Elevator group control method and system based on deep learning

Abstract

The invention relates to the technical field of elevator group control, in particular to an elevator group control method and system based on deep learning, and the method comprises the steps: building an elevator taking behavior time sequence data set through collecting historical elevator taking behavior data and combining with an elevator calling request and the number of people waiting for an elevator obtained through image recognition; the elevator taking frequency, the elevator calling interval and the passenger number fluctuation are modeled by adopting a long-short-term memory network, passenger flow periodical characteristics in workdays, holidays and peak periods are extracted, the passenger flow of each floor in the future is predicted, and pre-scheduling deployment of idle elevators is completed in advance. In the scheduling stage, state input containing the elevator running state and predicted flow is constructed, a reinforcement learning model is introduced for strategy optimization, a reward function is constructed based on passenger waiting time, response timeout and the service completion number, a scheduling strategy is driven to continuously converge, elevator response distribution is optimized, and position presetting and path adjustment are preset. And the scheduling intelligence and the service efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of elevator group control technology, and more particularly to an elevator group control method and system based on deep learning. Background Technology

[0002] As a core technology for ensuring the efficiency of vertical transportation within buildings, the level of intelligence in elevator group control systems directly impacts passenger waiting experience and energy consumption. During peak hours or specific concentrated travel periods, traditional elevator group control strategies often face problems such as untimely service response, long passenger waiting times, and significant waste from empty elevator operation. Especially in scenarios involving the coordinated scheduling of multiple elevators, existing algorithms generally lack a global optimal perspective and predictive capabilities, resulting in an inability to simultaneously achieve both scheduling efficiency and passenger satisfaction.

[0003] Currently, most common group control strategies are rule-based, such as "nearest elevator priority," "zonal scheduling," or "preset priority strategy." These methods have the advantages of simple implementation and fast response speed. However, when facing highly dynamic passenger flow, their static rules are often difficult to adapt to the changing operating conditions, exhibiting the following shortcomings: Existing strategies make scheduling decisions based on the current elevator call request and cannot predict the elevator demand of each floor in the future, resulting in insufficient scheduling preparation in the early stage of peak hours and scheduling delays when there is a sudden surge in passenger flow; Elevator demand is significantly affected by cyclical factors such as weekdays, time periods, and holidays. Traditional methods fail to use historical behavioral data to model the temporal patterns of passenger flow, and the scheduling process fails to introduce a feedback optimization mechanism, lacking the ability to dynamically weigh and adjust service quality indicators (passenger waiting time, response timeout). Summary of the Invention

[0004] To address the above problems, this invention provides an elevator group control method and system based on deep learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based elevator group control method includes: S1. Collect historical elevator behavior data in the target building and perform time series modeling to obtain a set of elevator behavior sequences; S2. Based on the set of elevator riding behavior sequences, input them into a long short-term memory network to predict the number of passengers on each floor within a future preset time window, and obtain the passenger flow prediction results for each floor. S3. Construct an elevator group scheduling state vector based on the current elevator location, operating status, load status, and passenger flow prediction results. S4. Using the elevator group scheduling state vector as the state input of the reinforcement learning model, the elevator scheduling strategy is forward reasoned through the reward value, and the output includes the scheduling actions of target elevator response allocation, idle elevator pre-adjustment position and in-transit elevator target adjustment; the reward value is calculated through the reward function, which is constructed by the number of passengers who complete the boarding and alighting services, the total waiting time and the service response timeout penalty term. S5. Generate corresponding control commands based on the scheduling actions to drive the elevator group to perform scheduling.

[0006] Furthermore, the historical elevator behavior data includes floor call request data and the number of people waiting on the floor when the call request occurred.

[0007] Furthermore, the number of people waiting on each floor is obtained through the following steps: Obtain floor call request data; Based on the elevator call request data, the location of cameras in the waiting area associated with each floor is determined; The waiting area camera acquires images of the waiting area; Based on image recognition algorithms, the number of people waiting on each floor is calculated from images of the waiting elevators.

