Underground cavern group gradient pressure regulation intelligent ventilation method and system based on hypergraph network

Abstract

The invention discloses an underground cavern group gradient pressure regulation intelligent ventilation method and system based on a hypergraph network. The method comprises the steps that a ventilation system hypergraph structure is established, a convolutional neural directed hypergraph aggregation method is adopted, monitoring data is converted into pipe section air volume requirement solving parameter input, and a model is endowed with more effective joint debugging features; solving an optimal wind network solving result by adopting a ventilation network solving technology and a response efficiency joint debugging method; for a joint debugging control strategy, ventilation control target curves under different target requirements are proposed, a control strategy based on a depth deterministic strategy gradient reinforcement learning algorithm is established, deviation between data and target control data is calculated in combination with each node, and a standard library, a criterion library, a rule library and the like are established. A fan-air valve-air door coupling hierarchical joint control instruction from power to the tail end is formed; according to the ventilation control system and equipment, an end-edge-cloud architecture is adopted, a ventilation monitoring system, a network analysis system and a control optimization system are included, and a dynamic regulation and control closed loop of the intelligent ventilation system is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent ventilation for tunnel engineering, and in particular relates to a hypergraph network-based gradient pressure-regulating intelligent ventilation method and system for underground cavern groups. Background Art

[0002] Intelligent on-demand ventilation is the basis for ensuring the environmental safety of underground caverns, and is also one of the core technologies for the high-quality construction and development of hydropower projects. With the improvement of the level of intelligent construction of projects, higher requirements are also put forward for intelligent ventilation and control under dynamic demand, involving multidisciplinary fields such as space science, information science, construction technology, mathematics and basic theory of fluid mechanics. Among them, ventilation network solution is the underlying core technology of intelligent ventilation and control. It is the key to accurately analyze monitoring data in real time and provide guidance for ventilation system fault diagnosis and disaster prediction based on spatial changes in construction progress, construction personnel and activities, abnormal fires, etc.

[0003] Theoretical research on ventilation networks: The Cross algorithm proposed in 1936 has developed to the present day, and the heat flow coupling ventilation theory, the concept of true and false topological networks, and fuzzy optimization mathematical models have emerged one after another, and are now relatively mature. In terms of solution methods, considering the influence of natural wind pressure, a real-time ventilation solution based on thermodynamics is proposed; the smoke distribution of subway stations and tunnels under emergency ventilation is analyzed, and the influence of fire wind pressure on ventilation solution and smoke control effectiveness under network-type smoke flow paths is analyzed; coupled ventilation efficiency and energy consumption and other multi-objectives are adjusted and controlled through genetic algorithms, particle swarm algorithms and other methods to optimize the energy consumption of the ventilation system, improve the overall convergence of the solution, and for the stability of the adjustment, a dynamic evaluation model of the stability and reliability of the mine ventilation system based on system dynamics and attribute mathematical theory is constructed. In terms of intelligent control, a multi-parameter environmental monitoring system is built, focusing on the combination with on-site environmental monitoring, improving the accuracy and engineering applicability of simulation results, establishing an air environment quality prediction based on CFD simulation, and building a three-dimensional visual emergency decision-making platform.

[0004] In summary, in recent years, domestic and foreign scholars have achieved certain results in solving the problem of dynamic changes in the required air volume of ventilation network branches by using ventilation network solution and intelligent technology, and gradually applied key intelligent technologies and platform facilities such as online monitoring, optimized layout, advance warning, fault diagnosis, and closed-loop control in engineering. However, with the increasing complexity of the ventilation system structure branches and the increase in the overall stability requirements during the ventilation system adjustment process, the traditional network solution cannot meet the accuracy requirements of ventilation parameters. The numerical simulation method has great limitations in computing resources in the real-time calculation of large-scale engineering ventilation parameters, and cannot provide accurate intelligent dynamic feedback control for the ventilation system. Therefore, there are still challenges in the real-time, stability and engineering applicability of ventilation system adjustment.

[0005] With the increasing demand for big data and real-time computing, graph theory and hypergraphs have been widely used in water conservancy, mining, power systems and other fields, making the modeling, storage and analysis of complex relational data more efficient and convenient. This includes the use of graph theory breadth-first algorithm and depth-first search to achieve parallelization of saturation sorting solutions. In the study of hypergraphs, a method for identifying hypernetwork nodes using information entropy was proposed, and a dynamic hypergraph representation method for wind turbines based on wake association along the time dimension was proposed for power system power prediction. However, there are still few studies on the use of hypergraph structures to solve ventilation regulation problems, including the selection of nodes, edges and hyperedges, and the determination of characteristic parameters. No relevant results have been reported on the use of hypergraph structures to complete real-time solution and joint control of ventilation systems. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a gradient pressure-regulated intelligent ventilation method and system for underground cavern groups based on a hypergraph network. Research is carried out on issues such as in-depth utilization of environmental monitoring data of complex structures of underground engineering cavern groups, fan and air valve coupling closed-loop control technology, matching and judgment strategy for on-demand air supply, and inspection of engineering air regulation effects guided by demand and energy consumption. The principles of on-demand supply of air volume load, precise solution of zoning regulation, and reinforcement learning closed-loop feed-control are clarified. A ventilation system diagram and hypergraph structure, a response efficiency indicator joint adjustment method, a deep deterministic strategy gradient reinforcement learning feed-control algorithm, and research and development of intelligent ventilation system devices and platforms based on "end-edge-cloud" are established.

[0007] The technical solution of the present invention is:

[0008] A hypergraph network-based intelligent ventilation method for underground cavern groups with gradient pressure regulation includes the following steps:

[0009] (1) Construction of ventilation system diagram and hypergraph structure, including the following steps:

[0010] S1. Construction of hypergraph structure: First, construct the basic graph structure of the ventilation system, then introduce the hypergraph concept to characterize the development of spatial fluid associations in the ventilation regulation process, and use directed hyperedges to describe the directed relationship between two node sets;

[0011] S2. Hypergraph matrix: A hypergraph matrix is ​​used to define the relationship between a node set and a hyperedge set. The expression of the hypergraph matrix structure is as follows:

[0012] G={V(H),E(H),V(D),E(D),X,W},

[0013] Where V(D) represents the node set of the general graph D, and V(D) = {v1, ..., v n} contains n nodes in the graph D, E(D) represents the edge set of the general graph D, and E(D) = {e1, ..., e m} contains m edges in the graph D, V(H) represents the node set of the hypergraph H and V(H) = {v1, ..., v n}, E(H) is the hyperedge set of the hypergraph H and E(H) = {E1, ..., E k}, X represents the feature matrix of the node, and W represents the weight matrix;

[0014] Each hyperedge in the hyperedge set E(H) describes T k and H k The relationship between the two subsets: E k =(T k , H k ),

[0015] Among them, T k represents the tail set of directed edge superedges and T k = {v T1 , ..., v Tk}, H k represents the head set of the edge hyperedge and H k = {v H1 ,...,v Hk};

[0016] The T k and H k is a subset of E(D), and the hypergraph association matrix is ​​as follows:

[0017] B=(b ij ) m*n

[0018]

[0019] Where B represents the hypergraph association matrix, b ij Represents node v i With the hyperedge E j The relationship between them, m represents the number of nodes in the hypergraph H, n represents the number of hyperedges in the hypergraph H, and v i represents the i-th node in the hypergraph H, H j Represents the hyperedge E j The head set, T j Represents the hyperedge E j The last set, E j represents the jth hyperedge set in the hypergraph H;

[0020] S3, convolution algorithm and hypergraph aggregation: the convolution algorithm is combined with the hypergraph aggregation process. The aggregation process includes: node information aggregation and hyperedge information aggregation. The node information aggregation is specifically as follows: according to the joint adjustment sensitivity, the feature vectors of all neighboring nodes of each node are weighted summed to form the head set and tail set representation vectors, and the node information is aggregated through the gate mechanism to finally obtain the representation vector of the directed hyperedge; the hyperedge information aggregation is specifically as follows: all directed hyperedges with the same head set are merged to obtain a new directed hyperedge, and the representation vectors of all directed hyperedges with the same head set are weighted summed to obtain the final output vector of the corresponding node. Finally, a ventilation adjustment hypergraph model is constructed;

[0021] (2) The optimized monitoring layout of the collective coverage model includes the following steps:

[0022] S4, the construction of the set coverage model, including: determining the target node set and building the response node combination, the determination of the target node set is specifically: by querying the performance attributes of all nodes in the ventilation adjustment map, determining the normal, abnormal, and warning categories, and screening out the target nodes to be adjusted that exceed the deviation range; the construction of the response node combination is specifically: selecting the air volume-wind resistance response efficiency in the data association between nodes as the adjustment basis, according to the output vector matrix of the corresponding node in the node information fusion process, using a multi-level relationship query method, retrieve the nodes with data association with the target node set, eliminate the nodes that do not meet the adjustable attributes and positive and negative correlation attributes, and build the response node set;

[0023] S5. Optimization of the layout of monitoring points, including the following steps:

[0024] Step A: By setting the release point where the concentration of harmful gases suddenly changes, the node range affected by the release point within the monitoring response time is calculated to form a release set;

[0025] Step B: define potential monitoring points, calculate the effective monitoring range of potential monitoring points, establish a set of effective ranges of potential monitoring points, and construct monitoring point hypergraph elements;

[0026] Step C: Use the greedy algorithm to solve the set coverage model and select the minimum number of monitoring points to ensure that there is at least one monitoring point within the effective response range of each release point;

[0027] (3) The ventilation network solution and response efficiency joint adjustment method includes the following steps:

[0028] S6, gradient voltage regulation model construction, including the following steps:

[0029] Step D: define the air volume-wind resistance response efficiency attenuation rate, quantify the effect of wind resistance change on air volume, fit the air volume-wind resistance response efficiency curve through a power function, and calculate the response efficiency attenuation rate;

[0030] Step E: Set the threshold of wind resistance value, determine the sensitive area of ​​air volume adjustment, and refine the adjustment plan to specific branches according to the air adjustment plan;

[0031] S7, air adjustment scheme optimization, including the following steps:

[0032] Step F: Filter the response nodes, and select the set of nodes participating in the adjustment according to the ranking, and establish the air adjustment plan within the response efficiency constraint;

[0033] Step G: If the solution does not meet the target requirements, more response nodes are selected to update the air adjustment solution until the target node air volume deviation is reduced to within the deviation threshold;

[0034] Step H: Each iteration selects new nodes to participate in the adjustment, forms a new air adjustment plan, and finally outputs a plan that meets the joint adjustment sensitivity requirements;

[0035] (4) Control strategy based on deep deterministic policy gradient reinforcement learning algorithm: Based on the control principle of on-demand air volume supply and personalized partitioning, a control strategy is constructed using a deep deterministic policy gradient reinforcement learning algorithm. By simulating various areas in the environment, personalized control is performed on different air supply demands.

