Energy-saving intelligent control method for coupling of tunnel ventilation and drainage system

Abstract

This invention discloses a coupled energy-saving intelligent control method for tunnel ventilation and drainage systems, belonging to the field of tunnel engineering. Addressing the high energy consumption problem in existing technologies where ventilation and drainage systems are controlled independently and the coupling mechanism of airflow-water vapor phase change-surrounding rock seepage is neglected, this method constructs and dynamically calibrates a digital twin model comprising a static geometric layer, a parameter model layer, and a physical constraint layer. Based on the digital twin model, a coupled state prediction model is constructed, including a spatiotemporal graph neural network module, a physical information neural network module, a coupling interaction layer, and a prediction output layer, used to predict ventilation and drainage load trends, latent heat of phase change terms, and seepage replenishment terms. Finally, a collaborative optimization control model is constructed to solve for collaborative control commands. This method can be used for collaborative energy-saving control of ventilation and drainage systems in tunnel operation.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology. Specifically, it relates to an energy-saving intelligent control method for coupled tunnel ventilation and drainage systems. Background Technology

[0002] As a crucial component of transportation infrastructure, tunnel operation safety and energy consumption control have always been key concerns in the industry. Tunnel ventilation systems primarily dilute pollutants such as carbon monoxide emitted by vehicles and ensure visibility and temperature / humidity levels within the tunnel, making them one of the main sources of energy consumption in tunnel operation. Tunnel drainage systems, on the other hand, remove seepage from surrounding rock and runoff from the road surface, ensuring driving safety and structural stability. For a long time, tunnel ventilation and drainage systems have been considered independent subsystems in design and operation, managed by different control logics and maintenance systems, lacking effective information exchange and coordination mechanisms between them.

[0003] In existing tunnel operation control technologies, ventilation systems typically rely on feedback control based on environmental indicators such as traffic flow, carbon monoxide concentration, and visibility. Methods include threshold triggering, fuzzy control, or lookup tables based on offline simulation results to adjust fan speed and valve opening. Drainage systems primarily depend on sump water level thresholds to control pump start-up and shutdown, employing simple upper and lower limit water level control strategies. While this independent control model meets basic safety requirements at the individual system level, it neglects the physical coupling between ventilation and drainage. Water evaporation within the tunnel alters the air humidity field, increasing the ventilation and dehumidification load; changes in ventilation airflow affect the water evaporation rate, altering the inflow load to the drainage system; and seepage from the surrounding rock, as background inflow to the drainage system, directly impacts drainage energy consumption through its spatiotemporal distribution characteristics, while also indirectly affecting ventilation demand through the humidity field. The multi-physical coupling mechanism of wind, humidity, and seepage has not been effectively modeled in existing technologies, leading to mutual constraints and energy consumption offsetting between the two systems during actual operation.

[0004] In existing technologies, a few studies have attempted to achieve unified management of ventilation and drainage data through data acquisition and centralized monitoring platforms. However, these studies essentially remain at the level of data aggregation and have not established predictive models reflecting the coupling mechanism of the two systems, let alone achieved coordinated optimization control. For example, some tunnel monitoring systems display real-time ventilation and drainage data on the same interface, allowing manual judgment based on experience as to whether coordinated adjustments are needed; or they attempt to treat ventilation and drainage as multi-objective optimization problems, solving them separately and then superimposing them. However, this approach ignores the bidirectional impact of coupling effects and makes it difficult to guarantee global optimality. Furthermore, the application of existing digital twin technology in the tunnel field is mostly concentrated on structural health monitoring or single-system operation simulation. A unified digital twin framework capable of simultaneously integrating static geometric information, dynamic evolution laws, and physical conservation constraints has not yet been formed, and digital twins and real-time control have not been integrated into a closed-loop iteration.

[0005] In summary, existing tunnel ventilation and drainage control technologies suffer from the following shortcomings: First, they lack accurate modeling of the multi-physics coupling mechanism of airflow, water vapor phase change, and surrounding rock seepage, making it impossible to quantify the mutual influence between the two systems in real time. Second, existing prediction models struggle to simultaneously handle the rapid dynamic changes of the ventilation system and the slow-changing background characteristics of the drainage system, resulting in insufficient accuracy in cross-timescale joint predictions. Third, under the dual pressure of safety constraints and energy efficiency targets, there is a lack of collaborative optimization control methods capable of achieving global energy optimization for both ventilation and drainage, often requiring the sacrifice of energy efficiency in one system to meet the safety requirements of the other. Therefore, how to achieve collaborative prediction and optimization control of the two systems while considering their physical coupling mechanism is a pressing technical challenge in this field. Summary of the Invention

[0006] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0007] Another objective of this invention is to provide a coupled energy-saving intelligent control method for tunnel ventilation and drainage systems. By constructing a digital twin model driven by physical coupling, it enables collaborative prediction and optimization control of the ventilation and drainage systems, breaking the current fragmented state of the systems, effectively reducing the overall energy consumption of tunnel operation, and achieving optimal global energy efficiency.

[0008] To achieve these objectives and other advantages according to the present invention, a coupled energy-saving intelligent control method for a tunnel ventilation and drainage system is provided, comprising: S1. Construct a digital twin model of the tunnel environment, the digital twin model including a static geometry layer, a parameter model layer and a physical constraint layer; inject real-time acquired multi-source heterogeneous sensing data into the digital twin model, and dynamically calibrate the real-time state variables in the digital twin model; S2. Based on the digital twin model, construct a coupled state prediction model, and under the unified framework of the digital twin model, predict the ventilation load change trend and drainage load change trend, phase change latent heat term and seepage supply term within the future preset time domain; S3. Based on the digital twin model, the ventilation load change trend and drainage load change trend predicted in step S2, the latent heat of phase change term and seepage supply term, construct a collaborative optimization control model, use a multi-objective optimization algorithm to solve the ventilation and drainage collaborative optimization control model, generate collaborative control commands for the ventilation system and drainage system in the current operating condition and future prediction time domain, and issue them for execution. The coupling state prediction model mentioned in step S2 includes: The spatiotemporal graph neural network module is used to extract the evolution law of each element in the ventilation system topology, drainage system topology and real-time state vector over time based on the digital twin model, and output the ventilation system state feature vector and drainage system state feature vector. The physical information neural network module uses the physical constraint layer of the digital twin model as the constraint condition, and the ventilation system state feature vector and drainage system state feature vector output by the spatiotemporal graph neural network module as the input boundary conditions, and outputs the latent heat of phase change term and seepage supply term. A coupling interaction layer connects the spatiotemporal graph neural network module and the physical information neural network module. It inputs the ventilation system state feature vector and the drainage system state feature vector as boundary conditions into the physical information neural network module, and feeds back the latent heat of phase change term and seepage supply term to the spatiotemporal graph neural network module to correct the temporal update of the ventilation system state feature vector and the drainage system state feature vector. The prediction output layer decodes and maps the ventilation system state feature vector and drainage system state feature vector after the coupling interaction layer correction, transforming them into the ventilation load change trend and drainage load change trend in the future preset time domain, and simultaneously outputs the phase change latent heat term and seepage supply term.

[0009] Preferably, the static geometric layer in step S1 includes at least the tunnel geometry, the ventilation system topology, and the drainage system topology; the parameter model layer includes at least the evolution law model of the dynamic parameters of traffic flow in the tunnel and the spatiotemporal distribution model of the seepage field in the surrounding rock; and the physical constraint layer includes at least the coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage and the corresponding boundary conditions and initial conditions.

