indicating air exhaust. By adjusting the control input Um(t), J(t) is minimized to achieve a dynamic balance between "safety and energy saving".

[0222] It should be noted that MPC only executes the optimized control input for the current moment. In the next control cycle, it will re-predict and optimize based on the new system state, continuously rolling the process, i.e., rolling optimization.

[0223] 6.2. Constraints:

[0224] State constraints: C pred_CO,i The concentration threshold is ≤110ppm. This threshold is based on a 100ppm safety threshold and allows for a 10% fluctuation to ensure control flexibility, but the ultimate goal is to keep the concentration below 100ppm. The pollution index is calculated using a strict safety threshold of 100ppm, and the control constraint uses an operational threshold of 110ppm to deal with instantaneous fluctuations.

[0225] Control constraints: Fan power range 0 ≤ U m ≤100%, power change rate |ΔU m | ≤ 20% / s change threshold to avoid frequent start-stop of the fan.

[0226] 6.3. Emergency Control for Extreme Operating Conditions:

[0227] When extreme conditions such as fires are triggered (DPI approaches infinity):

[0228] Forced weight switching: λ safe =1,λ energy =0, safety is the absolute priority;

[0229] Control strategy switching: Trigger gradient negative pressure smoke exhaust: Upstream fan at fire point rotates at full power (Um =100%), the downstream fan reverses, and the speed is set to 50% of the rated speed;

[0230] Firefighting coordination: The sprinkler system is activated upstream of the fire point, with a water curtain thickness of 5cm, increasing smoke blocking efficiency by 60%.

[0231] S7. Control Execution, Model Update and Feedback Correction: Issue fan control commands and use measured data to correct model parameters and error covariance online, forming a closed-loop feedback.

[0232] As another embodiment, the pollutant diffusion model corrects the diffusion coefficient using measured data, the graph neural network model is retrained weekly using newly added vehicle trajectory data, and the extended Kalman filter error covariance is updated using observation residuals.

[0233] The control execution, model update, and feedback correction methods are as follows:

[0234] 7.1 Control Execution: The fan power command is sent to the fan controller via the PLC, with a response delay of ≤1s.

[0235] 7.2. Model Update:

[0236] Pollutant diffusion model: The diffusion coefficient k is corrected using measured data after control. For example, if the diffusion velocity is 20% slower than predicted under full power of the fan, then k is corrected to 0.8 times the original value.

[0237] GNN model: Retrained weekly with 100,000 new vehicle trajectory data points, iterated 1000 times, maintaining prediction accuracy ≥90%.

[0238] 7.3. Parameter Correction: EKF covariance corrected based on observation residuals.

[0239] Observation residual calculation:

[0240] ;

[0241] Where Residual is the observation residual, representing the difference between the actual sensor observation and the theoretical observation based on the model prediction; Z(t+1) is the sensor's measured data at time t+1; and H is the observation matrix, representing the system state. Mapped to the sensor observation dimension; Let t be the prior state estimate at time t.

[0242] EKF covariance update:

[0243] ;

[0244] Where: P(t+1) is the posterior covariance at time t+1, which is the uncertainty measure of the state estimate after fusing observations. The smaller the covariance, the more reliable the estimate. K(t+1) is the prior covariance at time t; K(t+1) is the Kalman gain at time t+1, which balances the weights of model prediction and sensor observation. The larger the gain, the stronger the influence of the observation.

[0245] Based on the same inventive concept as the highway tunnel ventilation control method provided in the embodiments of this application, the embodiments of this application also provide a highway tunnel ventilation control system. If there is anything unclear about the content of the system embodiment, please refer to the corresponding content in the method embodiment.