[0008] Further, S1 includes the following steps: The elevator call request data for each floor is time-aligned and serialized with the number of people waiting for elevators on each floor during the corresponding time period to construct the original elevator data sequence. Based on the original elevator riding behavior sequence, elevator riding frequency, call interval and number of people fluctuation are extracted to generate a set of elevator riding behavior sequences.

[0009] Furthermore, the Long Short-Term Memory network is trained through the following steps: Based on the elevator behavior sequence set, the elevator frequency, call interval and number of people fluctuation features in each time slice are extracted, and the corresponding time context information, including weekday identifier, hour segment identifier and whether it is a holiday, are extracted simultaneously to construct a joint input feature vector for training. Based on the joint input feature vector and the actual passenger volume of each floor in adjacent time slices, supervised learning sample pairs are constructed to form a training dataset for time series prediction. Based on the training dataset, normalization and time-label encoding are performed on the joint input feature vector to obtain a standardized sample sequence. The standardized sample sequence is input into the long short-term memory network structure, and the temporal modeling and feature state propagation are performed in sequence to output the predicted passenger flow of each floor within the future preset time window; Based on the difference between the predicted output and the actual passenger volume label, the loss value is calculated using the mean squared error loss function, and the network weights are optimized iteratively using the backpropagation algorithm until the error on the validation set converges to the set threshold.

[0010] Furthermore, the mean squared error loss function is as follows: ; in, This is the loss value; The number of time slices for training samples; Total number of floors; For the first The time slice for the first Predicted passenger volume for each floor; For the first The first time in the film The actual number of passengers on each floor.

[0011] Furthermore, the reward function is as follows: ; in, As a reward value; The number of passengers who complete pick-up and drop-off services within the scheduling cycle; The total waiting time for all waiting passengers; The total duration of the response timeout; , and These are the weighting coefficients.

[0012] Furthermore, the reinforcement learning model is trained through the following steps: Based on historical scheduling process records, the elevator group scheduling state vector at each moment is paired with the corresponding scheduling action execution result. Combined with the actual generated reward value, a triple sequence of state, action and reward is constructed as a training sample set. Based on the training sample set, the elevator group scheduling state vector is input into the policy network in the reinforcement learning model to output the probability distribution of scheduling actions. At the same time, the elevator group scheduling state vector is input into the value network in the reinforcement learning model to calculate the predicted reward value. The deviation between the predicted return value and the corresponding reward value is calculated to obtain the strategy advantage value. Based on the strategy advantage value, gradient backpropagation of the policy network is performed to update the action output distribution. Based on the deviation between the predicted return value and the corresponding reward value, the loss calculation and backpropagation optimization of the value network are performed to complete the training of the reinforcement learning model.

[0013] Further, S5 includes the following steps: Based on the scheduling actions, the target elevator response allocation, the pre-adjusted position of the idle elevator and the target adjustment of the elevator in transit are analyzed respectively, and the corresponding elevator number and the scheduling target floor are extracted to generate a set of real-time control instructions. The set of real-time control commands is sent to the elevator control terminal to drive each elevator to complete the scheduling task.

[0014] A deep learning-based elevator group control system, applied to any of the aforementioned deep learning-based elevator group control methods, includes: The data acquisition and storage module is used to collect historical elevator behavior data in the target building for time series modeling to obtain an elevator behavior sequence set. The passenger flow prediction module is used to input the elevator behavior sequence set into the long short-term memory network to predict the passenger volume of each floor within a future preset time window, and obtain the passenger flow prediction results for each floor. The state construction module is used to construct the elevator group scheduling state vector based on the current elevator location, operating status, load status and the passenger flow prediction results. The reinforcement learning module uses the elevator group scheduling state vector as the state input of the reinforcement learning model. It performs forward reasoning on the elevator scheduling strategy through a reward function and outputs scheduling actions including target elevator response allocation, pre-adjustment of idle elevator positions, and target adjustment of in-transit elevators. The reward function is constructed by the total waiting time and the service response timeout penalty term. The elevator group scheduling module is used to generate corresponding control commands based on the scheduling actions, and drive the elevator group to perform scheduling.

