Elevator group control energy-saving scheduling method and system based on load prediction

Abstract

This invention relates to the field of elevator group control technology, specifically to an elevator group control energy-saving scheduling method and system based on load prediction. Step A: During elevator door opening, the drive current signal of the door operator motor and the infrared occlusion signal of the door zone light curtain are collected in real time. Step B: Current time-domain features are extracted from the drive current signal, and occlusion timing event features are extracted from the infrared occlusion signal. Step C: The current time-domain features and occlusion timing event features are input into a pre-trained time-series neural network model, outputting the final load prediction value for this door opening process. Step D: Before the door closing action is completed, the final load prediction value is sent to the group control scheduler. Step E: The group control scheduler dynamically adjusts the stopping allocation strategy of this elevator and other related elevators based on the final load prediction value. Compared to existing technologies that can only passively adjust subsequent elevators after door closing, this invention achieves feedforward active scheduling, solving the problem of scheduling information lagging behind load changes.

Description

Technical Field

[0001] This invention relates to the field of elevator group control technology, and specifically to an elevator group control energy-saving scheduling method and system based on load prediction. Background Technology

[0002] Elevator group control systems are the core transportation hubs of vertical transportation in modern high-rise buildings, and their scheduling efficiency directly affects passenger experience and building energy consumption. A typical elevator scheduling process includes: demand perception (capturing passenger requests through floor buttons and floor selection signals inside the car), dynamic allocation (assigning requests to the optimal elevator based on the elevator's current position, direction of travel, load status, and predicted traffic flow using scheduling algorithms), path planning (adjusting the stopping sequence and merging requests in the same direction), and real-time adjustment (recalculating priorities in case of sudden demand). Load information is one of the key input parameters for group control decisions.

[0003] Currently, most elevator group control systems only obtain the actual load through weighing sensors at the bottom of the car after the doors close, and then use this information to make decisions on dispatching subsequent elevators. This "load detection after door closing" mechanism has an inherent information lag problem: the time from elevator door opening to closing is usually 5 to 10 seconds, while the group control algorithm needs to make stop assignments at least 5 seconds in advance. When a large number of passengers rush in during the opening period (such as the lobby floor during the morning rush hour or the cafeteria floor during lunchtime), the system cannot know the final load of the elevator before the doors close. It can only passively adjust subsequent elevators based on historical statistics or the load after door closing, resulting in invalid dispatch (elevators that have already been dispatched do not actually need to stop at this floor) and overlapping stops (multiple elevators responding to the same floor at the same time), generating additional empty run energy consumption and start-stop energy consumption.

[0004] Numerous studies have attempted to introduce intelligent algorithms to optimize elevator group control scheduling. For example, Chinese invention patent CN120964536A discloses a dynamic scheduling method for elevator group control based on reinforcement learning. This method collects elevator request timestamps, source floors, and target floors to construct spatiotemporal characteristics of passenger behavior (such as a passenger flow direction matrix between floors at different times). It then uses reinforcement learning models (DDQN, PPO, Actor-Critic, etc.) to generate dynamic scheduling strategies, switching between pre-deployment and real-time calculation strategies based on the periodic characteristics of passenger flow to reduce average waiting time and energy consumption. This scheme achieves autonomous evolution of the scheduling strategy through a multi-objective reward function (waiting time penalty, energy consumption penalty, load balancing reward, and emergency demand reward). However, the input data for this method only includes the timestamp and floor information of the elevator call request, without considering real-time sensing signals during the elevator door opening process. Its "prediction" is essentially a macro-trend prediction based on historical passenger flow statistics (such as passenger flow on each floor in the next 15 minutes), rather than a real-time prediction of the final load of the elevator in the current opening cycle. Therefore, this solution still cannot solve the information lag problem caused by "knowing the load only after the doors close" - in the few seconds when passengers rush in, the system still cannot know in advance whether the elevator will be full, and thus cannot proactively adjust the stop allocation before the doors close.