[0036] (5) Response efficiency analysis and adjustment, including the following steps:

[0037] S8. Construction of the air volume-wind resistance response efficiency matrix includes the following steps:

[0038] Step I: Define the air volume-wind resistance response efficiency, calculate the relationship between the air volume change and wind resistance disturbance between nodes, form a response efficiency matrix, use an iterative method to continuously approach the optimal solution, and finally obtain the air volume-wind resistance response efficiency matrix of the entire ventilation network;

[0039] Step J: construct an iterative sequence and recalculate the response efficiency after each air volume adjustment until the calculation accuracy requirement is met;

[0040] S9, wind turbine power and wind resistance response efficiency, including the following steps:

[0041] Step K: Analyze the response relationship between fan power and wind resistance, define the sensitivity of fan power change to wind resistance change, form a fan power-wind resistance response efficiency matrix, and select a branch with low power sensitivity as a wind resistance adjustment branch;

[0042] Step L: Analyze the response efficiency of the second-order wind network, quantify the mutual influence between the adjustment branches, construct a three-dimensional matrix, calculate the mutual relationship between different branches, and ensure the coordination of each branch during the adjustment process.

[0043] Furthermore, the method steps for constructing the basic graph structure of the ventilation system are:

[0044] First, a ventilation adjustment graph database is constructed based on Python and Neo4j. The data interface API is called by Python program to obtain ventilation data, and the data is preprocessed into a "unit-association-feature" triple form. The physical space connection formed by the fluid inflow section or outflow section is represented by the flow path relationship, that is, the basic association; the main association derived from the spatial relationship that reflects the coupling influence law of ventilation characteristics is represented by the data relationship;

[0045] 2. Use py2neo to establish a connection between the Python program and Neo4j, batch process the construction instructions of the graph database, use Cypher query statements to define associations and assign parameters, use any node as the starting point l1 of the loop in the Neo4j graph database, find all relationship units that have basic associations with l1, and use the graph matching capability to find the shortest path in the relationship unit as the next loop node l2;

[0046] 3. Use Cypher statements to find the shortest path among the nodes that have a basic association with l2, and use it as the next loop node l3. Repeat the query process, traverse the nodes in turn, and build the loop set L.

[0047] Furthermore, the head set representation vector h in step S3 H The expression is:

[0048] h H =(H H ⊙W H )X,

[0049] Among them, h H The characterization vector representing the head set of the hyperedge, H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, and X represents the feature matrix of the node;

[0050] The tail set representation vector h of step S3 T The expression is:

[0051] h T =(H T ⊙W T )X,

[0052] Among them, h T The characterization vector representing the tail set of the hyperedge, H T represents the characteristic matrix of the tail node in the hyperedge, W T represents the weight matrix of the tail set, and X represents the feature matrix of the node;

[0053] The characterization vector h of the directed hyperedge in step S3 εi The expression is:

[0054] h ε =σ(h H )⊙h T ,

[0055] Among them, h ε represents the characterization vector of the directed hyperedge, h H The characterization vector representing the head set of the hyperedge, h T represents the characterization vector of the tail set of the hyperedge, σ represents the activation function;

[0056] The expression of the final output vector of step S3 is:

[0057] h=(H H ⊙W H )h ε ,

[0058] Among them, h represents the node output vector after aggregating hyperedge information, H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, h ε Represents the characterization vector of the directed hyperedge;

[0059] The expression of the ventilation adjustment hypergraph model in step S3 is:

[0060] H ε = {V(H), (H H ⊙W H )h ε},

[0061] Among them, H ε represents the basic node set whose joint debugging sensitivity is greater than ε, ε represents the sensitivity threshold, V(H) represents the node set of the hypergraph H, and H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, h ε Represents the characterization vector of a directed hyperedge.

[0062] Furthermore, the release point in step S5 is a node that has the requirement to install harmful gas concentration monitoring, and the expression of the release set in step S5 is:

[0063]

[0064]

[0065]

[0066] in, represents the collection involved, represents the set of nodes affected by the released concentration wave, represents the set of release points, v j represents node j, t in the hypergraph H max represents the response time, v n Indicates the nth release point;

[0067] If there are k nodes with sudden changes in the concentration of released gas in the node set V(D) of the general graph D, then the release point hyperedge set is constructed:

[0068]

[0069] Among them, E 0 (H) represents the hyperedge set of the release point, represents the edge set associated with the i-th release point, represents the edge set associated with the kth release point, k represents the total number of release points where the concentration of harmful gases suddenly changes; the potential monitoring point in step S5 is a node installed with a wind pressure monitoring device, and the expression of the effective monitoring range of the potential monitoring point is:

[0070]

[0071] in, represents the effective monitoring range set of potential monitoring points, v k Indicates the nodes within the valid range, N j represents a potential monitoring point, t max Indicates monitoring response time;

[0072] The expression of the monitoring point hypergraph element in step S5 is:

[0073]

[0074] Among them, E t (H) represents the hyperedge set of potential monitoring points, represents the effective monitoring range edge set of the jth potential monitoring point, and k represents the total number of potential monitoring points;

[0075] According to the response requirements, there is at least one potential monitoring point within the effective response range of the release point, so the goal of minimizing the total number of monitoring points under the constraints of monitoring point layout is established:

[0076]

[0077]

[0078]

[0079] Among them, z represents the number of monitoring deployments, x j Represents the layout variable, that is, the potential monitoring point N j Whether to set up monitoring deployment, n represents the total number of potential monitoring points, represents the potential monitoring point N j The set of nodes that can be covered;

[0080] The greedy algorithm for solving the set coverage model is specifically as follows: using Python programming language to compile and run on the Pycharm platform to achieve the transformation of the above target expression; first, from the hyperedge set E of potential monitoring points t (H) Find the set E of nodes most affected by the release concentration 0 Elements in (H) Corresponding to the set The relationship between the head set and the tail set will release the point set Monitoring point {v H1 , ..., v Hi} as the optimal monitoring layout position, and put it into the response node set whose original value is an empty set; secondly, delete the node element {v H1 , ..., v Hi}, from the hyperedge set E of potential monitoring points t (H) Delete the elements related to the head set representation vector in, Represents the hyperedge set E of potential monitoring points t (H) The edge set affected by the point i in the experiment; third, the hyperedge set E of the remaining potential monitoring points t (H) Find the set of nodes most affected by the release concentration, and repeat the above steps until the hyperedge set E of the release point is 0 (H) element is an empty set; finally, the response node set contains elements representing the optimized layout quantity and position of potential monitoring points.

[0081] Further, the wind volume-wind resistance response efficiency attenuation rate in step S6 is defined as: the degree of change in the branch wind resistance causing the branch wind volume-wind resistance response efficiency to change itself, and the expression of the wind volume-wind resistance response efficiency attenuation rate is:

[0082]

[0083] Among them, t ij represents the air volume-wind resistance response efficiency attenuation rate, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, △r j Indicates wind resistance rj A small perturbation, q i represents the wind volume at the response node j;

[0084] The expression of the power function fitting wind volume-wind resistance response efficiency curve in step S6 is:

[0085]

[0086] Among them, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, a, b and c are all constants greater than zero;

[0087] The calculation formula of the attenuation rate of the response efficiency in step S6 is:

[0088]

[0089] Among them, t ij represents the wind volume-wind resistance response efficiency attenuation rate and t ij Always negative, that is, d ij With the j The value increases and decreases, r j represents the wind resistance value of the response node j, and both a and b are constants greater than zero;

[0090] The expression of the threshold value of the wind resistance value in step S6 is:

[0091]

[0092] Among them, r jmax The wind resistance threshold, r j0 represents the initial value of wind resistance, a, b and c are all constants greater than zero, α represents the sensitivity threshold and 0<α<0.05.

[0093] Furthermore, the method steps for constructing the control strategy using the deep deterministic policy gradient reinforcement learning algorithm are as follows:

[0094] Step 1: Define the state space: The state space is composed of the target branch parameter T t , adjust the branch parameter R t , Power branch parameter P t The fan operating voltage, current and frequency are determined, and the function expression of the state space is as shown in the following formula:

[0095] s t =[T t , R t , P t ],

[0096] Among them, T t = {q1，t ,q 2，t ,...q n，t , p 1，t , p 2，t , ...p n，t} represents the air volume and air pressure parameters of the target branch at time t, R t ={θ 1，t ,θ 2，t , ...θ n，t} represents the opening of the stroke valve in the regulating branch at time t, P t = {I 1，t , I 2，t ,...I n，t , f 1，t , f 2，t , ...f n，t} represents the fan current and frequency parameters of the power branch at time t;

[0097] Step 2: Define the action space: Execute corresponding actions according to the observed environmental parameters. The action space is determined by the air valve opening and fan current and frequency that need to be adjusted. The function representation of the action space is as shown in the following formula:

[0098] a t =[Δθ t , ΔI t , Δf t ],

[0099] Among them, Δθ t ={Δθ 1，t , Δθ 2，t , ...Δθ n，t} represents the change in the opening of the wind valve in the regulating branch at time t, ΔI t ={ΔI 1，t , ΔI 2，t , ...ΔI n，t , Δf 1，t , Δf 2，t , ...Δf n，t} represents the change in the fan current and frequency parameters of the power branch at time t;

[0100] Step 3: Define RL rewards: Minimize ventilation energy consumption while meeting the target air volume load. Reduce the deviation between the supply air volume and the required air volume in terms of demand-based ventilation revenue rewards to meet:

[0101]

[0102] Among them, Q n，t is the air supply volume of branch n at time t, is the required air volume, d Q，t is the deviation between the air supply volume and the required air volume, dQ，t,max The maximum threshold for the deviation between the supply air volume and the required air volume;

[0103] Set the long-term benefit ratio to λ1; for energy consumption benefit rewards, control the air valve opening to be greater than 70%, and the fan operates in a stable working area, that is, the working pressure is less than 90% of the maximum pressure, and the operating efficiency is greater than 60%, meeting:

[0104] θ t ≥θ t，min , Q f，min ≤Q f ≤Q f，max ,

[0105] Among them, θ t is the air valve opening, θ t，min is the minimum threshold of air valve opening, Q f is the fan air volume, Q f,min and Q f,max They are the minimum and maximum thresholds of the fan air volume range respectively;

[0106] If the long-term profit ratio is set to λ2 and λ3, the reward function is as follows:

[0107] r t (s t , a t )=[R Q , R θ , R f ],

[0108] Among them, R Q =λ1c r，Q , c r，Q Represents the wind volume bonus coefficient, R θ =λ2c r，θ , c r，θ Indicates the air valve bonus coefficient, R θ =λ3c r，f , c r，f Indicates the wind turbine bonus coefficient;

[0109] The expression of the wind volume bonus coefficient is:

[0110]

[0111] The expression of the wind valve bonus coefficient is:

[0112]

[0113] The expression of the wind turbine bonus coefficient is:

[0114]