[0015] The beneficial effects of this invention are as follows: By collecting historical elevator behavior data in the target building and combining it with the image recognition results of floor call requests and waiting passengers, a time-series dataset reflecting the patterns of elevator behavior is constructed. A Long Short-Term Memory (LSTM) network is used to perform time-series modeling of elevator frequency, call intervals, and passenger fluctuations, capturing the periodic changes in passenger flow during weekdays, holidays, and commuting hours. This enables accurate prediction of passenger arrivals on each floor within a preset time window, allowing for proactive pre-deployment of idle elevators before peak hours. A reinforcement learning model is introduced in the scheduling decision-making stage. The current operating state of the elevator group and the predicted flow are used to construct the state input. A reward function is designed based on the total passenger waiting time, service response timeout duration, and the number of passengers boarding and alighting, driving the policy network to continuously optimize the scheduling strategy output. Through the interactive training process of reinforcement learning, autonomous optimization of key scheduling actions such as elevator response allocation, pre-deployment of idle elevators, and target adjustment of in-transit elevators is achieved. It solves the technical pain points of existing technologies, such as the inability to predict passenger demand, rigid scheduling rules, and lack of strategy feedback optimization. When facing complex and high-density passenger flow, it has scheduling agility, predictive foresight, and strategy convergence capabilities, which can significantly reduce average waiting time, reduce response timeout frequency, and improve the overall efficiency and service experience of multiple elevators operating in coordination. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of an elevator group control method based on deep learning in this invention.

[0017] Figure 2 This is a flowchart of the training steps of the Long Short-Term Memory Network in this invention. Detailed Implementation

[0018] Please see Figures 1-2 As shown, this invention relates to a deep learning-based elevator group control method, comprising: S1. Collect historical elevator behavior data in the target building and perform time series modeling to obtain a set of elevator behavior sequences; S2. Based on the set of elevator riding behavior sequences, input them into a long short-term memory network to predict the number of passengers on each floor within a future preset time window, and obtain the passenger flow prediction results for each floor. S3. Construct an elevator group scheduling state vector based on the current elevator location, operating status, load status, and passenger flow prediction results. S4. Using the elevator group scheduling state vector as the state input of the reinforcement learning model, the elevator scheduling strategy is forward reasoned through the reward value, and the output includes the scheduling actions of target elevator response allocation, idle elevator pre-adjustment position and in-transit elevator target adjustment; the reward value is calculated through the reward function, which is constructed by the number of passengers who complete the boarding and alighting services, the total waiting time and the service response timeout penalty term. S5. Generate corresponding control commands based on the scheduling actions to drive the elevator group to perform scheduling.

[0019] In some embodiments, firstly, the system continuously collects historical elevator behavior data for the target building over the past 90 days through a data acquisition module deployed in the elevator control center. The data includes structured information such as floor call request timestamps, passenger floors, responding elevator numbers, car load changes, operating paths, and operating directions. Furthermore, by interfacing with the video surveillance system, it automatically identifies images of waiting areas and extracts the number of people waiting at the time of a call. In the data preprocessing stage, all data is aggregated using a sliding window at a 1-minute time granularity. A multi-dimensional feature sequence is constructed by combining time context labels (such as weekday / holiday, hourly segment codes), and then organized into multi-channel time series inputs according to each floor. Subsequently, the aforementioned elevator behavior sequence set is input into a multi-layered stacked Long Short-Term Memory (LSTM) network model. Unlike existing methods that predict passenger flow using static averages or rule mappings, the LSTM model in this embodiment can automatically capture the upward and downward correlations and periodic fluctuation characteristics between floor levels. For example, during training, the model automatically identified a significant time shift between the elevator call volume on the B2 level parking garage and the call density on the 1st floor lobby during the morning peak hours from Monday to Friday. Therefore, when predicting passenger distribution within the next 15 minutes, it prioritizes outputting a prediction structure that emphasizes lower-level clustering and an upward trend. Actual comparisons show that the prediction accuracy is approximately 28% lower than the baseline statistical method during peak hours. In the scheduling decision-making phase, the system constructs a scheduling state vector by combining the real-time operating status of each elevator, including its location floor, direction of travel, load status, and target floor. Using this vector as input, the reinforcement learning model employs an Actor-Critic structure for policy updates during training. The policy network outputs the distribution of scheduling actions, while the value network evaluates the state rewards. The model uses an improved reward function during training: a positive incentive term represents the number of passengers boarding and alighting per unit time, while a negative penalty term represents the cumulative passenger waiting time and the number of services exceeding the preset response time limit. This weighted combination of three terms achieves a balanced optimization between passenger experience and scheduling efficiency. Unlike traditional rule-driven "nearest elevator priority" strategies, this solution can pre-allocate idle elevators to floors with high future call volumes based on predictive information and dynamically adjust the operation strategies of elevators in transit to achieve optimal allocation of building scheduling resources. Differentiating itself from existing technologies, this invention uses LSTM for future traffic prediction and inputs the prediction results as part of the scheduling state into a reinforcement learning model, achieving future-oriented scheduling optimization. It also designs a composite reward mechanism that integrates service quantity and latency penalties, enabling the scheduling strategy to have self-evaluation capabilities and truly achieve continuous optimization guided by service quality.