[0005] Another Chinese invention patent, CN121201931A, proposes an elevator group control method and system based on deep learning. It uses a long short-term memory network (LSTM) to perform time-series modeling on historical elevator behavior data (including elevator call requests and the number of people waiting for elevators obtained through image recognition), predicts the passenger flow of each floor within a preset time window in the future, and then uses the prediction results as one of the state inputs of the reinforcement learning model to realize the pre-adjustment and deployment of idle elevators. Compared to pure rule-based scheduling, this scheme has stronger predictive capabilities, but it still has the following shortcomings: First, its prediction input relies on image recognition to obtain the number of people waiting for the elevator. Image recognition is easily affected by factors such as lighting, occlusion, and camera angle, and it is prone to missed detection or double counting when multiple people enter and exit in parallel. Second, this scheme does not utilize the door operator motor current signal—the door operator current can directly reflect the resistance of passengers pushing the door and is a unique physical quantity for judging the degree of congestion. However, in the existing technology, the door operator current is mainly used for door operator fault detection (such as door obstruction, motor overload), and has not yet been used for load prediction. Third, the prediction output of this scheme is the macroscopic floor passenger flow (e.g., how many passengers will arrive at a certain floor in the next 15 minutes), rather than the final load of this elevator in the current door opening cycle. Therefore, it cannot solve the technical problem of "real-time prediction of the load of this elevator during the door opening process". Fourth, the loss function used by this scheme is the symmetrical mean square error (MSE), which does not consider the asymmetrical impact of the load prediction error direction on energy consumption costs (the energy loss caused by underestimating the load and resulting in invalid elevator dispatch is much higher than the passenger waiting loss caused by overestimating the load).

[0006] Furthermore, existing technologies suffer from the following common shortcomings: using a single sensor makes it difficult to distinguish between different congestion patterns—weighing sensors cannot reflect the dynamic influx during door opening, light curtains (infrared beams) are prone to pulse loss when multiple people enter and exit in parallel, and door operator currents cannot distinguish between "multiple people crowding and pushing the door" and "single person forcefully pushing the door." Mechanical aging (such as increased friction coefficient of the door operator and elastic drift of the wire rope) also causes the accuracy of prediction models with fixed parameters to decline rapidly over time. Existing elevator group control systems lack a fusion sensing method that can accurately predict the final load of the elevator in real time during door opening, lack an optimization mechanism for the asymmetry of load prediction error direction, and lack online learning capabilities that can adapt to mechanical aging and dynamic operating conditions.

[0007] To address this, this invention proposes a load feedforward prediction and group control energy-saving scheduling method and system based on the fusion of door operator current and light curtain occlusion timing. Unlike existing technologies, this invention simultaneously acquires the drive current signal of the door operator motor and the infrared occlusion signal of the door zone light curtain during elevator door opening. It extracts time-domain features such as peak value, integral, and slope from the current signal, and extracts timing event features such as occlusion frequency, longest continuous occlusion time, and number of occlusion-non-occlusion switching times from the light curtain signal. These two types of features are input into a lightweight LSTM model, which outputs the final load prediction value for this elevator trip during the door opening process (e.g., at the 3rd second). This prediction value is sent to the group control scheduler before the door closing action is completed, allowing for dynamic adjustment of the stopping allocation strategy for this elevator and other related elevators (e.g., canceling other elevators' pre-stopping on this floor, allowing this elevator to skip floors, and waking up backup elevators in advance). Meanwhile, this invention employs an asymmetric loss function to train the prediction model, imposing a higher penalty weight on under-prediction (predicted load is lower than actual load), and dynamically adjusting this weight based on the elevator's direction of travel (up / down) and floor congestion levels. Through online incremental updates and recursive identification of door operator mechanical parameters, the model adapts to the mechanical aging of older elevators. This invention can advance the acquisition time of load information from "after the door closes" to "3 seconds after the door opens," obtaining a feedforward time window of at least 2-7 seconds, effectively reducing invalid dispatches and overlapping stops, achieving significant energy-saving effects. Summary of the Invention

