Rail transit plane line-station collaborative optimization method and system based on hybrid reinforcement learning

Abstract

The application provides a rail transit plane line-station collaborative optimization method and system based on hybrid reinforcement learning, belongs to the technical field of rail transit operation control, constructs a graph structure of candidate stations, and takes comprehensive urban space information data of grid points as attributes of nodes in the graph; a pre-trained hierarchical perception multi-target graph pointer network is used to process the graph structure, the selection probability of the next station in each step is calculated through iteration, and multiple candidate station sequences are generated under the guidance of multi-target preferences; for each group of station sequences, a connection path between stations is generated, and these paths are fitted to obtain a global line scheme. The application accurately models the spatial heterogeneity between stations and the influence of future potential stations on the current selection; avoids the high computational burden of multiple model training, improves the optimization efficiency and the generalization ability of the model; significantly improves the efficiency of path generation and the speed of feasibility judgment, and adapts to the path planning between stations under complex constraint conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit operation control technology, and particularly relates to a rail transit plane line-station collaborative optimization method and system based on hybrid reinforcement learning. BACKGROUND

[0002] With the rapid development of urban rail transit, the imbalance between supply and demand within the network is increasingly prominent. Local network optimization through extension lines or connecting lines has become a key strategy for major cities to fully tap the potential of existing rail transit systems. This not only improves the service quality for existing passengers, but also helps to enhance the attractiveness of the rail system to potential users. Therefore, under the constraints of the existing rail transit network, optimizing the subway line, including the optimization of line orientation and station location, is crucial for traffic planners, operators, and governments, and can maximize network efficiency and long-term sustainable development.

[0003] However, the design of rail transit line orientation and station location needs to consider multiple factors, such as urban spatial structure, urban land use conditions, rail network topology, and planned infrastructure; this process also needs to balance multiple conflicting optimization objectives. In addition, the optimization of station location and line orientation is interdependent, and the complex interaction between the two also poses higher challenges to the design process. Currently, the design of urban rail transit lines still mainly relies on the experiential "trial and error" method, and designers design multiple alternative schemes based on design specifications and local conditions, and then evaluate the schemes according to engineering cost and system performance indicators. This process is time-consuming when faced with large-scale search space and complex constraint conditions, and it is difficult to ensure that the optimal solution is obtained. Therefore, the collaborative optimization of rail transit line orientation and station location is a highly complex and time-consuming task.

[0004] Route optimization: Since the 1960s, researchers have invested a great amount of effort in the field of route optimization for both rail and road networks. This field has introduced several mathematical optimization methods, such as dynamic programming, mixed-integer linear programming, mesh adaptive direct search (MADS), sequential quadratic programming, and quasi-network flow models. However, these mathematical methods often require over-simplification of the optimization problem, limiting their applicability in real-world cases. In recent years, heuristic algorithms have received increasing attention due to their flexibility and efficiency in handling complex optimization problems. The most representative of these include genetic algorithms (GAs) and particle swarm optimization (PSO). Additionally, simulated annealing and ant colony optimization have also been applied to route optimization. However, in heuristic algorithms, the number of intersections (PIs) is usually pre-determined and remains constant throughout the optimization process, which may affect the quality of the solution. Grid-based methods, such as the distance transform (DT) algorithm, have also been used to search for optimal paths by abstracting the study area into a series of grids. These DT paths, which are typically piecewise straight lines, can serve as initial solutions for further optimization using heuristic algorithms. Although grid-based methods are practical in practice, they are highly sensitive to resolution, with higher resolutions significantly increasing computational costs. In recent years, tree-based search methods, such as Dijkstra's algorithm, Monte Carlo tree search, sequential search algorithm, and rapidly-exploring random tree (RRT), have emerged as promising alternatives for fast route optimization. Unlike grid-based methods, these methods generate feasible paths by expanding a search tree through random sampling, balancing efficiency with optimization flexibility.

[0005] Station optimization: Research on station location optimization in transportation networks can be broadly categorized into two groups. The first group focuses on determining the optimal station locations based on a given route layout, but this approach often overlooks the potential impact of station locations on the overall route, potentially missing critical transportation nodes. The second group selects a number of stations from a pre-determined set of candidate points and determines their arrangement order. In terms of optimization objectives, existing research has considered net benefits, user travel costs, operating and construction costs, accessibility, demand coverage, and total system cost. To achieve these objectives, researchers have employed various optimization techniques, including genetic algorithms (GA), adaptive large neighborhood search (ALNS) algorithm, mixed-integer optimization method, tabu search (TS), memetic algorithm, max-min ant system (MMAS), clustering heuristic, meta-heuristic method, and ε-constraint method. Although these studies provide valuable insights for practical station location design, these methods often simplify the connections between stations as straight lines. In high-density urban areas, route geometry is often a decisive factor, which limits the practical application of these methods.

[0006] Coordinated optimization: In addition to the separate optimization of route alignment and station location, some studies have begun to focus on the simultaneous optimization of the two elements. Lai and Schonfeld et al. considered the passenger demand and system cost of rail transit, and used genetic algorithm to optimize the route alignment and station location simultaneously. However, the optimization results are highly sensitive to the initial configuration. Pu et al. proposed a decision tree (DT) algorithm combined with backtracking strategy to realize the simultaneous optimization of route alignment and station location, aiming at the coupling constraints between the two elements. Subsequently, Song et al. focused on the simultaneous optimization of route alignment and station location in high-speed rail systems, taking into account investment and revenue. However, the search range of station location is limited to specific line sections, excluding many possible better station layout schemes. To solve the above problems, Wan et al. proposed a distance transformation algorithm based on event tree, and designed an operator for station search, thus realizing the simultaneous optimization of route alignment and station location, and enhancing the diversity of solution space. These methods mostly follow the step-by-step optimization strategy, i.e., gradually advancing the calculation and adjustment, which easily overemphasizes the local area and lacks systematic understanding of the global situation, thus increasing the risk of falling into suboptimal solutions.

[0007] The simultaneous optimization of route alignment and station location in urban rail transit needs to integrate the comprehensive effects of multiple influencing factors, and is therefore a complex task that requires a comprehensive and global perspective. Under such high complexity, accurately grasping global information and continuously exploring valuable solutions from the overall perspective is the key to achieving optimal results. In recent years, deep reinforcement learning methods have been applied to route alignment optimization, and have shown their feasibility in global perception in terms of evaluating the value of local design in the optimization process. However, these methods generally ignore the optimization of station location. A deep reinforcement learning framework based on PointerNetwork has been used to solve the node sequence optimization problem, which selects the optimal node sequence by evaluating the global node state. This method has been extended to a number of real-world tasks, such as the traveling salesman problem (TSP), vehicle routing, and computer disk defragmentation. Station location selection is essentially the selection of the optimal sequence from candidate locations, so PointerNetwork has potential application prospects in this task.

