Construction method of new energy cellular network user access decision-making system and user access method
By constructing a new energy cellular network user access decision system through multi-agent deep reinforcement learning, the problem of user access strategies being unable to adapt to the fluctuations of renewable energy has been solved, achieving rapid response and efficient user access adjustment, thereby improving network performance and green emission reduction effects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-03-10
AI Technical Summary
Existing communication network user access strategies cannot quickly and accurately adapt to the spatial fluctuations of renewable energy, leading to a decline in network performance.
By employing a multi-agent deep reinforcement learning approach and through collaborative training of base station agents and central agents, a new energy cellular network user access decision system is constructed. Deep reinforcement learning technology is used to optimize user access strategies, enabling minute-level response to spatiotemporal fluctuations in renewable energy and service load.
It enables rapid and accurate user access adjustments, improves the performance of communication networks and green emission reduction effects, and solves the problems of policy lag and high computational complexity in traditional methods. It is suitable for urban cellular networks, suburban coverage and communication base stations in remote areas.
Smart Images
Figure CN121645280A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of communication network optimization, and more specifically, relates to a method for constructing a user access decision system for a new energy cellular network and a user access method. Background Technology
[0002] Unlike traditional cellular networks, renewable energy supply is random. Therefore, in order to maintain stable cellular network services, it is necessary to redesign the user access method of cellular network that is aware of renewable energy status. By dynamically adjusting the user-base station association, the service load can be guided to migrate to base stations with sufficient renewable energy, so as to achieve spatial matching between network energy consumption and renewable energy power generation, thereby reducing dependence on grid power and improving the green energy consumption rate.
[0003] Traditional user access optimization methods include setting exponential bias functions based on energy state, establishing topological potential attraction models, and setting latency-aware association rules. These methods rely on preset mathematical relationships or empirical thresholds to map parameters such as base station renewable energy reserves, user distance, and service load to access priority weights. However, traditional methods suffer from model rigidity and scenario mismatch. Heuristic rules lack theoretical optimality proof and cannot adapt to complex scenarios where energy and load are coupled. Static parameters cannot respond to minute-level renewable energy fluctuations, resulting in policy lag.
[0004] With the development of deep learning, distributed multi-agent Q-network schemes are now being used, such as using users as agents to decide which base stations to associate with or using base stations as agents to design access strategies. However, the current multi-agent Q-network schemes, where each agent is trained independently, suffer from drawbacks such as slow convergence and inconsistent strategies, and cannot accurately adapt to the spatial fluctuations of renewable energy.
[0005] Therefore, current communication network user access strategies cannot quickly and accurately adapt to the spatial fluctuations of renewable energy, which in turn affects the performance of communication networks. Summary of the Invention
[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for constructing a new energy cellular network user access decision system and a user access method, the purpose of which is to quickly and accurately adapt to the spatial fluctuations of renewable energy and improve the performance of communication networks.
[0007] To achieve the above objectives, the following technical solution is proposed.
[0008] According to a first aspect of the present invention, a method for constructing a new energy cellular network user access decision system is provided, comprising: S1. The base station agent interacts with the environment through its main network to generate a quadruple. And put it into the playback buffer, where t is time, and S(t) is the sum of the times for each base station. status The global state obtained by splicing is A(t), which represents the state of each base station. action The global action obtained by splicing is r(t), which is the global action reward. Let i represent the i-th base station in the k-th type of base station. Each base station is deployed with an independent base station agent. The state of the base station agent includes the renewable energy power that the corresponding base station can consume and the number of connected users. The actions of the base station agent include regulating user access to the corresponding base station. The reward function for calculating the reward represents the optimization objective of optimal network performance. S2. The network management unit obtains randomly sampled batches of quadruplets from the playback buffer as training quadruplets; each base station Obtain the corresponding state from the training quadruple. , As training samples for base stations and for obtaining them And use its base station agent's target network to select the state The action with the largest Q value The network management unit obtains the actions selected by its target network from each base station. After being concatenated into A(t+1), it forms the central training sample along with the corresponding training quadruples. ; S3. After obtaining the central training sample, the central agent deployed in the network management unit first calculates the Q value of performing action A(t+1) in state S(t+1) through its target network and calculates the target Q value at time t in combination with the reward r(t), which is used as the central training label of the central training sample at time t. Then, it uses its main network to calculate the Q value of performing A(t) in state S(t), which is used as the central prediction result at time t. The training objective is to minimize the deviation between the central prediction result and the central training label. The main network of the trained central agent is then used to calculate the Q value of performing A(t) in state S(t), which is used as the base station training label of the base station training sample at time t and sent to each base station. S4. After each base station agent obtains the base station training samples, it calculates the Q value of A(t) in state S(t) through its main network as the base station prediction result at time t. The main network is trained with the goal of minimizing the deviation between the base station prediction result and the base station training label. S5. Repeat S1 to S4 until the preset number of iterations is reached. The main network of the trained base station agent is the new energy cellular network user access decision system of the corresponding base station.