[0115] Step 4: Establish an Actor-Critic architecture, which includes: parameters θ u Actor current network function: u = (s, a | θ u ), where θ u Represents the parameter set of the Actor network, s represents the current state of the network, a represents the action selected based on the current state; the parameter is θ Q Critic current network function: Q = (s, a | θ Q ), where θ Q represents the parameter set of the Critic network, s represents the state received by the action network, a represents a deterministic action based on the received state; and the parameter is θ u →θ u′ and θ Q →θ Q′ The Actor-Critic corresponds to the target network, where θ u′ Represents the parameter set of the target Actor network, θ Q′ Represents the parameter set of the target Critic network;

[0116] Step 5: Actor-Critic network initialization:

[0117] Set the learning rate of the Actor network and the Critic network, set the update smoothing coefficient of the target Q network; set the number of neurons in the Actor network and the Critic network architecture, set the experience pool capacity and the number of sampling pools used for parameter updates;

[0118] Step 6: Actor-Critic network update loop:

[0119] Based on the current environment state output, the DDPG agent performs actions and obtains the next environment state and RL reward, while adding action perturbation N t , environmental state s t , perform action a t , RL reward r t and the next environment state s t+1 The samples are stored in the experience pool, and data {s t , a t , r t ,s t+1} to the Actor-Critic current network and update the network parameters:

[0120] θ Q′ =τθ Q +(1-τ)θ Q′ ,

[0121] Among them, θ Q Represents the parameter set of the Critic network, θ Q′ represents the parameter set of the target Critic network, and τ represents the factor that controls the update speed of the target network parameters;

[0122] θ u′ =τθ u +(1-τ)θ u′ ,

[0123] Among them, θ u Represents the parameter set of the Actor network, θ u′ represents the parameter set of the target Actor network, and τ represents the factor that controls the update speed of the target network parameters;

[0124] Step 7: Continue to update the process in step 6 until the policy network learns a good enough policy.

[0125] Furthermore, the definition of the air volume-wind resistance response efficiency in step S8 is: the influence degree of branch wind resistance on node pressure or branch air volume, and the expression of the air volume-wind resistance response efficiency is:

[0126]

[0127] Among them, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, q i represents the wind volume at the target node i, △r j Indicates wind resistance r j A small perturbation, △q i represents the wind volume change at the target node i;

[0128] The expression of the air volume-wind resistance response efficiency matrix in step S8 is:

[0129]

[0130] Among them, D Q represents the air volume-wind resistance response efficiency matrix, d ij represents the air volume-wind resistance response efficiency, q k represents the wind volume at the response node j, r n represents the wind resistance value of the nth response node, d jn represents the wind resistance r of the response node n n The wind volume q of the response node j j The degree of influence, n represents the number of responding nodes;

[0131] The expression of the iterative sequence in step S8 is:

[0132]

[0133]

[0134] in, represents the wind volume-wind resistance response efficiency in the kth iteration, represents the wind volume value of target node i at the kth iteration, represents the wind volume value of target node i at the k-1th iteration, r j (k) represents the wind resistance value of the response node j at the kth iteration, r j (k-1) represents the wind resistance value of the response node j at the k-1th iteration, dr j (k) represents the wind resistance disturbance, that is, the change in wind resistance of the response node j in the kth iteration, dr j (k-1) represents the change in wind resistance of the response node j in the k-1th iteration, and γ represents the acceleration factor;

[0135] The expression of the calculation accuracy is:

[0136]

[0137] in, represents the wind volume-wind resistance response efficiency in the kth iteration, represents the wind volume-wind resistance response efficiency in the k-1th iteration, and ε represents the calculation accuracy.

[0138] Further, the fan power-wind resistance response efficiency in step S9 is defined as: the degree of influence of the fan output power change on the branch wind resistance change. The expression of the fan power-wind resistance response efficiency in step S9 is:

[0139]

[0140] Among them, dN ij Indicates the wind turbine power-wind resistance response efficiency, ΔN fi represents the wind turbine output power change at target node i, Δr j Represents the change in wind resistance at the response node j, N fi represents the wind turbine output power at the target node i, r j represents the wind resistance value at the response node j;

[0141] For wind turbine power-wind resistance response efficiency dN ij , fan output power N fi Satisfy the product relationship between fan pressure and fan air volume:

[0142] N fi =h f q i ,

[0143] Among them, N fi represents the wind turbine output power at the target node i, h f Indicates the fan pressure, q i represents the wind volume value at the target node i;

[0144] Fan air pressure h f The least squares fitting relationship of a univariate polynomial is satisfied, and the degree of the polynomial is 3. The output power characteristic curve of the fan is as shown in the following formula:

[0145]

[0146] Among them, N fi represents the wind turbine output power at the target node i, q i represents the wind volume value at the target node i, a0 represents the constant term, that is, when the wind volume value q i is the fan power change rate when is zero, a1, a2 and a3 are the coefficients of the polynomial, describing the nonlinear effect of air volume on fan power;

[0147] Air volume q i Find the partial derivative:

[0148]

[0149] Among them, N fi represents the wind turbine output power at the target node i, q i represents the wind volume value at the target node i, a0 represents the constant term, that is, when the wind volume value q i is the fan power change rate when is zero, a1, a2 and a3 are the coefficients of the polynomial, describing the nonlinear effect of air volume on fan power;

[0150] The expression of the wind turbine power-wind resistance response efficiency matrix in step S9 is:

[0151]

[0152]

[0153] Among them, D N represents the wind turbine power-wind resistance response efficiency matrix, dN ij Indicates the wind turbine power-wind resistance response efficiency, N fn represents the wind turbine output power at the response node n, r n represents the wind resistance value of the nth response node, d nn represents the wind resistance r of the response node nn The wind turbine output power N for the response node n fn The degree of influence, ΔN fi represents the wind turbine output power change at target node i, Δr j represents the change in wind resistance at the response node j, d ij It indicates the air volume-wind resistance response efficiency;

[0154] The second-order wind network response efficiency in step S9 is defined as: adjusting branch e k Effect of wind resistance change on other branches of wind network i The influence of air volume, the expression of the second-order wind network response efficiency is:

[0155]

[0156] Among them, dd ijk represents the second-order wind network response efficiency, d ij represents the air volume-wind resistance response efficiency, r k represents the wind resistance at node k, q i represents the wind volume value at the target node i, r j represents the wind resistance value at the response node j;

[0157] The three-dimensional matrix in step S9 is a second-order wind network response efficiency matrix, and the expression of the second-order wind network response efficiency matrix is:

[0158] DD(i,j,k)=dd ijk ,

[0159] Among them, DD(i,j,k) is the second-order wind network response efficiency matrix, dd ijk It represents the second-order wind network response efficiency. When j=k, it represents the influence of branch regulation on its own wind volume, that is, the attenuation rate of the effect efficiency of the regulation branch. When j≠k, it represents the mutual influence between branches. If two regulation branches e k and e j , satisfying the conditions: d ij >0,d ik >0,d ijk >0, it means that the two adjustment branches are related to branch e i Same promotion, if the conditions are met: d ij >0,d ik >0,d ijk <0, it means that the two adjustment branches are related to branch e i Same inhibition;

[0160] Take the second-order derivative of the loop wind pressure equation vector with respect to the wind resistance vector:

[0161]

[0162] Among them, (R, Q y (R)) indicates that the free variables of the loop wind pressure equations are the wind resistance vector R and the residual wind volume vector Q y , The Jacobian matrix of the circuit wind pressure equations for wind resistance is denoted as f1′. The Jacobian matrix of the loop wind pressure equations for the residual tree wind volume is denoted by f2′. is the Jacobian gradient algorithm matrix of the wind volume to wind resistance of the remaining tree, denoted as f3′;

[0163] make The analytical expression of the three-dimensional matrix of the second-order wind network response efficiency is as follows:

[0164]

[0165] in, Represents the residual tree wind volume vector Q y The second derivative of the drag vector R, f 11 ″ represents the second-order Jacobian matrix of the loop wind pressure equation for wind resistance, f 12 ″ represents the mixed second-order Jacobian matrix of the loop wind pressure equation for wind resistance and residual wind volume, f 21 ″ represents the mixed second-order Jacobian matrix of the wind pressure equation for the residual wind volume and wind resistance, f 22 ″ represents the second-order Jacobian matrix of the loop wind pressure equation for the residual tree wind volume.

[0166] A gradient pressure regulating intelligent ventilation system for underground cavern groups, comprising a ventilation monitoring system, a network analysis system, a control optimization system and a visualization system, wherein the network analysis system is respectively connected to the ventilation monitoring system and the control optimization system, and the visualization system is connected to the network analysis system. The ventilation monitoring system comprises: an integrated air environment monitoring device, an air volume and air pressure monitoring device and a fan operation parameter monitoring device, wherein the integrated air environment monitoring device is arranged in a tunnel, the air volume and air pressure monitoring device is arranged in a ventilation duct, and the fan operation parameter monitoring device is arranged near the fan and connected to the network analysis system; the network analysis system comprises: an integrated intelligent fan control cabinet and a cloud server, wherein the integrated intelligent fan control cabinet is respectively connected to the ventilation monitoring system The system is connected with the control optimization system, the cloud server is connected with the integrated intelligent fan control cabinet and the visualization system respectively, the integrated intelligent fan control cabinet comprises: a data acquisition module, a data processing module, a calculation and analysis module and a communication module, the data acquisition module is connected with the calculation and analysis module, the calculation and analysis module is connected with the data processing module, and the communication module is connected with the data processing module and the cloud server respectively; the control optimization system comprises: an intelligent air valve and a variable frequency fan control cabinet, the intelligent air valve is arranged on the ventilation duct, the variable frequency fan control cabinet is arranged near the fan, the intelligent air valve and the variable frequency fan control cabinet are both connected with the network analysis system, the opening range of the intelligent air valve is not less than 70%, and the operation efficiency of the fan is greater than 60%.

[0167] The beneficial effects of the present invention are as follows: a hypergraph network-based gradient pressure-regulated intelligent ventilation method and system for underground cavern groups of the present invention adopts regional model theory, proposes node extraction principles and methods for one-dimensional tube bundle fluid and three-dimensional space flow field unit structure, and realizes the construction of different ventilation fluid regional graphs and hypergraph network structures of the ventilation system by establishing a field network node coupling solution of virtual branches; on this basis, a set coverage model is established to optimize the monitoring layout, revealing the coupling linkage effect between ventilation branches and loops, and a resistance regulation solution method based on joint regulation sensitivity is proposed. With the target unit air regulation efficiency and overall stability as core indicators, threshold parameters are set and the regulation scheme is optimized; a reinforcement learning closed-loop feed-back control regulation algorithm based on DDPG is established to obtain the optimal air volume regulation strategy in a dynamic environment; an intelligent ventilation system platform based on the "end-edge-cloud" architecture is developed, which has the functions of real-time solution, remote control and data analysis, and a multi-parameter integrated air environment monitoring device and an intelligent fan control cabinet are developed, which realizes the comprehensive perception, analysis, control and 3D display and early warning of air environment parameters and ventilation system operation status, and improves the intelligent regulation and management capabilities of construction ventilation. BRIEF DESCRIPTION OF THE DRAWINGS

[0168] Figure 1It is a structural schematic diagram of a gradient pressure regulating intelligent ventilation system for underground cavern groups based on a hypergraph network of the present invention.