[0008] To optimize the line layout and station location simultaneously, there are still several key issues to be solved. First, the selection of the current station sequence is affected by the characteristics of all candidate stations. Specifically, the selection of the current station is not only affected by the adjacent stations, but also varies among different stations, reflecting spatial heterogeneity. In addition, the possible future station sequence also significantly affects the selection process of the current station, and existing research still has deficiencies in dealing with this influence. Second, the selection of station location and the generation of line are interdependent processes, and there are multiple coupling constraints between them. Therefore, efficient generation of high-quality lines during the station selection process is the key to achieving the coordinated optimization of the two. Finally, the existing multi-objective Pointer Network method usually needs to train multiple reinforcement learning models for different single-objective sub-problems, which not only brings a large computational burden, but also may cause the problem of mismatch between the preset weight vector and the Pareto frontier, affecting the optimization effect. SUMMARY

[0009] The present application aims to provide a hybrid reinforcement learning-based rail transit plane line-station collaborative optimization method and system to solve at least one of the technical problems in the background art. First, a hierarchical perception multi-objective graph pointer network (HA-MOGPN) is designed to generate a station sequence. This method introduces a graph attention network (GAT) in the encoding layer to capture the spatial context influence between stations; a hierarchical structure is introduced in the decoding layer to simulate the dynamic influence of future potential stations on the current selection. At the same time, an adaptive weight sharing mechanism is adopted, making it possible to achieve multi-objective optimization in a single network. Then, a parallel multi-branch RRT-Connect algorithm is proposed, which enhances the bidirectional search capability through a multi-branch strategy and constructs a rich node reserve. On this basis, combined with GPU parallel computing, an efficient station-to-station path solution can be quickly generated. Finally, a hierarchical Actor-Critic framework is constructed to train the HA-MOGPN, in which a reward function is constructed based on the line-station combination solution to guide the optimization process of network parameters, thereby achieving the overall optimization goal.

[0010] To achieve the above-mentioned purposes, the present application adopts the following technical solutions:

[0011] In a first aspect, the present application provides a hybrid reinforcement learning-based rail transit plane line-station collaborative optimization method, comprising:

[0012] First, a graph structure of candidate stations is constructed, and the comprehensive urban spatial information data of grid points is taken as the attributes of nodes in the graph;

[0013] The pre-trained hierarchical perception multi-objective graph pointer network is used to process the graph structure, and a plurality of candidate station sequences are generated under the guidance of multi-objective preferences by iteratively calculating the selection probability of the next station in each step; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is used to encode the geographical graph by using a graph attention network to capture the spatial heterogeneity influence of adjacent stations on station sequence selection, and node attributes are mapped to different sub-problem spaces through adaptive weighting to reflect various types of objective preferences; the bottom layer network of the hierarchical context perception decoder provides the possibility of future potential paths through backtracking, and the upper layer network is used to iteratively select the next station by fusing historical state and future state information, and finally generate a complete station sequence.

[0014] For each set of station sequences, a parallel multi-branch RRT-connect algorithm is used to generate connection paths between stations, and these paths are fitted to obtain a global line scheme.

[0015] As a further limitation of the first aspect of the application, training the hierarchical perception multi-objective graph pointer network comprises: first training the lower layer network and fixing its optimal parameters; then, the upper layer network receives the latent variables from the lower layer to learn its own strategy; considering that the optimization model simultaneously contains two objectives of passenger flow intensity and construction cost, the actual reward function should take these two indicators into account, and a reward item related to the feasibility of the solution is introduced: a positive reward is given to the solution that meets all the constraints, and a negative reward is given if any constraint is violated.

[0016] As a further limitation of the first aspect of the application, the multi-objective graph encoder comprises a target encoder and a weight encoder; the purpose of the target encoder is to obtain the node information of each target in the graph, and the weight encoder adaptively and dynamically generates a set of most effective weight vectors to aggregate the embedding of each target; the target encoder is composed of m graph attention network submodules, which encode the node feature vector of the geographical graph into m target embedding spaces; for each target, the graph is input into the corresponding graph attention network submodule to obtain the target embedding of all nodes; the graph attention network submodule nonlinearly updates the node state by iteratively aggregating the information of the node neighborhood using a self-attention mechanism; the weight encoder adopts a two-layer linear layer, and the input is N weight vectors and the output is the predicted most suitable weight matrix.

[0017] As a further limitation of the first aspect of the application, the hierarchical context-aware decoding layer comprises two layers of networks, both of which are constructed based on the neighborhood masking GRU-attention module: (1) the lower layer decoder performs a backward search to evaluate the feasibility of the future; (2) the upper layer decoder performs a forward search to encode historical trajectory information; wherein the neighborhood masking GRU-attention module is used to solve the node sequence selection problem.

[0018] As a further limitation of the first aspect of the application, the parallel multi-branch RRT-connect algorithm comprises: initializing search trees for all stations simultaneously, with each station as the root node of a tree, and expanding multiple branches from the root node concurrently; each tree attempts to grow towards its adjacent two trees, tree B finds the nearest nodes of its newly generated nodes in adjacent trees A and C, and makes these nearest nodes extend along the corresponding new node direction by a fixed length, this process continues until an obstacle is encountered or a connection with the adjacent tree is made; once two trees meet, a connecting path is formed, and the process is iterated until all trees corresponding to all stations are connected, forming a complete path; all possible global paths are generated and evaluated according to construction cost, and the path with the lowest cost is selected as the initial global route.

[0019] As a further limitation of the first aspect of the application, the path fitting comprises: ensuring that each station and its bidirectional adjacent vertices are on the same straight line; if not on a straight line, extending a new vertex by a fixed length along the direction connecting the station and the adjacent vertex; then selecting the adjustment mode with the lowest construction cost; Step 2: fix the bidirectional adjacent vertices of each station as the starting point or ending point of the path fitting between stations; then calculate the distance from each vertex on the path to the initial straight line connecting the starting point and the ending point, and the vertex with the maximum distance is considered as an intersection point; based on the intersection point, divide the initial edge into two sub-edges, then configure the minimum circular curve radius for the intersection point, and calculate the circular curve length and the tangent length between the adjacent two circular curves; if both are greater than the minimum limit length, the intersection point is feasible, otherwise it is not feasible; Step 4: repeat step 3 until there are no new intersection points in the path; finally complete the fitting of the path to generate the route.

[0020] In a second aspect, the application provides a track transportation plane line-station collaborative optimization system based on hybrid reinforcement learning, comprising:

[0021] The construction module is configured to construct a graph structure of candidate stations, and to use the integrated urban space information data of the grid points as the attributes of the nodes in the graph;

[0022] The processing module is configured to process the graph structure by using a pre-trained hierarchical perception multi-objective graph pointer network, to generate a plurality of candidate station sequences by iteratively calculating a selection probability of a next station in each step under the guidance of multi-objective preferences; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is configured to encode a geographical graph by using a graph attention network, to capture the spatial heterogeneity influence of adjacent stations on station sequence selection, and to map node attributes to different sub-problem spaces by adaptive weighting, so as to reflect various types of objective preferences; the bottom layer network of the hierarchical context perception decoder is configured to provide the possibility of a future potential path by reverse search, and the upper layer network is configured to iteratively select a next station by fusing historical state and future state information, and finally generate a complete station sequence.