[0009] According to a second aspect of the present invention, a method for user access in a new energy cellular network is provided, comprising: base station Get the state observed locally at time t. It also provides access to the user decision-making system, whereby the main network of the intelligent agent corresponding to the user access decision-making system selects the state... The action with the largest Q value To control user access to the base station; The access decision system for each base station is constructed using the aforementioned method.
[0010] Overall, compared with the prior art, the technical solutions conceived in this invention have the following beneficial effects.
[0011] 1. This invention employs a multi-agent deep reinforcement learning method to construct a new energy cellular network user access decision system. Unlike traditional multi-agent deep reinforcement learning methods, this invention deploys base station agents and a central agent during the agent training phase. In each iteration of training, the base station agent and the central agent cooperate with each other. The central agent is trained based on the experience data of the base station agent, and then transfer learning is performed. The trained central agent generates training labels for the base station agent, enabling the base station agent to complete training locally. The above training process is repeated to finally obtain a trained base station agent. Each base station can quickly and accurately adjust user access when facing spatial fluctuations in renewable energy based on its deployed base station agents. In this invention, a hybrid learning architecture as described above is adopted. A centralized trainer coordinates the experience of all base stations to optimize the global policy. The trained general policy is then quickly adapted to the local environment of each base station through transfer learning. This not only ensures the global synergy of the policy, but also greatly improves the learning efficiency and environmental adaptability of individual agents. It overcomes the shortcomings of slow convergence and inconsistent policies in fully distributed learning, and realizes a minute-level granularity "moving with green electricity" access policy. It can efficiently respond to the spatiotemporal fluctuations of renewable energy and service load, and can be widely used in urban cellular networks, suburban coverage and communication base stations in remote areas.
[0012] 2. In a specific embodiment, the action is set as a base station access bias, thereby reducing the complexity of the action space from the combinatorial level (exponentially increasing with the number of users) of traditional methods to the linear level. This overcomes the problem of high computational complexity and poor real-time performance caused by the exponential expansion of the action space with the number of users, and solves the inherent "action space explosion" problem of distributed reinforcement learning in dense user scenarios, making real-time online decision-making possible in large-scale networks.
[0013] 3. In a specific embodiment, by taking "carbon efficiency" as the key optimization objective and deeply integrating it into the reward function design, this algorithm differs from traditional optimization schemes that take energy consumption or network throughput as a single objective. While maximizing the use of renewable energy, it strictly guarantees the service quality requirements of users, truly achieving a win-win situation for green emission reduction and communication performance. Attached Figure Description
[0014] Figure 1 This is a flowchart of the steps in constructing a new energy cellular network user access decision system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of data interaction between a base station, a network management unit, and a playback buffer in one embodiment of the present invention; Figure 3 This is a schematic diagram comparing the performance of the user access method (IFMAD) of this invention with existing multi-agent deep reinforcement learning schemes (MADQN) and existing maximum receive power user access methods (Max receive). Detailed Implementation
[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0016] In a first aspect, the present invention claims a method for constructing a decision system for user access in a new energy cellular network.