[0169] Figure 2 It is a response efficiency wind control joint adjustment optimization step diagram of a gradient pressure regulation intelligent ventilation method for underground cavern groups based on a hypergraph network of the present invention.

[0170] Figure 3 It is a schematic diagram of the field network structure of a hypergraph network-based intelligent ventilation method for gradient pressure regulation of underground cavern groups of the present invention.

[0171] Figure 4 It is a basic graph structure construction and Cypher statement query flow chart of a ventilation system of a gradient pressure regulation intelligent ventilation method for underground cavern groups based on a hypergraph network of the present invention.

[0172] Figure 5 It is a directed hypergraph structure aggregation schematic diagram of a hypergraph network-based underground cavern group gradient pressure regulation intelligent ventilation method of the present invention.

[0173] Figure 6 It is a schematic diagram of the user interface of an intelligent ventilation software platform of an underground cavern group gradient pressure regulation intelligent ventilation system based on a hypergraph network of the present invention. DETAILED DESCRIPTION

[0174] like Figure 1-6 As shown, a gradient pressure regulating intelligent ventilation method for underground cavern groups based on a hypergraph network comprises the following steps:

[0175] (1) Construction of ventilation system diagram and hypergraph structure, including the following steps:

[0176] S1. Construction of hypergraph structure: First, construct the basic graph structure of the ventilation system, then introduce the hypergraph concept to characterize the development of spatial fluid associations in the ventilation regulation process, and use directed hyperedges to describe the directed relationship between two node sets;

[0177] S2. Hypergraph matrix: A hypergraph matrix is ​​used to define the relationship between a node set and a hyperedge set. The expression of the hypergraph matrix structure is as follows:

[0178] G={V(H),E(H),V(D),E(D),X,W},

[0179] Where V(D) represents the node set of the general graph D, and V(D) = {v1, ..., v n} contains n nodes in the graph D, E(D) represents the edge set of the general graph D, and E(D) = {e1, ..., e m} contains m edges in the graph D, V(H) represents the node set of the hypergraph H and V(H) = {v1, ..., v n}, E(H) is the hyperedge set of the hypergraph H and E(H) = {E1, ..., E k}, X represents the feature matrix of the node, and W represents the weight matrix;

[0180] Each hyperedge in the hyperedge set E(H) describes T k and H k The relationship between the two subsets: E k =(T k , H k ),

[0181] Among them, T k represents the tail set of directed edge superedges and T k = {v T1 , ..., v Tk}, H k represents the head set of the edge hyperedge and H k = {v H1 ,...,v Hk};

[0182] The T k and H k is a subset of E(D), and the hypergraph association matrix is ​​as follows:

[0183] B=(b ij ) m*n

[0184]

[0185] Where B represents the hypergraph association matrix, b ij Represents node v i With the hyperedge E j The relationship between them, m represents the number of nodes in the hypergraph H, n represents the number of hyperedges in the hypergraph H, and v i represents the i-th node in the hypergraph H, H j Represents the hyperedge E j The head set, T j Represents the hyperedge E j The last set, E j represents the jth hyperedge set in the hypergraph H;

[0186] S3, convolution algorithm and hypergraph aggregation: the convolution algorithm is combined with the hypergraph aggregation process. The aggregation process includes: node information aggregation and hyperedge information aggregation. The node information aggregation is specifically as follows: according to the joint adjustment sensitivity, the feature vectors of all neighboring nodes of each node are weighted summed to form the head set and tail set representation vectors, and the node information is aggregated through the gate mechanism to finally obtain the representation vector of the directed hyperedge; the hyperedge information aggregation is specifically as follows: all directed hyperedges with the same head set are merged to obtain a new directed hyperedge, and the representation vectors of all directed hyperedges with the same head set are weighted summed to obtain the final output vector of the corresponding node. Finally, a ventilation adjustment hypergraph model is constructed;

[0187] (2) The optimized monitoring layout of the collective coverage model includes the following steps:

[0188] S4, the construction of the set coverage model, including: determining the target node set and building the response node combination, the determination of the target node set is specifically: by querying the performance attributes of all nodes in the ventilation adjustment map, determining the normal, abnormal, and warning categories, and screening out the target nodes to be adjusted that exceed the deviation range; the construction of the response node combination is specifically: selecting the air volume-wind resistance response efficiency in the data association between nodes as the adjustment basis, according to the output vector matrix of the corresponding node in the node information fusion process, using a multi-level relationship query method, retrieve the nodes with data association with the target node set, eliminate the nodes that do not meet the adjustable attributes and positive and negative correlation attributes, and build the response node set;

[0189] S5. Optimization of the layout of monitoring points, including the following steps:

[0190] Step A: By setting the release point where the concentration of harmful gases suddenly changes, the node range affected by the release point within the monitoring response time is calculated to form a release set;

[0191] Step B: define potential monitoring points, calculate the effective monitoring range of potential monitoring points, establish a set of effective ranges of potential monitoring points, and construct monitoring point hypergraph elements;

[0192] Step C: Use the greedy algorithm to solve the set coverage model and select the minimum number of monitoring points to ensure that there is at least one monitoring point within the effective response range of each release point;

[0193] (3) The ventilation network solution and response efficiency joint adjustment method includes the following steps:

[0194] S6, gradient voltage regulation model construction, including the following steps:

[0195] Step D: define the air volume-wind resistance response efficiency attenuation rate, quantify the effect of wind resistance change on air volume, fit the air volume-wind resistance response efficiency curve through a power function, and calculate the response efficiency attenuation rate;

[0196] Step E: Set the threshold of wind resistance value, determine the sensitive area of ​​air volume adjustment, and refine the adjustment plan to specific branches according to the air adjustment plan;

[0197] S7, air adjustment scheme optimization, including the following steps:

[0198] Step F: Filter the response nodes, and select the set of nodes participating in the adjustment according to the ranking, and establish the air adjustment plan within the response efficiency constraint;

[0199] Step G: If the solution does not meet the target requirements, more response nodes are selected to update the air adjustment solution until the target node air volume deviation is reduced to within the deviation threshold;

[0200] Step H: Each iteration selects new nodes to participate in the adjustment, forms a new air adjustment plan, and finally outputs a plan that meets the joint adjustment sensitivity requirements;

[0201] (4) Control strategy based on deep deterministic policy gradient reinforcement learning algorithm: Based on the control principle of on-demand air volume supply and personalized partitioning, a control strategy is constructed using a deep deterministic policy gradient reinforcement learning algorithm. By simulating various areas in the environment, personalized control is performed on different air supply demands.

[0202] (5) Response efficiency analysis and adjustment, including the following steps:

[0203] S8. Construction of the air volume-wind resistance response efficiency matrix includes the following steps:

[0204] Step I: Define the air volume-wind resistance response efficiency, calculate the relationship between the air volume change and wind resistance disturbance between nodes, form a response efficiency matrix, use an iterative method to continuously approach the optimal solution, and finally obtain the air volume-wind resistance response efficiency matrix of the entire ventilation network;

[0205] Step J: construct an iterative sequence and recalculate the response efficiency after each air volume adjustment until the calculation accuracy requirement is met;

[0206] S9, wind turbine power and wind resistance response efficiency, including the following steps:

[0207] Step K: Analyze the response relationship between fan power and wind resistance, define the sensitivity of fan power change to wind resistance change, form a fan power-wind resistance response efficiency matrix, and select a branch with low power sensitivity as a wind resistance adjustment branch;

[0208] Step L: Analyze the response efficiency of the second-order wind network, quantify the mutual influence between the adjustment branches, construct a three-dimensional matrix, calculate the mutual relationship between different branches, and ensure the coordination of each branch during the adjustment process.

[0209] Preferably, the method steps for constructing the basic graph structure of the ventilation system are:

[0210] First, a ventilation adjustment graph database is constructed based on Python and Neo4j. The data interface API is called by Python program to obtain ventilation data, and the data is preprocessed into a "unit-association-feature" triple form. The physical space connection formed by the fluid inflow section or outflow section is represented by the flow path relationship, that is, the basic association; the main association derived from the spatial relationship that reflects the coupling influence law of ventilation characteristics is represented by the data relationship;

[0211] 2. Use py2neo to establish a connection between the Python program and Neo4j, batch process the construction instructions of the graph database, use Cypher query statements to define associations and assign parameters, use any node as the starting point l1 of the loop in the Neo4j graph database, find all relationship units that have basic associations with l1, and use the graph matching capability to find the shortest path in the relationship unit as the next loop node l2;

[0212] 3. Use Cypher statements to find the shortest path among the nodes that have a basic association with l2, and use it as the next loop node l3. Repeat the query process, traverse the nodes in turn, and build the loop set L.

[0213] Furthermore, the head set representation vector h in step S3 H The expression is:

[0214] h H =(H H ⊙W H )X,

[0215] Among them, h H The characterization vector representing the head set of the hyperedge, H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, and X represents the feature matrix of the node;

[0216] The tail set representation vector h of step S3 T The expression is:

[0217] h T =(H T ⊙W T )X,

[0218] Among them, h T The characterization vector representing the tail set of the hyperedge, H T represents the characteristic matrix of the tail node in the hyperedge, W T represents the weight matrix of the tail set, and X represents the feature matrix of the node;

[0219] The characterization vector h of the directed hyperedge in step S3 εi The expression is:

[0220] h ε =σ( h H )⊙h T ,

[0221] Among them, h ε represents the characterization vector of the directed hyperedge, h H represents the characterization vector of the head set of the hyperedge, hT represents the characterization vector of the tail set of the hyperedge, and σ represents the activation function;

[0222] The expression of the final output vector of step S3 is:

[0223] h=(H H ⊙W H )h ε ,

[0224] Among them, h represents the node output vector after aggregating hyperedge information, H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, h ε Represents the characterization vector of the directed hyperedge;

[0225] The expression of the ventilation adjustment hypergraph model in step S3 is:

[0226] H ε = {V(H), (H H ⊙W H )h ε},

[0227] Among them, H ε represents the basic node set whose joint debugging sensitivity is greater than ε, ε represents the sensitivity threshold, V(H) represents the node set of the hypergraph H, and H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, h ε Represents the characterization vector of a directed hyperedge.