[0008] Existing elevator group control and scheduling systems only obtain the actual load of the car through weighing sensors after the door closing action, resulting in the following technical defects: scheduling information lags behind load changes (a 5-10 second information window is wasted), a single sensor cannot distinguish congestion patterns (weighing lacks dynamics, light curtain pulses are easily lost, and current makes it difficult to identify congestion types), prediction error direction leads to asymmetrical energy consumption costs (under-prediction costs are much higher than over-prediction costs), and mechanical aging and dynamic operating conditions affect prediction reliability. Therefore, this invention provides a load feedforward prediction and group control energy-saving scheduling method and system based on the fusion of door operator current and light curtain obstruction timing.

[0009] To achieve the above objectives, the present invention provides the following technical solution: An elevator group control energy-saving scheduling method based on load prediction includes: Step A, during elevator door opening, real-time acquisition of the drive current signal of the door operator motor and the infrared occlusion signal of the door zone light curtain; Step B, extraction of current time-domain features from the drive current signal and extraction of occlusion timing event features from the infrared occlusion signal; Step C, inputting the current time-domain features and occlusion timing event features into a pre-trained time-series neural network model, and outputting the final load prediction value of this door opening process; Step D, before the door closing action is completed, sending the final load prediction value to the group control scheduler; Step E, the group control scheduler dynamically adjusts the stopping allocation strategy of this elevator and other related elevators according to the final load prediction value.

[0010] Through the above scheme, this invention advances the acquisition time of load information from "after the door closes" to "during the door opening process" (e.g., at the 3rd second), obtaining a feedforward time window of at least 2 to 7 seconds. Actual measurements show that this scheme can reduce invalid elevator dispatch by approximately 65%, reduce empty-run energy consumption by 25% to 30%, reduce overlapping stops by approximately 40%, and correspondingly reduce start-stop energy consumption by 18%. Compared to existing technologies that can only passively adjust subsequent elevators after the door closes, this invention achieves feedforward active scheduling, fundamentally solving the problem of scheduling information lagging behind load changes.

[0011] The current time-domain characteristics include the current peak value, current integral value, and current rise slope within the first N seconds after the door is opened; the occlusion timing event characteristics include the occlusion frequency, the longest continuous occlusion time, and the number of occlusion-non-occlusion switching, where N is 3 to 5 seconds.

[0012] The aforementioned feature extraction has clear physical meaning: peak current reflects the maximum congestion thrust, current integral reflects the accumulated congestion energy, and current slope reflects the influx velocity; occlusion frequency and switching frequency reflect the activity level of people entering and exiting, and the longest continuous occlusion time reflects wheelchair or multiple people entering and exiting in parallel. These features provide physically interpretable inputs for subsequent temporal neural networks, enabling the model to learn the essential differences between different congestion patterns. Experiments show that using these 6-dimensional features can reduce the prediction error (RMSE) by 59.5%, which is a significant advantage compared to using the original signal or single sensor features.

[0013] The temporal neural network model is a Long Short-Term Memory (LSTM) network. Its loss function during the training phase is an asymmetric loss function. This asymmetric loss function applies a higher penalty weight to under-predicted samples whose predicted values ​​are lower than the true values ​​than to over-predicted samples whose predicted values ​​are higher than the true values.

[0014] This solution addresses the characteristic in elevator dispatching where "underestimating the load leads to a much higher energy cost for invalid elevator dispatches than overestimating the load," employing an asymmetric loss function. Compared to traditional symmetric loss methods (such as Mean Squared Error, MSE), this solution actively biases the model towards higher predictions, thereby reducing the incidence of costly under-predicted events. An unexpected effect: experiments show that using the asymmetric loss function reduces the variance of model predictions by approximately 35% (improved prediction stability), which in turn reduces the frequency of command changes by the group control dispatcher and saves about 15% on communication bus load and controller computational overhead. This indirect energy saving benefit was completely unforeseen during the design phase.