[0017] like Figure 1 The diagram shows a flowchart of the construction method of a new energy cellular network user access decision system according to an embodiment of the present invention. The order of steps in the diagram is only one possible approach and is not limited thereto. Figure 2 The diagram shown is a data interaction schematic of the base station, network management unit, and playback buffer in one embodiment of the present invention. The following is a summary of the data interaction. Figure 1 , Figure 2 This method will be described in detail.
[0018] New energy cellular networks comprise multiple base stations, each capable of adjusting user access strategies. This invention employs deep reinforcement learning to construct a user access decision system for new energy cellular networks, enabling base stations to optimize user access policies. The invention includes base station agents deployed on each base station, a central agent deployed in the network management unit, and a replay buffer.
[0019] The main symbols involved in this invention will be explained first.
[0020] , represents the i-th base station in the k-th type of base station; , indicating base station The base station status at time t; S(t) represents the global state at time t, which is formed by splicing together the states of all base station agents participating in the training at time t; , indicating base station The action performed at time t; A(t) represents the global action at time t, which is composed of the actions of all base station agents participating in the training at time t. r(t) represents the global action reward. The Q value represents the expected cumulative reward for performing an action in a given state.
[0021] Define the state, action space, and reward function involved in utilizing deep reinforcement learning techniques in the scenario of this invention.
[0022] Status. In this invention, a single base station... The state of the base station agent at time t This includes the energy consumption of the base station at time t and the number of users connected to the base station. The global state space S(t) is a concatenation of the states of all base stations.
[0023] Action space. In this invention, a single base station The action space of the base station agent includes the operable space for the corresponding base station to regulate user access.
[0024] Reward function. In this invention, the reward function is used to achieve optimal network performance.
[0025] Understandably, the state, action space, and reward function can be flexibly adjusted according to the network optimization objective, so that the user access strategy output by the trained agent can be as close as possible to the optimization objective.
[0026] In a specific embodiment, the action space of the base station agent is the selectable range of the base station access bias. The range of selectable access bias is: ; In the formula, Indicates the first Maximum access offset for base stations.
[0027] Each base station action Access bias selected by the base station at time t , can be represented as: ; Correspondingly, the global action space is: ; In the formula, This represents the number of base stations included in the Kth type of base station.
[0028] When the base station agent selects the base station access bias at time t, the users in the network will use the maximum bias received power access mechanism to select a base station to access. That is, the user will select the base station with the maximum access bias received power to access, and then obtain the state of the agent at time t+1.
[0029] definition to indicate Time users and The relationship between them, namely: ; In the formula, No. The transmit power of base stations This refers to the base station index number under the corresponding base station type. It is a base station and users The channel gain between them consists of small-scale fading and large-scale fading. It is a base station Access bias at time t This represents the set of base stations of type K.
[0030] In a specific embodiment, the optimization objective is to maximize the network's carbon efficiency under user service quality constraints, and the reward function can be expressed as: ; In the formula, It is the carbon efficiency of the network at time t. It is the throughput of user j at time t. It is the user throughput requirement. It is the penalty function when the user service quality constraints are not met. When the condition is met, the penalty function is 1; otherwise, the penalty function is 0. It is the set penalty weight.
[0031] Among them, carbon efficiency The ratio of network throughput to carbon emissions can be expressed as: ; In the formula, This represents the total network throughput at time t. Let t be the total carbon emissions of the entire network at time t.
[0032] In a specific embodiment, the state Including base stations The renewable energy power available for consumption at time t and base stations Number of users accessing the system at time t Furthermore, when the optimization objective is to maximize the network's carbon efficiency under user service quality constraints, the state also includes user throughput requirements. At this time, the base station The state at time t It can be represented as: .
[0033] Accordingly, the global state S(t) can be expressed as: ; In the formula, K represents the number of base station types. This represents the number of base stations included in the Kth type of base station.
[0034] Based on the actions, states, and rewards defined above, the following deep reinforcement learning process is executed.