[0228] Furthermore, the release point in step S5 is a node that has the requirement to install harmful gas concentration monitoring, and the expression of the release set in step S5 is:

[0229]

[0230]

[0231]

[0232] in, represents the set involved, T i 0 represents the set of nodes affected by the released concentration wave, represents the set of release points, v j represents node j, t in the hypergraph H max represents the response time, v n Indicates the nth release point;

[0233] If there are k nodes with sudden changes in the concentration of released gas in the node set V(D) of the general graph D, then the release point hyperedge set is constructed:

[0234]

[0235] Among them, E 0 (H) represents the hyperedge set of the release point, represents the edge set associated with the i-th release point, represents the edge set associated with the kth release point, k represents the total number of release points where the concentration of harmful gases suddenly changes; the potential monitoring point in step S5 is a node installed with a wind pressure monitoring device, and the expression of the effective monitoring range of the potential monitoring point is:

[0236]

[0237] in, represents the effective monitoring range set of potential monitoring points, v k Indicates the nodes within the valid range, N j represents a potential monitoring point, t max Indicates monitoring response time;

[0238] The expression of the monitoring point hypergraph element in step S5 is:

[0239]

[0240] Among them, E t (H) represents the hyperedge set of potential monitoring points, represents the effective monitoring range edge set of the jth potential monitoring point, and k represents the total number of potential monitoring points;

[0241] According to the response requirements, there is at least one potential monitoring point within the effective response range of the release point, so the goal of minimizing the total number of monitoring points under the constraints of monitoring point layout is established:

[0242]

[0243]

[0244]

[0245] Among them, z represents the number of monitoring deployments, x j Represents the layout variable, that is, the potential monitoring point N j Whether to set up monitoring deployment, n represents the total number of potential monitoring points, represents the potential monitoring point N j The set of nodes that can be covered;

[0246] The greedy algorithm for solving the set coverage model is specifically as follows: using Python programming language to compile and run on the Pycharm platform to achieve the transformation of the above target expression; first, from the hyperedge set E of potential monitoring points t (H) Find the set E of nodes most affected by the release concentration 0 Elements in (H) Corresponding to the set The relationship between the head set and the tail set will release the point set Monitoring point {v H1 , ..., v Hi} as the optimal monitoring layout position, and put it into the response node set whose original value is an empty set; secondly, delete the node element {v H1 , ..., v Hi}, from the hyperedge set E of potential monitoring points t (H) Delete the elements related to the head set representation vector in, Represents the hyperedge set E of potential monitoring points t (H) The edge set affected by the point i in the experiment; third, the hyperedge set E of the remaining potential monitoring points t (H) Find the set of nodes most affected by the release concentration, and repeat the above steps until the hyperedge set E of the release point is 0 (H) element is an empty set; finally, the response node set contains elements representing the optimized layout quantity and position of potential monitoring points.

[0247] Preferably, the air volume-wind resistance response efficiency attenuation rate in step S6 is defined as: the degree of change in the branch wind resistance causing the branch wind volume-wind resistance response efficiency to change itself, and the expression of the air volume-wind resistance response efficiency attenuation rate is:

[0248]

[0249] Among them, t ij represents the air volume-wind resistance response efficiency attenuation rate, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, △rj Indicates wind resistance r j A small perturbation, q i represents the wind volume at the response node j;

[0250] The expression of the power function fitting wind volume-wind resistance response efficiency curve in step S6 is:

[0251]

[0252] Among them, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, a, b and c are all constants greater than zero;

[0253] The calculation formula of the attenuation rate of the response efficiency in step S6 is:

[0254]

[0255] Among them, t ij represents the wind volume-wind resistance response efficiency attenuation rate and t ij Always negative, that is, d ij With the j The value increases and decreases, r j represents the wind resistance value of the response node j, and both a and b are constants greater than zero;

[0256] The expression of the threshold value of the wind resistance value in step S6 is:

[0257]

[0258] Among them, r jmax The wind resistance threshold, r j0 represents the initial value of wind resistance, a, b and c are all constants greater than zero, α represents the sensitivity threshold and 0<α<0.05.

[0259] Preferably, the method steps for constructing a control strategy using the deep deterministic policy gradient reinforcement learning algorithm are:

[0260] Step 1: Define the state space: The state space is composed of the target branch parameter T t , adjust the branch parameter R t , Power branch parameter P t The fan operating voltage, current and frequency are determined, and the function expression of the state space is as shown in the following formula:

[0261] s t =[T t , R t , P t ],

[0262] Among them, Tt = {q 1，t ,q 2，t , ...q n，t , p 1，t , p 2，t , ...p n，t} represents the air volume and air pressure parameters of the target branch at time t, R t ={θ 1，t ,θ 2，t , ...θ n，t} represents the opening of the stroke valve in the regulating branch at time t, P t = {I 1，t , I 2，t ,…I n，t , f 1，t , f 2，t , …f n，t} represents the fan current and frequency parameters of the power branch at time t;

[0263] Step 2: Define the action space: Execute corresponding actions according to the observed environmental parameters. The action space is determined by the air valve opening and fan current and frequency that need to be adjusted. The function representation of the action space is as shown in the following formula:

[0264] a t =[Δθ t , ΔI t , Δf t ]

[0265] Among them, Δθ t ={Δθ 1，t , Δθ 2，t , ...Δθ n，t} represents the change in the opening of the wind valve in the regulating branch at time t, ΔI t ={ΔI 1，t , ΔI 2，t , ...ΔI n，t , Δf 1，t , Δf 2，t , ...Δf n，t} represents the change in the fan current and frequency parameters of the power branch at time t;

[0266] Step 3: Define RL rewards: Minimize ventilation energy consumption while meeting the target air volume load. Reduce the deviation between the supply air volume and the required air volume in terms of demand-based ventilation revenue rewards to meet:

[0267]

[0268] Among them, Q n，t is the air supply volume of branch n at time t, is the required air volume, d Q，tis the deviation between the air supply volume and the required air volume, d Q，t,max The maximum threshold for the deviation between the supply air volume and the required air volume;

[0269] Set the long-term benefit ratio to λ1; for energy consumption benefit rewards, control the air valve opening to be greater than 70%, and the fan operates in a stable working area, that is, the working pressure is less than 90% of the maximum pressure, and the operating efficiency is greater than 60%, meeting:

[0270] θ t ≥θ t，min , Q f，min ≤Q f ≤Q f，max ,

[0271] Among them, θ t is the air valve opening, θ t，min is the minimum threshold of air valve opening, Q f is the fan air volume, Q f，min and Q f，max They are the minimum and maximum thresholds of the fan air volume range respectively;

[0272] If the long-term profit ratio is set to λ2 and λ3, the reward function is as follows:

[0273] r t (s t , a t )=[R Q ,R θ , R f ],

[0274] Among them, R Q =λ1c r，Q , c r,Q Represents the wind volume bonus coefficient, R θ =λ2c r，θ , c r，θ Indicates the air valve bonus coefficient, R θ =λ3c r，f , c r，f Indicates the wind turbine bonus coefficient;

[0275] The expression of the wind volume bonus coefficient is:

[0276]

[0277] The expression of the wind valve bonus coefficient is:

[0278]

[0279] The expression of the wind turbine bonus coefficient is:

[0280]

[0281] Step 4: Establish an Actor-Critic architecture, which includes: parameters θ u Actor current network function: u = (s, a | θ u ), where θ u Represents the parameter set of the Actor network, s represents the current state of the network, a represents the action selected based on the current state; the parameter is θ Q Critic current network function: Q = (s, a | θ Q ), where θ Q represents the parameter set of the Critic network, s represents the state received by the action network, a represents a deterministic action based on the received state; and the parameter is θ u →θ u′ and θ Q →θ Q′ The Actor-Critic corresponds to the target network, where θ u′ Represents the parameter set of the target Actor network, θ Q′ Represents the parameter set of the target Critic network;

[0282] Step 5: Actor-Critic network initialization:

[0283] Set the learning rate of the Actor network and the Critic network, set the update smoothing coefficient of the target Q network; set the number of neurons in the Actor network and the Critic network architecture, set the experience pool capacity and the number of sampling pools used for parameter updates;

[0284] Step 6: Actor-Critic network update loop:

[0285] Based on the current environment state output, the DDPG agent performs actions and obtains the next environment state and RL reward, while adding action perturbation N t , environmental state s t , perform action a t , RL reward r t and the next environment state s t+1 The samples are stored in the experience pool, and data {s t ,a t ,r t ,s t+1} to the Actor-Critic current network and update the network parameters:

[0286] θ Q′ =τθ Q +(1-τ)θ Q′ ,

[0287] Among them, θ Q Represents the parameter set of the Critic network, θ Q′ represents the parameter set of the target Critic network, and τ represents the factor that controls the update speed of the target network parameters;

[0288] θ u′ =τθ u +(1-τ)θ u ′,

[0289] Among them, θ u Represents the parameter set of the Actor network, θ u′ represents the parameter set of the target Actor network, and τ represents the factor that controls the update speed of the target network parameters;

[0290] Step 7: Continue to update the process in step 6 until the policy network learns a good enough policy.

[0291] Preferably, the air volume-wind resistance response efficiency in step S8 is defined as: the influence degree of branch wind resistance on node pressure or branch air volume, and the expression of the air volume-wind resistance response efficiency is:

[0292]

[0293] Among them, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, q i represents the wind volume at the target node i, △r j Indicates wind resistance r j A small perturbation, △q i represents the wind volume change at the target node i;

[0294] The expression of the air volume-wind resistance response efficiency matrix in step S8 is:

[0295]

[0296] Among them, D Q represents the air volume-wind resistance response efficiency matrix, d ij represents the air volume-wind resistance response efficiency, q j represents the wind volume at the response node j, r n represents the wind resistance value of the nth response node, d jn represents the wind resistance r of the response node n n The wind volume q of the response node j j The degree of influence, n represents the number of responding nodes;

[0297] The expression of the iterative sequence in step S8 is:

[0298]

[0299] dr j (k) =γdr j (k-1) ,

[0300] in, represents the wind volume-wind resistance response efficiency in the kth iteration, represents the wind volume value of target node i at the kth iteration, represents the wind volume value of target node i at the k-1th iteration, r j (k) represents the wind resistance value of the response node j at the kth iteration, r j (k-1) represents the wind resistance value of the response node j at the k-1th iteration, dr j (k) represents the wind resistance disturbance, that is, the change in wind resistance of the response node j in the kth iteration, dr j (k-1) represents the change in wind resistance of the response node j in the k-1th iteration, and γ represents the acceleration factor;

[0301] The expression of the calculation accuracy is:

[0302]

[0303] in, represents the wind volume-wind resistance response efficiency in the kth iteration, represents the wind volume-wind resistance response efficiency in the k-1th iteration, and ε represents the calculation accuracy.