[0020] Furthermore, the historical elevator behavior data includes floor call request data and the number of people waiting on the floor when the call request occurred.

[0021] Furthermore, the number of people waiting on each floor is obtained through the following steps: Obtain floor call request data; Based on the elevator call request data, the location of cameras in the waiting area associated with each floor is determined; The waiting area camera acquires images of the waiting area; Based on image recognition algorithms, the number of people waiting on each floor is calculated from images of the waiting elevators.

[0022] Specifically, when the elevator system receives a call request from a specific floor, it first locates the camera number or IP address pre-configured in the waiting area of ​​that floor based on the floor's identification number, thus establishing a one-to-one mapping between the floor and the image acquisition terminal. Then, it automatically calls the image acquisition module corresponding to that camera to capture the current waiting image of that floor in real time. To ensure the temporal consistency between the extracted image and the call behavior, the image acquisition operation must be completed within a preset time window (e.g., within 1 second) after receiving the call request, and precisely aligned with the timestamp of the call request. The acquired image is then input into an image recognition model deployed on an edge server or in the cloud (e.g., a detection network based on YOLOv5, EfficientDet, or CenterNet) to extract identifiable human target bounding boxes from the image. Post-processing techniques such as non-maximum suppression and scale filtering are used to remove false detections and overlapping results, ultimately obtaining a statistical value of the actual number of passengers waiting in the current waiting area.

[0023] Further, S1 includes the following steps: The elevator call request data for each floor is time-aligned and serialized with the number of people waiting for elevators on each floor during the corresponding time period to construct the original elevator data sequence. Based on the original elevator riding behavior sequence, elevator riding frequency, call interval and number of people fluctuation are extracted to generate a set of elevator riding behavior sequences.

[0024] In some embodiments, firstly, elevator call request data for each floor within the target building and statistical data on the number of people waiting for elevators on each floor obtained through image recognition are collected over a long period. Considering the differences in sampling frequency and recording granularity of different data sources, the data needs to be aligned using a uniform time step (e.g., every minute or every five minutes). Specifically, a main timeline is constructed based on the elevator call request timestamps, and the number of people waiting for elevators on each floor obtained from the most recent image recognition is associated within the corresponding time period, forming an original elevator usage data sequence based on the unit of "time step-floor-number of requests-number of people waiting for elevators". After completing the time-series alignment, feature extraction is performed on the original elevator usage data sequence: for each time window, the number of elevator call requests is counted to measure the elevator usage frequency, the time interval between two consecutive requests is calculated as the elevator call interval feature, and simultaneously, moving average and variance analysis are performed on the sequence of people waiting for elevators to extract fluctuation amplitude indicators reflecting the characteristics of sudden changes in passenger flow. The features of these multiple dimensions collectively reflect the user's elevator usage demand patterns and behavioral rhythms. Independent feature sequences are constructed for different floors and different time periods (such as weekdays and holidays, peak commuting hours and off-peak hours) to serve as training inputs for subsequent long short-term memory network models, thereby capturing the periodic patterns and sudden anomalies in elevator riding behavior.