[0010] Preferably, the multi-source heterogeneous sensing data in step S1 includes: Traffic flow data, including traffic volume, vehicle speed, and vehicle type composition within the tunnel; Ventilation environment data, including wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system; Drainage system data, including water levels in sump pits at each node of the drainage system, operating status of drainage pumps, and drainage flow rate; Surrounding rock seepage data, including pore water pressure in the surrounding rock; The calibrated real-time state vector includes at least the following: wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system; water level in the sump and drainage flow rate at each node of the drainage system; and pore water pressure in the surrounding rock.

[0011] Preferably, the dynamic calibration of the real-time state variables in the digital twin model described in step S1 specifically includes: Traffic flow data, ventilation environment data, drainage system data, and surrounding rock seepage data from real-time acquired multi-source heterogeneous sensing data are mapped to the corresponding ventilation system nodes and drainage system nodes in the digital twin model, which serve as the observation vectors for the Kalman filter algorithm. The Kalman filter algorithm is used to linearize the evolution model of the dynamic parameters of traffic flow in the tunnel in the parameter model layer, and then use it as the state transition matrix to perform iterative optimization of the real-time state vector in the digital twin model in two steps: prediction and update. Prediction steps: Based on the real-time state vector and state transition matrix calibrated at the previous time step, predict the predicted state vector and its covariance matrix at the current time step. The predicted state vector is the estimated state value at the current time step without being corrected by the observed data. Update steps: The multi-source heterogeneous sensing data mapped to the digital twin model at the current moment is used as the observation vector. Combined with the predicted state vector and its covariance matrix obtained in the prediction step, the Kalman gain is calculated, and the predicted state vector is weighted and corrected to obtain the calibrated real-time state vector. The real-time state vector includes wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system, as well as water level in the sump, drainage flow rate, and pore water pressure in the surrounding rock at each node of the drainage system. The spatiotemporal distribution model of the surrounding rock seepage field serves as a slowly varying background field. It does not participate in the state transition during the Kalman filter prediction step. It is only used to verify and correct the state estimates of the drainage system nodes after calibration. The correction of its model parameters is updated over a long period of time by the feedback results after the execution of the collaborative control command in step S3.

[0012] Preferably, the prediction of the future ventilation load change trend and drainage load change trend, phase change latent heat term and seepage supply term within a preset time domain in step S2 specifically includes: Step S21: Input the real-time state vector obtained after dynamic calibration in step S1 into the spatiotemporal graph neural network module. The spatiotemporal graph neural network module constructs a heterogeneous spatiotemporal graph based on the topology of the ventilation system and the topology of the drainage system in the static geometric layer. Each element in the real-time state vector is used as the initial feature of the heterogeneous spatiotemporal graph node. The spatial topological features of the ventilation system and the drainage system are extracted through the graph convolutional network. The temporal evolution law of the state variables of each heterogeneous spatiotemporal graph node with time is extracted through the gated recurrent unit. The ventilation system state feature vector and the drainage system state feature vector are output. S22. Input the ventilation system state feature vector and the drainage system feature vector into the physical information neural network module. The physical information neural network module uses the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations in the physical constraint layer as physical constraints, the surrounding rock seepage field spatiotemporal distribution model in the parameter model layer as initial seepage boundary conditions, and the ventilation system state feature vector as the boundary conditions of the energy equation and phase change equation. The module solves the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations through an automatic differentiation mechanism and outputs the phase change latent heat term and seepage supply term. S23. Achieve bidirectional information interaction between the spatiotemporal graph neural network module and the physical information neural network module through a coupling interaction layer, including: inputting the state feature vector of the ventilation system and the state feature vector of the drainage system as boundary conditions into the physical information neural network module, and feeding back the latent heat of phase change and seepage supply terms output by the physical information neural network module to the spatiotemporal graph neural network module to correct the state feature vector update of the spatiotemporal graph neural network module in the next time step, so that the state prediction of the ventilation system and the drainage system evolves in a coordinated manner under the physical coupling mechanism; S24. Input the ventilation system state feature vector and drainage system state feature vector after the coupling interaction layer correction into the prediction output layer; the prediction output layer decodes and maps the high-dimensional feature vector into a time-series prediction result with physical meaning through a fully connected network, outputs the ventilation load change trend and drainage load change trend in the future preset time domain, and simultaneously outputs the prediction sequence of the phase change latent heat term and seepage supply term.

[0013] Preferably, the coupled partial differential equation set of air flow-water vapor phase change-surrounding rock seepage includes the mass conservation equation, momentum conservation equation, energy conservation equation, carbon monoxide transport equation, water vapor phase change equation, and Darcy flow equation. The latent heat of phase transition is calculated from the phase transition source term in the energy conservation equation, where the energy conservation equation is: ; Phase transition latent heat term Q latent for: ; L is the latent heat of vaporization of water, 2.5 × 10⁻⁶. 6 J / kg; The mass flow rate of water vapor phase change per unit volume per unit time; kg / (m³) 3 •s); described by the water vapor phase transition equation: ; Or engineering model: ; k c is the mass transfer coefficient, with a value ranging from 0.001 to 0.01; A is the water-air contact area; C vapor C represents the concentration of water vapor in the air. sat (T) represents the saturated water vapor concentration at the current temperature; RH represents the relative humidity; A' represents the contact area between water and air per unit volume of tunnel. The seepage supply term Q seepage The seepage flow rate at the tunnel lining boundary is obtained by integrating: ; denoted as the contact surface between the tunnel lining and the surrounding rock; n is the normal vector of the contact surface. q is the seepage velocity, in m / s, which is described by Darcy's seepage equation: ; c p ρ1 is the specific heat capacity of air at constant pressure; J / (kg·K); ρ1 is the density of air; kg / m³ 3 ; T is the air temperature, K; u is the air velocity vector, m / s; k is the thermal conductivity of air, W / (m·K); K is the permeability of the surrounding rock, m 2 μ is the dynamic viscosity of water, Pa·s; p is the pore water pressure in Pa; ρ2 is the density of water; z is the position head.

[0014] Preferably, the collaborative optimization control model in step S3 uses the ventilation system topology and drainage coefficient topology provided by the digital twin model as the spatial constraint boundary of the optimization problem; uses the airflow-water vapor phase change-surrounding rock seepage coupled partial differential equations provided by the digital twin model as the derivation basis for the safety limits and physical constraints of equipment operation; uses the ventilation load change trend and drainage load change trend predicted in step S2 as the dynamic input of the optimization problem; uses the latent heat of phase change term and seepage supply term as coupling constraint parameters; and aims to minimize the overall energy consumption of the tunnel operation. The latent heat of phase change term is used to quantify the influence of the drainage system's water accumulation state on the thermal and moisture load of the ventilation system, and the seepage supply term is used to quantify the rigid demand of surrounding rock seepage on the inflow load of the drainage system.

[0015] Preferably, the collaborative control commands in step S3 include combinations of fan speeds and valve openings in the ventilation system; pump start / stop sequences and pump frequency combinations in the drainage system; and the feedback results after the collaborative control commands are issued and executed are synchronously updated to the parameter model layer of the digital twin model to correct the evolution law model of traffic flow dynamic parameters in the tunnel and the spatiotemporal distribution model of the surrounding rock seepage field.

