Double-elevator group control scheduling method and system based on scene self-adaption and machine learning

Abstract

The invention discloses a double-elevator group control scheduling method and system based on scene self-adaption and machine learning, and belongs to the technical field of intelligent building control. In order to solve the problems that an existing double-elevator system is lagged in response in a non-peak period and low in peak period transport capacity matching degree, full-floor balanced response is achieved by monitoring the elevator state in real time, controlling double elevators to stop at the bottom layer in an idle mode and calculating a middle floor through a weighted gravity center method; in the peak mode, differentiated directional elevator parking strategies are executed according to building types, and the up-down tidal passenger flow of morning and evening peaks is accurately adapted. In addition, the system introduces a machine learning module, and dynamically updates a peak time window based on historical data by using a time sequence density detection algorithm. According to the method, the average waiting time is effectively shortened, the carrying efficiency in the peak period is improved, and the elevator system is endowed with the self-evolution ability adapting to the flow change of building personnel.

Description

A method and system for dual-elevator group control and scheduling based on scene adaptation and machine learning Technical Field

[0001] This invention relates to the field of intelligent building and automatic control technology, and more specifically, to an elevator dual-elevator group control scheduling method and system based on scene adaptation and machine learning. Background Technology

[0002] With the acceleration of urbanization and the increasing number of high-rise buildings, the operational efficiency of elevator systems directly affects the building's user experience and energy consumption. Existing dual-elevator or multi-elevator group control systems commonly suffer from the following technical challenges:

[0003] The problem of "ineffective waiting for two elevators" during off-peak hours: At night or outside of working hours, traditional elevators often adopt a "wait nearby" or "wait at the last position" strategy. This may result in both elevators stopping at the same time on a high floor or a non-hotspot floor. If a user calls from the ground floor lobby, or if there is a need to cross floors in the middle, the elevator response time will be too long.

[0004] Rigid Peak-Hour Strategies: Existing crowd control algorithms typically use fixed schedules (e.g., 7:00-9:00 AM and 5:00-7:00 PM) to trigger peak-hour patterns. However, passenger flow patterns vary significantly across different building types. For example, the evening peak in residential buildings is characterized by "a large flow of people moving upwards from the ground floor," while in office buildings it's characterized by "a large flow of people moving downwards from the upper floors." If a simple "zoned flow control" or "base station return" strategy is used uniformly without considering the building context, it can easily lead to wasted capacity (e.g., in office buildings, elevators run empty back to the ground floor during the evening peak, while the upper floors are crowded with people waiting to leave work).

[0005] Lack of adaptive evolutionary capability: The personnel structure of buildings is dynamic (such as the arrival of new companies and changes in work and rest schedules due to seasonal changes). Once the parameters of traditional algorithms (such as standby floor location and peak start and end times) are written into the program chip, they are often difficult to change, resulting in a significant decline in system efficiency after several years of operation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for dual-elevator group control and scheduling based on scene adaptation and machine learning.

[0007] The core technical solution of this invention is as follows:

[0008] An elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning includes the following steps:

[0009] Step S1 (Status Awareness): Real-time monitoring of the operating status data (position, direction, load, door status) of two elevators in the elevator group and the building's call command flow data.

[0010] Step S2 (Intelligent Idle): Determine whether the system meets the idle mode entry conditions. If it does, control the two elevators to stop at the base station and the dynamic intermediate floor respectively according to the asymmetric distributed standby strategy. If it is determined that the idle mode conditions are not met (i.e. the system is in "busy mode"), the system responds to the call command according to the standard minimum waiting time group control strategy and does not perform forced return or distributed standby actions; thereby minimizing the average response distance of the entire floor.

[0011] The logic for determining the "Idle Mode Entry Condition" is as follows: When the following conditions are met simultaneously and the preset threshold Tidle is continuously set, the system is determined to enter the idle mode: (1) The load in both elevator cars is zero; (2) Both elevator door operators are closed; (3) There are currently no unresponsive external call commands or internal selection commands in the system.

[0012] The “asymmetric distributed standby strategy” is as follows: control the first elevator E1 to automatically run to the ground floor base station 1F of the building and keep it in standby mode; control the second elevator E2 to automatically run to the calculated dynamic intermediate floor Fmid and keep it in standby mode; if the difference between the current floor of the second elevator and the target Fmid is less than the preset fine-tuning threshold Nth, then it will stay in place and not perform the re-leveling action.