[0015] The penalty weight of the asymmetric loss function is dynamically adjusted according to the current direction of elevator operation and the degree of congestion on each floor: the degree of congestion on each floor is determined based on the number of elevator call requests in the past statistical window or the number of people waiting for the elevator obtained through image recognition; during peak hours or on congested floors (defined as the number of elevator calls exceeding a threshold or the number of people waiting for the elevator exceeding 50% of the rated passenger capacity in the past 5 minutes), the penalty weight for under-prediction is increased to 2 to 5 times that for over-prediction; during off-peak hours, the penalty weight for under-prediction is reduced to 1 to 1.5 times that for over-prediction.

[0016] The dynamic adjustment mechanism enables the model to adapt to differences in energy consumption costs under different scenarios. During peak upward travel times, underprediction can lead to a large number of passengers being stranded in the lobby and causing multiple elevators to be dispatched ineffectively, resulting in extremely high energy costs. Therefore, a high penalty weight (3:1) is used. During off-peak downward travel times, the impact of underprediction is smaller, and a low penalty weight (1.2:1) is used. This adaptive mechanism further improves energy-saving performance. Real-world testing shows that dynamic adjustment can save an additional 8% to 12% of energy compared to fixed weights.

[0017] The stop allocation strategy in step E includes: if the final load prediction value exceeds the first threshold, a "this elevator will be full" signal is sent to the group control scheduler, and the pre-stop instructions of other elevators to this floor are cancelled; if the final load prediction value is lower than the second threshold, the elevator is allowed to cross at least two pre-stop floors consecutively; if the final load prediction value is in the middle range and the deviation from the historical pattern exceeds the set range, the standby elevator is woken up in advance to stand by on the adjacent floor.

[0018] The three feedforward scheduling actions described above correspond to heavy load, light load, and abnormal fluctuation scenarios, respectively, forming a complete closed loop of the elevator stop allocation strategy. Specifically, "canceling other elevator pre-stops" directly eliminates the energy consumption source of ineffective elevator dispatch; "no stopping across floors" reduces unnecessary start-stop energy consumption; and "awakening backup elevators" improves the responsiveness to sudden passenger flow. These three actions work synergistically, enabling the system to maintain efficient and energy-saving operation under different load conditions. Actual measurements show that this strategy reduces elevator empty run mileage by 29% and the number of start-stops by 19%.

[0019] Steps A to C are repeated multiple times within the same door opening cycle: the intermediate load prediction values ​​are output at the 1st, 2nd, and 3rd seconds respectively. When the difference between the subsequent prediction value and the previous prediction value exceeds the correction threshold, the group control scheduler performs incremental correction on the issued docking allocation instructions.

[0020] Rolling prediction and instruction correction ensure the system can handle sudden changes in passenger flow late in the door opening period (e.g., only one person initially, then 10 people suddenly enter after 5 seconds). Synergistic effect: Rolling prediction is a common technique, and feedforward scheduling is a known concept, but by using the output change rate of rolling prediction as a trigger condition and combining it with the incremental correction mechanism of group control scheduling instructions, the system can achieve a closed loop of "prediction → decision → withdrawal → re-decision" within the same door opening cycle. This dynamic correction capability is unprecedented in the field of elevator group control and is only possible when the prediction latency is extremely low (<100ms). The lightweight LSTM of this invention (2.3ms per inference) precisely meets this latency requirement, a result of the collaboration between the algorithm and hardware.

[0021] The time-series neural network model automatically distinguishes between the following two congestion modes by jointly learning the current time-domain features and the occlusion time-series event features: high current integral and low occlusion switching frequency are determined to be a multi-person parallel congestion influx mode; low current integral and high occlusion switching frequency are determined to be a single person frequent entry and exit mode; and outputs the corresponding load prediction confidence level according to different modes.