[0016] The present invention has at least the following beneficial effects: Firstly, this invention constructs a digital twin model that includes a static geometry layer, a parametric model layer, and a physical constraint layer, integrating the ventilation system and the drainage system into a unified description framework. This breaks the technical bias of separating the two systems in traditional control, providing a foundation for information consistency for subsequent coupled prediction and collaborative optimization, and effectively improving the integrity and maintainability of the tunnel operation control system. Secondly, this invention introduces the evolution law model of traffic flow dynamic parameters in the tunnel and the spatiotemporal distribution model of the seepage field in the surrounding rock into the parameter model layer, and embeds the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equation set into the physical constraint layer, realizing explicit modeling of the multi-physical field coupling mechanism of wind-humidity-seepage, providing a computable mathematical expression for accurately quantifying the physical interaction relationship between ventilation and drainage, and overcoming the technical limitation of traditional decoupling models that cannot describe the coupled dynamics. Thirdly, this invention uses the Kalman filter algorithm to dynamically calibrate real-time sensing data and digital twin models, and synchronously updates the feedback results after the execution of cooperative control commands to the parameter model layer, forming a closed-loop iterative mechanism of model-prediction-optimization-execution-update. This enables the digital twin model to continuously track the time-varying characteristics of the actual working conditions of the tunnel, effectively suppressing the impact of model drift on control accuracy. Fourth, this invention adopts a bidirectional coupled prediction architecture that combines spatiotemporal graph neural networks and physical information neural networks. Through the coupling interaction layer, it realizes the closed-loop feedback of ventilation state characteristics with latent heat of phase change and seepage supply. The latent heat of phase change and seepage supply are used as coupling constraint parameters of the collaborative optimization control model, so that the optimization decision can be directly based on the physical coupling relationship between the two systems to weigh energy consumption, providing a feasible technical path for reducing the overall energy consumption of tunnel operation.

[0017] Other advantages, objectives and features of the present invention will be apparent in part from the following description, and in part from the understanding of those skilled in the art through study and practice of the invention. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the energy-saving intelligent control method for coupled tunnel ventilation and drainage systems described in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0020] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0021] like Figure 1 As shown, the present invention provides an energy-saving intelligent control method for coupled tunnel ventilation and drainage systems, comprising: S1. Construct a digital twin model of the tunnel environment, the digital twin model including a static geometry layer, a parameter model layer and a physical constraint layer; inject real-time acquired multi-source heterogeneous sensing data into the digital twin model, and dynamically calibrate the real-time state variables in the digital twin model; S2. Based on the digital twin model, construct a coupled state prediction model, and under the unified framework of the digital twin model, predict the ventilation load change trend and drainage load change trend, phase change latent heat term and seepage supply term within the future preset time domain; S3. Based on the digital twin model, the ventilation load change trend and drainage load change trend predicted in step S2, the latent heat of phase change term and seepage supply term, construct a collaborative optimization control model, use a multi-objective optimization algorithm to solve the ventilation and drainage collaborative optimization control model, generate collaborative control commands for the ventilation system and drainage system in the current operating condition and future prediction time domain, and issue them for execution. The coupling state prediction model mentioned in step S2 includes: The spatiotemporal graph neural network module is used to extract the evolution law of each element in the ventilation system topology, drainage system topology and real-time state vector over time based on the digital twin model, and output the ventilation system state feature vector and drainage system state feature vector. The physical information neural network module uses the physical constraint layer of the digital twin model as the constraint condition, and the ventilation system state feature vector and drainage system state feature vector output by the spatiotemporal graph neural network module as the input boundary conditions, and outputs the latent heat of phase change term and seepage supply term. A coupling interaction layer connects the spatiotemporal graph neural network module and the physical information neural network module. It inputs the ventilation system state feature vector and the drainage system state feature vector as boundary conditions into the physical information neural network module, and feeds back the latent heat of phase change term and seepage supply term to the spatiotemporal graph neural network module to correct the temporal update of the ventilation system state feature vector and the drainage system state feature vector. The prediction output layer decodes and maps the ventilation system state feature vector and drainage system state feature vector after the coupling interaction layer correction, transforming them into the ventilation load change trend and drainage load change trend in the future preset time domain, and simultaneously outputs the phase change latent heat term and seepage supply term.

[0022] The specific method for achieving coupled energy-saving intelligent control of the tunnel ventilation and drainage system using the above technical solution is as follows: A digital twin model is constructed, which includes a static geometric layer (tunnel geometry, ventilation and drainage system topology), a parameter model layer (traffic flow evolution law, spatiotemporal distribution of surrounding rock seepage), and a physical constraint layer (coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage). Real-time multi-source heterogeneous sensing data (such as wind speed, water level, pore water pressure, etc.) is used to dynamically calibrate the real-time state variables through Kalman filtering, thereby solving the technical problem in the existing technology that the digital twin model is disconnected from real-time control and cannot accurately reflect the current state of the system. Based on this, a coupled state prediction model is constructed using this digital twin model: its spatiotemporal graph neural network module constructs a heterogeneous spatiotemporal graph based on the topological structure in the static geometric layer, uses each element in the calibrated real-time state vector as the initial feature of the graph node, extracts spatial topological features through a graph convolutional network, and extracts temporal evolution laws through gated recurrent units, outputting the state feature vector of the ventilation and drainage system, thus solving the problem of cross-timescale prediction that traditional models cannot simultaneously capture rapid traffic flow changes and slow-varying seepage backgrounds; the physical information neural network module uses the coupled partial differential equations in the physical constraint layer as constraints and the state feature vector as boundary conditions, and solves the problem through an automatic differentiation mechanism. Solving the system of equations and outputting the latent heat of phase change (quantifying the impact of water accumulation on ventilation heat and humidity load) and the seepage recharge (quantifying the rigid demand of surrounding rock seepage on drainage inflow) solves the problem of inaccurate modeling of strong coupling mechanisms of multi-physics fields. The coupling interaction layer realizes bidirectional feedback, inputting the state feature vector into the physical information neural network, and feeding back the latent heat of phase change and seepage recharge to the spatiotemporal graph neural network to correct the temporal update of the state feature vector, so that the ventilation and drainage system can evolve together under the physical coupling mechanism. Finally, the prediction output layer decodes and maps the corrected state feature vector into a prediction sequence of ventilation load, drainage load, latent heat of phase change, and seepage recharge in the future time domain. Furthermore, based on the digital twin model and prediction results, a collaborative optimization control model is constructed with the goal of minimizing global energy consumption while taking into account environmental and water level safety constraints. The collaborative control commands such as fan speed, valve opening, pump start-stop sequence and frequency combination are obtained through multi-objective optimization algorithm and then issued for execution. The execution feedback is synchronously updated to the parameter model layer to form a closed-loop iteration, thereby breaking through the bottleneck of independent control of two systems and mutual cancellation of energy consumption in the existing technology.

[0023] The above technical solution, through the deep integration of digital twin and hybrid-driven coupled state prediction model, has for the first time achieved real-time quantification and collaborative prediction of the multi-physics field coupling mechanism of wind-humidity-seepage in tunnel ventilation and drainage control, effectively improving the decision bias problem caused by model fragmentation. By utilizing the bidirectional interaction of spatiotemporal graph neural network and physical information neural network, it can simultaneously adapt to the rapid dynamics of traffic flow and the slow-changing characteristics of surrounding rock seepage, improving the reliability and accuracy of cross-timescale joint prediction. By introducing the latent heat of phase change and seepage supply as coupling constraint parameters into the collaborative optimization model, the optimization decision can quantify the impact of drainage water accumulation on ventilation energy consumption and the rigid demand of seepage on drainage load, continuously reducing the overall energy consumption of tunnel operation while meeting safety constraints.