[0013] The calculation formula for the dynamic intermediate layer Fmid is a weighted centroid model:

[0014]

[0015] Where H is the total number of floors in the building, L i W is the physical layer number of the i-th layer. i Let W be the call weight for the i-th layer; i for:

[0016]

[0017] Among them, C up (i) and C down (i) represents the number of uplink calls and downlink calls at layer ii within the statistical period, respectively, and α and β are preset coefficients that cause the uplink and downlink weights to differ.

[0018] Step S3 (Scenario-based Peak): Determine if the current time is within a peak time window. If it is, identify the preset building scenario type and execute a differentiated directional elevator parking strategy accordingly. If it is a "residential building," execute an upward peak strategy prioritizing the ground floor; if it is an "office building," execute a downward peak strategy prioritizing the upper floors. If it is detected that the current time is not within a peak time window (i.e., in "off-peak or low-peak mode"), the system jumps to step S2 based on the real-time call status to execute idle standby determination or regular scheduling. The "differentiated directional elevator parking strategy" specifically includes:

[0019] Step S31): Read the building type identifier from the system configuration file;

[0020] Step S32): If the building is identified as a "residential building", during the evening peak time window, when any elevator completes its service task and becomes idle, a scheduling instruction to return to the bottom base station is forcibly generated, unless the elevator has been assigned to respond to other calls.

[0021] Step S33): If the building is identified as an "office building", during the evening peak hours, when any elevator completes its service task and becomes idle, a scheduling instruction is forcibly generated to move upwards to the preset high-rise interval Ftop−KFtop−K to FtopFtop.

[0022] Step S4 (Closed-loop evolution): Based on historical operating data, the location parameters of the dynamic intermediate layer and the time parameters of the peak time window are periodically updated using a machine learning model to obtain the specific floor number of the "intermediate hotspot layer" and the start and end times of the "peak time window". If the analysis results show that the passenger flow characteristics have not changed significantly or the data sample size is insufficient, the current parameters are kept unchanged and no update operation is performed.

[0023] The update of the "peak time window" adopts a time series density detection algorithm: Define a time sliding window Δt, and calculate the call density ρ(t) within the window; When it is detected that ρ(t) is higher than the high threshold θhigh for M consecutive periods, mark the current time as the peak start time; When it is detected that ρ(t) is lower than the low threshold θlow for N consecutive periods, mark the current time as the peak end time; The update operation smooths the newly identified time points through moving average filtering, and the time adjustment range of a single update is limited to within the maximum step size ΔTmax.

[0024] An elevator control system includes: a data acquisition module configured to acquire Hall sensor signals from all floors, car weighing signals, and button trigger signals; a central processing unit (CPU) configured with scene recognition logic and a mode switching state machine for executing the scheduling method according to any one of claims 1 to 6; a storage module for storing building type configuration files and historical operation logs for at least the most recent 30 days; and a cloud or local computing module for running machine learning algorithms, periodically reading data from the storage module and calculating update parameters, and sending the results back to the central processing unit.

[0025] The central processing unit also includes a hysteresis comparator, which introduces a time delay when switching between idle and peak modes to prevent frequent mode oscillations in critical states.

[0026] The beneficial effects of this invention are as follows:

[0027] Lightning-fast response: Through a "bottom and middle" standby distribution, the system can dispatch the nearest elevator to respond regardless of whether the call occurs on the bottom floor or any intermediate floor, significantly shortening the average waiting time during off-peak hours.

[0028] Precise capacity matching: A reverse peak strategy is designed to address the tidal differences between office buildings and residential buildings, which significantly reduces the ineffective empty runs of elevators and improves the transportation capacity during peak hours.

[0029] Self-growth capability: The system can automatically adjust parameters according to the building's usage, requiring no frequent manual maintenance and always maintaining optimal operating status. Attached Figure Description

[0030] Figure 1 is a block diagram of the overall system architecture of an embodiment of the present invention, showing the interaction of the three layers of data acquisition, core control and intelligent decision-making.

[0031] Figure 2 is a flowchart of the main control logic of an embodiment of the present invention, which shows in detail the state machine switching of the system between normal, idle and peak modes.

[0032] Figure 3 is a logic diagram of the specific strategies for idle and peak periods in an embodiment of the present invention, showing the judgment logic for the "residential / office building" branch.

[0033] Figure 4 is a flowchart of machine learning adaptive optimization according to an embodiment of the present invention, showing the closed-loop process of parameter update. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. A method for scheduling dual-elevator group control based on scene adaptation and machine learning includes the following steps:

[0035] Step S1): Real-time monitoring of the operating status data of two elevators in the elevator group and the building's call command flow data.

