A Method for Establishing a Multi-Agent-Based Elevator Group Control and Scheduling Model and an Elevator Group Control System

Abstract

This invention discloses a method for establishing a multi-agent elevator group control scheduling model and an elevator group control system, relating to the field of intelligent elevator technology. The method includes: establishing an agent for each elevator in the elevator group control system, and constructing a global state space for the elevator group control system based on elevator data and the expected number of waiting passengers on each floor; constructing a collaboratively optimized reward function for each agent; and performing offline collaborative training on all agents based on the global state space and the reward function to obtain a policy model for formulating output action strategies for each agent. This invention deeply integrates a multi-agent reinforcement learning framework with the elevator group control energy scheduling problem, achieving multi-dimensional performance improvements in energy efficiency, equipment lifespan, and overall efficiency of the elevator group control system.

Description

Technical Field

[0001] This invention relates to the field of intelligent elevator technology, and in particular to a method for establishing a multi-agent-based elevator group control and scheduling model and an elevator group control system. Background Technology

[0002] In terms of energy scheduling, some elevator group control systems face a series of interconnected technical challenges, including insufficient internal coordination, rigid multi-objective trade-offs, poor real-time decision-making, and high wear and tear on auxiliary equipment.

[0003] Traditional elevator group control and scheduling methods cannot simultaneously achieve dynamic optimization of multiple objectives, such as efficient energy coordination and energy saving, equipment loss control, energy storage life maintenance, and passenger experience assurance, while meeting real-time requirements. Specifically, centralized optimization algorithms are too computationally complex to respond in real time; while simple distributed rules, although fast in decision-making, cannot achieve global energy coordination and multi-objective balance, resulting in the system losing sight of the balance between energy consumption, equipment life, and service quality. Summary of the Invention

[0004] The main objective of this invention is to provide a method for establishing a multi-agent-based elevator group control and scheduling model and an elevator group control system. This aims to solve the technical problem that traditional methods often focus on the energy-saving operation of a single elevator or simple load balancing, lacking a mechanism for real-time coordinated scheduling and direct matching of energy across multiple elevators at the system level.

[0005] To achieve the above objectives, this invention proposes a method for establishing a multi-agent-based elevator group control and scheduling model, wherein the elevator group control system includes multiple elevators; the method includes: An intelligent agent is established for each elevator in the elevator group control system. The intelligent agent is a computational model with an embedded reinforcement learning policy, and each intelligent agent is uniquely bound to one elevator. The global state space of the elevator group control system is constructed based on the elevator data of each elevator and the expected number of people waiting on each floor. A collaboratively optimized reward function is constructed for each agent, wherein the reward function is constructed as a weighted sum of multiple reward items, specifically including: an energy balance reward item aimed at achieving internal energy balance by promoting the matching of power generation and power consumption; an equipment loss reward item aimed at encouraging the reduction of equipment loss by associating with the real-time operating loss of elevators; an energy storage usage reward item aimed at extending the service life of energy storage equipment by associating with the usage behavior of energy storage equipment; and an elevator experience reward item aimed at improving service quality by associating with the average waiting time of the system. The weight coefficients corresponding to each reward item can be adjusted according to the system operating conditions. Based on the global state space and the reward function, offline collaborative training is performed on all agents to obtain a policy model, and output action strategies are formulated for each agent. The action strategies include assigning the elevator the roles of generator, power consumer, or backup agent, and determining the target floor of the elevator.

[0006] Furthermore, the energy balance reward is calculated as follows: R_energy=1-2×|P_gen-P_con| / (P_gen+P_con+ε); Where R_energy is the energy balance reward, P_gen is the sum of the absolute values ​​of the downward regenerative braking power of all elevators assigned as power generators, P_con is the sum of the absolute values ​​of the upward traction power of all elevators assigned as power consumers, and ε is a minimum constant. If |P_gen-P_con|≤α×(P_gen+P_con), where α is a preset energy difference coefficient, then a first preset positive reward value is added to R_energy.