[0023] The computing module is configured to, for each group of station sequences, generate connection paths between stations by using a parallel multi-branch RRT-connect algorithm, and fit the paths to obtain a global line scheme.

[0024] In a third aspect, the present application provides a non-transitory computer readable storage medium for storing computer instructions, which, when executed by a processor, implement the mixed reinforcement learning based rail transit plane line-station collaborative optimization method according to the first aspect.

[0025] In a fourth aspect, the present application provides a computer device comprising a memory and a processor, the processor and the memory being in communication with each other, the memory storing program instructions executable by the processor, and the processor invoking the program instructions to execute the mixed reinforcement learning based rail transit plane line-station collaborative optimization method according to the first aspect.

[0026] In a fifth aspect, the present application provides an electronic device comprising a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes instructions for implementing the mixed reinforcement learning based rail transit plane line-station collaborative optimization method according to the first aspect.

[0027] The application has the advantages of: accurately modeling spatial heterogeneity between stations and the influence of future potential stations on current selection; adopting an adaptive weight sharing mechanism to realize multi-objective optimization in a single network, avoiding the high computational burden of multiple model training in the prior art, improving optimization efficiency and model generalization ability; significantly improving the efficiency and feasibility judgment speed of path generation, adapting to inter-station path planning under complex constraint conditions; effectively guiding the collaborative optimization of station location and line design, and generating high-quality solutions that meet actual engineering requirements.

[0028] The advantages of the additional aspects of the application will be more apparent from the following description section or will be understood through the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0030] Figure 1 The urban space information synthesis model described in the embodiments of the application is shown in the schematic diagram.

[0031] Figure 2 The subway line described in the embodiments of the application is shown in the schematic diagram.

[0032] Figure 3 The hierarchical perception multi-objective graph pointer network structure described in the embodiments of the application is shown in the schematic diagram.

[0033] Figure 4 The station selection process described in the embodiments of the application is shown in the schematic diagram.

[0034] Figure 5 The line search process described in the embodiments of the application is shown in the schematic diagram.

[0035] Figure 6 The path fitting process described in the embodiments of the application is shown in the schematic diagram. DETAILED DESCRIPTION

[0036] The embodiments of the application will be described in detail below, and the examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below through the drawings are exemplary and are only used to explain the application, and cannot be interpreted as a limitation on the application.

[0037] As used herein, the terms "have," "has," "have," "having," "include," "includes," "including," "indude," "indudes," "induding," or "indudeing" are used to indicate one or more of the stated features, elements, steps, or components or groups thereof; are used to indicate that a stated feature, element, step or component is among the ones listed; and are used to indicate that a stated feature, element, step or component is present with possible embodiments or examples. Further, it will be understood that any listed ranges are inclusive of the beginning and ending range values, and further that the ranges are to be understood to encompass ranges from the lower value to the upper value, unless expressly stated otherwise or indicated by context to the extent that the use of include and / or including would render the statement ambiguously broad.

[0038] It should also be understood that any listed ranges are intended to include all sub-ranges within the range, and that the disclosure should be interpreted to include ranges that are narrower than the ranges expressly stated. For example, a range of "1 to 10" is intended to include all sub-ranges between and including the recited minimum value of 1 and the recited maximum value of 10, that is, all sub-ranges beginning with a minimum of 1 or more and ending with a maximum of 10 or less, as well as all sub-ranges beginning with a minimum of 1 or more and ending with a maximum of 10 or less, and including the minimum and maximum values.

[0039] As used herein, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise.

[0040] In this description and in the claims that follow, reference can be made to terms such as "one embodiment," "some embodiments," "an example," "a specific example," or "some examples" meant to introduce a described feature, structure, material, or characteristic. However, the

[0041] To facilitate an understanding of this application, many aspects are described in specific examples of embodiments, and specific language is used herein for the purpose of facilitating an understanding. The specific examples of embodiments described herein are not intended to be limiting of the scope of the present application, but are for the purpose of illustrating the features of the present application.

[0042] Those skilled in the art will understand that the drawings described are merely illustrative of embodiments of the application and that the size of the various elements in the figures can not be necessarily to scale.

[0043] Under the constraints of existing urban rail transit network, the collaborative optimization of metro line and station location is of great significance to fully tap the potential of existing rail transit system and improve service efficiency and sustainability. However, due to the complex coupling constraints such as space and planning, as well as the mutual influence between line and station layout, it is extremely challenging to perform line-station collaborative optimization in a large search space. To address the above problems, the present invention proposes a hierarchical deep reinforcement learning (HDRL) method to realize the collaborative optimization of station layout and line alignment. Specifically, first, a hierarchical aware multi-objective graph pointer network (HA-MOGPN) is designed to generate station sequence. Based on the graph self-attention network and adaptive weight sharing mechanism, a multi-objective graph encoder is designed. A hierarchical context-aware decoder is integrated to capture the spatial heterogeneity of neighborhood influence and the potential influence of future candidate stations on the selection of the current station sequence. Then, a parallel multi-branch RRT-connect algorithm is designed to generate the line alignment between stations. These modules are integrated into a hierarchical Actor-Critic training framework. Based on the reward function calculated by the line-station joint scheme, the optimization of network parameters is guided, and the agent training is realized. The trained agent can intelligently generate a line-station collaborative scheme based on the given state of the study area.

[0044] Example 1

[0045] In this embodiment 1, first provide a track traffic plane line-station collaborative optimization system based on hybrid reinforcement learning, comprising: a construction module for constructing a graph structure of candidate stations, and taking the integrated urban space information data of the grid points as the attributes of the nodes in the graph; a processing module for processing the graph structure by using a pre-trained hierarchical perception multi-objective graph pointer network, generating multiple candidate station sequences under the guidance of multi-objective preference by iteratively calculating the selection probability of the next station in each step; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is used to encode the geographic graph by using a graph attention network to capture the spatial heterogeneity influence of adjacent stations on station sequence selection, and the node attributes are mapped to different sub-problem spaces through adaptive weighting to reflect various target preferences; the bottom network of the hierarchical context perception decoder provides the possibility of future potential paths through backtracking, and the upper network is used to iteratively select the next station by fusing historical state and future state information, and finally generate a complete station sequence; a calculation module for generating connection paths between stations by using a parallel multi-branch RRT-connect algorithm for each set of station sequences, and fitting these paths to obtain a global line scheme.

[0046] In this embodiment, the above system is used to realize a track traffic plane line-station collaborative optimization method based on hybrid reinforcement learning, comprising: constructing a graph structure of candidate stations, and taking the integrated urban space information data of the grid points as the attributes of the nodes in the graph; processing the graph structure by using a pre-trained hierarchical perception multi-objective graph pointer network, generating multiple candidate station sequences under the guidance of multi-objective preference by iteratively calculating the selection probability of the next station in each step; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is used to encode the geographic graph by using a graph attention network to capture the spatial heterogeneity influence of adjacent stations on station sequence selection, and the node attributes are mapped to different sub-problem spaces through adaptive weighting to reflect various target preferences; the bottom network of the hierarchical context perception decoder provides the possibility of future potential paths through backtracking, and the upper network is used to iteratively select the next station by fusing historical state and future state information, and finally generate a complete station sequence; for each set of station sequences, a parallel multi-branch RRT-connect algorithm is used to generate connection paths between stations, and these paths are fitted to obtain a global line scheme.