[0246] A ventilation control system for highway tunnels, comprising:

[0247] Data acquisition module: Integrates a heterogeneous sensor array deployed in the tunnel to collect environmental data, traffic data, vehicle characteristic data, fire alarm signals and vehicle data in real time, forming a multi-source raw dataset;

[0248] Data fusion module: Cleans, spatiotemporally aligns, and performs radar-visual fusion processing on the original dataset to generate a tunnel grid dataset with a unified spatiotemporal reference;

[0249] State estimation module: Based on the extended Kalman filter algorithm, it estimates the pollutant concentration field, traffic flow state and extreme conditions in real time, and outputs gridded pollution distribution and condition judgment results;

[0250] Traffic flow prediction module: Employs a graph neural network model to predict traffic density, average vehicle speed, and the proportion of heavy vehicles in future time periods;

[0251] Dynamic pollution index calculation module: calculates the node pollution index and dynamic pollution index, and dynamically allocates the weights of safety and energy-saving targets through Nash equilibrium;

[0252] Optimized control module: Based on the model predictive control framework, the wind turbine power is optimized to minimize the weighted sum objective function of safety and energy saving while satisfying state constraints and control constraints;

[0253] Execution feedback module: issues fan control commands and uses measured data to correct model parameters and error covariance online, forming a closed-loop feedback control.

[0254] As another embodiment, in the data acquisition module, the heterogeneous sensor array includes a carbon monoxide / visibility detector, an anemometer, a camera, a radar, and a vehicle-to-everything (V2X) unit. The vehicle data is exchanged in real time through at least one of Dedicated Short Range Communication (DSRC) or a 5G network.

[0255] In the data fusion module, the cleaning process removes outliers and missing values, the spatiotemporal alignment uses linear interpolation to unify the sampling frequency, and the radar-visual fusion integrates radar point clouds and visual features through a target association algorithm.

[0256] In the state estimation module, the input of the extended Kalman filter is the fused grid dataset, and the output includes pollutant concentration field, traffic flow state parameters, and judgment indicators for extreme conditions such as fire / congestion.

[0257] In the traffic flow prediction module, the graph neural network uses the tunnel topology as graph nodes and historical traffic flow data as input features to predict traffic parameters for the next 5 to 15 minutes.

[0258] In the dynamic pollution index calculation module, the node pollution index is calculated based on the ratio of pollutant concentration to safety threshold. The dynamic pollution index is generated through spatiotemporal weighted aggregation, and the Nash equilibrium dynamically adjusts the weights according to the real-time threat level.

[0259] In the optimization control module, the state constraints include the upper limit of pollutant concentration and the range of fan power, the control constraint is the fan start-stop frequency, and the objective function is the weighted sum of safety and energy saving.

[0260] In the execution feedback module, the model correction uses the recursive least squares method to update parameters, and the error covariance is adjusted online by the residual between the measured value and the predicted value.

[0261] Specifically, the highway tunnel ventilation control system achieves closed-loop intelligent control through multi-module collaboration:

[0262] First, the data acquisition module captures environmental parameters (such as carbon monoxide concentration and visibility), traffic flow data (vehicle volume and speed), vehicle characteristics (proportion of heavy vehicles) and fire alarm signals in the tunnel in real time, forming a raw dataset and transmitting it to the data fusion module.

[0263] The data fusion module cleans, aligns, and integrates the data with radar and visual data to generate a unified gridded dataset, which is then synchronously output to the state estimation module and the traffic flow prediction module.

[0264] The state estimation module analyzes the pollutant diffusion trend and traffic conditions in real time based on grid data, outputs pollutant concentration distribution map, traffic condition judgment results and extreme condition warning signals, and transmits them to the dynamic pollution index calculation module.

[0265] The dynamic pollution index calculation module combines pollutant concentration and traffic density to calculate the dynamic pollution index, quantifies safety risks, and dynamically allocates safety and energy-saving weights. The results are then sent to the optimization control module. Simultaneously, the traffic flow prediction module uses grid data to predict future traffic flow trends and outputs congestion risk signals to the optimization control module.

[0266] The optimized control module integrates dynamic pollution index, traffic flow prediction results and fan operation constraints, and generates fan control commands (including start / stop, power and wind direction of jet fans and emergency fans) through collaborative decision-making. The commands are then sent to the execution feedback module to drive the fan operation.