[0025] Furthermore, the Long Short-Term Memory network is trained through the following steps: Based on the elevator behavior sequence set, the elevator frequency, call interval and number of people fluctuation features in each time slice are extracted, and the corresponding time context information, including weekday identifier, hour segment identifier and whether it is a holiday, are extracted simultaneously to construct a joint input feature vector for training. Based on the joint input feature vector and the actual passenger volume of each floor in adjacent time slices, supervised learning sample pairs are constructed to form a training dataset for time series prediction. Based on the training dataset, normalization and time-label encoding are performed on the joint input feature vector to obtain a standardized sample sequence. The standardized sample sequence is input into the long short-term memory network structure, and the temporal modeling and feature state propagation are performed in sequence to output the predicted passenger flow of each floor within the future preset time window; Based on the difference between the predicted output and the actual passenger volume label, the loss value is calculated using the mean squared error loss function, and the network weights are optimized iteratively using the backpropagation algorithm until the error on the validation set converges to the set threshold.

[0026] In some embodiments, firstly, based on the constructed set of elevator riding behavior sequences, multi-dimensional features are extracted for each time slice, including: elevator riding frequency per unit time (i.e., number of elevator call requests), the mean and standard deviation of the time interval between consecutive elevator calls, and the fluctuation index of the number of people waiting for elevators on each floor during the corresponding time period. Simultaneously, corresponding contextual label information is extracted from the time dimension, such as weekday / weekend identifiers, hourly segment identifiers (e.g., 08:00–09:00), and whether it is a statutory holiday, etc., and added to the feature vector through one-hot encoding to enhance the model's ability to identify periodic fluctuations and sudden surges in passenger flow. Secondly, based on the above joint input features, supervised sample pairs are constructed together with the actual passenger volume on each floor of the next time slice. Each sample pair is in the form of:<Xt, Yt+1> Here, Xt is the joint feature vector at time t, and Yt+1 is the passenger volume label at time t+1. To improve the model's generalization ability, the training set also includes sample instances from multiple weekdays and holidays, different time periods, and different floors. In the input preprocessing stage, normalization operations (such as Z-score standardization or Min-Max scaling) are performed on all joint input features to eliminate the influence of dimensions. At the same time, time labels (such as hourly segments and weekday indices) are encoded into continuous-time embedding vectors and concatenated with behavioral features to obtain a standardized sample sequence with consistent structure. Next, a multi-layer stacked LSTM network structure is constructed, using the standardized sample sequence as input, to transmit state information and capture short-term and long-term dependent features time-slice by time. The model output is the predicted passenger flow value for each floor within a future preset time window (such as the number of passengers every 5 minutes in the next 15 minutes). Finally, the model calculates the deviation between the predicted output and the label value based on the mean squared error (MSE) loss function, and optimizes the weight parameters of each LSTM layer through the backpropagation algorithm. After each training round, an independent validation set is used to evaluate the error. Training is terminated when the error is below the preset convergence threshold or the maximum number of training rounds is reached.

[0027] Furthermore, the mean squared error loss function is as follows: ; in, This is the loss value; The number of time slices for training samples; Total number of floors; For the first The time slice for the first Predicted passenger volume for each floor; For the first The first time in the film The actual number of passengers on each floor.

[0028] It should be noted that the loss function is based on the difference between the predicted and actual values. First, the difference between the predicted passenger volume and the corresponding actual passenger volume for each floor within each time slice is calculated. Then, this difference is squared to amplify the impact of the prediction error on model training. Subsequently, the squared error values ​​for each floor across all time slices are summed and normalized according to the number of time slices and the total number of floors to obtain the final average loss value. This loss value comprehensively reflects the degree of prediction deviation of the model throughout the entire time period and the floor spatial dimension. During training, the backpropagation algorithm is used to update weights with this loss value as the target, continuously optimizing the generalization ability and fitting effect of the prediction model until the set error convergence threshold is reached on the validation dataset. This mechanism ensures high accuracy of passenger flow prediction results across different floors and time periods, providing a reliable input basis for the subsequent construction of the scheduling state of the reinforcement learning model, effectively distinguishing it from existing strategies that do not systematically optimize the prediction error.