[0047] In this embodiment, a city space information synthesis model (CUSIM) is used to represent the urban space information of the study area, which is crucial for optimizing the station sites and lines of rail transit. As shown in FIG. 1, the CUSIM is composed of three parts: the grid points, the nodes and the edges. The grid points represent the basic unit of the urban space, and the nodes represent the candidate stations. The edges represent the connection between the nodes, and the weight of the edge represents the cost of the connection path between the nodes. Figure 1As shown, CUSIM consists of a finite number of square grid cells, each ranging from 10 m to 1 km. All spatial information can be stored in these grids, including whether a grid is in the forbidden zone, whether a grid is a candidate station, centrality, road density, land use characteristics, and unit cost. The model is implemented by the following steps:

[0048] (1) Make a GIS map of the study area, perform rasterization, and form a grid layer.

[0049] (2) Superimpose the land use type layer, assign unit cost to each grid, and form a cost layer.

[0050] (3) Define the forbidden zone, superimpose with the cost layer, and form a feasible area layer.

[0051] (4) Combine the road network with the feasible area layer, and determine the grid within the 200-meter road buffer as a potential station location.

[0052] (5) Superimpose the existing rail transit network, and designate the grid containing the existing station as a candidate station.

[0053] (6) Summarize the spatial attributes of land use POI, population, house price, road density, etc. within the 800m buffer of the candidate station grid.

[0054] According to these steps, a candidate library of potential station locations can be generated, as well as the spatial information required for optimizing station locations and alignment.

[0055] A subway horizontal alignment can be defined by a set of horizontal plane intersections (HPIs), and station locations can be determined by the coordinates (x si ,y si ) of the station center, as shown in Figure 2 . The data of HPIs include intersection coordinates (x i ,y i ), horizontal curve radius r i , and easement curve length L Ti , where the easement curve length is not a design variable because it can be automatically calculated according to the curve radius r i and the design speed V. Therefore, in the collaborative optimization problem of rail transit line alignment and station location, all design variables can be represented by five column vectors:

[0056] x = [x1, x2, …, x m ] T

[0057] y = [y1, y2, …, y m ] T

[0058] r = [r1, r2, …, m ] T

[0059] x s = [x s1 , x s2 , …, x sn ] T

[0060] y s = [y s1 , y s2 , …, y sn ] T

[0061] The objective of the model is to maximize the passenger intensity of the new optimized metro line, which ensures that the line optimization focuses more on the efficiency of attracting passengers rather than simply relying on the extension of the line length to increase the total passenger flow.

[0062] The calculation method of passenger intensity is where L represents the length of the new line, and Q represents the total passenger flow attracted by the optimized line, which is calculated as Q = q1 + q2 + q3 + q4; where q1 represents internal passenger flow, representing passengers traveling the entire new line; q2 represents out-of-station transfer passenger flow, representing passengers transferring from the new line to other lines; q3 represents in-station transfer passenger flow, representing passengers transferring from other lines to the new line; and q4 represents through passenger flow, representing passengers passing through the new line.

[0063] The calculation of through passenger flow requires both OD flow prediction and network-wide flow distribution, which is very costly. To simplify this, the following steps are implemented: (1) Calculate the shortest paths for all OD pairs in the existing rail transit network using Dijkstra's algorithm. (2) If the shortest path to the starting or ending point of the new line exceeds 20 kilometers or involves more than two transfers, the od pairs related to the existing stations will be deleted. (3) Integrate the optimized line into the network and recalculate the shortest paths for the remaining OD pairs between existing stations, filtering out OD pairs with unchanged shortest paths. (4) Predict the OD flow between new stations, from new stations to existing stations, from existing stations to new stations, and the remaining OD flow between existing stations, which correspond to internal passenger flow, out-of-station transfer passenger flow, in-station transfer passenger flow, and through passenger flow, respectively.

[0064] Past studies use machine learning algorithms to predict the passenger volume of a station by featuring the land properties around the station. In this embodiment, the station-related attributes (such as centrality, population, road density, house price, and land use POI) are combined with the trip-related attributes (including travel distance, transfer times, and detour index) to predict the OD volume. The machine learning model is built based on a cascade framework. First, the inbound and outbound passenger flow predictors are trained based on station-related attributes, and then the predicted inbound and outbound passenger flows, together with station-related attributes and trip-related attributes, are input into the OD predictor for OD volume prediction. The LightGBM cascade model is trained based on historical data, and once a new line is generated, the trained model is directly used to predict the OD passenger flow and calculate the passenger flow intensity.

[0065] The present object aims to minimize the construction cost of a new subway line, which includes three parts: the cost related to the length of the line, the cost related to the value of the land, and the construction cost of the station.

[0066] (1) The cost related to the length of the line: Most of the construction cost of a subway line is proportional to the length of the line, including civil engineering, track, electrical equipment, and other auxiliary facilities: l = (U c + U e + U r + U s ) × L; where U c , U e , U r , and U s represent the cost of civil engineering, track, electrical equipment, and auxiliary facilities per unit length (yuan / m), respectively, and L represents the total length of the line (m).

[0067] (2) The cost related to the value of the land: If the subway line encroaches on the safety range of existing buildings, the relevant buildings may need to be demolished. Even for underground lines, due to the prohibition of deep foundation construction and excavation, the land along the line will face development restrictions, resulting in a loss of land value. The cost per unit length of different types of land is different, and the total land value-related cost is the sum of the costs of all land blocks crossed by the line:

[0068] where T k is the land type of the land block, g(·) is the unit length cost of a specific land type, and l k is the length of the subway line passing through the land block.

[0069] (3) Station construction cost: The construction cost c s (τ) of each station is related to its scale level τ. Stations are divided into three levels according to passenger demand Q s (passengers / day): Level 1: Q s≥ 15000, Secondary: 5000 ≤ Q s < 15000, Tertiary: Q s < 5000. The total station construction cost is the sum of the cost of all stations along the line: where Qn is the inbound passenger flow of the nth station, d(·) is the level of the nth station, c(·) is the construction cost of the line segment between the nth and (n+1)th stations, and d(n) is the level of the nth station. sn s (τ) is the construction cost of the scale level τ.

[0070] The constraints include:

[0071] (1) Forbidden zone constraint: CUSIM has determined the grids with the attribute of "forbidden zone". The Euclidean distance between the points on the newly generated line and the grids with the attribute of "forbidden zone" should be greater than the minimum allowed distance.

[0072]

[0073] (2) Station spacing constraint: The distance between stations should be greater than the minimum station spacing to ensure train operation efficiency and avoid frequent stops. In addition, it should also be less than the maximum station spacing to maintain optimal service coverage and meet the accessibility needs of passengers. D min ≤ D k ≤ D max .