[0304] Preferably, the fan power-wind resistance response efficiency in step S9 is defined as: the degree of influence of the fan output power change on the branch wind resistance change. The expression of the fan power-wind resistance response efficiency in step S9 is:

[0305]

[0306] Among them, dN ij Indicates the wind turbine power-wind resistance response efficiency, ΔN fi represents the wind turbine output power change at target node i, Δr j Represents the change in wind resistance at the response node j, N fi represents the wind turbine output power at the target node i, r j represents the wind resistance value at the response node j;

[0307] For wind turbine power-wind resistance response efficiency dN ij , fan output power N fi Satisfy the product relationship between fan pressure and fan air volume:

[0308] N fi =h f q i ,

[0309] Among them, N fi represents the wind turbine output power at the target node i, h f Indicates the fan pressure, q i represents the wind volume value at the target node i;

[0310] Fan air pressure h f The least squares fitting relationship of a univariate polynomial is satisfied, and the degree of the polynomial is 3. The output power characteristic curve of the fan is as shown in the following formula:

[0311]

[0312] Among them, N fi represents the wind turbine output power at the target node i, q i represents the wind volume value at the target node i, a0 represents the constant term, that is, when the wind volume value q i is the rate of change of fan power when is zero, a1, a2 and a3 are coefficients of the polynomial, describing the nonlinear effect of air volume on fan power;

[0313] Air volume q i Find the partial derivative:

[0314]

[0315] Among them, N fi represents the wind turbine output power at the target node i, q i represents the wind volume value at the target node i, a0 represents the constant term, that is, when the wind volume value q i is the fan power change rate when is zero, a1, a2 and a3 are the coefficients of the polynomial, describing the nonlinear effect of air volume on fan power;

[0316] The expression of the wind turbine power-wind resistance response efficiency matrix in step S9 is:

[0317]

[0318]

[0319] Among them, D N represents the wind turbine power-wind resistance response efficiency matrix, dN ij Indicates the wind turbine power-wind resistance response efficiency, Nfn represents the wind turbine output power at the response node n, r n represents the wind resistance value of the nth response node, d nn represents the wind resistance r of the response node n n The wind turbine output power N for the response node n fn The degree of influence, ΔN fi represents the wind turbine output power change at target node i, Δr j represents the change in wind resistance at the response node j, d ij It indicates the air volume-wind resistance response efficiency;

[0320] The second-order wind network response efficiency in step S9 is defined as: adjusting branch e k Effect of wind resistance change on other branches of wind network i The influence of air volume, the expression of the second-order wind network response efficiency is:

[0321]

[0322] Among them, dd ijk represents the second-order wind network response efficiency, d ij represents the air volume-wind resistance response efficiency, r k represents the wind resistance at node k, q i represents the wind volume value at the target node i, r j represents the wind resistance value at the response node j;

[0323] The three-dimensional matrix in step S9 is a second-order wind network response efficiency matrix, and the expression of the second-order wind network response efficiency matrix is:

[0324] DD(i,j,k)=dd ijk ,

[0325] Among them, DD(i, j, k) is the second-order wind network response efficiency matrix, dd ijk It represents the second-order wind network response efficiency. When j=k, it represents the influence of branch regulation on its own wind volume, that is, the attenuation rate of the effect efficiency of the regulation branch. When j≠k, it represents the mutual influence between branches. If two regulation branches e k and e j , satisfying the conditions: d ij >0,d ik >0,d ijk >0, it means that the two adjustment branches are related to branch e i Same promotion, if the conditions are met: d ij >0,d ik >0,d ijk <0, it means that the two adjustment branches are related to branch e i Same inhibition;

[0326] Take the second-order derivative of the loop wind pressure equation vector with respect to the wind resistance vector:

[0327]

[0328] Among them, (R, Q y (R)) indicates that the free variables of the loop wind pressure equations are the wind resistance vector R and the residual wind volume vector Q y , The Jacobian matrix of the circuit wind pressure equations for wind resistance is denoted as f1′. The Jacobian matrix of the loop wind pressure equations for the residual tree wind volume is denoted by f2′. is the Jacobian gradient algorithm matrix of the wind volume to wind resistance of the remaining tree, denoted as f3′;

[0329] make The analytical expression of the three-dimensional matrix of the second-order wind network response efficiency is as follows:

[0330]

[0331] in, Represents the residual tree wind volume vector Q y The second derivative of the drag vector R, f 11 ″ represents the second-order Jacobian matrix of the loop wind pressure equation for wind resistance, f 12 ″ represents the mixed second-order Jacobian matrix of the loop wind pressure equation for wind resistance and residual wind volume, f 21 ″ represents the mixed second-order Jacobian matrix of the wind pressure equation for the residual wind volume and wind resistance, f 22 ″ represents the second-order Jacobian matrix of the loop wind pressure equation for the residual tree wind volume.

[0332] A gradient pressure regulating intelligent ventilation system for underground caverns, comprising a ventilation monitoring system, a network analysis system (analyzing control parameters such as ventilation network branch air volume, wind pressure, wind resistance, fan operation characteristic curve, etc.), a control optimization system (responsible for adjusting the operation parameters and resistance of ventilation equipment according to the calculated data, and realizing real-time control of air state and wind pressure in the tunnel. Combining the calculation results of the data analysis module with the operation parameters of the ventilation system equipment in the tunnel, the control module issues instructions according to the decision results, adjusts the fan and air valve to perform actions through data transmission, and feeds back the execution data at the same time) and a visualization system (the visualization display system is based on the three-dimensional modeling of the engineering cavern group, and visualizes the parameter data such as ventilation environment, ventilation fluid, ventilation equipment and facility operation, and displays system energy consumption optimization, ventilation environment warning, and ventilation equipment failure information). The network analysis system is connected to the ventilation monitoring system and the control optimization system respectively, and the visualization system is connected to the network analysis system. The ventilation monitoring system comprises: an integrated air environment monitoring device (collecting pollutant concentration data such as temperature, humidity, carbon dioxide, carbon monoxide, oxygen, dust and harmful gases), an air volume and wind pressure monitoring device and a fan An operating parameter monitoring device (reads the state parameters of the fan operating current, voltage, power, frequency, control mode, etc., to realize real-time collection and uploading of air environment and ventilation system parameters), the integrated air environment monitoring device is arranged in the tunnel, the air volume and air pressure monitoring device is arranged inside the ventilation duct, and the fan operating parameter monitoring device is arranged near the fan and connected to the network analysis system; the network analysis system includes: an integrated intelligent fan control cabinet and a cloud server, the integrated intelligent fan control cabinet is respectively connected to the ventilation monitoring system and the control optimization system (the integrated intelligent fan control cabinet adopts a wired or wireless connection method of an industrial general standard interface), the cloud server is respectively connected to the integrated intelligent fan control cabinet and the visualization system, the integrated intelligent fan control cabinet includes: a data acquisition module, a data processing module (responsible for processing the collected data, including data cleaning, conversion, integration, storage, transmission and other tasks), a calculation and analysis module (responsible for real-time calculation of the air volume and air pressure of each branch in the tunnel, and the estimated calculation of the ventilation load according to the parameter data of the database, and the real-time solution of the parameters of each branch and node of the ventilation network based on the principle of ventilation on demand.At the same time, it has functions such as abnormal warning of air environment data, energy consumption diagnosis of ventilation system, and diagnosis and warning of fan operation faults) and a communication module (transmitting the collected data to the network solution and display platform through 5G wireless transmission or wired transmission technology), the data acquisition module is connected to the calculation and analysis module, the calculation and analysis module is connected to the data processing module, and the communication module is connected to the data processing module and the cloud server respectively; the control optimization system includes: an intelligent air valve and a variable frequency fan control cabinet (receiving instructions formed by the cloud ventilation network solution results, and completing the change of the air valve opening and the fan operation state through the proportional-integral-differential controller PID to ensure that the deviation between the required air volume and the target air volume is less than 5%), the intelligent air valve is arranged on the ventilation duct, and the variable frequency fan control cabinet is arranged near the fan, the intelligent air valve and the variable frequency fan control cabinet are both connected to the network analysis system, the opening range of the intelligent air valve is not less than 70%, and the operation efficiency of the fan is greater than 60%.

[0333] Although the specific implementation of the present invention is introduced above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that those non-creative modifications or equivalent replacements of the technical solution of the present invention, which do not depart from the protection scope of the technical solution, should be included in the scope of the claims of the present invention.

Claims

1. A gradient pressure regulating intelligent ventilation method for underground caverns based on hypergraph network, characterized in that: The following steps are involved: (1) Construction of ventilation system diagram and hypergraph structure, including the following steps: S1. Construction of hypergraph structure: First, construct the basic graph structure of the ventilation system, then introduce the hypergraph concept to characterize the development of spatial fluid associations in the ventilation regulation process, and use directed hyperedges to describe the directed relationship between two node sets; S2. Hypergraph matrix: A hypergraph matrix is ​​used to define the relationship between a node set and a hyperedge set. The expression of the hypergraph matrix structure is as follows: G={V(H),E(H),V(D),E(D),X,W}, Where V(D) represents the node set of the general graph D, and V(D) = {v1, ..., v n } contains n nodes in the graph D, E(D) represents the edge set of the general graph D, and E(D) = {e1, ..., e m } contains m edges in the graph S, V(H) represents the node set of the hypergraph H and V(H) = {v1, ..., v n }, E(H) is the hyperedge set of the hypergraph H and E(H) = {E1, ..., E k }, X represents the feature matrix of the node, and W represents the weight matrix; Each hyperedge in the hyperedge set E(H) describes T k and H k The relationship between the two subsets: E k =(T k , H k ), where T k represents the tail set of directed edge superedges and T k = {v T1 , ..., v Tk }, H k represents the head set of the edge hyperedge and H k = {v H1 , ..., v Hk }; The T k and H k is a subset of E(D), and the hypergraph association matrix is ​​as follows: B=(b ij ) m*n Where B represents the hypergraph association matrix, b ij Represents node v i With the hyperedge E j The relationship between them, m represents the number of nodes in the hypergraph H, n represents the number of hyperedges in the hypergraph H, and v i represents the i-th node in the hypergraph H, H j Represents the hyperedge E j The head set, T j Represents the hyperedge E j The last set, E j represents the jth hyperedge set in the hypergraph H; S3, convolution algorithm and hypergraph aggregation: the convolution algorithm is combined with the hypergraph aggregation process. The aggregation process includes: node information aggregation and hyperedge information aggregation. The node information aggregation is specifically as follows: according to the joint adjustment sensitivity, the feature vectors of all neighboring nodes of each node are weighted summed to form the head set and tail set representation vectors, and the node information is aggregated through the gate mechanism to finally obtain the representation vector of the directed hyperedge; the hyperedge information aggregation is specifically as follows: all directed hyperedges with the same head set are merged to obtain a new directed hyperedge, and the representation vectors of all directed hyperedges with the same head set are weighted summed to obtain the final output vector of the corresponding node. Finally, a ventilation adjustment hypergraph model is constructed; (2) The optimized monitoring layout of the collective coverage model includes the following steps: S4, the construction of the set coverage model, including: determining the target node set and building the response node combination, the determination of the target node set is specifically: by querying the performance attributes of all nodes in the ventilation adjustment map, determining the normal, abnormal, and warning categories, and screening out the target nodes to be adjusted that exceed the deviation range; the construction of the response node combination is specifically: selecting the air volume-wind resistance response efficiency in the data association between nodes as the adjustment basis, according to the output vector matrix of the corresponding node in the node information fusion process, using a multi-level relationship query method, retrieve the nodes with data association with the target node set, eliminate the nodes that do not meet the adjustable attributes and positive and negative correlation attributes, and build the response node set; S5. Optimization of the layout of monitoring points, including the following steps: Step A: By setting the release point where the concentration of harmful gases suddenly changes, the node range affected by the release point within the monitoring response time is calculated to form a release set; Step B: define potential monitoring points, calculate the effective monitoring range of potential monitoring points, establish a set of effective ranges of potential monitoring points, and construct monitoring point hypergraph elements; Step C: Use the greedy algorithm to solve the set coverage model and select the minimum number of monitoring points to ensure that there is at least one monitoring point within the effective response range of each release point; (3) The ventilation network solution and response efficiency joint adjustment method includes the following steps: S6, gradient voltage regulation model construction, including the following steps: Step D: define the air volume-wind resistance response efficiency attenuation rate, quantify the effect of wind resistance change on air volume, fit the air volume-wind resistance response efficiency curve through a power function, and calculate the response efficiency attenuation rate; Step E: Set the threshold of wind resistance value, determine the sensitive area of ​​air volume adjustment, and refine the adjustment plan to specific branches according to the air adjustment plan; S7, air adjustment scheme optimization, including the following steps: Step F: Screen the response nodes, and select the set of nodes participating in the adjustment according to the ranking, and establish the air adjustment plan within the response efficiency constraint; Step G: If the solution does not meet the target requirements, more response nodes are selected to update the air adjustment solution until the target node air volume deviation is reduced to within the deviation threshold; Step H: Each iteration selects new nodes to participate in the adjustment, forms a new air adjustment plan, and finally outputs a plan that meets the joint adjustment sensitivity requirements; (4) Control strategy based on deep deterministic policy gradient reinforcement learning algorithm: Based on the control principle of on-demand air volume supply and personalized partitioning, a control strategy is constructed using a deep deterministic policy gradient reinforcement learning algorithm. By simulating various areas in the environment, personalized control is performed on different air supply demands. (5) Response efficiency analysis and adjustment, including the following steps: S8. Construction of the air volume-wind resistance response efficiency matrix includes the following steps: Step I: Define the air volume-wind resistance response efficiency, calculate the relationship between the air volume change and wind resistance disturbance between nodes, form a response efficiency matrix, use an iterative method to continuously approach the optimal solution, and finally obtain the air volume-wind resistance response efficiency matrix of the entire ventilation network; Step J: construct an iterative sequence and recalculate the response efficiency after each air volume adjustment until the calculation accuracy requirement is met; S9, wind turbine power and wind resistance response efficiency, including the following steps: Step K: Analyze the response relationship between fan power and wind resistance, define the sensitivity of fan power change to wind resistance change, form a fan power-wind resistance response efficiency matrix, and select a branch with low power sensitivity as a wind resistance adjustment branch; Step L: Analyze the response efficiency of the second-order wind network, quantify the mutual influence between the adjustment branches, construct a three-dimensional matrix, calculate the mutual relationship between different branches, and ensure the coordination of each branch during the adjustment process.