[0007] Furthermore, the calculation method for the equipment loss reward item is as follows: R_loss = 0.5 - L_i / L_threshold; Where R_loss is the equipment loss reward item, L_i is the current loss value of the elevator corresponding to agent i, and L_threshold is a preset loss threshold greater than zero. If L_i ≥ L_threshold, then subtract the first preset negative penalty value from R_loss; if L_i ≤ β × L_threshold, where β is a loss coefficient less than 1, then add the second preset positive reward value to R_loss.

[0008] Furthermore, the calculation method for the energy storage usage reward is as follows: R_storage = -k × N_cd; Where R_storage is the energy storage usage reward, k is the preset penalty coefficient, and N_cd is the number of times the energy storage device is charged and discharged within a single scheduling cycle; If the energy storage device does not charge or discharge during the scheduling cycle, a third preset positive reward value is added to R_storage; If the depth of charge and discharge of the energy storage device exceeds a preset threshold for its state of charge, then a second preset negative penalty value is subtracted from R_storage.

[0009] Furthermore, the calculation method for the elevator experience reward items is as follows: R_experience=0.2-(T_wait / T_threshold)×0.2; Where R_experience is the elevator experience reward item, T_wait is the average waiting time of the elevator group control system, and T_threshold is the preset waiting time threshold. If T_wait≤T_threshold, then add a fourth preset positive reward value to R_experience.

[0010] Furthermore, the dynamic adjustment of weighting coefficients based on system operating conditions includes: When the system is operating during peak passenger flow periods, the weighting coefficient of the elevator experience reward items will be increased; When the health status of the energy storage device is detected to be lower than a preset health threshold, the weight coefficient of the energy storage usage reward item is increased.

[0011] Furthermore, the strategy model is configured to execute a time-segmented pairing strategy: When the global state space indicates peak passenger flow mode, the control agent operates according to building floor zones and prioritizes matching power generators and power consumers within the same zone; When the global state space indicates a peak or off-peak passenger flow pattern, the control agents autonomously pair up to form a mutual supply unit.

[0012] Furthermore, the first preset positive reward value, the second preset positive reward value, the third preset positive reward value, the fourth preset positive reward value, the first preset negative penalty value, and the second preset negative penalty value are all preset constants, and their values ​​are all between 0.1 and 0.5.

[0013] This invention also proposes a multi-agent-based elevator group control and scheduling method. The method employs a multi-agent-based elevator group control and scheduling model established using a multi-agent-based elevator group control and scheduling model establishment method. The method is characterized by comprising: In each scheduling cycle, the real-time operating status data of the elevator group control system is acquired; The real-time operating status data is input into the strategy model; based on the action strategy output by the strategy model, a real-time scheduling instruction is generated, which includes the roles assigned to each elevator and their target operating floors. The real-time scheduling instructions are executed to control the operation of the elevator group control system and perform energy scheduling and optimization.

[0014] The present invention also proposes an elevator group control system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the elevator group control and scheduling method based on any one of claims 1 to 9.

[0015] This invention deeply integrates a multi-agent reinforcement learning framework with the energy scheduling problem of elevator group control, solving the group control problems of insufficient energy coordination, difficulty in dynamically balancing multiple objectives, and complex centralized decision-making, and realizing multi-dimensional performance improvement of elevator group control system in terms of energy efficiency, equipment life and efficiency. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the multi-agent-based elevator group control and scheduling method of the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of the present invention and are not intended to limit the present invention.

[0021] To better understand the technical solution of the present invention, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, Figure 1 This is a flowchart provided in Embodiment 1 of the elevator group control and scheduling method based on multi-agent systems of the present invention.

[0023] A method for establishing a multi-agent-based elevator group control and scheduling model, characterized in that the method includes: S10 establishes an intelligent agent for each elevator in the elevator group control system. The intelligent agent is a computational model that embeds a reinforcement learning strategy, and each intelligent agent is uniquely bound to one elevator. S20, construct the global state space of the elevator group control system based on the elevator data of each elevator and the expected number of people waiting on each floor; S30, construct a collaboratively optimized reward function for each agent, wherein the expression of the reward function is: R=ω1×R_energy+ω2×R_loss+ω3×R_storage+ω4×R_experience; where ω1,ω2,ω3,ω4 are weight coefficients dynamically adjusted according to the system operating conditions, and R_energy, R_loss, R_storage, and R_experience are the energy balance reward, equipment loss reward, energy storage usage reward, and elevator experience reward, respectively. S40, based on the global state space and the reward function, perform offline collaborative training on all agents to obtain a policy model, and formulate output action policies for each agent; the action policies include assigning the elevator the roles of generator, power consumer or backup agent, and determining the target floor of the elevator.