[0029] Furthermore, the reward function is as follows: ; in, As a reward value; The number of passengers who complete pick-up and drop-off services within the scheduling cycle; The total waiting time for all waiting passengers; The total duration of the response timeout; , and These are the weighting coefficients.

[0030] It should be noted that the reward function is based on the actual scheduling performance within the scheduling cycle, mainly considering three core indicators: the number of passengers who completed boarding and alighting services, the cumulative waiting time of all waiting passengers, and the total timeout duration of scheduling responses. The number of passengers who completed boarding and alighting services reflects the service coverage capacity of the elevator scheduling within that cycle; a higher number indicates a more efficient scheduling strategy, thus contributing positively to the reward value. The cumulative waiting time measures the total waiting time from when a passenger calls the elevator until they actually board. A longer waiting time indicates that the current scheduling strategy has failed to effectively respond to passenger needs, thus negatively impacting the reward value. The total timeout duration of scheduling responses represents the total duration during which elevator calls fail to receive service within the set time limit; this quantifies extreme scheduling failures and also negatively impacts the reward value. By weighting and combining these three indicators, the weight allocation can be flexibly adjusted according to actual needs, forming a reward feedback mechanism with multi-objective optimization capabilities.

[0031] Furthermore, the reinforcement learning model is trained through the following steps: Based on historical scheduling process records, the elevator group scheduling state vector at each moment is paired with the corresponding scheduling action execution result. Combined with the actual generated reward value, a triple sequence of state, action and reward is constructed as a training sample set. Based on the training sample set, the elevator group scheduling state vector is input into the policy network in the reinforcement learning model to output the probability distribution of scheduling actions. At the same time, the elevator group scheduling state vector is input into the value network in the reinforcement learning model to calculate the predicted reward value. The deviation between the predicted return value and the corresponding reward value is calculated to obtain the strategy advantage value. Based on the strategy advantage value, gradient backpropagation of the policy network is performed to update the action output distribution. Based on the deviation between the predicted return value and the corresponding reward value, the loss calculation and backpropagation optimization of the value network are performed to complete the training of the reinforcement learning model.

[0032] In some embodiments, the training process is based on constructing a state-action-reward triple sequence and guides the policy convergence direction potential through policy advantage. First, based on historical scheduling data of the target building, the elevator group scheduling state vector is extracted for each time step, including information such as the current floor position, direction of travel, load status, and predicted floor passenger flow of all elevators. Simultaneously, the actual scheduling actions taken in this state (such as elevator allocation decisions, pre-set position settings, and target adjustments for elevators en route) and the reward value generated within the corresponding period after the action are recorded. The aforementioned state, action, and reward triples form a training sample set, constituting the input basis of the model. During the training phase, the scheduling state vector is first input into the policy network. The policy network outputs the probability distribution of each scheduling action in the action space through a multi-layer feedforward structure, representing the probabilities of different scheduling schemes in the current state. Simultaneously, this state vector is input into the value network, which outputs a scalar reward prediction value to evaluate the overall long-term value of the current state. Subsequently, the difference between the predicted reward value and the actual reward value is calculated to form the policy advantage value, reflecting the expected reward gain of the current action in this state. The advantage value serves as a gradient signal, guiding the policy network to perform gradient backpropagation updates, thereby gradually shifting the network towards outputting the probability distribution of actions that yield higher actual rewards. To avoid high variance in policy updates, methods such as Generalized Advantage Estimation (GAE) can be used to smooth the advantage value. Based on the error between the predicted and actual rewards, the value network calculates the mean squared error loss function and performs backpropagation optimization accordingly, making its output state value estimate closer to the true reward and enhancing the stability of the policy update process. Through this dual-network collaborative optimization strategy, the reinforcement learning model can efficiently learn dynamic scheduling strategies that meet the global goals of elevator scheduling (such as reducing waiting time and improving service coverage), overcoming the technical bottlenecks of traditional rule-based scheduling strategies' poor adaptability to complex elevator behavior and lack of adaptive optimization capabilities.