[0036] Step S2): When the system determines that it has entered idle mode, it executes the "asymmetric distributed standby strategy". It controls the first elevator to automatically return to the bottom base station to wait, and controls the second elevator to automatically run to the dynamically calculated "intermediate hotspot floor" to wait, thereby minimizing the average response distance of the entire floor; if it determines that the idle mode conditions are not met (i.e., the system is in "busy mode"), the system responds to the call command according to the standard minimum waiting time group control strategy, and does not execute the forced return or distributed standby actions.

[0037] The judgment logic of "Idle mode entry condition" in step S2) is as follows: when the following conditions are met and the preset threshold Tidle is continuously preset, the idle mode is determined: (1) the load in both elevator cars is zero; (2) the door operators of both elevators are closed; (3) there are no unresponsive external call commands or internal selection commands in the system.

[0038] The “asymmetric distributed standby strategy” in step S2) is as follows: control the first elevator E1 to automatically run to the base station 1F on the ground floor of the building and keep it in standby mode; control the second elevator E2 to automatically run to the calculated dynamic intermediate floor Fmid and keep it in standby mode; wherein, if the difference between the current floor of the second elevator and the target Fmid is less than the preset fine-tuning threshold Nth, it will stay in place and not perform the re-leveling action.

[0039] The formula for calculating the dynamic intermediate layer Fmid is the weighted centroid model:

[0040]

[0041] Where H is the total number of floors in the building, Li is the physical floor number of the i-th floor, and Wi is the call weight of the i-th floor; the call weight Wi is:

[0042]

[0043] Where Cup(i) and Cdown(i) are the number of uplink calls and downlink calls in layer ii within the statistical period, respectively, and α and β are preset coefficients that cause the uplink and downlink weights to differ.

[0044] Step S3): Determine whether the current time is within the peak time window. If it is within the peak period, identify the preset building scenario type and execute a differentiated directional elevator parking strategy accordingly. If it is a "residential building", execute the upward peak strategy of prioritizing the ground floor; if it is an "office building", execute the downward peak strategy of prioritizing the upper floors. If it is detected that the current time is not within the peak time window (i.e., in "off-peak or low-peak mode"), the system jumps to step S2 to perform idle standby determination or regular scheduling based on the real-time call status.

[0045] The "differentiated directional elevator strategy" in step S3) specifically includes:

[0046] Step S31): Read the building type identifier from the system configuration file;

[0047] Step S32): If the building is identified as a "residential building", during the evening peak time window, when any elevator completes its service task and becomes idle, a scheduling instruction to return to the bottom base station is forcibly generated, unless the elevator has been assigned to respond to other calls.

[0048] Step S33): If the building is identified as an "office building", during the evening peak hours, when any elevator completes its service task and becomes idle, a scheduling instruction is forcibly generated to move upwards to the preset high-rise interval Ftop−KFtop−K to FtopFtop.

[0049] Step S4): Based on historical operational data, the location parameters of the dynamic intermediate layer and the time parameters of the peak time window are periodically updated using a machine learning model to obtain the specific floor number of the "intermediate hotspot layer" and the start and end times of the "peak time window". If the analysis results show that the passenger flow characteristics have not changed significantly or the data sample size is insufficient, the current parameters are kept unchanged and no update operation is performed.

[0050] In step S4), the update of the "peak time window" adopts a time series density detection algorithm: a time sliding window Δt is defined, and the call density ρ(t) within the window is calculated; when ρ(t) is detected to be higher than the high threshold θhigh for M consecutive periods, the current time is marked as the peak start time; when ρ(t) is detected to be lower than the low threshold θlow for N consecutive periods, the current time is marked as the peak end time; the update operation smooths the newly identified time points through a moving average filter MovingAverage, and the time adjustment range of a single update is limited to within the maximum step size ΔTmax.

[0051] An elevator control system includes: a data acquisition module configured to acquire Hall sensor signals from all floors, car weighing signals, and button trigger signals; a central processing unit (CPU) configured with scene recognition logic and a mode switching state machine for executing the scheduling method; a storage module for storing building type configuration files and historical operation logs for at least the last 30 days; and a cloud or local computing module for running machine learning algorithms, periodically reading data from the storage module and calculating and updating parameters, and sending the results back to the CPU. The CPU further includes a hysteresis comparator (HysteresisComparator) to introduce a time delay when switching between idle and peak modes, preventing frequent mode oscillations in critical states.

[0052] Example 1: System Overall Architecture

[0053] As shown in Figure 1, this system mainly includes a physical layer sensing module, a core control unit (CPU), and an intelligent decision-making module.