[0024] This invention deeply integrates a multi-agent reinforcement learning framework with the energy scheduling problem of elevator group control, solving the group control problems of insufficient energy coordination, difficulty in dynamically balancing multiple objectives, and complex centralized decision-making, and realizing multi-dimensional performance improvement of elevator group control system in terms of energy efficiency, equipment life and efficiency.

[0025] Taking a 20-story office building with 2 basement levels, equipped with 6 elevators (numbered E1-E6) with a rated load capacity of 1600kg and a speed of 2.5m / s, as an example, this system is equipped with a lithium-ion battery energy storage device with a rated capacity of 50kWh. This embodiment will elaborate in detail on how to apply the method of the claims to achieve energy optimization scheduling of this elevator group control system.

[0026] S10 establishes an intelligent agent for each elevator in the elevator group control system. The intelligent agent is a computational model that embeds a reinforcement learning strategy, and each intelligent agent is uniquely bound to one elevator. In this embodiment, the six elevators (E1-E6) are modeled as six independent agents (Agent1-Agent6). Each agent has an independent policy network, which is responsible for learning and outputting the operation policy of its corresponding elevator.

[0027] In this embodiment, the global state space construction includes: predicting the number of people waiting for elevators on each floor in the next 3 minutes (time period t) using visual sensors deployed in the elevator lobby and historical data analysis (dimension: 1x22); collecting the load, direction of operation, speed, current floor, motor current and temperature of each elevator in real time, and calculating the real-time power P_i (positive for upward power consumption and negative for downward power generation) and the comprehensive loss index L_i obtained based on the harmonics of motor current and the spectrum analysis of bearing vibration; obtaining the current state of charge (SOC), state of health (SOH), and number of charge and discharge cycles (N_cd) of the energy storage device in the previous scheduling cycle; and defining the system operation mode G (peak / off-peak / low-peak) based on the passenger flow data.

[0028] In this embodiment, the global state space S_t, which is updated every 3 seconds, can be represented as: S_t=[D_t,(P_i,L_i,Floor_i)_{i=1...6},SOC,SOH,N_cd,G].

[0029] S20, construct a collaboratively optimized reward function for each agent, where the expression of the reward function is: R=ω1×R_energy+ω2×R_loss+ω3×R_storage+ω4×R_experience; where ω1,ω2,ω3,ω4 are weight coefficients dynamically adjusted according to the system operating conditions, and R_energy, R_loss, R_storage, and R_experience are the energy balance reward, equipment loss reward, energy storage usage reward, and elevator experience reward, respectively. In this embodiment, a reward function R as defined in the claims is constructed for each agent. The specific process includes: in a high-fidelity digital twin simulation environment, the MAPPO (Multi-Agent Proximal Policy Optimization) algorithm is used to conduct 300,000 steps of offline collaborative training on six agents. Each agent outputs role assignment and target floor based on its current state S_t. After the environment performs an action, it transitions to a new state S_{t+1} and receives a reward R_t. The agent uses this empirical data to update its policy network parameters. This process iterates continuously until the policy model converges. The optimal policy model that has converged is then deployed to the central controller of the actual building. During operation, the controller inputs the real-time collected S_t into the model every 3 seconds. The model instantly outputs scheduling instructions for E1-E6, including: assigning each elevator as a generator (F), a consumer (C), or a backup (S), and the target response floor for each elevator. The controller executes these instructions to complete one scheduling cycle.