[0022] This scheme achieves intelligent recognition of congestion patterns. Synergistic effect: A single sensor cannot achieve the above distinction. Current is sensitive to "force," while the light curtain is sensitive to "geometric timing," and their physical mechanisms are complementary. Through joint learning with LSTM, the model automatically emerges with the ability to classify the two patterns. This is not a simple feature concatenation, but rather the generation of a new classification ability—the ability to output confidence levels for different patterns, thereby guiding the crowd controller to take differentiated scheduling actions (e.g., immediately canceling pre-stopping for multi-person congestion patterns, while maintaining the original strategy for single-person frequent entry and exit patterns). This ability was not explicitly programmed in the original design, but rather a synergistic effect of the model's self-organizing emergence.

[0023] It also includes step F: extracting the internal state vector of the time-series neural network model, inputting the internal state vector into the gantry crane health assessment module, outputting the gantry crane guide rail lubrication or bearing wear index, and issuing a maintenance warning when the wear index exceeds the threshold.

[0024] The LSTM's internal state vector was originally used only to transmit timing information for load prediction. Experiments revealed that when the gantry crane's guide rails are poorly lubricated or the bearings are worn, the current waveform exhibits a specific high-frequency oscillation pattern. To accurately predict the load, the LSTM model automatically learns and encodes this oscillation pattern into its internal state vector. This state vector can then be used to output the gantry crane's health indicators at no additional cost, enabling predictive maintenance. In a real-world test, this module successfully provided an early warning of early bearing wear in a gantry crane (wear index increased from 0.32 to 0.78). After maintenance inspection and confirmation, the bearing was replaced, preventing a sudden failure. Those skilled in the art, when designing load prediction models, could not have foreseen that they would also receive free gantry crane health diagnostic functionality.

[0025] It also includes step G: during the actual operation of the elevator, the real load value obtained by the weighing sensor after the door is closed is used to perform online incremental updates on the time-series neural network model; at the same time, the actual friction coefficient and response delay of the door operator system are recursively identified, and the identification results are used as one of the input features of the model, so that the model can adapt to the mechanical parameter drift of the old elevator.

[0026] Online incremental updates enable the model to continuously adapt to changes in passenger flow patterns (such as new merchant onboarding or changes in floor functions); recursive identification of door operator parameters solves the problem of decreased prediction accuracy caused by mechanical aging. An unexpected cross-domain compensation effect: During testing on an old elevator (12 years old), it was found that the load cell itself also exhibited zero-point drift (up to ±8% of the rated load). The recursive identification algorithm of this invention, while fitting the current-load relationship, inadvertently identified the drift characteristics of the load cell. Using the identification results, the sensor readings were calibrated in reverse, restoring the weighing accuracy to ±2%, without requiring hardware replacement. This cross-domain compensation (mechanical aging identification simultaneously correcting electrical sensor drift) completely exceeded the original design expectations.

[0027] An elevator group control energy-saving scheduling system based on load prediction includes: a current acquisition module for real-time acquisition of the drive current signal of the door operator motor during elevator door opening; a light curtain signal acquisition module for real-time acquisition of the infrared blocking signal of the door zone light curtain during elevator door opening; an edge computing unit with a built-in temporal neural network model for outputting the final load prediction value of the current door opening process based on the drive current signal and the infrared blocking signal; and a group control scheduler for receiving the final load prediction value and dynamically adjusting the stopping allocation strategy of the current elevator and other related elevators based on the prediction value before the door closing action is completed; wherein, the edge computing unit and the group control scheduler are connected through a real-time communication bus, and the temporal neural network model is trained using an asymmetric loss function.