[0035] S1. The base station agent interacts with the environment through its main network to generate a quadruple. And put it into the playback buffer, where t is time, and S(t) is the sum of the times for each base station. status The global state obtained by splicing is A(t), which represents the state of each base station. action The global action obtained by splicing is r(t), which is the global action reward. Let i represent the i-th base station in the k-th type of base station. Each base station deployment has an independent base station agent. The state of the base station agent includes the renewable energy power that the corresponding base station can consume and the number of connected users. The action of the base station agent is the access bias of the corresponding base station. The reward function for calculating the reward is characterized with the optimization objective of optimal network performance.
[0036] In agent-based deep reinforcement learning, each agent has a deep Q-network module. Each deep Q-network module contains two neural networks: a main network and a target network. The main network selects the corresponding action for the agent at time t and evaluates the Q-value obtained after performing the action. The target network selects the action with the maximum Q-value for the agent at time t+1. The agent trains the main network and uses a soft update mechanism to periodically synchronize the network parameters of the main network to the target network parameters.
[0037] In this invention, each base station will deploy a corresponding base station agent to optimize user access decisions for the base station, and a central agent will be deployed in the network management unit to optimize user access decisions for the entire network.
[0038] In this process, each base station agent interacts with the main network and the environment to generate experiential data. Specifically, this is based on the current state. based on Strategy selection action That is, intelligent agents have The probability of randomly selecting an action is... Select the action with the highest Q value, then execute the selected action, and subsequently obtain the next state. .
[0039] The action with the largest Q value can be represented as: ; In the formula, These are the current network parameters of the main network for the base station agent. This indicates that the current network parameters are The main network in its current state Next action The Q value after that.
[0040] The current state of all base station agents By concatenating these values, we can obtain the global state S(t) at the current moment, and the actions of all base station agents at the current moment. By concatenating these values, we can obtain the global action A(t) at the current moment, and the states of all base station agents at the next moment. By concatenating these elements, we can obtain the global state S(t+1) for the next time step. We then calculate the reward for performing the global action A(t) in the current global state S(t), obtaining the global action reward r(t). This forms a quadruple. The data is then stored in the playback buffer D. Once the playback buffer D has sufficient data, subsequent steps are executed.
[0041] S2. The network management unit obtains randomly sampled batches of quadruplets from the playback buffer as training quadruplets; each base station Obtain the corresponding state from the training quadruple. , As training samples for base stations and for obtaining them And use its base station agent's target network to select the state The action with the largest Q value The network management unit obtains the actions selected by its target network from each base station. After being concatenated into A(t+1), it forms the central training sample along with the corresponding training quadruples. .
[0042] Specifically, when the playback buffer When there is sufficient data, random sampling is performed from the playback buffer. A training batch consists of four tuples. The network management unit obtains the complete four-tuple data from this training batch. Each base station only needs to obtain the current state, current action, and next state within each four-tuple in the training batch, and each base station only obtains its own data. Get each quadruple , and .
[0043] When each base station Obtain the training quadruple , and After that, with , The training samples are used as training examples for the base station to train the subsequent base station agents. This training process is described in step S4 below. Based on the acquired data... Base station Using its base station agent, the target network selects the state. The action with the largest Q value Specifically, it can be expressed as: ; In the formula, These are the current network parameters of the target network for the base station agent. This indicates that the current network parameters are The target network in the next state Next action The Q value after that.
[0044] For each collected quadruple Each base station Actions are selected through the target network of its base station agents. Then it is sent to the network management unit, which will then send it to each base station. of Concatenate the results to obtain A(t+1), then combine it with the corresponding quadruple. Combined, they form a set of central training samples. .
[0045] After obtaining N sets of central training samples, the central agent deployed in the network management unit can be trained, i.e., step S3 is executed.
[0046] S3. After obtaining the central training sample, the central agent deployed in the network management unit first calculates the Q value of performing action A(t+1) in state S(t+1) through its target network and calculates the target Q value at time t in combination with the reward r(t), which is used as the central training label of the central training sample at time t. Then, it uses its main network to calculate the Q value of performing A(t) in state S(t), which is used as the central prediction result at time t. The training objective is to minimize the deviation between the central prediction result and the central training label. The main network of the central agent is then trained. The Q value of performing A(t) in state S(t) is calculated using the trained central agent's main network, which is used as the base station training label of the base station training sample at time t and sent to each base station.