[0030] In this embodiment, within a certain training step or scheduling cycle, the system state is as follows: Policy model output, E1, E3, E5 are generators (F), E2, E4 are consumers (C), E6 is a reserve (S); power is: P1_gen = -10kW, P3_gen = -6kW, P5_gen = -4kW; P2_con = 12kW, P4_con = 5kW; loss is: L_2 (corresponding to Agent2) = 85 (unit: loss exponent), preset L_threshold = 100, β = 0.7; within this scheduling cycle, The energy storage device is charged and discharged N_cd=2 times, with one discharge depth of 35% and a preset depth threshold of 30%; the average waiting time for the elevator in the past 3 minutes T_wait=28 seconds and the national standard threshold T_threshold=30 seconds; preset constants: ε=1e-3, α=0.1, k=0.15; all reward / penalty values ​​are within the range [0.1,0.5]. In this embodiment: first positive reward=0.2, second positive reward=0.1, third positive reward=0.3, fourth positive reward=0.1, first negative penalty=0.5, second negative penalty=0.2; The energy difference coefficient α determines the leniency of matching power generation and consumption. A smaller α (e.g., 0.05-0.15) requires more precise matching, which is beneficial for extreme energy saving; a larger α gives the scheduling strategy more flexibility to take other objectives into account. The loss coefficient β serves as a baseline for low loss. A range of 0.6-0.8 indicates that a reward is given when the loss is below 60%-80% of the threshold, encouraging the strategy to maintain a low loss level. The penalty coefficient k is used to adjust the cost of charging and discharging behavior of energy storage devices. It needs to be set in conjunction with the cycle life curve of the energy storage battery so that the model can reasonably avoid harmful frequent shallow charging and discharging when making decisions. The dynamic weights ω1-ω4 are key to achieving dynamic trade-offs among multiple objectives, and their adjustment logic is directly related to system operating condition identification.

[0031] Specifically, the optimal combination of parameters is obtained through automatic parameter optimization in a simulation environment. The algorithms used include grid search, Bayesian optimization, or meta-optimization in parallel with reinforcement learning. The optimization objective is to achieve the optimal overall performance indicators in a representative, long-term simulation scenario (e.g., a simulation run of one month).

[0032] The following is an example of reward calculation for Agent2 (controlling the elevator E2): The energy balance bonus R_energy is calculated as follows: First, calculate P_gen and P_con, where P_gen = |-10| + |-6| + |-4| = 20kW, and P_con = |12| + |5| = 17kW; then, calculate the base value R_energy for the energy balance bonus using P_gen and P_con, where R_energy = 1 - 2 |20-17| / (20+17+0.001)≈0.838; Determine if |P_gen-P_con|≤ɑ×(P_gen+P_con) is true: |20-17|=3≤0.1 (20+17)=3.7, trigger condition, increase R_energy by 0.2 to get R_energy, R_energy=0.838+0.2=1.038.

[0033] This reward incentivizes the strategy model to match the total power generation with the total power consumption as much as possible when allocating roles for power generators and consumers. A high R_energy value means a high proportion of renewable energy is consumed locally within the system, directly reducing the need to purchase and feed power from the grid, thus achieving system-level energy savings. In this embodiment, the 3kW power difference accounts for only 8.1% of the total power flow, indicating that the scheduling strategy achieves efficient internal energy balance.

[0034] In this embodiment, the calculation process for the equipment loss reward R_loss (Agent2 individual value) is as follows: First, calculate the base value of the equipment loss reward: R_loss = 0.5 - / L_threshold=0.5-85 / 100=-0.35; Li=85, 85≥100 is not valid, 85≤70 is not valid, there is no additional reward or penalty, and the final result is R_loss=-0.35.

[0035] This reward is negatively correlated with the real-time wear and tear value L_i of the elevator. When the target floor or operating strategy selected by the model for elevator E2 leads to an increase in its wear and tear, R_loss will decrease, thus penalizing that strategy during training. Long-term training enables the model to learn to select a more wear-efficient operating curve for each elevator (such as smooth acceleration and reduced abrupt stops and starts), thereby actively controlling the wear index L_i of each elevator below the threshold L_threshold, achieving the effect of extending the life of the elevator's mechanical and electrical components.

[0036] In this embodiment, the calculation process for the energy storage usage reward R_storage (system global value) is as follows: First, calculate the base value of the energy storage usage reward: R_storage = -0.15 2 = -0.3; The system detects that it is charging and discharging, but does not trigger the zero charge / discharge reward, and the discharge depth is 35% > 30%, triggering the depth penalty, and finally obtaining R_storage = -0.3 - 0.2 = -0.5.