2. According to claim 1, a hypergraph network-based gradient pressure regulation intelligent ventilation method for underground caverns, characterized in that: The method steps for constructing the basic graph structure of the ventilation system are as follows: First, a ventilation adjustment graph database is constructed based on Python and Neo4j. The data interface API is called by Python program to obtain ventilation data, and the data is preprocessed into a "unit-association-feature" triple form. The physical space connection formed by the fluid inflow section or outflow section is represented by the flow path relationship, that is, the basic association; the main association derived from the spatial relationship that reflects the coupling influence law of ventilation characteristics is represented by the data relationship; 2. Use py2neo to establish a connection between the Python program and Neo4j, batch process the construction instructions of the graph database, use Cypher query statements to define associations and assign parameters, use any node as the starting point l1 of the loop in the Neo4j graph database, find all relationship units that have basic associations with l1, and use the graph matching capability to find the shortest path in the relationship unit as the next loop node l2; 3. Use Cypher statements to find the shortest path among the nodes that have a basic association with l2, and use it as the next loop node l3. Repeat the query process, traverse the nodes in turn, and build the loop set L.

3. According to claim 1, a hypergraph network-based gradient pressure regulation intelligent ventilation method for underground caverns, characterized in that: The head set representation vector h of step S3 H The expression is: h H =(N H ⊙W H )X, Among them, h H The characterization vector representing the head set of the hyperedge, H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, and X represents the feature matrix of the node; The tail set representation vector h of step S3 T The expression is: h T =(H T ⊙W T )X, Among them, h T The characterization vector representing the tail set of the hyperedge, H T represents the characteristic matrix of the tail node in the hyperedge, W T represents the weight matrix of the tail set, and X represents the feature matrix of the node; The characterization vector h of the directed hyperedge in step S3 εi The expression is: h ε =σ(h H )⊙h T , Among them, h ε represents the characterization vector of the directed hyperedge, h H The characterization vector representing the head set of the hyperedge, h T represents the characterization vector of the tail set of the hyperedge, σ represents the activation function; The expression of the final output vector of step S3 is: h=(H H ⊙W H )h ε , Among them, h represents the node output vector after aggregating hyperedge information, H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, h ε Represents the characterization vector of the directed hyperedge; The expression of the ventilation adjustment hypergraph model in step S3 is: H ε ={V(H),(H H ⊙W H )h ε }, Among them, H ε represents the basic node set whose joint debugging sensitivity is greater than ε, ε represents the sensitivity threshold, V(H) represents the node set of the hypergraph H, and H H represents the feature matrix of the head node in the hyperedge, W H represents the weight matrix of the head set, h ε Represents the characterization vector of a directed hyperedge.

4. According to claim 1, a hypergraph network-based gradient pressure regulation intelligent ventilation method for underground cavern groups is characterized in that: The release point in step S5 is a node that has the requirement to install harmful gas concentration monitoring. The expression of the release set in step S5 is: in, represents the collection involved, represents the set of nodes affected by the released concentration wave, represents the set of release points, v j represents node j, t in the hypergraph H max represents the response time, v n Indicates the nth release point; If there are k nodes with sudden changes in the concentration of released gas in the node set V(D) of the general graph D, then the release point hyperedge set is constructed: Among them, E 0 (H) represents the hyperedge set of the release point, represents the edge set associated with the i-th release point, represents the edge set associated with the kth release point, k represents the total number of release points where the concentration of harmful gases suddenly changes; the potential monitoring point in step S5 is a node installed with a wind pressure monitoring device, and the expression of the effective monitoring range of the potential monitoring point is: in, represents the effective monitoring range set of potential monitoring points, v k Indicates the nodes within the effective range, N j represents a potential monitoring point, t max Indicates monitoring response time; The expression of the monitoring point hypergraph element in step S5 is: Among them, E t (H) represents the hyperedge set of potential monitoring points, represents the effective monitoring range edge set of the jth potential monitoring point, and k represents the total number of potential monitoring points; According to the response requirements, there is at least one potential monitoring point within the effective response range of the release point, so the goal of minimizing the total number of monitoring points under the constraints of monitoring point layout is established: Among them, z represents the number of monitoring deployments, x j Represents the layout variable, that is, the potential monitoring point N j Whether to set up monitoring deployment, n represents the total number of potential monitoring points, represents the potential monitoring point N j The set of nodes that can be covered; the greedy algorithm solves the set coverage model specifically as follows: use Python programming language to compile and run on the Pycharm platform to achieve the conversion of the above target expression; first, from the hyperedge set E of potential monitoring points t (H) Find the set E of nodes most affected by the release concentration 0 Elements in (H) Corresponding to the set The relationship between the head set and the tail set will release the point set Monitoring point {v H1 , ..., v Hi } as the optimal monitoring layout position, and put it into the response node set whose original value is an empty set; secondly, delete the node element {v H1 , ..., v Hi }, from the hyperedge set E of potential monitoring points t (H) Delete the elements related to the head set representation vector in, Represents the hyperedge set E of potential monitoring points t (H) The edge set affected by the point i in the experiment; third, the hyperedge set E of the remaining potential monitoring points t (H) Find the set of nodes most affected by the release concentration, and repeat the above steps until the hyperedge set E of the release point is 0 (H) element is an empty set; finally, the response node set contains elements representing the optimized layout quantity and position of potential monitoring points.

5. According to claim 1, a hypergraph network-based gradient pressure regulation intelligent ventilation method for underground caverns, characterized in that: The wind volume-wind resistance response efficiency attenuation rate in step S6 is defined as: the degree of change in the branch wind resistance caused by the branch wind volume-wind resistance response efficiency itself, and the expression of the wind volume-wind resistance response efficiency attenuation rate is: Among them, t ij represents the air volume-wind resistance response efficiency attenuation rate, d ij represents the air volume-wind resistance response efficiency, r j Represents the wind resistance value of the response node j, Δr j Indicates wind resistance r j A small perturbation, q i represents the wind volume at the response node j; the expression of the power function fitting wind volume-wind resistance response efficiency curve in step S6 is: Among them, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, a, b and c are all constants greater than zero; The calculation formula of the attenuation rate of the response efficiency in step S6 is: Among them, t ij represents the wind volume-wind resistance response efficiency attenuation rate and t ij Always negative, that is, d ij With the j The value increases and decreases, r j represents the wind resistance value of the response node j, and both a and b are constants greater than zero; The expression of the threshold value of the wind resistance value in step S6 is: Among them, r jmax The wind resistance threshold, r j0 represents the initial value of wind resistance, a, b and c are all constants greater than zero, α represents the sensitivity threshold and 0<α<0.

05.