[0047] This step involves training a central agent using central training samples, and then obtaining the training labels for the base station based on the trained central agent.
[0048] Specifically, for any set of center training samples at time t The central agent first calculates the Q-value of performing action A(t+1) in state S(t+1) through its target network, and then calculates the target Q-value at time t by combining it with the reward r(t). This target Q-value is used as the central training label y(t) of the central training sample at time t. Specifically, it can be expressed as follows: ; In the formula, The learning rate is set. The current network parameters of the target network for the central agent. This indicates that the current network parameters are The Q value of the target network after performing action A(t+1) in the next state S(t+1).
[0049] After obtaining the training labels for each central training sample, the main network of the central agent is trained using the central training samples and their labels. The training loss can be expressed as: ; In the formula, This indicates calculating the mean. The current network parameters of the main network for the central agent. This indicates that the current network parameters are The Q value of the main network after performing action A(t) in the current state S(t) is the center prediction result at time t.
[0050] Training ends when the training loss converges to a preset level. Subsequently, for each training quadruple, the Q-value of the trained central agent main network performing action A(t) in state S(t) is calculated and used as the base station training label for the base station training samples at time t, and sent to each base station. The training labels for each base station are the same and can all be represented as: .
[0051] Subsequently, the training process for the base station agent can be executed as follows.
[0052] S4. After obtaining the base station training samples, each base station agent calculates the Q value of A(t) in state S(t) through its main network as the base station prediction result at time t. The main network is trained with the goal of minimizing the deviation between the base station prediction result and the base station training label.
[0053] Specifically, the training loss of each base station agent can be expressed as: ; In the formula, This indicates that the mean value is being calculated.
[0054] When the training loss converges to a preset level, the training ends, and thus the base station agent and the central agent complete one training iteration.
[0055] S5. Repeat S1 to S4 until the preset number of iterations is reached. The main network of the trained base station agent forms a new energy cellular network user access decision system.
[0056] Repeat the above process until the preset number of iterations is reached, for example, 30,000 iterations, then end the iterative training.
[0057] After the iterative training is completed, the main network of the base station agent is used as the new energy cellular network user access decision system for the base station.
[0058] Secondly, the present invention also relates to a method for user access in a new energy cellular network.
[0059] This access method includes: base station Get the state observed locally at time t. It also provides access to the user decision-making system, whereby the main network of the intelligent agent corresponding to the user access decision-making system selects the state... The action with the largest Q value To control user access to the base station; The access decision system for each base station is constructed using the method described above.
[0060] Specifically, the local master network parameters that have been trained are loaded, the base station observes the local state, selects the optimal action based on the local master network, and executes the access action, such as adjusting the access bias.
[0061] like Figure 3 The diagram shows a comparison of the performance of the user access method (IFMAD) of this invention with existing multi-agent deep reinforcement learning schemes (MADQN) and existing maximum receive power user access methods (Max receive). It can be seen that when the optimization goal is to improve "carbon efficiency", the carbon efficiency of this scheme (MADQN) is higher than that of the traditional scheme in all hours, and it maintains high stability in differentiated service areas (peak / valley). This shows that the method proposed in this invention has strong adaptability to service load fluctuations.
[0062] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" in this invention are intended to illustrate the invention and are not intended to limit the invention.