[0037] This energy storage system uses rewards to directly penalize the number of times the energy storage device operates and the depth of charge and discharge. The strong negative rewards force the strategy model to prioritize balancing instantaneous power by finely scheduling the F / C matching between elevators when making decisions, and to use energy storage as a last resort. This can significantly reduce the number of energy storage cycles, avoid harmful deep discharge, and thus effectively delay battery capacity decay and extend the overall lifespan of the energy storage system.

[0038] In this embodiment, the calculation process for the elevator experience reward R_experience (system global value) is as follows: First, calculate the base value of the elevator experience reward: R_experience = 0.2 - (28 / 30) 0.2≈0.013; if T_wait=28 seconds≤30 seconds, the reward condition is triggered, and 0.1 is added, resulting in R_experience=0.013+0.1=0.113.

[0039] The elevator experience reward is directly linked to service quality. In order to obtain a higher R_experience, the strategy model must optimize scheduling to reduce the average passenger waiting time T_wait. This reward allows the model to learn during training how to quickly respond to the call requests of each floor while saving energy and protecting equipment, thus ensuring and improving the passenger's elevator experience.

[0040] Furthermore, the weighting coefficients are dynamically adjusted based on the system's operating conditions, including: When the system is operating during peak passenger flow periods, increase the weight coefficient ω4 for the passenger experience reward items; When the health status of an energy storage device is detected to be lower than the preset health threshold, the weight coefficient ω3 of the energy storage usage reward item is increased.

[0041] S30, based on the global state space and the reward function, perform offline collaborative training on all agents to obtain a policy model, and formulate output action policies for each agent; the action policies include assigning the elevator the roles of generator, power consumer or backup agent, and determining the target floor of the elevator. In this embodiment, the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm is selected, employing a centralized training and distributed execution framework. Within this framework, each elevator agent possesses an independent policy network. Its input consists of features relevant to itself in the global state space, such as its own elevator's real-time power, losses, location, and passenger flow information on associated floors. The output is a well-defined action policy, comprising two parts: first, a probability distribution for assigning its own elevator the roles of generator, consumer, or backup; and second, a probability distribution for selecting the target operating floor. All agents share a centralized value evaluation network, which evaluates the overall performance of the global state system. The agents are trained in a high-fidelity digital twin simulation environment for elevator group control, accurately simulating elevator dynamics, energy consumption models, and random passenger flow. During the training cycle, each agent interacts in the simulation environment according to the current policy, collecting a large amount of state, action, reward, and new state experience data. The reward is calculated in real time by the hierarchical weighted reward function defined above. By using this experience data to calculate the advantage function of each agent's actions to evaluate their merits and update the parameters of the policy network and value network, the expected value of the long-term cumulative reward of each agent can be maximized under the constraint of stable policy update magnitude.

[0042] S40, based on the global state space and the reward function, perform offline collaborative training on all agents to obtain a policy model, and formulate output action policies for each agent; the action policies include assigning the elevator the roles of generator, power consumer or backup agent, and determining the target floor of the elevator.

[0043] Furthermore, through hundreds of thousands to millions of iterative training steps, the strategy model gradually learns to internalize complex optimization logic, enabling it to understand how to match power generators and power consumers to achieve efficient energy balance, proactively avoid high-loss operation, reduce unnecessary use of energy storage devices, and dynamically adjust the optimization focus according to passenger flow patterns (peak / off-peak), so that more emphasis is placed on the elevator experience during peak hours.

[0044] Furthermore, the convergence of training is marked by the stability of policy performance and the achievement of preset requirements on multi-objective test metrics; the final obtained policy model, namely the set of parameters of each agent's policy network after training, can achieve millisecond-level inference, thereby transforming the complex multi-objective online optimization problem into an efficient mapping of the real-time global state and directly outputting the optimal role allocation and floor operation instructions.

[0045] Furthermore, the strategy model is configured to execute a time-segmented pairing strategy: When the global state space indicates peak passenger flow mode, the control agent operates according to building floor zones and prioritizes matching power generators and power consumers within the same zone. When the global state space indicates that the elevator passenger flow is in a non-peak or low-peak mode, the control agents autonomously pair up to form a mutual supply unit.