[0033] Further, S5 includes the following steps: Based on the scheduling actions, the target elevator response allocation, the pre-adjusted position of the idle elevator and the target adjustment of the elevator in transit are analyzed respectively, and the corresponding elevator number and the scheduling target floor are extracted to generate a set of real-time control instructions. The set of real-time control commands is sent to the elevator control terminal to drive each elevator to complete the scheduling task.

[0034] Specifically, after the reinforcement learning model outputs scheduling actions, the action vector is first structured and analyzed to identify three core scheduling decisions: First, target elevator response allocation, which selects the optimal target elevator to respond to calls from each floor; second, pre-positioning of idle elevators, which pre-assigns idle elevators to potentially high-demand floors based on predicted traffic distribution; and third, target adjustment of elevators in transit, which dynamically modifies the target floors of elevators already in operation based on the latest predictions and scheduling results to improve overall service response efficiency. After extracting the above scheduling strategies, the elevator numbers, target floors, and service types involved are further extracted and combined to form a set of real-time control instructions containing control objectives, instruction types, and execution parameters. Each instruction clearly defines the elevator ID, the action to be performed (responding to a call, pre-positioning, or changing the target), and its target floor, ensuring the resolvability and effective execution of the control instructions.

[0035] The present invention also includes a deep learning-based elevator group control system, applied to any of the foregoing deep learning-based elevator group control methods, comprising: The data acquisition and storage module is used to collect historical elevator behavior data in the target building for time series modeling to obtain an elevator behavior sequence set. The passenger flow prediction module is used to input the elevator behavior sequence set into the long short-term memory network to predict the passenger volume of each floor within a future preset time window, and obtain the passenger flow prediction results for each floor. The state construction module is used to construct the elevator group scheduling state vector based on the current elevator location, operating status, load status and the passenger flow prediction results. The reinforcement learning module uses the elevator group scheduling state vector as the state input of the reinforcement learning model. It performs forward reasoning on the elevator scheduling strategy through a reward function and outputs scheduling actions including target elevator response allocation, pre-adjustment of idle elevator positions, and target adjustment of in-transit elevators. The reward function is constructed by the total waiting time and the service response timeout penalty term. The elevator group scheduling module is used to generate corresponding control commands based on the scheduling actions, and drive the elevator group to perform scheduling.

[0036] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning-based elevator group control method, characterized by, The method comprises the following steps: S1, collecting historical elevator behavior data in a target building for time series modeling to obtain an elevator behavior sequence set; S2, inputting the elevator behavior sequence set into a long short-term memory network to predict the passenger volume of each floor in a future preset time window and obtain a passenger flow prediction result of each floor; S3, constructing an elevator group scheduling state vector according to the elevator position, running state and load state at the current time and the passenger flow prediction result; S4, inputting the elevator group scheduling state vector as the state input of a reinforcement learning model, performing forward reasoning on the elevator scheduling strategy through a reward value, and outputting scheduling actions including target elevator response allocation, idle elevator pre-adjustment position and in-transit elevator target adjustment; the reward value is calculated through a reward function, and the reward function is constructed by the number of passengers completing boarding and alighting services, the total value of waiting time and a service response timeout penalty term; S5, generating corresponding control instructions according to the scheduling actions to drive the elevator group to execute scheduling. 2.The elevator group control method based on deep learning according to claim 1, characterized in that, The historical elevator behavior data includes floor call request data and the number of people waiting for elevators at the floor when the call request occurs. 3.The elevator group control method based on deep learning according to claim 2, characterized in that, The number of people waiting for elevators at the floor is obtained by the following steps: Obtain floor call request data; Locate and position the camera for the waiting area associated with the floor according to the floor call request data; Obtain the waiting area image in the waiting area through the waiting area camera; Based on the image recognition algorithm, calculate the number of people waiting for elevators at the floor based on the waiting area image.

4. The deep learning-based elevator group control method according to claim 3, characterized by, The S1 comprises the following steps: Align the floor call request data with the number of people waiting for elevators at the floor in the corresponding time period in time and sequence to construct an original elevator data sequence; Based on the original elevator behavior sequence, extract the elevator frequency, call interval and number of people fluctuation to generate and construct the elevator behavior sequence set.