[0074] (3) Station line constraint: Stations are usually located on roads with limited construction space, and drivers need clear vision when entering and exiting the station to ensure safety. Therefore, the coordinates of the left and right HPIs should be aligned with the straight lines extending from the station center.

[0075] (4) Curve radius constraint: The curve radius should exceed the minimum allowed radius to ensure running speed and minimize the wear and tear of the rail caused by horizontal curves. In addition, it should not exceed the maximum allowed radius to maintain appropriate track geometry for regular track maintenance. R min ≤ R i ≤ R max . In the formula: R i represents the curve radius at the ith plane intersection point.

[0076] (5) Length constraint: The length of a plane circular curve should not be less than the minimum circular curve length, and the length of a straight line between curves should be greater than the minimum allowed straight line length to maintain smooth driving. L Ci ≥ L Cmin , L Ti ≥ L Tmin .

[0077] ​In this embodiment, a hybrid deep reinforcement learning method framework is proposed to solve the problem of collaborative optimization of station location and line alignment under the constraint of existing rail transit network. The framework mainly includes the following three core components:

[0078] (1) Hierarchical perception multi-objective graph pointer network (HA-MOGPN): First, a graph structure of candidate stations is constructed, and the comprehensive urban spatial information data of grid points is taken as the attribute of nodes in the graph. Then the graph is input into the HA-MOGPN network, which generates multiple candidate station sequences by iteratively calculating the selection probability of the next station in each step under the guidance of multi-objective preference.

[0079] (2) Parallel multi-branch RRT-connect algorithm (PMB-RRT-connect): For each set of station sequences, a parallel multi-branch RRT-connect algorithm is used to quickly generate connection paths between stations, and these paths are fitted to obtain the global line scheme (alignment).

[0080] (3) Hierarchical Actor-Critic training strategy: HA-MOGPN is trained under the Actor-Critic framework, and an evaluation network (Critic) is constructed to guide the optimization process of the policy network. The reward function is customized according to the station-line combination scheme. At the same time, a strategy of optimizing from bottom to top is designed to realize the step-by-step optimization of network strategy. After training, a set of Pareto-optimal solutions can be obtained, in which the station sequence is determined by HA-MOGPN, and the optimal line alignment is generated by the parallel multi-branch RRT-connect algorithm.

[0081] Inspired by the existing pointer network structure, a hierarchical perception multi-objective graph pointer network (HA-MOGPN) is designed to handle the problem of station sequence generation under multi-objective preference in a single model. The network consists of three main modules, as shown in Figure 3 .

[0082] Input layer (Input Layer): A geographic graph structure is constructed, in which candidate stations are represented as nodes in the graph, and node attributes are corresponding quantitative spatial data. Random weights are generated for each candidate node and input into the proposed pointer network.

[0083] Multi-Objective Graph Encoder: A graph attention network (GAT) is used to encode the geographical graph to capture the spatial heterogeneity of the influence of neighboring stations on station sequence selection. Node attributes are mapped to different sub-problem spaces through adaptive weighting to reflect various target preferences.

[0084] Hierarchical Context-Aware Decoder: In this embodiment, a hierarchical context-aware decoder is designed by combining a GRU module and an attention mechanism. The bottom layer network provides the likelihood of future potential paths through backward search; the upper layer network iteratively selects the next station by fusing historical and future state information, and finally generates a complete station sequence.

[0085] For the input layer, in order to explore the spatial heterogeneity of the influence of urban conditions on station selection, a geographical graph G(V, E, A) is constructed with candidate stations as nodes, where the spatial features of candidate stations are used as the feature vectors of nodes in the graph S = {s1,..., s n} T A is used to represent the geographical adjacency matrix of the graph, which is defined as follows:

[0086]

[0087] where geo(·) represents the Euclidean distance from node v i to v j , SD g is the standard deviation of the Euclidean distance, and η min and η max represent the minimum and maximum geographical distance thresholds, which are determined through parameter sensitivity analysis.

[0088] Weight setting is crucial in sub-problem generation. The weight vector represents the weight distribution of the m objectives in the tth sub-problem. N weight vectors are randomly generated, each with its own weight distribution for generating N sub-problems. W = {w 1 ,..., w N} T

[0089] The multi-objective graph encoding layer encoder consists of two parts: 1) objective encoder; 2) weight encoder. The purpose of the objective encoder is to obtain the node information of each objective in the graph, and the weight encoder adaptively and dynamically generates a set of the most effective weight vectors to aggregate the embedding of each objective, and then realizes the embedding of sub-problems.

[0090] Objective Encoder: This module consists of m Graph Attention Network (GAT) sub-modules that encode the node feature vectors H of the geographical graph G(V, E, A) into m objective embedding spaces. For each objective, the graph is input into the corresponding GAT to obtain the objective embedding of all nodes. GAT updates the node state nonlinearly with a self-attention mechanism by iteratively aggregating the information of the node neighborhood. The attention weight is crucial for aggregating the neighborhood messages, which is determined by the features of the connecting nodes and the features of the edges.

[0091] Let the state of node v i at the l-th layer of GAT be The state of the edge connecting node v i and v j is ε ij . GAT starts the process by applying a linear transformation to the input vector.

[0092]

[0093] where, and V (l) ∈R τ×1 denote trainable parameter matrices, k and τ are the embedding dimensions of the node state and edge state, is the message vector passed to the neighboring sites. Before aggregating these message vectors, the attention score of each edge is calculated This score represents the importance of the features of site v j to site v i : where || denotes the concatenation operation, a (l) ∈R (2k+ τ )×1 is a trainable parameter vector that maps the concatenation of messages to a scalar value, and σ(·) is an activation function (e.g., ReLU, Sigmoid). Then, the attention score is normalized by the softmax function:

[0094] where Φ(i) denotes the graph neighborhood of site v i . Let be a trainable parameter matrix, and the final output is aggregated from the neighborhood influence and self-influence.

[0095]

[0096] For the optimization objective k th , k th GAT operates based on the geographical graph G, and node vi Embedding vector of Expressed as:

[0097] Among them, k∈(1,…,m) represents the index of GAT, i∈(1,…,n) represents the index of node, and m e is the embedding dimension of GAT.

[0098] use To represent the embedding vectors of all nodes, for k th Optimization target. Then, for m optimization targets, the final output matrix is represented as:

[0099] The weight encoder uses a two-layer linear layer, with N weight vectors as input and the most appropriate weight matrix as output.

[0100]

[0101] After completing the above process, we get two embedding matrices and Represent the node embedding features of the m optimization objectives and the weights of the N sub-problems. By multiplying these two matrices, we get the weighted embedding vector In addition, use To express p th Weighted embeddings of subproblems.

[0102] For the hierarchical context-aware decoding layer: The decoder adopts a hierarchical architecture consisting of two layers, both built on the neighborhood masked GRU-attention module: (1) The lower decoder performs a backward search to evaluate future feasibility; (2) The upper decoder performs a forward search to encode historical trajectory information. The final decision is made in the upper layer by integrating past trajectories and future constraints. This hierarchical design ensures that each selected node is consistent in history and compatible in the future.