[0046] This invention also proposes a multi-agent-based elevator group control and scheduling method. The method employs a multi-agent-based elevator group control and scheduling model established using a multi-agent-based elevator group control and scheduling model establishment method. The method is characterized by comprising: In each scheduling cycle, real-time operating status data of the elevator group control system is acquired; the real-time operating status data is input into the strategy model; based on the action strategy output by the strategy model, a real-time scheduling instruction is generated, which includes the role assigned to each elevator and its target operating floor; the real-time scheduling instruction is executed to control the operation of the elevator group control system and perform energy scheduling and optimization.

[0047] In this embodiment, the converged policy model, i.e., the final policy network parameters of the six agents, is fixed and deployed in the scheduling server in the central control room of the office building. This server is connected in real-time to the controllers of the six elevators (E1-E6), the call panel sensors on each floor, and the central energy storage battery management system via an industrial communication network. Online scheduling operates in a fixed scheduling cycle (e.g., 3 seconds): at the beginning of each cycle, the scheduling server collects real-time data through the sensor network, including the instantaneous load, direction of travel, speed, and floor of each elevator; the current call signals on each floor and the number of people waiting for elevators estimated by visual sensors; and the real-time state of charge (SOC) and state of health (SOH) of the energy storage devices. This raw data is preprocessed and integrated into the current global state space S_t in a format completely consistent with the offline training phase. Subsequently, S_t is input into the loaded strategy model for forward inference; the model calculates the composite action output for the 6 elevators in parallel within milliseconds using forward propagation: for each elevator, the model outputs a three-dimensional vector representing the probability of the three roles of generator, consumer, and backup, and a probability distribution vector representing the selection tendency of 22 target floors.

[0048] Furthermore, in the online application of this embodiment, the scheduling algorithm adopts a deterministic greedy strategy, selecting the role and target floor with the highest probability for each elevator, thereby generating a set of clear and unambiguous real-time scheduling instructions. These instructions are immediately encapsulated and distributed to the independent controllers of each elevator: the instructions not only include the target floor, but more importantly, they clarify the energy role assigned to it in this scheduling cycle. For example, instruction E1: "Target floor B2, role: generator"; instruction E2: "Target floor 15F, role: power consumer". While executing the floor operation command, the elevator controller broadcasts its energy role status through the communication network, so that the regenerative energy generated by elevators designated as "generators" during braking can be preferentially directed to elevators designated as "power consumers" at the same time through a common DC bus or energy router, thereby achieving direct, real-time, and coordinated energy scheduling. The entire scheduling process repeats itself, and the strategy model continuously makes decisions based on the latest global state, enabling the system to dynamically respond to changes in passenger flow and adaptively protect equipment. It always operates in a multi-objective collaborative optimal direction of energy balance, controllable loss, and guaranteed experience, ultimately transforming the intelligence gained from offline training into tangible energy-saving benefits, extended equipment lifespan, and guaranteed service quality during online operation.

[0049] The present invention also proposes an elevator group control system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The feature is that when the processor executes the program, it implements an elevator group control scheduling method based on multiple agents.

[0050] The present invention also proposes a computer-readable storage medium storing a computer program, characterized in that the program, when executed by a processor, implements a multi-agent-based elevator group control and scheduling method.

[0051] The above are only some embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made under the technical concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for establishing a multi-agent-based elevator group control and scheduling model, wherein the elevator group control system includes multiple elevators; characterized in that, The method includes: An intelligent agent is established for each elevator in the elevator group control system. The intelligent agent is a computational model with an embedded reinforcement learning policy, and each intelligent agent is uniquely bound to one elevator. The global state space of the elevator group control system is constructed based on the elevator data of each elevator and the expected number of people waiting on each floor; A collaboratively optimized reward function is constructed for each agent, wherein the reward function is constructed as a weighted sum of multiple reward items, specifically including: an energy balance reward item for achieving internal energy balance by promoting the matching of power generation and power consumption; an equipment loss reward item for encouraging the reduction of equipment loss by associating with the real-time operating loss of elevators; an energy storage usage reward item for extending the service life of energy storage devices by associating with the usage behavior of energy storage devices; and an elevator experience reward item for improving service quality by associating with the average waiting time of the system. The weight coefficients corresponding to each reward item can be adjusted according to the system operating conditions. Based on the global state space and the reward function, offline collaborative training is performed on all agents to obtain a policy model, which is used to formulate output action policies for each agent. The action policies include assigning the elevator the roles of generator, power consumer, or backup agent, and determining the target floor of the elevator.