[0063] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. A method for constructing a new energy cellular network user access decision system, characterized in that, Comprise: S1, the base station agent interacts with the environment through its main network to generate a four-tuple and put into the playback buffer, t is time, S(t) is the global state obtained by splicing the state of each base station A(t) is the global action obtained by splicing the action of each base station r(t) is the global action reward; represents the i-th base station in the k-th type of base station, and each base station is deployed with an independent base station agent; the state of the base station agent includes the renewable energy power that can be consumed by the corresponding base station and the number of access users, the action of the base station agent includes regulating the user access to the corresponding base station, and the reward function for calculating the reward represents the optimization target of optimizing the network performance; S2, the network management unit obtains a randomly sampled quadruple batch from the playback buffer as a training quadruple; each base station obtains the corresponding state in the training quadruple , as a base station training sample and obtains the action selected by its base station agent from the target network in the state with the maximum Q value ; the network management unit obtains the action selected by each base station from its target network and splices it into A(t+1) to form a center training sample with the corresponding training quadruple ; S3, the central agent deployed in the network management unit obtains the central training sample, then calculates the Q value of performing action A(t+1) in state S(t+1) through its target network, combines with the reward r(t) to calculate the target Q value at time t, as the central training label of the central training sample at time t, and then calculates the Q value of performing A(t) in state S(t) by using its master network, as the central prediction result at time t, and the deviation between the central prediction result and the central training label is minimized as the training target to train the master network of the central agent; the master network of the trained central agent is used to calculate the Q value of performing A(t) in state S(t) as the base station training label of the base station training sample at time t and sent to each base station; S4, each base station agent obtains the base station training sample, then calculates the Q value of performing A(t) in state S(t) through its master network as the base station prediction result at time t, and the deviation between the base station prediction result and the base station training label is minimized as the training target to train the master network of the base station agent; S5, repeat S1 to S4 until a preset iteration number is reached, and the master network of the trained base station agent is the new energy cellular network user access decision system of the corresponding base station.
2. The method for constructing a new energy cellular network user access decision system as described in claim 1, characterized in that, State including the base station renewable energy power consumable at time t and the base station number of users accessing at time t and the user throughput requirement ; The reward function is: wherein, is the carbon efficiency of the network at time t, is the throughput of user j at time t, and U(t) is the total number of users accessing the network, is a penalty function for not meeting the throughput requirement of user j, which takes a positive number when and 0 or a negative number otherwise, is a set penalty weight.
3. The method of claim 2, wherein the method further comprises: determining a new energy cell network user access decision system based on the new energy cell network user access decision system model. When the penalty function takes the value 1, otherwise the penalty function takes the value 0.
4. The method for constructing a new energy cellular network user access decision system as described in claim 1, characterized in that, The action space of the base station agent is the selectable range corresponding to the base station access bias, and each base station action Access bias selected by the base station at time t : ； In the formula, represents the maximum access bias of the base station of the first type. represents the maximum access bias of the base station of the first type.
5. The method of claim 1, wherein the method further comprises: determining whether the new energy cellular network is a new energy cellular network of a same operator as the cellular network; and if the new energy cellular network is a new energy cellular network of a same operator as the cellular network, determining whether the new energy cellular network is a new energy cellular network of a same operator as the cellular network. When the master network of the central agent is trained with the deviation between the central prediction result and the central training label as the training target, the corresponding loss function is: ； In the formula, y(t) is the center training label of the center training sample at t moment, is the current network parameter of the master network of the center agent, represents a loss function, represents averaging, represents that the current network parameter is the Q value of the master network after performing the action A(t) at the current state S(t), which is also the center prediction result at t moment.
6. The method of claim 1, wherein the method further comprises: determining whether the new energy cellular network is a new energy cellular network of a first type or a new energy cellular network of a second type; and selecting the new energy cellular network access decision system based on the determination. For the base station When the base station main network is trained with the training target of minimizing the deviation between the base station prediction result and the base station training label, the loss function corresponding to the setting is: ； In the formula, is the base station training label of the base station training sample at time t, is the current network parameter of the main network of the base station agent, represents a loss function, represents averaging, represents that the current network parameter is the main network in the current state executes an action after the Q value, which is also the base station prediction result at time t.
7. The method of claim 1, wherein the method further comprises: determining a new energy cell network user access decision system based on the new energy cell network user access decision system model. Repeat S1 to S4 until a preset iteration number is reached, and the preset iteration number is greater than 30000 times.
8. A new energy cellular network user access method, characterized in that, Comprise: Base station Obtaining the state observed locally at the current time t And provide its user access decision system, the user access decision system corresponding to the agent master network selects the action with the maximum Q value in the state At the current time t To control the user to access the base station; Wherein, the access decision system of each base station is composed by using the construction method of any one of claims 1 to 7.