[0267] The execution feedback module collects actual ventilation status and environmental change data after simultaneous execution, compares it with the predicted values, outputs correction parameters, and sends them back to the state estimation module and traffic flow prediction module to form a closed-loop optimization. The entire process improves perception accuracy through multi-source data fusion, and combines forward-looking traffic prediction with dynamic risk weight allocation to achieve adaptive adjustment of ventilation intensity, reducing energy consumption while ensuring safety.

[0268] Based on the same inventive concept as the highway tunnel ventilation control method provided in the embodiments of this application, the embodiments of this application also provide a computing device cluster, which includes at least one computing device, each computing device including a processor and a memory; the processor of the at least one computing device is used to execute instructions stored in the memory of the at least one computing device, so that the computing device cluster executes the aforementioned highway tunnel ventilation control method. The computing device cluster can form a distributed network system, and the highway tunnel ventilation control method provided in this application may not be a node and / or a central server in the distributed network system.

[0269] Based on the same inventive concept as the highway tunnel ventilation control method provided in the embodiments of this application, the embodiments of this application also provide a computer-readable storage medium storing computer instructions for execution by a computer to implement the aforementioned highway tunnel ventilation control method.

[0270] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage medium described above can be referred to the corresponding process in the foregoing method embodiments, and therefore will not be repeated here.

[0271] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method of controlling ventilation in a highway tunnel, characterized in that The method comprises the following steps: S1, collecting environmental data, traffic data, vehicle characteristic data, fire alarm signals and vehicle-mounted data in the tunnel in real time through heterogeneous sensors to form a multi-source original data set; S2, performing data cleaning, time-space alignment and radar-visual fusion on the original data set to generate a tunnel grid data set with a unified time-space reference; S3, estimating the pollutant concentration field, traffic flow state and extreme working condition based on an extended Kalman filter, outputting the current grid pollutant distribution and working condition determination result, and using a pollutant diffusion model when estimating the pollutant concentration field; S4, predicting future traffic flow parameters including traffic density, average vehicle speed and heavy vehicle proportion using a graph neural network model, and verifying the accuracy of the prediction result; S5, obtaining the full-tunnel dynamic pollution index by weighting fusion of node pollution indexes, and dynamically distributing the weights of safety and energy-saving targets through Nash equilibrium; S6, optimizing the fan power based on model predictive control, dynamically adjusting the fan power to balance the safety and energy-saving targets under the premise of meeting state constraints and control constraints, performing rolling optimization, and triggering emergency control strategies in extreme working conditions; S7, issuing fan control instructions, and using measured data to correct model parameters and error covariance online to form a closed-loop feedback.

2. The method of claim 1, wherein, The environmental data includes carbon monoxide concentration, visibility, wind speed, wind direction, temperature and humidity; the traffic data includes traffic flow, average vehicle speed and heavy vehicle proportion; the vehicle characteristic data includes vehicle position, speed and vehicle type, which are collected by millimeter wave radar and camera.

3. The method of claim 1, wherein, The data cleaning includes removing carbon monoxide concentration, visibility outliers and stationary vehicle data; the time-space alignment maps sensor data to the tunnel control grid based on the inverse distance weighting method; the radar-visual fusion is realized through time-space calibration, feature matching and outlier removal.

4. The method of claim 1, wherein, The pollutant concentration field includes the spatial distribution of carbon monoxide concentration and visibility, and the extended Kalman filter prediction includes a prediction stage and a correction stage, the prediction stage derives a state transition equation based on Fick's diffusion law and pollutant emission model, and the correction stage fuses sensor data through an observation equation to output the pollutant concentration field estimate; the extreme working condition includes new energy vehicle fire, heavy vehicle congestion and longitudinal slope accumulation.

5. The method of claim 1, wherein, The input of the graph neural network model includes radar-visual fusion data, current traffic state, tunnel topology information and historical disturbance data, in the graph structure, the node is the tunnel control grid, the node feature includes the vehicle type, average speed and position, the edge is the directed connection between nodes, and the weight is determined by the node distance, a spatio-temporal graph convolution network is used, the spatial convolution aggregates the adjacent node features, and the time convolution extracts the time sequence features.