6. According to claim 1, a hypergraph network-based gradient pressure regulation intelligent ventilation method for underground caverns, characterized in that: The method steps of constructing the control strategy by the deep deterministic policy gradient reinforcement learning algorithm are as follows: Step 1: Define the state space: The state space is composed of the target branch parameter T t , adjust the branch parameter R t , Power branch parameter P t The fan operating voltage, current and frequency are determined, and the function expression of the state space is as shown in the following formula: s t =[T t ,R t ,P t ], Among them, T t = {q 1，t ,q 2，t ,...q n，t , p 1，t , p 2，t , ...p n，t } represents the air volume and air pressure parameters of the target branch at time t, R t ={θ 1，t ,θ 2，t , ...θ n，t } represents the opening of the stroke valve in the branch at time t, P t = {I 1，t , I 2，t ,...I n，t , f 1，t , f 2，t , ...f n，t } represents the fan current and frequency parameters of the power branch at time t; Step 2: Define the action space: Execute corresponding actions according to the observed environmental parameters. The action space is determined by the air valve opening and fan current and frequency that need to be adjusted. The function representation of the action space is as shown in the following formula: t =[Δθ t , ΔI t , Δf t ], Among them, Δθ t ={Δθ 1，t , Δθ 2，t , ...Δθ n，t } represents the change in the opening of the wind valve in the regulating branch at time t, ΔI t ={ΔI 1，t , ΔI 2，t , ...ΔI n，t , Δf 1，t , Δf 2，t , ...Δf n，t } represents the change in the fan current and frequency parameters of the power branch at time t; Step 3: Define RL rewards: Minimize ventilation energy consumption while meeting the target air volume load. Reduce the deviation between the supply air volume and the required air volume in terms of demand-based ventilation revenue rewards to meet: Among them, Q n，t is the air supply volume of branch n at time t, is the required air volume, d Q，t is the deviation between the air supply volume and the required air volume, d Q，t，max The maximum threshold for the deviation between the supply air volume and the required air volume; Set the long-term benefit ratio to λ1; for energy consumption benefit rewards, control the air valve opening to be greater than 70%, and the fan operates in a stable working area, that is, the working pressure is less than 90% of the maximum pressure, and the operating efficiency is greater than 60%, meeting: i t ≥θ t，min ,Q f，min ≤Q f ≤Q f，max , Among them, θ t is the air valve opening, θ t，min is the minimum threshold of air valve opening, Q f is the fan air volume, Q f，min and Q f，max They are the minimum and maximum thresholds of the fan air volume range respectively; If the long-term profit ratio is set to λ2 and λ3, the reward function is as follows: r t (s t ,a t )=[R Q ,R θ ,R f ], Among them, R Q =λ1c r，Q ,c r，Q Represents the wind volume bonus coefficient, R θ =λ2c r，θ ,c r，θ Indicates the air valve bonus coefficient, R θ =λ3c r，f ,c r，f Indicates the wind turbine bonus coefficient; The expression of the wind volume bonus coefficient is: The expression of the wind valve reward coefficient is: The expression of the wind turbine bonus coefficient is: Step 4: Establish an Actor-Critic architecture, which includes: parameters θ u Actor current network function: u = (s, a | θ u ), where θ u Represents the parameter set of the Actor network, s represents the current state of the network, a represents the action selected based on the current state; the parameter is θ Q Critic current network function: Q = (s, a | θ Q ), where θ Q represents the parameter set of the Critic network, s represents the state received by the action network, a represents a deterministic action based on the received state; and the parameter is θ u →θ u′ and θ Q →θ Q′ The Actor-Critic corresponds to the target network, where θ u′ Represents the parameter set of the target Actor network, θ Q′ Represents the parameter set of the target Critic network; Step 5: Actor-Critic network initialization: Set the learning rate of the Actor network and the Critic network, set the update smoothing coefficient of the target Q network; set the number of neurons in the Actor network and the Critic network architecture, set the experience pool capacity and the number of sampling pools used for parameter updates; Step 6: Actor-Critic network update loop: Based on the current environment state output, the DDPG agent performs actions and obtains the next environment state and RL reward, while adding action perturbation N t , environmental state s t , perform action a t , RL reward r t and the next environment state s t+1 The samples are stored in the experience pool, and data {s t , a t , r t ,s t+1 } to the Actor-Critic current network and update the network parameters: i Q′ =tθ Q +(1-τ)θ Q′ , Among them, θ Q Represents the parameter set of the Critic network, θ Q′ represents the parameter set of the target Critic network, and τ represents the factor that controls the update speed of the target network parameters; i u′ =tθ u +(1-τ)θ u′ , Among them, θ u Represents the parameter set of the Actor network, θ u′ represents the parameter set of the target Actor network, and τ represents the factor that controls the update speed of the target network parameters; Step 7: Continue to update the process in step 6 until the policy network learns a good enough policy.

7. The method for gradient pressure regulation intelligent ventilation of underground caverns based on hypergraph network according to claim 1 is characterized in that: The definition of the air volume-wind resistance response efficiency in step S8 is: the influence degree of branch wind resistance on node pressure or branch air volume. The expression of the air volume-wind resistance response efficiency is: Among them, d ij represents the air volume-wind resistance response efficiency, r j represents the wind resistance value of the response node j, q i represents the wind volume at the target node i, Δr j Indicates wind resistance r j A small perturbation, Δq i represents the wind volume change at the target node i; The expression of the air volume-wind resistance response efficiency matrix in step S8 is: Among them, D Q represents the air volume-wind resistance response efficiency matrix, d ij represents the air volume-wind resistance response efficiency, q j represents the wind volume at the response node j, r n represents the wind resistance value of the nth response node, d jn represents the wind resistance r of the response node n n The wind volume q of the response node j j The degree of influence, n represents the number of responding nodes; The expression of the iterative sequence in step S8 is: in, represents the wind volume-wind resistance response efficiency in the kth iteration, represents the wind volume value of target node i at the kth iteration, represents the wind volume value of target node i at the k-1th iteration, r j (k) represents the wind resistance value of the response node j at the kth iteration, r j (k-1) represents the wind resistance value of the response node j at the k-1th iteration, dr j (k) represents the wind resistance disturbance, that is, the change in wind resistance of the response node j in the kth iteration, dr j (k-1) represents the change in wind resistance of the response node j in the k-1th iteration, and γ represents the acceleration factor; The expression of the calculation accuracy is: in, represents the wind volume-wind resistance response efficiency in the kth iteration, represents the wind volume-wind resistance response efficiency in the k-1th iteration, and ε represents the calculation accuracy.

8. The method for gradient pressure regulation intelligent ventilation of underground caverns based on hypergraph network according to claim 1 is characterized in that: The definition of the fan power-wind resistance response efficiency in step S9 is: the influence degree of the fan output power change on the branch wind resistance change. The expression of the fan power-wind resistance response efficiency in step S9 is: Among them, dN ij Indicates the wind turbine power-wind resistance response efficiency, ΔN fi represents the wind turbine output power change at target node i, Δr j Represents the change in wind resistance at the response node j, N fi represents the wind turbine output power at the target node i, r j represents the wind resistance value at the response node j; For wind turbine power-wind resistance response efficiency dN ij , fan output power N fi Satisfy the product relationship between fan pressure and fan air volume: N fi =h f q i , Among them, N fi represents the wind turbine output power at the target node i, h f Indicates the fan pressure, q i represents the wind volume value at the target node i; Fan air pressure h f The least squares fitting relationship of a univariate polynomial is satisfied, and the degree of the polynomial is 3. The output power characteristic curve of the fan is as shown in the following formula: Among them, N fi represents the wind turbine output power at the target node i, q i represents the wind volume value at the target node i, a0 represents the constant term, that is, when the wind volume value q i is the fan power change rate when is zero, a1, a2 and a3 are the coefficients of the polynomial, describing the nonlinear effect of air volume on fan power; Air volume q i Find the partial derivative: Among them, N fi represents the wind turbine output power at the target node i, q i represents the wind volume value at the target node i, a0 represents the constant term, that is, when the wind volume value q i is the fan power change rate when is zero, a1, a2 and a3 are the coefficients of the polynomial, describing the nonlinear effect of air volume on fan power; The expression of the wind turbine power-wind resistance response efficiency matrix in step S9 is: Among them, D N represents the wind turbine power-wind resistance response efficiency matrix, dN ij Indicates the wind turbine power-wind resistance response efficiency, N fn represents the wind turbine output power at the response node n, r n represents the wind resistance value of the nth response node, d nn represents the wind resistance r of the response node n n The wind turbine output power Nf for the response node n n The degree of influence, ΔN fi represents the wind turbine output power change at target node i, Δr j represents the change in wind resistance at the response node j, d ij It indicates the air volume-wind resistance response efficiency; The second-order wind network response efficiency in step S9 is defined as: adjusting branch e k Effect of wind resistance change on other branches of wind network i The influence of air volume, the expression of the second-order wind network response efficiency is: Among them, dd ijk represents the second-order wind network response efficiency, d ij represents the air volume-wind resistance response efficiency, r k represents the wind resistance at node k, q i represents the wind volume value at the target node i, r j represents the wind resistance value at the response node j; The three-dimensional matrix in step S9 is a second-order wind network response efficiency matrix, and the expression of the second-order wind network response efficiency matrix is: DD(i,j,k)=dd ijk , Among them, DD(i, j, k) is the second-order wind network response efficiency matrix, dd ijk It represents the second-order wind network response efficiency. When j=k, it represents the influence of branch regulation on its own wind volume, that is, the attenuation rate of the effect efficiency of the regulation branch. When j≠k, it represents the mutual influence between branches. If two regulation branches e k and e j , satisfying the conditions: d ij >0,d ik >0,d ijk >0, it means that the two adjustment branches are related to branch e i Same promotion, if the conditions are met: d ij >0,d ik >0,d ijk <0, it means that the two adjustment branches are related to branch e i Same suppression; calculate the second-order derivative of the loop wind pressure equation vector to the wind resistance vector: Among them, (R, Q y (R)) indicates that the free variables of the loop wind pressure equations are the wind resistance vector R and the residual wind volume vector Q y , The Jacobian matrix of the circuit wind pressure equations for wind resistance is denoted as f1′. The Jacobian matrix of the loop wind pressure equations for the residual tree wind volume is denoted by f2′. is the Jacobian gradient algorithm matrix of the wind volume to wind resistance of the remaining tree, denoted as f3′; make The analytical expression of the three-dimensional matrix of the second-order wind network response efficiency is as follows: in, Represents the residual tree wind volume vector Q y The second derivative of the drag vector R, f″ 11 The second-order Jacobian matrix of the loop wind pressure equation for wind resistance, f″ 12 The mixed second-order Jacobian matrix of the loop wind pressure equation for wind resistance and residual wind volume, f″ 21 The mixed second-order Jacobian matrix of the wind pressure equation for the residual wind volume and wind resistance, f″ 22 The second-order Jacobian matrix representing the loop wind pressure equation for the residual tree wind volume.

9. An intelligent ventilation system for gradient pressure regulation of underground caverns, characterized in that: It includes a ventilation monitoring system, a network analysis system, a control optimization system and a visualization system. The network analysis system is connected to the ventilation monitoring system and the control optimization system respectively, and the visualization system is connected to the network analysis system. The ventilation monitoring system includes: an integrated air environment monitoring device, an air volume and air pressure monitoring device and a fan operation parameter monitoring device. The integrated air environment monitoring device is arranged in the tunnel, the air volume and air pressure monitoring device is arranged inside the ventilation duct, and the fan operation parameter monitoring device is arranged near the fan and connected to the network analysis system; the network analysis system includes: an integrated intelligent fan control cabinet and a cloud server, and the integrated intelligent fan control cabinet is connected to the ventilation monitoring system and the control optimization system respectively. The cloud server is connected to the integrated intelligent fan control cabinet and the visualization system respectively, the integrated intelligent fan control cabinet includes: a data acquisition module, a data processing module, a calculation and analysis module and a communication module, the data acquisition module is connected to the calculation and analysis module, the calculation and analysis module is connected to the data processing module, and the communication module is connected to the data processing module and the cloud server respectively; the control optimization system includes: an intelligent air valve and a variable frequency fan control cabinet, the intelligent air valve is arranged on the ventilation duct, the variable frequency fan control cabinet is arranged near the fan, the intelligent air valve and the variable frequency fan control cabinet are both connected to the network analysis system, the opening range of the intelligent air valve is not less than 70%, and the operation efficiency of the fan is greater than 60%.

Images