2. The method according to claim 1, characterized in that, The energy balance reward is calculated as follows: R_energy=1-2×|P_gen-P_con| / (P_gen+P_con+ε); Where R_energy is the energy balance reward, P_gen is the sum of the absolute values ​​of the downward regenerative braking power of all elevators assigned as power generators, P_con is the sum of the absolute values ​​of the upward traction power of all elevators assigned as power consumers, and ε is a minimum constant. If |P_gen-P_con|≤α×(P_gen+P_con), where α is a preset energy difference coefficient, then a first preset positive reward value is added to R_energy.

3. The method according to claim 1, characterized in that, The calculation method for the equipment loss reward is as follows: R_loss = 0.5 - L_i / L_threshold; Where R_loss is the equipment loss reward item, L_i is the current loss value of the elevator corresponding to agent i, and L_threshold is a preset loss threshold greater than zero. If L_i ≥ L_threshold, then subtract the first preset negative penalty value from R_loss; if L_i ≤ β × L_threshold, where β is a loss coefficient less than 1, then add the second preset positive reward value to R_loss.

4. The method according to claim 1, characterized in that, The calculation method for the energy storage usage reward is as follows: R_storage = -k × N_cd; Where R_storage is the energy storage usage reward, k is the preset penalty coefficient, and N_cd is the number of times the energy storage device is charged and discharged within a single scheduling cycle; If the energy storage device does not charge or discharge during the scheduling cycle, a third preset positive reward value is added to R_storage; If the depth of charge and discharge of the energy storage device exceeds a preset threshold for its state of charge, then a second preset negative penalty value is subtracted from R_storage.

5. The method according to claim 1, characterized in that, The calculation method for the elevator experience rewards is as follows: R_experience=0.2-(T_wait / T_threshold)×0.2; Where R_experience is the elevator experience reward item, T_wait is the average waiting time of the elevator group control system, and T_threshold is the preset waiting time threshold. If T_wait≤T_threshold, then add a fourth preset positive reward value to R_experience.

6. The method according to claim 1, characterized in that, The dynamic adjustment of weighting coefficients based on system operating conditions includes: When the system is operating during peak passenger flow periods, the weighting coefficient of the elevator experience reward items will be increased; When the health status of the energy storage device is detected to be lower than a preset health threshold, the weight coefficient of the energy storage usage reward item is increased.

7. The method according to claim 1, characterized in that, The strategy model is configured to execute a time-segmented pairing strategy: When the global state space indicates peak passenger flow mode, the control agent operates according to building floor zones and prioritizes matching power generators and power consumers within the same zone; When the global state space indicates a peak or off-peak passenger flow pattern, the control agents autonomously pair up to form a mutual supply unit.

8. The method according to any one of claims 2 to 5, characterized in that, The first preset positive reward value, the second preset positive reward value, the third preset positive reward value, the fourth preset positive reward value, the first preset negative penalty value, and the second preset negative penalty value are all preset constants, and their values ​​are all between 0.1 and 0.

5.

9. A multi-agent-based elevator group control and scheduling method, wherein the method employs a multi-agent-based elevator group control and scheduling model established by the multi-agent-based elevator group control and scheduling model establishment method as described in any one of claims 1 to 8, characterized in that, The method includes: acquiring real-time operating status data of the elevator group control system in each scheduling cycle; inputting the real-time operating status data into a strategy model; generating real-time scheduling instructions containing the roles assigned to each elevator and their target operating floors according to the action strategies output by the strategy model; and executing the real-time scheduling instructions to control the operation of the elevator group control system and perform energy scheduling and optimization.

10. A group control system for elevators, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the elevator group control and scheduling method based on any one of claims 1 to 9.

Images