[0024] In one of the technical solutions, the static geometric layer in step S1 includes at least the tunnel geometry, the ventilation system topology, and the drainage system topology; the parameter model layer includes at least the evolution law model of the dynamic parameters of traffic flow in the tunnel and the spatiotemporal distribution model of the seepage field in the surrounding rock; and the physical constraint layer includes at least the coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage and the corresponding boundary conditions and initial conditions.

[0025] In the above technical solution, the evolution law model of the dynamic parameters of traffic flow in the tunnel in the parameter model layer is obtained in the following way: historical traffic flow data in the tunnel is collected, including traffic flow, vehicle speed, vehicle type composition and timestamps; the historical traffic flow data is used to train and establish the evolution law model of the dynamic parameters of traffic flow. The evolution law model of the dynamic parameters of traffic flow is used to describe the dynamic characteristics of traffic flow parameters changing over time and serves as the basis for constructing the state transition matrix in the Kalman filter algorithm. Specifically, the process involves: collecting historical traffic flow data within the tunnel, including timestamps, traffic volume, vehicle speed, and vehicle type composition, with a sampling period of 1-5 minutes; training the historical traffic flow data using gated recurrent units or long short-term memory networks to establish a time-series prediction model for traffic flow dynamic parameters. The model's input is traffic flow data from the past N time steps, and its output is traffic flow data from the future M time steps. The structure of the time-series prediction model includes: an input layer for receiving normalized historical traffic flow data; a hidden layer containing at least two layers of gated recurrent units, each containing 64-256 hidden units, used to extract the time-series features of the traffic flow data; and an output layer, a fully connected layer, used to map the hidden layer features to the predicted traffic flow parameters. After training, the time-series prediction model is used as an evolutionary model of the traffic flow dynamic parameters within the tunnel, describing the dynamic characteristics of traffic flow parameters changing over time, and serving as the basis for constructing the state transition matrix in the Kalman filter algorithm.

[0026] In the above technical solution, the relationship between the static geometry layer, parameter model layer, physical constraint layer, and real-time state vector of the digital twin model is as follows: The static geometry layer provides spatial structure information that does not change with time and state, including tunnel geometric dimensions, ventilation system duct connection relationships, and drainage system pipe network connection relationships, serving as a fixed foundation for the spatiotemporal graph neural network module to construct a heterogeneous spatiotemporal graph; The parameter model layer provides a priori models describing the system evolution law, including the evolution law model of traffic flow dynamic parameters within the tunnel and the spatiotemporal distribution model of the surrounding rock seepage field. The evolution law model serves as the state transition prior for the Kalman filter algorithm, and the surrounding rock seepage field... The empty distribution model serves as the initial seepage boundary condition for the physical information neural network module; the physical constraint layer provides a set of partial differential equations describing the physical conservation laws of the system, including the mass conservation equation, momentum conservation equation, energy conservation equation, carbon monoxide transport equation, water vapor phase change equation, and Darcy seepage equation, which serve as physical constraints for the physical information neural network module; the real-time state vector is a dynamic variable independent of the three-layer structure, composed of the state variables of each node in the ventilation system and each node in the drainage system, and interacts with the information in the three-layer structure through the dynamic calibration in step S1, serving as the initial condition for the coupled state prediction model in step S2.

[0027] In the above technical solution, the spatiotemporal distribution model of the surrounding rock seepage field is obtained through the following method: Geological survey data of the surrounding rock along the tunnel route was collected, including rock type, fracture distribution, and permeability range; hydrological monitoring data was collected, including historical data on groundwater level, rainfall, and pore water pressure; a numerical simulation model of the seepage field of the surrounding rock was established, and the Darcy equation was solved using the finite element method or finite difference method to obtain the spatial distribution law of pore water pressure in the surrounding rock. The Darcy equation is as follows: ; The spatiotemporal distribution model of the surrounding rock seepage field is stored in the form of a three-dimensional grid. Each grid node contains spatial coordinates (x, y, z) and the corresponding permeability, initial hydraulic head h0, and pore water pressure p, which serve as the initial seepage boundary conditions for the physical information neural network module. K x K y and K z The anisotropic hydraulic conductivity coefficient is determined based on the surrounding rock type and fracture development direction provided in the geological survey report, and is taken as 10. -8 ~10 -5 m / s; S s The water storage coefficient is determined based on the compressibility and porosity of the surrounding rock, and is taken as 10. -7 ~10 -3 m -1 ; h is the total head, m, h = p / ρ²g + z.

[0028] By combining real-time monitoring data of pore water pressure in the surrounding rock, ensemble Kalman filtering or particle filtering algorithms are used to invert and correct the permeability parameters of the numerical simulation model, thus obtaining a spatiotemporal distribution model of the seepage field in the surrounding rock.

[0029] In the above technical solution, the spatiotemporal distribution model of the surrounding rock seepage field is updated long-term using a Bayesian update algorithm. The update cycle is set according to the hydrogeological change rate of the surrounding rock, and its parameter correction is driven by the cumulative feedback results after the execution of the collaborative control command in step S3. The evolution law model of the dynamic parameters of traffic flow in the tunnel is updated online using the recursive least squares method. The above technical solution provides a clear and comprehensive unified foundation for the coupled control of the ventilation and drainage system by clearly defining the three-layer structure of the digital twin model—the static geometric layer (tunnel geometry, ventilation and drainage system topology), the parameter model layer (dynamic evolution law of traffic flow, spatiotemporal distribution of surrounding rock seepage), and the physical constraint layer (coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage and their boundary / initial conditions). Among them, the static geometry layer ensures that the subsequent spatiotemporal graph neural network can construct heterogeneous spatiotemporal graphs based on accurate spatial topology. The parameter model layer provides updatable prior knowledge for dynamic calibration of Kalman filter and initial boundary of physical information neural network. The physical constraint layer embeds the multi-physics coupling mechanism into the prediction model in the form of partial differential equations, enabling the digital twin model to simultaneously support rapid state calibration, cross-timescale coupling prediction and explicit expression of physical constraints within the same framework. This effectively avoids control decision bias caused by information dispersion or model fragmentation, laying a structurally consistent and mechanistically transparent model foundation for collaborative optimization control.

[0030] In one of the technical solutions, the multi-source heterogeneous sensing data mentioned in step S1 includes: Traffic flow data, including traffic volume, vehicle speed, and vehicle type composition within the tunnel; Ventilation environment data, including wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system; Drainage system data, including water levels in sump pits at each node of the drainage system, operating status of drainage pumps, and drainage flow rate; Surrounding rock seepage data, including pore water pressure in the surrounding rock; The calibrated real-time state vector includes at least the following: wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system; water level in the sump and drainage flow rate at each node of the drainage system; and pore water pressure in the surrounding rock.

[0031] The aforementioned technical solution clearly defines four components of multi-source heterogeneous sensing data: traffic flow data (flow rate, vehicle speed, vehicle type composition), ventilation environment data (wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node), drainage system data (water level in each node's catchment well, drainage pump status, and drainage flow rate), and surrounding rock seepage data (pore water pressure). It also provides the specific physical quantities of the calibrated real-time state vector, offering a clearly structured and physically meaningful data foundation for the dynamic calibration of the digital twin model and the coupled state prediction model. This also enables Kalman filtering to accurately map sensing data to corresponding nodes as observation vectors, while ensuring that the spatiotemporal graph neural network uses a unified and complete sequence of node features for topological feature extraction and temporal evolution law learning. This avoids inconsistencies in model input caused by missing data types or ambiguous correspondences, thereby improving the accuracy of state estimation and the reliability of the prediction model.