6. The method of claim 1, wherein, The node pollution index integrates the carbon monoxide concentration exceeding amount and the visibility deterioration degree, the dynamic pollution index is obtained by weighted summation of the node pollution indexes, and the weight is dynamically adjusted according to the pollution degree; The safety target is to minimize the maximum carbon monoxide concentration exceeding amount, the energy-saving target is to minimize the total energy consumption of the fan, and the weight is dynamically distributed by the dynamic pollution index.

7. The method of claim 1, wherein, The objective function of the rolling optimization is the weighted sum of safety and energy saving, the state constraint is that the predicted value of carbon monoxide concentration is less than or equal to the concentration threshold, and the control constraint is that the fan power range is 0-100% and the power change rate is less than or equal to the change threshold; under extreme conditions, trigger gradient negative pressure smoke exhaust, fan full power positive rotation upstream of fire point, fan reverse downstream, and at the same time, spray system forms water curtain.

8. The method of claim 1, wherein, The pollutant diffusion model corrects the diffusion coefficient through measured data, the graph neural network model is retrained every week with new vehicle trajectory data, and the extended Kalman filter error covariance is updated through observation residuals.

9. A system for the control of the ventilation of a highway tunnel according to the method of any one of claims 1 to 8, characterised in that, It comprises: a data acquisition module: a heterogeneous sensor array integrated and deployed in the tunnel, which acquires environmental data, traffic data, vehicle feature data, fire alarm signals and vehicle data in real time to form a multi-source raw data set; a data fusion module: cleaning, spatio-temporal alignment and radar-visual fusion processing of the raw data set to generate a tunnel grid data set with a unified spatio-temporal reference; a state estimation module: based on the extended Kalman filter algorithm, real-time estimation of pollutant concentration field, traffic flow state and extreme conditions, output of grid-based pollution distribution and condition determination results; a traffic flow prediction module: using a graph neural network model to predict traffic density, average speed and heavy vehicle proportion in the future period; a dynamic pollution index calculation module: calculation of node pollution index and dynamic pollution index, and dynamic allocation of weights of safety and energy saving through Nash equilibrium; an optimal control module: based on the model predictive control framework, optimization of fan power to minimize the weighted sum objective function of safety and energy saving under the condition of meeting state constraints and control constraints; an execution feedback module: issuing fan control instructions and using measured data to correct model parameters and error covariance online to form a closed-loop feedback control.

10. The system of claim 9, wherein, In the data acquisition module, the heterogeneous sensor array includes carbon monoxide / visibility detectors, anemometers, cameras, radars and vehicle-road cooperation units, and vehicle data is interacted in real time through at least one of dedicated short-range communication or 5G network; In the data fusion module, the cleaning process removes outliers and missing values, the spatio-temporal alignment uses linear interpolation method to unify the sampling frequency, and the radar-visual fusion integrates radar point cloud and visual features through target association algorithm; In the state estimation module, the input of the extended Kalman filter is the fused grid data set, and the output includes pollutant concentration field, traffic flow state parameters and determination flags of extreme conditions such as fire / congestion; In the traffic flow prediction module, the graph neural network takes the tunnel topology as the graph node and the historical traffic flow data as the input feature to predict future traffic parameters; In the dynamic pollution index calculation module, the node pollution index is calculated based on the ratio of pollutant concentration to safety threshold, the dynamic pollution index is generated by spatio-temporal weighted aggregation, and the Nash equilibrium dynamically adjusts the weight according to the real-time threat level; In the optimal control module, the state constraint includes the upper limit of pollutant concentration and the range of fan power, the control constraint is the fan start-stop frequency, and the objective function is the weighted sum of safety and energy saving. In the feedback executing module, the model correction adopts the recursive least square method to update parameters, and the error covariance is adjusted on line through the residual of the measured value and the predicted value.

Images