[0103] Neighborhood Masked GRU-Attention Module: We propose a Neighborhood Masked GRU-Attention Module (NMGA) based on a Pointer Network Decoder to solve the node sequence selection problem. This module summarizes information about previously selected nodes and calculates the probability of the next selection. By integrating a GRU module, neighborhood templates, and an attention mechanism, the NMGA module ensures that the node selection process is both spatially constrained and temporally consistent.

[0104] GRU (Gated Recurrent Unit) is used to capture temporal dependencies and maintain historical trajectory information. Specifically, in decoding step t, suppose the node selected from step t-1 is xt , the weighted embedding of all nodes is E = {e1,..., en} n} T , the hidden state d t-1 of the decoder stores the knowledge of the previously selected path. Then, the weighted e t embedding of node x t is passed into the GRU module to compute the hidden state d t as:

[0105]

[0106] where the update gate z t is defined as: z t = σ(W z e t + U z h t-1 + b z )

[0107] The candidate hidden state is computed as:

[0108] The reset gate r t is given as: r t = σ(W r e t + U r h t-1 + b r ).

[0109] For the neighborhood template: in the decoding step, a neighborhood template centered on the selected node x t is defined, including only the nodes within the desired range of inter-site distances. Then a mask M t is constructed, assigning a value of zero to the nodes that are outside the neighborhood template, effectively excluding them from the selection process.

[0110]

[0111] For the attention layer, in the decoding step, the attention scores quantify the relevance of each node to the current state:

[0112] After applying the neighborhood mask, the attention scores are modified as:

[0113] Finally, the probability of selecting each node is computed using the softmax function:

[0114] Based on the GRU-attention module with neighborhood templates, a hierarchical architecture is designed to solve the station selection problem, which consists of two layers. The upper decoder: performs a forward search from the start node S, the forward search selects the most relevant node according to the historical trajectory information encoded in the upper hidden state and the future feasibility information captured by the lower decoder, which ensures that the path is consistent in history and feasible in the future. The lower decoder: each time the upper decoder makes a selection, the lower decoder performs a backward search from the destination node E to the selected node of the upper layer. This helps to explore the feasibility of potential future paths in the station selection process. As shown in Fig. Figure 4 illustrates the station selection process.

[0115] ①The upper decoder performs a forward search from the start node S. In the decoding step t, the node selected by the upper layer at step t-1 is denoted as

[0116] ②In the lower layer, a backward search from the destination node E is performed based on the NMGA module. In step the weighted embedding of the selected node is concatenated with the weighted embedding of the selected node of the upper layer and input to the NMGA module. The input is:

[0117] This ensures that the node selection of the upper network dynamically affects the backward search of the lower network, thereby adaptively determining the future feasibility. The backward search will continue until the distance between the selected node and the node satisfies the distance constraint:

[0118] where is the last selected node in the lower backward search process. The final hidden state of the last step of the lower layer is represented as:

[0119] ③The weighted embedding of the node is then passed to the GRU module of the upper layer, where its hidden state is calculated as:

[0120] where denotes the hidden state of the upper layer at step t-1.

[0121] ④For each candidate node, additive attention is applied to measure its relevance to the lower last hidden state d t (L) and the upper hidden state :

[0122] where W o, W (L) and W (U) are trainable weight matrices that map different hidden states to a common attention space, v o is a learnable vector that scores interactions.

[0123] ⑤Finally, the probability of selecting each node at the t-th step is calculated by integrating the forward and backward hidden states at the upper layer, ensuring historical consistency and future feasibility:

[0124] ⑥The node of the decoding step is selected based on the calculated probability. Steps ①-⑤ will be repeated until the destination node E is reached, completing the station selection process.

[0125] For the multi-branch RRT-connect algorithm, the station layout and line alignment are incorporated into the calculation of the reward function during the training process of the actor-critic network. To ensure the efficiency of the training process, it is crucial to quickly search for the line alignment between stations. RRT-Connect has been proven to be effective in fast path searching and has been successfully applied to the generation of urban rail transit plane lines. On this basis, we propose a parallel multi-branch RRT-Connect algorithm. Unlike traditional methods that only search between a pair of stations, our method simultaneously extends multiple branches between all stations and grows bidirectionally between adjacent stations. By introducing multi-branch searching, this method can ensure a rich supply of nodes, thereby speeding up the bidirectional search process. In addition, with the help of GPU parallel computing technology, parallel processing of alignment search between multiple stations is realized. The specific calculation steps are shown in Figure 5 .

[0126] ①Multi-branch expansion: initialize the search tree for all stations simultaneously, with each station as the root node of a tree. From the root node, multiple branches are concurrently expanded. Specifically, for each tree, multiple random sampling points (v rand ) are generated, and the nearest existing node (v near ) to these sampling points in the current tree is found. Then, each nearest node extends a fixed length in the direction of its corresponding sampling point, generating a new node (v new ), as shown in Figure 5 a.

[0127] ②Bidirectional growth: each tree attempts to grow towards its adjacent two trees. As shown in Figure 5 b, tree B finds the nearest nodes (v a-new , v c-new ) of its newly generated node (v b-near ) in adjacent trees A and C.), and make these recent nodes extend a fixed length along the corresponding new node direction. This process continues until an obstacle is encountered or a connection with an adjacent tree is made.

[0128] ③ Tree merging: Once two trees meet, a segment of the connection path is formed. This process iterates until all trees corresponding to stations are connected, forming a complete path. All possible global paths are generated and evaluated according to construction cost, and the path with the lowest cost is selected as the initial global route, as shown in Figure 5 .

[0129] For path fitting, the initial global path generated by the PMB-RRT-connect algorithm consists of a series of segmented linear tangents. To ensure the smoothness of the path and meet the relevant constraints, these line segments need to be further fitted into a continuous route, as shown in Figure 6 .

[0130] Step 1: Ensure that each station and its bidirectional adjacent vertices (v bl ,v br ) are on the same straight line. If not, extend a fixed length along the direction connecting the station and adjacent vertices to generate new vertices. Then select the adjustment method with the lowest construction cost, as shown in Figure 6 a.

[0131] Step 2: Fix the bidirectional adjacent vertices (v bl ,v br ) of each station as the starting or ending point (e a-b ) of path fitting between stations. Then, calculate the distance from each vertex on the path to the initial straight line connecting the starting and ending points. The vertex with the maximum distance (PD max ) is considered as the intersection point (PI), as shown in Figure 6 b.

[0132] Step 3: Based on the intersection point PI, divide the initial edge into two sub-edges. Then configure the minimum circular curve radius (R min ) for this PI, and calculate the circular curve length (L C ) and the tangent length (L T ) between the adjacent two circular curves. If both are greater than the minimum limit length (L Cmin ,L Tmin ), then the PI is feasible (as shown in Figure 6 c); otherwise, it is not feasible (as shown in Figure 6 d).

[0133] Step 4: Repeat Step 3 until there are no new PIs in the path. Finally, complete the fitting of the path to generate the route, as shown in Figure 6 e.