[0032] In one of the technical solutions, the dynamic calibration of the real-time state variables in the digital twin model in step S1 specifically includes: Traffic flow data, ventilation environment data, drainage system data, and surrounding rock seepage data from real-time acquired multi-source heterogeneous sensing data are mapped to the corresponding ventilation system nodes and drainage system nodes in the digital twin model, which serve as the observation vectors for the Kalman filter algorithm. The Kalman filter algorithm is used to linearize the evolution model of the dynamic parameters of traffic flow in the tunnel in the parameter model layer, and then use it as the state transition matrix to perform iterative optimization of the real-time state vector in the digital twin model in two steps: prediction and update. Prediction steps: Based on the real-time state vector and state transition matrix calibrated at the previous time step, predict the predicted state vector and its covariance matrix at the current time step. The predicted state vector is the estimated state value at the current time step without being corrected by the observed data. Update steps: The multi-source heterogeneous sensing data mapped to the digital twin model at the current moment is used as the observation vector. Combined with the predicted state vector and its covariance matrix obtained in the prediction step, the Kalman gain is calculated, and the predicted state vector is weighted and corrected to obtain the calibrated real-time state vector. The real-time state vector includes wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system, as well as water level in the sump, drainage flow rate, and pore water pressure in the surrounding rock at each node of the drainage system. The spatiotemporal distribution model of the surrounding rock seepage field serves as a slowly varying background field. It does not participate in the state transition during the Kalman filter prediction step. It is only used to verify and correct the state estimates of the drainage system nodes after calibration. The correction of its model parameters is updated over a long period of time by the feedback results after the execution of the collaborative control command in step S3.

[0033] In the above technical solution, the linearization of the evolution law model of the dynamic parameters of traffic flow in the tunnel in the parameter model layer, after which it is used as the state transition matrix, specifically includes: representing the evolution law model of the dynamic parameters of traffic flow in the tunnel in discrete-time state-space form: X k =f(X k-1 )+W k-1 , where X k Let W be the real-time state vector of k out of control, f be the nonlinear state transition function, and W be the nonlinear state transition function. k-1 To address process noise, a first-order Taylor expansion is used to linearize the nonlinear function f: , F is the real-time state vector calibrated at the previous time step. k-1 Jacobian matrix Its elements are: The Jacobian matrix F k-1 As the state transition matrix in the Kalman filter algorithm, it realizes the linearization approximation of the nonlinear traffic flow evolution model within the Kalman filter framework.

[0034] The aforementioned technical solution maps multi-source heterogeneous sensing data into Kalman-filtered observation vectors and linearizes the state transition matrix using a model of the evolution of traffic flow dynamic parameters, achieving recursive optimal estimation of real-time state vectors. Simultaneously, the spatiotemporal distribution model of the surrounding rock seepage field is used as a slowly varying background field, only employed after Kalman filter calibration to verify and correct the state estimates of drainage system nodes. Its model parameters are updated by long-term feedback rather than participating in high-frequency state transitions. This differentiated processing mechanism effectively avoids model jitter and overfitting issues caused by frequently correcting slowly varying seepage parameters with rapidly changing observation data. It maintains the long-term stability of the parameter model layer while ensuring the accuracy of state estimation, providing a reliable state benchmark for subsequent coupled prediction and collaborative optimization.

[0035] In one of the technical solutions, the prediction of the future ventilation load change trend and drainage load change trend, phase change latent heat term and seepage supply term within a preset time domain in step S2 specifically includes: Step S21: Input the real-time state vector obtained after dynamic calibration in step S1 into the spatiotemporal graph neural network module. The spatiotemporal graph neural network module constructs a heterogeneous spatiotemporal graph based on the topology of the ventilation system and the topology of the drainage system in the static geometric layer. Each element in the real-time state vector is used as the initial feature of the heterogeneous spatiotemporal graph node. The spatial topological features of the ventilation system and the drainage system are extracted through the graph convolutional network. The temporal evolution law of the state variables of each heterogeneous spatiotemporal graph node with time is extracted through the gated recurrent unit. The ventilation system state feature vector and the drainage system state feature vector are output. S22. Input the ventilation system state feature vector and the drainage system feature vector into the physical information neural network module. The physical information neural network module uses the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations in the physical constraint layer as physical constraints, the surrounding rock seepage field spatiotemporal distribution model in the parameter model layer as initial seepage boundary conditions, and the ventilation system state feature vector as the boundary conditions of the energy equation and phase change equation. The module solves the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations through an automatic differentiation mechanism and outputs the phase change latent heat term and seepage supply term. S23. Achieving bidirectional information interaction between the spatiotemporal graph neural network module and the physical information neural network module through a coupling interaction layer includes: inputting the state feature vector of the ventilation system and the state feature vector of the drainage system as boundary conditions into the physical information neural network module, and feeding back the latent heat of phase change and seepage replenishment terms output by the physical information neural network module to the spatiotemporal graph neural network module to correct the state feature vector update of the spatiotemporal graph neural network module in the next time step. This enables the state prediction of the ventilation system and the drainage system to evolve collaboratively under a physical coupling mechanism; S24. Input the ventilation system state feature vector and drainage system state feature vector after the coupling interaction layer correction into the prediction output layer; the prediction output layer decodes and maps the high-dimensional feature vector into a time-series prediction result with physical meaning through a fully connected network, outputs the ventilation load change trend and drainage load change trend in the future preset time domain, and simultaneously outputs the prediction sequence of the phase change latent heat term and seepage supply term.

[0036] In the above technical solution, step S22, which involves solving the coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage using an automatic differentiation mechanism, specifically includes: The total loss function L for constructing the physical information neural network module total : . λ PDE , λ BC , λ IC and λ data These are the weighting coefficients for the partial differential equation residual loss term, boundary condition loss term, initial condition loss term, and data fitting loss term, respectively, all with a value of 10. -2 ~10 2 .

[0037] Among them, L PDE This is the residual loss term in the partial differential equation. M is the number of partial differential equations, which is taken as 6, and N m Let u be the residual operator for the m-th partial differential equation. i This represents the network output, and θ represents the network parameters.

[0038] L BC For boundary condition loss terms, N BC This represents the number of boundary condition sampling points, with a value of 10. 2 ~10 4 u(x) j ,t j ) represents the network sampling points at the boundary (x) j ,t j The output value at (); u BC (x j ,t j () sets the boundary conditions. Boundary condition types include inlet boundary: given wind speed u=u in (t) or given pressure p=p in ;Exit boundary: Given pressure p=p out Wall boundary: no-slip condition u=0; Lining boundary: given pore water pressure p=p rock (x,y,z,t) or a given seepage rate.

[0039] L IC For the initial condition loss term, N IC This represents the number of initial condition sampling points, with a value of 10. 2 ~10 4 u(x) k (t0) represents the network at initial time t0 and spatial point x. k Output value at u; IC (x k () represents the initial condition setting value, and the source of the initial condition is: u IC (x k The real-time state vector is directly taken from the Kalman filter calibration in step S1 to ensure that the prediction model is consistent with the current system state.

[0040] L data For data fitting loss term, N data This represents the number of sensor measurement points, with a value of 10. 2 ~10 4 u(x) l ,t l ) represents the network at sensor location x l Measurement time t l Output value at u; obs (x l ,t l The value represents the actual measurement from the sensor. Sensor measurements need to be normalized before input, typically using Min-Max normalization or Z-Score normalization.