5. The deep learning-based elevator group control method according to claim 4, characterized by, The long short-term memory network is trained by the following steps: Based on the elevator behavior sequence set, extract the elevator frequency, call interval and number of people fluctuation features in each time slice, and synchronously extract the corresponding time context information, including the week identifier, hour segment identifier and whether it is a holiday, to construct a joint input feature vector for training; According to the joint input feature vector and the actual passenger volume of each floor in the adjacent time slice, construct a supervised learning sample pair to form a training data set for time series prediction; Based on the training data set, perform normalization processing and time label encoding on the joint input feature vector to obtain a standardized sample sequence; Input the standardized sample sequence into the long short-term memory network structure, sequentially perform time series modeling and feature state propagation, and output the passenger flow prediction value of each floor in the future preset time window; According to the difference between the prediction output and the actual passenger volume label, calculate the loss value through the mean square error loss function, and optimize and iterate the network weight through the back propagation algorithm until the error on the validation set converges to a set threshold.

6. The deep learning-based elevator group control method according to claim 5, characterized by, The mean square error loss function is as follows: ； wherein, is a loss value; is a number of time slices of the training sample; is a total number of floors; is a predicted passenger volume for the th floor in the th time slice; is an actual passenger volume of the th floor in the th time slice. 7.The elevator group control method based on deep learning according to claim 1, characterized in that, The reward function is as follows: ； wherein, is a reward value; is the number of passengers who complete the boarding and alighting service within the dispatch cycle; is the cumulative waiting time of all waiting passengers; is the total duration of response timeout; , and are weight coefficients. 8.The elevator group control method based on deep learning according to claim 7, characterized in that, The reinforcement learning model is trained by the following steps: Based on historical scheduling process records, the elevator group scheduling state vector at each time is paired with the corresponding scheduling action execution result, and combined with the actual generated reward value, a triple sequence of state, action and reward is constructed as a training sample set; According to the training sample set, the elevator group scheduling state vector is input into the policy network in the reinforcement learning model, and the probability distribution of the scheduling action is output, and at the same time, the elevator group scheduling state vector is input into the value network in the reinforcement learning model, and the return prediction value is calculated; The deviation between the return prediction value and the corresponding reward value is calculated to obtain the policy advantage value, and the gradient back propagation of the policy network is performed based on the policy advantage value to update the action output distribution; Based on the deviation between the return prediction value and the corresponding reward value, the loss calculation and back propagation optimization of the value network are performed to complete the reinforcement learning model training. 9.The elevator group control method based on deep learning according to claim 1, characterized in that, The S5 includes the following steps: According to the scheduling action, the target elevator response allocation, the idle elevator pre-adjustment position and the in-transit elevator target adjustment are analyzed respectively, the corresponding elevator number and scheduling target floor are extracted, and a real-time control instruction set is generated; The real-time control instruction set is sent to the elevator control terminal to drive each elevator to complete the scheduling task.

10. A deep learning-based elevator group control system applied to the deep learning-based elevator group control method of any one of claims 1-9, characterized in that, It includes: A data acquisition and storage module is used to collect historical elevator behavior data in the target building for time series modeling to obtain an elevator behavior sequence set; A flow prediction module is used to input the elevator behavior sequence set into a long short-term memory network to predict the passenger volume of each floor in a future preset time window and obtain passenger flow prediction results for each floor; A state construction module is used to construct an elevator group scheduling state vector based on the elevator position, running state, load state at the current time and the passenger flow prediction results; A reinforcement learning module is used to input the elevator group scheduling state vector as the state input of the reinforcement learning model, perform forward reasoning on the elevator scheduling strategy through a reward function, and output scheduling actions including target elevator response allocation, idle elevator pre-adjustment position and in-transit elevator target adjustment; The reward function is constructed by the total value of waiting time and service response timeout penalty term; An elevator group scheduling module is used to generate corresponding control instructions according to the scheduling action to drive the elevator group to execute scheduling.

Images