[0134] In this example, the HA-MOGPN network is trained under the Actor-Critic framework, and a layer-by-layer policy training method is designed to optimize the hierarchical network. First, the lower layer network is trained, and its optimal parameters are fixed; then, the upper layer network receives the latent variables from the lower layer to learn its own policy. During the training of each layer, a critic network with the same encoder structure as the HA-MOGPN is constructed to evaluate the expected reward based on the candidate station state, and its output is called the state value. The actual reward is calculated based on the station-line solution determined by the station sequence generated by the HA-MOGPN and the line layout generated by the PMB-RRT-connect.

[0135] Specifically, considering that the optimization model contains both passenger flow intensity and construction cost as objectives, the actual reward function should take these two indicators into account. In addition, considering the strict design requirements, it is difficult to find a station-line solution that fully satisfies all constraints, so we introduce a reward item related to the feasibility of the solution: a positive reward is given to the solution that satisfies all constraints, and a negative reward is given if any constraint is violated. The reward function is expressed as follows:

[0136]

[0137] where Iqand Cqare the passenger flow intensity and total construction cost of the qthsub-problem, respectively, and Iqis the feasibility reward item of the qthsub-problem. q ,C q th are the random weighting coefficients of the two objectives of the qthsub-problem. It should be noted that since the optimization objective is to minimize the total construction cost, which is reflected in the reward function as the higher the reward value, the lower the cost, we express as a negative value.

[0138] The optimization objective of the Actor network is to improve the generation probability of the actual reward station sequence, while the optimization objective of the policy network is to make the expected reward as close to the actual reward as possible. The parameters of both networks are trained using the Adam optimizer. Algorithm 1 shown in Table 1 below demonstrates the detailed pseudo-code flow.

[0139]

[0140]

[0141] The flow is described as follows:

[0142] ​​​Input data: M candidate station scheme sets (selected from the city space information synthesis model (CUSIM), simulating the process of manual candidate scheme), maximum training times Iter max , learning rate for two-layer actor network

[0143] 1. Initialize network parameters, including lower actor network and upper actor network.

[0144] 2. Generate N initial weight vectors.

[0145] 3. Train the lower network and the upper network in turn, first when k = 0, the lower network is trained, first for each iteration (step 4), first select the i-th candidate scheme from the candidate station scheme set, and input the city space feature information of each candidate station as the feature (step 5).

[0146] 4. Then according to the current training strategy of the policy network constantly output the next selected station scheme wherein, represents the station number selected at the t-th step, represents the hidden state of the selected GRU module at the t-th step (steps 6-10) (when k = 0, step 8 is skipped, when k = 1, step 8 is executed). Constant iteration until the starting point is selected to the terminal point, and the station sequence PS is obtained k (step 11).

[0147] 5. After generating the station sequence, the inter-station line scheme is obtained based on the RRT related algorithm (step 12).

[0148] 6. The station sequence and the inter-station line scheme together, the reward function is calculated (step 13).

[0149] 7. Calculate the descending gradient of the actor network and the critic network (steps 14-15).

[0150] 8. Calculate the network parameters after updating the actor network and the critic network (steps 16-17).

[0151] 9. Reach the maximum iteration times, and the training is terminated (step 18)

[0152] 10. Train the lower network and then the upper network in turn, and the training is terminated (step 19)

[0153] 11. Output the trained strategy of the actor network

[0154] Example 3

[0155] The embodiment 3 provides a non-transitory computer readable storage medium for storing computer instructions, the computer instructions being executed by a processor to implement the track traffic plane line-station collaborative optimization method based on hybrid reinforcement learning as described above, the method comprising:

[0156] constructing a graph structure of the candidate stations, and taking the comprehensive urban space information data of the grid points as attributes of nodes in the graph;

[0157] processing the graph structure by using a pre-trained hierarchical perception multi-objective graph pointer network, generating a plurality of candidate station sequences under the guidance of multi-objective preferences by iteratively calculating selection probabilities of next stations in each step; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is used to encode the geographic graph by using a graph attention network, so as to capture the spatial heterogeneity influence of adjacent stations on station sequence selection, and node attributes are mapped to different sub-problem spaces through adaptive weighting, so as to reflect various types of objective preferences; the bottom network of the hierarchical context perception decoder provides the possibility of future potential paths through back search, and the upper network is used to iteratively select the next station by fusing historical state and future state information, and finally generate a complete station sequence;

[0158] for each group of station sequences, a parallel multi-branch RRT-connect algorithm is used to generate connection paths between stations, and the paths are fitted to obtain a global line scheme.

[0159] Embodiment 4

[0160] The embodiment 4 provides a computer device comprising a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the track traffic plane line-station collaborative optimization method based on hybrid reinforcement learning as described above, the method comprising:

[0161] constructing a graph structure of the candidate stations, and taking the comprehensive urban space information data of the grid points as attributes of nodes in the graph;

[0162] The pre-trained hierarchical perception multi-objective graph pointer network is used to process the graph structure, a selection probability of a next station in each step is iteratively calculated, and multiple candidate station sequences are generated under the guidance of multi-objective preferences; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is used to encode a geographic graph by using a graph attention network, so as to capture a spatial heterogeneity influence of adjacent stations on station sequence selection, and node attributes are mapped to different sub-problem spaces through adaptive weighting, so as to reflect various types of objective preferences; a bottom network of the hierarchical context perception decoder provides a possibility of a future potential path through reverse search, and an upper network is used to iteratively select a next station by fusing historical state and future state information, and finally generate a complete station sequence.

[0163] For each group of station sequences, a parallel multi-branch RRT-connect algorithm is used to generate connection paths between stations, and the paths are fitted to obtain a global line scheme.

[0164] Embodiment 5

[0165] The embodiment 5 provides an electronic device, comprising a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes instructions for implementing a track traffic plane line-station collaborative optimization method based on hybrid reinforcement learning as described above, the method comprising:

[0166] A graph structure of candidate stations is constructed, and comprehensive urban space information data of grid points are taken as attributes of nodes in the graph;

[0167] The pre-trained hierarchical perception multi-objective graph pointer network is used to process the graph structure, a selection probability of a next station in each step is iteratively calculated, and multiple candidate station sequences are generated under the guidance of multi-objective preferences; the hierarchical perception multi-objective graph pointer network comprises a multi-objective graph encoder and a hierarchical context perception decoder; the multi-objective graph encoder is used to encode a geographic graph by using a graph attention network, so as to capture a spatial heterogeneity influence of adjacent stations on station sequence selection, and node attributes are mapped to different sub-problem spaces through adaptive weighting, so as to reflect various types of objective preferences; a bottom network of the hierarchical context perception decoder provides a possibility of a future potential path through reverse search, and an upper network is used to iteratively select a next station by fusing historical state and future state information, and finally generate a complete station sequence;

[0168] For each group of station sequences, a parallel multi-branch RRT-connect algorithm is used to generate connection paths between stations, and the paths are fitted to obtain a global line scheme.

[0169] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In a software embodiment, the methods can be tangibly embodied in a machine-readable storage medium having stored thereon instructions that can be used to program a computer to perform any of the methods. The software implementation can be for example, in a modulated data signal such as a carrier wave or other transport mechanism, or a computer readable storage medium.