[0041] By minimizing the total loss function using the backpropagation algorithm and gradient descent optimizer, the network parameters of the physical information neural network module are trained, enabling the network to fit the dynamic characteristics of the actual system while satisfying the physical constraints of the coupled partial differential equations. After training, the state feature vectors of the ventilation system and the drainage system are input into the trained physical information neural network module, and the spatial and temporal distributions of the latent heat of phase change and the seepage replenishment term are directly calculated and output through forward propagation of the network.

[0042] In the above technical solution, the spatiotemporal graph neural network module (ST-GNN) includes a graph convolutional network (GCN) for extracting spatial topological features and a gated recurrent unit (GRU) or LSTM for extracting temporal evolution patterns. A coupling bias term is added to the GRU update equation of each node in the ST-GNN, that is, the standard form of the GRU update equation is modified to h. t =h t base +W fb ·c t-1 c t-1 The feedback vector is composed of the physical quantities from the previous time step. W fb For the feedback weight matrix, map the feedback vector to h. t Same dimensions.

[0043] The aforementioned technical solution achieves coupled state prediction of ventilation and drainage systems within a unified digital twin framework by sequentially inputting the dynamically calibrated real-time state vector into a spatiotemporal graph neural network (constructing a heterogeneous spatiotemporal graph and extracting the temporal evolution laws of spatial topology and gated cyclic units), a physical information neural network (solving latent heat of phase change and seepage supply terms with coupled partial differential equations as constraints), a coupled interaction layer (bidirectional feedback to correct the temporal update of state feature vectors), and a prediction output layer (decoding and mapping to prediction sequences of ventilation load, drainage load, and coupling parameters). This not only utilizes graph convolutional networks and gated cyclic units to handle spatial topology and multi-timescale dynamics respectively, overcoming the shortcomings of traditional models that struggle to simultaneously adapt to rapid changes in traffic flow and slow-changing characteristics of surrounding rock seepage, but also enables the state prediction of ventilation and drainage systems to co-evolve under a physical coupling mechanism through bidirectional information interaction in the coupled interaction layer. This avoids prediction bias caused by system fragmentation and provides accurate coupled parameter inputs, including latent heat of phase change and seepage supply, for subsequent collaborative optimization control, thereby improving the effectiveness and reliability of global energy consumption optimization.

[0044] In one of the technical solutions, the coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage include the mass conservation equation, momentum conservation equation, energy conservation equation, carbon monoxide transport equation, water vapor phase change equation, and Darcy flow equation. The latent heat of phase transition is calculated from the phase transition source term in the energy conservation equation, where the energy conservation equation is: ; Phase transition latent heat term Q latent for: ; L is the latent heat of vaporization of water, 2.5 × 10⁻⁶. 6 J / kg; The mass flow rate of water vapor phase change per unit volume per unit time; kg / (m³) 3 •s); described by the water vapor phase transition equation: ; ; k c is the mass transfer coefficient, with a value ranging from 0.001 to 0.01; A is the water-air contact area; C vapor C represents the concentration of water vapor in the air. sat (T) represents the saturated water vapor concentration at the current temperature; RH represents the relative humidity; A' represents the contact area between water and air per unit volume within the tunnel; The seepage supply term Q seepage The seepage flow rate at the tunnel lining boundary is obtained by integrating: ; denoted as the contact surface between the tunnel lining and the surrounding rock; n is the normal vector of the contact surface. q is the seepage velocity, in m / s, which is described by Darcy's seepage equation: ; c p ρ1 is the specific heat capacity of air at constant pressure; J / (kg·K); ρ1 is the density of air; kg / m³ 3 ; T is the air temperature, K; u is the air velocity vector, m / s; k is the thermal conductivity of air, W / (m·K); K is the permeability of the surrounding rock, m 2 μ is the dynamic viscosity of water, Pa·s; p is the pore water pressure in Pa; ρ2 is the density of water; z is the position head.

[0045] In the above technical solution, the mass conservation equation is: t represents the time variable in the control period, in seconds. The momentum conservation equation is: p1 is air pressure, Pa; μ1 is aerodynamic viscosity, Pa·s; QUOTE For the Laplace term of velocity, m / s 2 , representing viscous diffusion, calculated through numerical difference; f is the volume force vector, N / m 3 This mainly includes gravity. The carbon monoxide transport equation is: C represents the carbon monoxide concentration, mg / m³. 3 D is the diffusion coefficient of carbon monoxide in air, with a value of 1.6 × 10⁻⁶. -5 m 2 / s;S C For pollutant source items, mg / (m 3 ·s), determined by traffic flow data, N i For the number of car models, E i This represents the CO emission factor for this vehicle model.

[0046] The above technical solution clearly defines the coupled partial differential equations of air flow, water vapor phase change, and surrounding rock seepage (including mass conservation, momentum conservation, energy conservation, carbon monoxide transport, water vapor phase change, and Darcy's seepage equation), and provides specific mathematical expressions and key parameters (such as latent heat of vaporization L and mass transfer coefficient k) for the latent heat of phase change and seepage recharge terms. c The physical meaning and typical values ​​of parameters such as permeability K of the surrounding rock provide a source of calculable physical constraints and coupling parameters for the physical information neural network module. This enables the quantitative expression of heat and moisture exchange and seepage replenishment mechanisms between ventilation and drainage systems, avoiding the problems of ambiguous coupling relationships or arbitrary setting of empirical parameters caused by the lack of explicit physical equations in traditional methods. This enhances the physical consistency and interpretability of the prediction model and provides coupling constraint parameters with clear physical basis for energy consumption trade-off decisions in collaborative optimization control.

[0047] In one of the technical solutions, the collaborative optimization control model in step S3 uses the ventilation system topology and drainage coefficient topology provided by the digital twin model as the spatial constraint boundary of the optimization problem; uses the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations provided by the digital twin model as the derivation basis for the safety limits and physical constraints of equipment operation; uses the ventilation load change trend and drainage load change trend predicted in step S2 as the dynamic input of the optimization problem; uses the latent heat of phase change term and seepage supply term as coupling constraint parameters; and aims to minimize the overall operating energy consumption of the tunnel; wherein, the latent heat of phase change term is used to quantify the influence of the water accumulation state of the drainage system on the heat and moisture load of the ventilation system, and the seepage supply term is used to quantify the rigid demand of surrounding rock seepage on the inflow load of the drainage system.

[0048] In the above technical solution, the specific objective function of the collaborative optimization control model in step S3 is: Pvent (t) represents the energy consumption of the ventilation system at time t, P drain (t) represents the energy consumption of the drainage system at time t, where T is the prediction time domain. Constraints include environmental indicators: C CO (i,t)≤C CO max VI(i,t)≥VI min RH(i,t)≤RH max (H(j,t)), where i is the node number of the ventilation system, C CO max VI is the safe upper limit for carbon monoxide concentration. min As the lower limit of visibility safety, RH max (H) is a function that dynamically adjusts the upper limit of relative humidity based on the water level in the collection well. ; The maximum permissible relative humidity is 95%, and RH0 is the upper limit, which can be 80%; Rg is the sensitivity coefficient, with a value of 0.5~5.0% relative humidity / meter water level; H threashold The water level threshold for initiating regulation; water level safety constraint: H(j,t) ≤ H j max , j is the node number of the drainage system, H j max Q is the upper limit of the safe water level in the collection well. inflow (t) represents the inflow of road surface water due to traffic flow, Q seepage (t) represents the seepage recharge term, Q pump (t) represents the pump's discharge flow rate; equipment physical constraints: w k min ≤w k (t)≤w k max θ k min ≤θ k (t)≤θ k max f l min ≤f l (t)≤f l max N l on,min ≤N l (t)≤N l max w k For the fan speed, θ k f represents the valve opening. l N is the pump frequency. l For the number of pump start-stop cycles; Coupling constraint: The latent heat term Q of the phase change is...latent (t) and seepage supply term Q seepage (t) is used as a known input parameter to correct the energy trade-off between ventilation load and drainage load.