[0028] This system employs an edge computing architecture, deploying a lightweight LSTM on an STM32H7 microcontroller (400MHz). A single inference operation requires only 2.3ms, significantly less than the sampling interval during door opening (100ms), supporting embedded real-time deployment. The system can be flexibly added to new elevators or used to retrofit older elevators without replacing existing door operators or light curtain hardware. Real-world testing shows that in commercial complex environments, this system achieves an average daily energy saving of 22.3% and a 32.5% reduction in average waiting time. It also provides additional functions such as door operator health warnings and weighing sensor self-calibration, demonstrating significant economic benefits and technological advancement.

[0029] Compared with the prior art, the present invention has the following beneficial effects: This invention integrates the timing of door operator current and light curtain obstruction to advance elevator load prediction from "after door closing" to "during door opening" (outputting the predicted value at the 3rd second), obtaining a feedforward time window of 2-7 seconds, effectively solving the problem of delayed scheduling information. Actual measurements show that it can reduce invalid elevator dispatches by approximately 65%, reduce empty-run energy consumption by 25%-30%, reduce overlapping stops by approximately 40%, correspondingly reduce start-stop energy consumption by 18%, and decrease the total system energy consumption by over 22%.

[0030] This invention produces several synergistic and unexpected technical effects: the physical complementarity of current and light curtain enables the model to automatically distinguish between "multiple people crowding in" and "single people frequently entering and exiting" modes, guiding differentiated scheduling; the combination of rolling prediction and low-latency inference enables incremental correction of instructions within the same door opening cycle, responding to sudden changes in passenger flow; the LSTM's internal state vector can output door operator health indicators without additional cost, achieving predictive maintenance; the asymmetric loss function unexpectedly improves prediction stability and saves controller energy consumption while reducing under-prediction energy consumption; the online identification mechanism automatically compensates for weighing sensor drift, restoring accuracy without replacing hardware.

[0031] Compared with existing technologies (CN120964536A, CN121201931A), this invention has outstanding and substantial features and significant progress in terms of prediction real-time performance, sensor fusion depth, asymmetric optimization and adaptive capabilities, and also has multiple benefits such as energy saving, cost reduction and predictive maintenance. Attached Figure Description

[0032] Figure 1 This is an overall flowchart of the elevator group control energy-saving scheduling method based on load prediction in an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the acquisition and feature extraction of the door operator current and light curtain signal in an embodiment of the present invention.

[0034] Figure 3 This is a schematic diagram of the LSTM model in an embodiment of the present invention.

[0035] Figure 4 This is a logic diagram of the group control scheduler adjusting dock allocation based on predicted load in an embodiment of the present invention.

[0036] Figure 5 This is a timing diagram of rolling prediction and instruction correction in an embodiment of the present invention.

[0037] Figure 6 This is an architecture diagram of an elevator group control energy-saving scheduling system based on load prediction in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0039] Example 1: An Energy-Saving Elevator Group Control Scheduling Method Based on Load Prediction This embodiment provides an elevator group control energy-saving scheduling method based on load prediction, such as... Figure 1 As shown, it includes the following steps.

[0040] Step S1: Signal Acquisition. Timing begins when the elevator reaches a floor and initiates the door opening action. During the elevator door opening period, the instantaneous value of the three-phase drive current output by the door operator servo driver is acquired in real time at a sampling frequency of 100Hz. Simultaneously, the occlusion status sequence of the infrared light curtain in the door area is read at a rate of one frame every 10ms. The acquisition duration is the first 5 seconds after door opening to cover most of the passenger entry and exit time. In this embodiment, the door operator current signal reflects the resistance generated by passengers pushing the door, and the light curtain occlusion sequence reflects the timing pattern of people entering and exiting the elevator car.

[0041] Step S2: Feature extraction. For example... Figure 2 As shown, three time-domain features are extracted from the acquired current signal: current peak value. (Maximum value within the first 3 seconds, in A), current integral value (The integral of current over time in the first 3 seconds reflects the accumulated congestion energy), current rise slope (Maximum rate of change of current within the first 3 seconds). Three temporal event features were extracted from the light curtain occlusion sequence: occlusion frequency. (Number of occlusion events per unit time, in Hz), longest continuous occlusion time (unit: seconds), and the number of times the occlusion-to-unocclusion switch occurs. (Number of switching times in the first 3 seconds). The above 6-dimensional features constitute the feature vector of the current door opening cycle.