[0054] Physical layer: Includes call buttons installed in each elevator lobby, weighing sensors inside the car (to determine if the elevator is empty), and door operator status sensors. Control unit: Responsible for executing real-time scheduling logic. Intelligent decision-making module: Can be deployed on a local server or in the cloud, responsible for storing long-term historical data (such as call logs for the past 30 days) and running machine learning algorithms.

[0055] Example 2: Control Logic for Idle Mode

[0056] As shown in Figures 2 and 3 (left side), the system continuously monitors the elevator status. Entry conditions: To prevent elevators from being forcibly reset during service, the system employs strict "idle determination logic." An elevator is considered idle only when it simultaneously meets the following conditions for a duration exceeding the threshold (TidleTidle, e.g., 180 seconds): the load sensor reading inside the car is zero (no one is present); the elevator doors are fully closed; and there are currently no primary or secondary task instructions assigned to that elevator. Execution action: Once idle mode is confirmed, the system looks up a table to assign a standby target:

[0057] Elevator A: The target floor is set to floor 1 (base station). This ensures that users entering the building receive a response within seconds.

[0058] Elevator B: The target floor is set to FmidFmid (a dynamic intermediate floor, which can initially be set to half of the total number of floors). This ensures that inter-floor movement requests within a floor (such as going from the 15th floor to the 5th floor) can be responded to quickly.

[0059] Fine-tuning energy logic: To avoid energy consumption caused by frequent micro-movements, the system introduces a fine-tuning threshold Nth (e.g., 2 floors). Assuming the optimal standby floor Fmid is 10 floors, if elevator B is currently stopped at the 11th or 9th floor (difference ≤ Nth), the system determines that it is "in position" and does not send a movement command, remaining in place and waiting.

[0060] Example 3: Scenario-based Peak Mode

[0061] A major innovation of this invention lies in "scene awareness". As shown on the right side of Figure 3: the Building_Type parameter is preset in the system configuration file.

[0062] Scenario A (Residential Building): During the evening rush hour, a large number of residents return home from get off work. At this time, the system's primary task is to quickly "reclaim" the elevators scattered across various floors and bring them back to the 1st floor. Therefore, the strategy mandates that as soon as an elevator becomes available, it should immediately descend to the 1st floor to queue, rather than remaining stationary.

[0063] Scenario B (Office Building): During the evening rush hour, a large number of employees leave work. At this time, the system's primary task is to "pick up" passengers on higher floors. The strategy mandates that once the elevator drops off passengers and becomes available on the ground floor, it should immediately ascend at full speed to the preset high-floor waiting area (e.g., floors 15-20) to reduce the waiting anxiety of employees on higher floors.

[0064] Example 4: Detailed Explanation of Machine Learning Algorithms (Step S4)

[0065] As shown in Figure 4, the system executes a parameter evolution task once every Sunday at 2:00 AM.

[0066] 1. Weighting and centrifugation algorithm for the intermediate standby layer (Fmid)

[0067] The system does not simply select the midpoint of the physical floor, but instead calculates the "center of gravity of passenger flow." The formula is as follows:

[0068]

[0069] in:

[0070] H represents the total number of floors.

[0071] L i is the physical layer number of the i-th layer.

[0072] W i This is the "heat weight" for this layer. The weight calculation fully considers directionality: W i =1.2×Cup(i)+1.0×Cdown(i)W i=1.2×Cup(i)+1.0×Cdown(i). For example, a building has a cafeteria on the upper floors (floors 18-20), with extremely high passenger flow between floors during lunchtime. The algorithm calculates that the center of gravity of passenger flow may shift from the 10th floor to the 16th floor. Starting next week, elevator B will automatically stop at the 16th floor when idle, thus getting closer to the high-frequency usage area.

[0073] 2. Density detection algorithm for peak time windows

[0074] The system uses a time-sliding window (e.g., window width Δt = 15 min) to scan historical data. It calculates the passenger flow density ρ(t) at any given time t.

[0075] Entry Decision: A hysteresis comparison strategy is employed. Only when ρ(t) exceeds the high threshold θ for M = 2 consecutive windows... high (For example, when it reaches 30 times / minute), the peak is considered to have begun.

[0076] Exit criterion: Only when ρ(t) is below the low threshold θ for N=3 consecutive windows. low The peak is considered to have ended when the frequency reaches 10 times per minute (e.g., 10 times per minute).