[0049] The above technical solution constructs a collaborative optimization control model with the goal of minimizing global energy consumption by using the static geometric layer (ventilation and drainage system topology) of the digital twin model as the optimization space boundary, the physical constraint layer (coupled partial differential equations) as the derivation basis for safety limits and physical constraints, and the load trend predicted in step S2 as the dynamic input. It also introduces latent heat of phase change and seepage replenishment terms as coupling constraint parameters. This allows the optimization decision to quantify the impact of drainage system water accumulation on ventilation heat and humidity loads and the rigid demand of surrounding rock seepage on drainage inflow loads. Under the premise of meeting environmental indicators and water level safety, it achieves a synergistic trade-off between ventilation and drainage energy consumption, avoiding the problems of mutual energy consumption cancellation or excessively conservative safety margins caused by neglecting coupling relationships in traditional independent control. This effectively improves the overall energy efficiency of tunnel operation.

[0050] In one of the technical solutions, the collaborative control commands in step S3 include combinations of fan speeds and valve openings in the ventilation system; pump start-stop sequences and pump frequency combinations in the drainage system; and the feedback results after the collaborative control commands are issued and executed are synchronously updated to the parameter model layer of the digital twin model, which is used to correct the evolution law model of traffic flow dynamic parameters in the tunnel and the spatiotemporal distribution model of the surrounding rock seepage field.

[0051] In the above technical solution, the collaborative control command is generated in a rolling manner through a model predictive control framework, specifically including: setting the prediction time domain T. p (Values ​​range from 30 to 120 minutes) and control time domain T c (Values ​​range from 10 to 30 minutes), where T c ≤T p The control period is Δt (values ​​range from 1 to 5 minutes); in each control period t k : Obtain the real-time state vector x(t) at the current moment after calibration in step S1. k ); Call the coupled state prediction model of step S2, with x(t) k Given initial conditions, predict the future T. p The trends of ventilation load variation, drainage load variation, latent heat of phase change, and seepage recharge term in the time domain are analyzed. The collaborative optimization control model from step S3 is invoked, using the prediction results as input, to solve the optimization problem and obtain the control time domain T. c The control command sequence within {u ∗ (t k ),u ∗ (t k+1 ),…,u∗ (t k+Tc Only the first instruction in the control instruction sequence is executed. ∗ (t k ), discard subsequent instructions; in the next control cycle t k+1 Repeat the above process; the feedback result is synchronously updated to the parameter model layer of the digital twin model, specifically including: updating the feedback result x after execution. feedback (t k ) and the QUOTE predicted in step S2 Compare the results and calculate the prediction deviation e(t). k ), The recursive least squares method or Bayesian update algorithm is used to correct the model parameters of the evolution law model of the dynamic parameters of traffic flow in the tunnel and the spatiotemporal distribution model of the seepage field in the surrounding rock in the parameter model layer according to the prediction deviation. θ old and θ new These are the model parameter vectors before and after the update, including the traffic flow evolution model parameters and the surrounding rock seepage field model parameters, and K. k An adaptive gain matrix is ​​calculated based on historical data to gradually converge the deviation between the digital twin model and the actual physical system.

[0052] The aforementioned technical solution concretizes the collaborative control commands into combinations of fan speed and valve opening in the ventilation system, and pump start-stop sequences and frequency combinations in the drainage system. The feedback results after execution are synchronously updated to the parameter model layer of the digital twin model. This is used to correct the evolution model of traffic flow dynamic parameters within the tunnel and the spatiotemporal distribution model of the surrounding rock seepage field. This closed-loop feedback mechanism enables the parameter model layer to continuously correct its prior knowledge based on actual operating data, gradually reducing the deviation between the digital twin model and the physical system. This enhances the model's adaptability to changes in traffic flow patterns, seasonal fluctuations in surrounding rock seepage, and equipment performance degradation. It provides more accurate prior input for state calibration and coupling prediction in subsequent control cycles, which is beneficial for steadily improving the effectiveness and energy-saving benefits of collaborative optimization control in long-term operation.

[0053] The number of devices and processing scale described herein are for simplification of the invention. Applications, modifications, and variations of the energy-saving intelligent control method for coupled tunnel ventilation and drainage systems of this invention will be readily apparent to those skilled in the art.

[0054] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for coupled energy-saving intelligent control of tunnel ventilation and drainage systems, characterized in that, include: S1. Construct a digital twin model of the tunnel environment, wherein the digital twin model includes a static geometry layer, a parametric model layer, and a physical constraint layer; The real-time acquired multi-source heterogeneous sensing data is injected into the digital twin model to dynamically calibrate the real-time state variables in the digital twin model. S2. Based on the digital twin model, construct a coupled state prediction model, and under the unified framework of the digital twin model, predict the ventilation load change trend and drainage load change trend, phase change latent heat term and seepage supply term within the future preset time domain; S3. Based on the digital twin model, the ventilation load change trend and drainage load change trend predicted in step S2, the latent heat of phase change term and seepage supply term, construct a collaborative optimization control model, use a multi-objective optimization algorithm to solve the ventilation and drainage collaborative optimization control model, generate collaborative control commands for the ventilation system and drainage system in the current operating condition and future prediction time domain, and issue them for execution. The coupling state prediction model mentioned in step S2 includes: The spatiotemporal graph neural network module is used to extract the evolution law of each element in the ventilation system topology, drainage system topology and real-time state vector over time based on the digital twin model, and output the ventilation system state feature vector and drainage system state feature vector. The physical information neural network module uses the physical constraint layer of the digital twin model as the constraint condition, and the ventilation system state feature vector and drainage system state feature vector output by the spatiotemporal graph neural network module as the input boundary conditions, and outputs the latent heat of phase change term and seepage supply term. A coupling interaction layer connects the spatiotemporal graph neural network module and the physical information neural network module. It inputs the ventilation system state feature vector and the drainage system state feature vector as boundary conditions into the physical information neural network module, and feeds back the latent heat of phase change term and seepage supply term to the spatiotemporal graph neural network module to correct the temporal update of the ventilation system state feature vector and the drainage system state feature vector. The prediction output layer decodes and maps the ventilation system state feature vector and drainage system state feature vector after the coupling interaction layer correction, transforming them into the ventilation load change trend and drainage load change trend in the future preset time domain, and simultaneously outputs the phase change latent heat term and seepage supply term.

2. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 1, characterized in that, The static geometric layer in step S1 includes at least the tunnel geometry, ventilation system topology, and drainage system topology; the parameter model layer includes at least the evolution law model of traffic flow dynamic parameters in the tunnel and the spatiotemporal distribution model of the surrounding rock seepage field; the physical constraint layer includes at least the coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage and the corresponding boundary conditions and initial conditions.

3. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 2, characterized in that, The multi-source heterogeneous sensing data mentioned in step S1 includes: Traffic flow data, including traffic volume, vehicle speed, and vehicle type composition within the tunnel; Ventilation environment data, including wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system; Drainage system data, including water levels in sump pits at each node of the drainage system, operating status of drainage pumps, and drainage flow rate; Surrounding rock seepage data, including pore water pressure in the surrounding rock; The calibrated real-time state vector includes at least the following: wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system; water level in the sump and drainage flow rate at each node of the drainage system; and pore water pressure in the surrounding rock.

4. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 3, characterized in that, The dynamic calibration of real-time state variables in the digital twin model described in step S1 specifically includes: Traffic flow data, ventilation environment data, drainage system data, and surrounding rock seepage data from real-time acquired multi-source heterogeneous sensing data are mapped to the corresponding ventilation system nodes and drainage system nodes in the digital twin model, which serve as the observation vectors for the Kalman filter algorithm. The Kalman filter algorithm is used to linearize the evolution model of the dynamic parameters of traffic flow in the tunnel in the parameter model layer, and then use it as the state transition matrix to perform iterative optimization of the real-time state vector in the digital twin model in two steps: prediction and update. Prediction steps: Based on the real-time state vector and state transition matrix calibrated at the previous time step, predict the predicted state vector and its covariance matrix at the current time step. The predicted state vector is the estimated state value at the current time step without being corrected by the observed data. Update steps: The multi-source heterogeneous sensing data mapped to the digital twin model at the current moment is used as the observation vector. Combined with the predicted state vector and its covariance matrix obtained in the prediction step, the Kalman gain is calculated, and the predicted state vector is weighted and corrected to obtain the calibrated real-time state vector. The real-time state vector includes wind speed, wind pressure, temperature, humidity, carbon monoxide concentration, and visibility at each node of the ventilation system, as well as water level in the sump, drainage flow rate, and pore water pressure in the surrounding rock at each node of the drainage system. The spatiotemporal distribution model of the surrounding rock seepage field serves as a slowly varying background field. It does not participate in the state transition during the Kalman filter prediction step. It is only used to verify and correct the state estimates of the drainage system nodes after calibration. The correction of its model parameters is updated over a long period of time by the feedback results after the execution of the collaborative control command in step S3.

5. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 4, characterized in that, Step S2, which involves predicting the future trends of ventilation load and drainage load changes within a preset time domain, as well as the latent heat of phase change and seepage replenishment, specifically includes: Step S21: Input the real-time state vector obtained after dynamic calibration in step S1 into the spatiotemporal graph neural network module. The spatiotemporal graph neural network module constructs a heterogeneous spatiotemporal graph based on the topology of the ventilation system and the topology of the drainage system in the static geometric layer. Each element in the real-time state vector is used as the initial feature of the heterogeneous spatiotemporal graph node. The spatial topological features of the ventilation system and the drainage system are extracted through the graph convolutional network. The temporal evolution law of the state variables of each heterogeneous spatiotemporal graph node with time is extracted through the gated recurrent unit. The ventilation system state feature vector and the drainage system state feature vector are output. S22. Input the ventilation system state feature vector and the drainage system feature vector into the physical information neural network module. The physical information neural network module uses the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations in the physical constraint layer as physical constraints, the surrounding rock seepage field spatiotemporal distribution model in the parameter model layer as initial seepage boundary conditions, and the ventilation system state feature vector as the boundary conditions of the energy equation and phase change equation. The module solves the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations through an automatic differentiation mechanism and outputs the phase change latent heat term and seepage supply term. S23. Achieve bidirectional information interaction between the spatiotemporal graph neural network module and the physical information neural network module through a coupling interaction layer, including: inputting the state feature vector of the ventilation system and the state feature vector of the drainage system as boundary conditions into the physical information neural network module, and feeding back the latent heat of phase change and seepage supply terms output by the physical information neural network module to the spatiotemporal graph neural network module to correct the state feature vector update of the spatiotemporal graph neural network module in the next time step, so that the state prediction of the ventilation system and the drainage system evolves in a coordinated manner under the physical coupling mechanism; S24. Input the ventilation system state feature vector and drainage system state feature vector after the coupling interaction layer correction into the prediction output layer; the prediction output layer decodes and maps the high-dimensional feature vector into a time-series prediction result with physical meaning through a fully connected network, outputs the ventilation load change trend and drainage load change trend in the future preset time domain, and simultaneously outputs the prediction sequence of the phase change latent heat term and seepage supply term.

6. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 5, characterized in that, The coupled partial differential equations of air flow-water vapor phase change-surrounding rock seepage include the mass conservation equation, momentum conservation equation, energy conservation equation, carbon monoxide transport equation, water vapor phase change equation, and Darcy flow equation. The latent heat of phase transition is calculated from the phase transition source term in the energy conservation equation, where the energy conservation equation is: ； Phase transition latent heat term Q latent for: ； L is the latent heat of vaporization of water, 2.5 × 10⁻⁶. 6 J / kg; The mass flow rate of water vapor phase change per unit volume per unit time; kg / (m³) 3 •s); described by the water vapor phase transition equation: ； Or engineering model: ； k c is the mass transfer coefficient, with a value ranging from 0.001 to 0.01; A is the water-air contact area; C vapor C represents the concentration of water vapor in the air. sat (T) represents the saturated water vapor concentration at the current temperature; RH represents the relative humidity; A' represents the contact area between water and air per unit volume of tunnel. The seepage supply term Q seepage The seepage flow rate at the tunnel lining boundary is obtained by integrating: ； denoted as the contact surface between the tunnel lining and the surrounding rock; n is the normal vector of the contact surface. q is the seepage velocity, in m / s, which is described by Darcy's seepage equation: (p+ gz); c p ρ1 is the specific heat capacity of air at constant pressure; J / (kg·K); ρ1 is the density of air; kg / m³ 3 ; T is the air temperature, K; u is the air velocity vector, m / s; k is the thermal conductivity of air, W / (m·K); K is the permeability of the surrounding rock, m 2 μ is the dynamic viscosity of water, Pa·s; p is the pore water pressure in Pa; ρ2 is the density of water; z is the position head.

7. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 6, characterized in that, The collaborative optimization control model described in step S3 uses the ventilation system topology and drainage coefficient topology provided by the digital twin model as the spatial constraint boundary of the optimization problem; it uses the air flow-water vapor phase change-surrounding rock seepage coupled partial differential equations provided by the digital twin model as the derivation basis for the safety limits and physical constraints of equipment operation; and it uses the ventilation load change trend and drainage load change trend predicted in step S2 as the dynamic input of the optimization problem. The latent heat of phase change and the seepage recharge term are used as coupling constraint parameters; the goal is to minimize the overall operating energy consumption of the tunnel; wherein, the latent heat of phase change term is used to quantify the degree of influence of the water accumulation state of the drainage system on the heat and humidity load of the ventilation system, and the seepage recharge term is used to quantify the rigid demand of the seepage of the surrounding rock on the inflow load of the drainage system.

8. The energy-saving intelligent control method for coupled tunnel ventilation and drainage systems as described in claim 7, characterized in that, The collaborative control commands mentioned in step S3 include combinations of fan speeds and valve openings in the ventilation system; pump start / stop sequences and pump frequency combinations in the drainage system; and the feedback results after the collaborative control commands are issued and executed are synchronously updated to the parameter model layer of the digital twin model, which is used to correct the evolution law model of traffic flow dynamic parameters in the tunnel and the spatiotemporal distribution model of the surrounding rock seepage field.

Images