[0042] Step S3: Load Prediction. Organize the 6-dimensional feature vector obtained in Step S2 by time steps (one time step every 0.1 seconds, for a total of 30 steps), and input it into the pre-trained LSTM model. For example... Figure 3 As shown, the LSTM model uses a single-layer structure with 32 hidden units, followed by a Dropout layer (dropout rate of 0.2) and a fully connected layer (outputting 1D). The model outputs the final predicted load value for this door-opening process. (Unit: kg). In this embodiment, the LSTM model is trained using an asymmetric loss function, defined as follows: in This represents the actual load value. For predicted values, The underprediction penalty coefficient, The over-prediction penalty coefficient, and In this embodiment, The value is dynamically adjusted according to the elevator's direction of travel and the degree of congestion on each floor: 3:1 during peak upward travel periods or on congested floors, and 1.2:1 during off-peak downward travel periods.

[0043] Step S4: Feedforward Transmission. Three seconds after the door opens (at which point the door is usually still open, with approximately 2-7 seconds remaining before closing), the edge computing unit transmits the predicted value. The data is sent to the group control scheduler via the CAN bus.

[0044] Step S5: Scheduling decision. For example... Figure 4 As shown, the group control scheduler according to With rated load ratio Implement the following strategy: like If the threshold is reached (first threshold), it is determined to be a heavy-load scenario. The dispatcher sends a "This elevator is about to be full" signal to the group control system and cancels the pre-stop instructions of other elevators to this floor, thereby preventing the dispatched elevators from running to this floor empty.

[0045] like If the second threshold is reached, it is determined to be a light-load scenario. The scheduler allows the elevator to skip at least two pre-stopped floors in a row (i.e., the floors originally set to stop are skipped) to reduce unnecessary starts and stops.

[0046] like Then, the deviation between the current forecast value and the historical average load for the same period is further determined. If the deviation exceeds 30%, a backup elevator is activated in advance to standby on an adjacent floor; otherwise, the original stop allocation strategy is maintained.

[0047] The above decision is executed before the door closing action is completed, thus achieving feedforward control.

[0048] Step S6 (optional): Rolling prediction and instruction correction. For example... Figure 5 As shown, within the same door opening cycle, this embodiment performs three predictions: the first second outputs... Output at 2 seconds Output at 3 seconds Let the correction threshold be... .like Then withdraw based on The issued instructions, and based on Recalculate the scheduling instructions; if If so, it will be revised again. This mechanism can cope with sudden changes in customer flow in the later stages of opening (for example, only 1 person at the beginning of opening, and then 10 people suddenly flooding in later).

[0049] Step S7 (Optional): Door Operator Health Diagnosis. Extract the internal state vector (32-dimensional) of the LSTM model during the door opening process, input it into a lightweight fully connected network (door operator health assessment module), and output the door operator guide rail lubrication score (0-100) and bearing wear index (0-1). When the wear index exceeds 0.7, a maintenance warning is issued through the elevator IoT platform. In this embodiment, this function does not require additional sensors and can be implemented using existing LSTM state vectors.

[0050] Step S8 (optional): Online adaptive update. After each door closure, obtain the actual load value from the load cell. The current and light curtain characteristics collected during the door opening process are compared with... New training samples are generated. Every 100 samples accumulated, incremental training (using mini-batch gradient descent with a batch size of 32) is triggered, updating the LSTM model weights. Simultaneously, the recursive least squares (RLS) algorithm is used to identify the actual friction coefficient μ and response delay τ of the gantry crane system online. The specific identification model is as follows: Establish the dynamic equations for the gantry crane motor: in, This is the motor drive current. The door leaf angle (measured by the door operator encoder). Let be the moment of inertia (a known constant). The viscous damping coefficient is... Coulomb friction torque (proportional to the coefficient of friction) ), For disturbances. Since the viscous damping changes relatively little, the main identification parameter is the equivalent friction coefficient. (K is the torque coefficient).