Claims

1. A method for scheduling dual-elevator group control based on scene adaptation and machine learning, characterized in that, Includes the following steps: Step S1): Real-time monitoring of the operating status data of two elevators in the elevator group and the building's call command flow data; Step S2): When the system determines that it has entered idle mode, it executes the "asymmetric distributed standby strategy"; it controls the first elevator to automatically return to the bottom base station to wait, and controls the second elevator to automatically run to the dynamically calculated "intermediate hotspot floor" to wait, thereby minimizing the average response distance of the entire floor; if it determines that the idle mode conditions are not met, that is, the system is in "busy mode", the system responds to the call command according to the standard minimum waiting time group control strategy, without executing forced return or distributed standby actions; Step S3): It determines whether the current time is within the peak time window. If it is within the peak period, it identifies the preset building scene type and executes a differentiated directional elevator parking strategy accordingly. If it is a "residential building", it executes the upward peak strategy of prioritizing the bottom floor; if it is an "office building", it executes the downward peak strategy of prioritizing the upper floors; if it is detected that the current time is not within the peak time window, that is, in "off-peak or low-peak mode", the system jumps to step S2 according to the real-time call status. Perform idle standby determination or regular scheduling; Step S4): Based on historical operation data, use a machine learning model to periodically update the position parameters of the dynamic intermediate layer and the time parameters of the peak time window to obtain the specific floor number of the "intermediate hotspot layer" and the start and end time of the "peak time window". If the analysis results show that the passenger flow characteristics have not changed significantly or the data sample size is insufficient, the current parameters are kept unchanged and no update operation is performed.

2. The elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning according to claim 1, characterized in that, The judgment logic of "idle mode entry condition" in step S2) is as follows: when the following conditions are met simultaneously and the preset threshold Tidle is continuously set, the idle mode is determined to be entered: (1) The load in both elevator cars is zero; (2) The door operators of both elevators are closed; (3) There are currently no unresponsive external call commands or internal selection commands in the system.

3. The elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning according to claim 1, characterized in that, The "asymmetric distributed standby strategy" in step S2) is as follows: control the first elevator E1 to automatically run to the ground floor base station 1F of the building and keep it in standby mode; control the second elevator E2 to automatically run to the calculated dynamic intermediate floor Fmid and keep it in standby mode; wherein, if the difference between the current floor of the second elevator and the target Fmid is less than the preset fine-tuning threshold Nth, it will remain in place and not perform the re-leveling action.

4. The elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning according to claim 3, characterized in that, The calculation formula for the dynamic intermediate layer Fmid is a weighted centroid model: Where H is the total number of floors in the building, Li is the physical floor number of the i-th floor, and Wi is the call weight of the i-th floor; the call weight Wi is: Where Cup(i) and Cdown(i) are the number of uplink calls and downlink calls in layer ii within the statistical period, respectively, and α and β are preset coefficients that cause the uplink and downlink weights to differ.

5. The elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning according to claim 1, characterized in that, The "differentiated directional elevator parking strategy" in step S3) specifically includes: Step S31): Reading the building type identifier in the system configuration file; Step S32): If the identifier is "residential building", during the evening peak time window, when any elevator completes its service task and becomes idle, a scheduling instruction to return to the bottom base station is forcibly generated, unless the elevator has been assigned to respond to other calls; Step S33): If the identifier is "office building", during the evening peak time window, when any elevator completes its service task and becomes idle, a scheduling instruction to run upwards to the preset high-rise interval Ftop−K to FtopFtop is forcibly generated.

6. The elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning according to claim 1, characterized in that, In step S4), the update of the "peak time window" adopts a time series density detection algorithm: Define a time sliding window Δt, calculate the call density ρ(t) within the window; When it is detected that ρ(t) is higher than the high threshold θhigh for M consecutive periods, mark the current time as the peak start time; When the detected ρ(t) is below the low threshold θlow for N consecutive periods, the current time is marked as the end time of the peak. The update operation smooths the newly identified time point through the MovingAverage filter, and the time adjustment range of a single update is limited to the maximum step size ΔTmax.

7. An elevator control system used in the elevator dual-elevator group control and scheduling method based on scene adaptation and machine learning as described in any one of claims 1-6, characterized in that, include: The data acquisition module is configured to collect Hall sensor signals from all floors, car weighing signals, and button trigger signals; the central processing unit (CPU) is configured with scene recognition logic and a mode switching state machine to execute the scheduling method described in any one of claims 1 to 6; the storage module is used to store building type configuration files and historical operation logs for at least the most recent 30 days; and the cloud or local computing power module is used to run machine learning algorithms, periodically read data from the storage module and calculate update parameters, and send the results back to the central processing unit.

8. The elevator control system according to claim 7, characterized in that, The central processing unit also includes a hysteresis comparator, which introduces a time delay when switching between idle and peak modes to prevent frequent mode oscillations in critical states.

Images