[0170] The present application is described in relation to flow diagrams and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It is understood that each block of the flow diagrams and / or block diagrams, and combinations of blocks in the flow diagrams and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks. Figure 1 means for performing each of the functions specified in the flow diagram and / or block diagram block or blocks.

[0171] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flow diagram and / or block diagram block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks. Figure 1 means for performing each of the functions specified in the flow diagram and / or block diagram block or blocks.

[0172] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow diagram and / or block diagram block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks. Figure 1 means for performing each of the functions specified in the flow diagram and / or block diagram block or blocks.

[0173] The above description is implemented with reference to the accompanying drawings, and is not intended to limit the scope of the present application. It should be understood by those skilled in the art that various modifications or changes can be made to the disclosed technical solutions without departing from the scope of the present application.

Claims

1. A rail transit line-station collaborative optimization method based on hybrid reinforcement learning, characterized in that: include: Construct a graph structure of candidate stations and use the comprehensive urban spatial information data of grid points as attributes of nodes in the graph; The graph structure is processed using a pre-trained hierarchical perceptual multi-objective graph pointer network. The probability of selecting the next station in each step is iteratively calculated, and multiple candidate station sequences are generated under the guidance of multi-objective preferences. The hierarchical perceptual multi-objective graph pointer network includes a multi-objective graph encoder and a hierarchical context-aware decoder; A multi-objective graph encoder is used to encode the geographic graph using a graph attention network to capture the spatial heterogeneity of the influence of neighboring stations on station sequence selection. Node attributes are adaptively weighted and mapped to different sub-problem spaces to reflect various objective preferences. The bottom layer of the hierarchical context-aware decoder provides the possibility of future potential paths through backward search. The upper layer is used to iteratively select the next station by fusing historical state information with future state information, ultimately generating a complete station sequence. For each set of station sequences, a parallel multi-branch RRT-connect algorithm is used to generate connection paths between stations, and these paths are fitted to obtain a global route plan.

2. The rail transit plane line-station collaborative optimization method based on hybrid reinforcement learning according to claim 1 is characterized in that: Training the hierarchical perceptual multi-objective graph pointer network includes: first training the lower-layer network and fixing its optimal parameters; then, the upper-layer network receives latent variables from the lower layer to learn its own strategy; considering that the optimization model includes both passenger flow intensity and construction cost objectives, the actual reward function should take these two indicators into consideration and introduce a reward item related to the feasibility of the solution: positive rewards are given to solutions that meet all constraints, and negative rewards are given if any constraints are violated.

3. The rail transit line-station collaborative optimization method based on hybrid reinforcement learning according to claim 1 is characterized in that: The multi-objective graph encoder consists of a target encoder and a weight encoder. The target encoder aims to obtain the node information of each target in the graph, and the weight encoder adaptively and dynamically generates a set of the most effective weight vectors to aggregate the embedding of each target. The target encoder consists of m graph attention network submodules, which encode the node feature vectors of the geographic graph into m target embedding spaces; For each target, the graph is input into the corresponding graph attention network submodule to obtain the target embedding of all nodes; the graph attention network submodule updates the node state nonlinearly using the self-attention mechanism by iteratively aggregating information from the node neighborhood; The weight encoder uses a two-layer linear layer, takes N weight vectors as input, and outputs the most appropriate weight matrix for prediction.

4. The rail transit line-station collaborative optimization method based on hybrid reinforcement learning according to claim 1 is characterized in that: The hierarchical context-aware decoding layer consists of two layers, both of which are built based on the neighborhood masked GRU-attention module: (1) the lower decoder performs backward search to evaluate future feasibility; (2) the upper decoder performs forward search to encode historical trajectory information; among them, the neighborhood masked GRU-attention module is used to solve the node sequence selection problem.

5. The rail transit line-station collaborative optimization method based on hybrid reinforcement learning according to claim 1 is characterized in that: The parallel multi-branch RRT-connect algorithm includes: simultaneously initializing search trees for all stations, with each station serving as the root node of a tree, and concurrently expanding multiple branches from the root node; each tree will attempt to grow toward its two adjacent trees, and tree B will find the nearest nodes of its newly generated node in adjacent trees A and C, and extend these nearest nodes along the direction of the corresponding new nodes for a fixed length. This process continues until it encounters an obstacle or connects with an adjacent tree; once two trees meet, a connecting path is formed, and the process is iterated continuously until all trees corresponding to all stations are connected to form a complete path; all possible global paths will be generated and evaluated based on construction cost, and finally the path with the lowest cost is selected as the initial global route direction.

6. The rail transit line-station collaborative optimization method based on hybrid reinforcement learning according to claim 1 is characterized in that: Path fitting includes: ensuring that each station and its two-way adjacent vertices are on the same straight line; if they are not on a straight line, extending them by a fixed length in the direction connecting the station and the adjacent vertices to generate new vertices; then selecting the adjustment method with the lowest construction cost; Step 2: Fix the two-way adjacent vertices of each station as the starting point or end point of the path fitting between stations; then, calculate the distance from each vertex on the path to the initial straight line connecting the starting point and the end point, and the vertex with the largest distance is regarded as the intersection; based on the intersection, the initial edge is divided into two sub-edges, and then the minimum circular curve radius is configured for the intersection, and the circular curve length and the tangent length between the two adjacent circular curves are calculated; if they are both greater than the minimum limit length, the intersection is feasible, otherwise it is infeasible; Step 4: Repeat step 3 until there are no new intersections in the path; finally, the path fitting is completed and the route direction is generated.

7. A rail transit line-station collaborative optimization system based on hybrid reinforcement learning, characterized in that: include: A construction module is used to construct a graph structure of candidate stations and use the comprehensive urban spatial information data of grid points as attributes of nodes in the graph; a processing module for processing the graph structure using a pre-trained hierarchical perceptual multi-objective graph pointer network, iteratively calculating the probability of selecting the next station in each step, and generating multiple candidate station sequences under the guidance of multi-objective preferences; The hierarchical perceptual multi-objective graph pointer network includes a multi-objective graph encoder and a hierarchical context-aware decoder; A multi-objective graph encoder is used to encode the geographic graph using a graph attention network to capture the spatial heterogeneity of the influence of neighboring stations on station sequence selection. Node attributes are adaptively weighted and mapped to different sub-problem spaces to reflect various objective preferences. The bottom layer of the hierarchical context-aware decoder provides the possibility of future potential paths through backward search. The upper layer is used to iteratively select the next station by fusing historical state information with future state information, ultimately generating a complete station sequence. The calculation module is used to generate connection paths between stations using a parallel multi-branch RRT-connect algorithm for each set of station sequences, and to fit these paths to obtain a global route plan.

8. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the rail transit plane line-station collaborative optimization method based on hybrid reinforcement learning as described in any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the rail transit plane line-station collaborative optimization method based on hybrid reinforcement learning as described in any one of claims 1 to 6.

10. An electronic device, characterized in that: include: A processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the rail transit plane line-station collaborative optimization method based on hybrid reinforcement learning as described in any one of claims 1 to 6.

Images