[0051] To simplify online identification, the following first-order discrete model is adopted: Among them, output quantity (i.e., the component of the current used to overcome friction and damping), input vector Parameters to be identified Obtained through RLS recursion and Then calculate the coefficient of friction. The response delay τ is obtained by measuring the time difference between the actual current reaching 90% of the command value after the current command jump, and is filtered using an exponentially weighted moving average.

[0052] The identified μ and τ are added as additional input features to the LSTM model, enabling the model to adapt to the drift of mechanical parameters in old elevators.

[0053] Example 2: Elevator Group Control Energy-Saving Scheduling System Based on Load Prediction like Figure 6 As shown, this embodiment provides an elevator group control energy-saving scheduling system based on load prediction, including: Current acquisition module: Connected to the current detection terminal of the gantry servo driver, it acquires the three-phase drive current signal at a sampling rate of 100Hz, and outputs a digital signal after filtering and analog-to-digital conversion.

[0054] Light curtain signal acquisition module: connected to the output of the gate area light curtain controller, reads infrared occlusion status data every 10ms frame, and generates occlusion timing sequence.

[0055] Edge computing unit: Employs an STM32H7 series microcontroller (400MHz clock speed, 2MB Flash, 1MB RAM), with a built-in pre-trained LSTM model (quantized model size approximately 200KB). The edge computing unit receives current and light curtain data, performs feature extraction and LSTM inference, and outputs the final load prediction value 3 seconds after the door opens. The edge computing unit connects to the group control scheduler via a CAN bus (500kbps baud rate).

[0056] Group control scheduler: Receives the predicted load value, executes the stop allocation strategy in Example 1, generates control commands (such as canceling pre-stop, skipping stops across floors, and waking up the standby elevator), and sends them to each elevator master controller via fieldbus.

[0057] Weighing sensor interface: Receives the actual load value after the door is closed, for use in online incremental training.

[0058] Health assessment module: Integrated in the edge computing unit, it receives the internal state vector of the LSTM and outputs the gate health score.

[0059] This system can be deployed in newly built elevators or upgraded in existing elevators by adding edge computing units, without the need to replace the door operator or light curtain hardware.

[0060] Example 3: Experimental Data and Comparative Examples To verify the technical effectiveness of this invention, a 30-day comparative test was conducted in a commercial complex (4 elevators, serving floors B2-25, rated load 1000kg, daily passenger flow approximately 8000 people). During the test, the elevator group control system operated in three different configurations: Comparative Example 1 (Traditional Solution): Weighing feedback scheduling is adopted after the door is closed, that is, the actual load is only obtained after the door is closed for subsequent elevator dispatching decisions, without using any prediction.

[0061] Comparative Example 2 (Single Sensor Solution): Only light curtain counting is used (without current fusion), and LSTM is used to predict the final load, but the loss function is symmetric MSE, and there is no rolling prediction or adaptive update.

[0062] The present invention employs a fusion of gantry current and light curtain, an asymmetric loss function (α / β=3:1), rolling prediction (seconds 1, 2, and 3), online incremental updates, and gantry parameter identification.

[0063] The test results are shown in Table 1.

[0064] As shown in Table 1, compared with Comparative Example 1, the present invention reduces the number of invalid elevator dispatches by 65.5%, the empty running mileage by 29.0%, the number of start-stop cycles by 19.0%, and the total energy consumption by 22.3%; compared with Comparative Example 2, the prediction error is reduced by 29.4% and the energy consumption is reduced by 10.6%.

[0065] Ablation experiment: To verify the contribution of each technical feature, an ablation experiment was conducted under the same test environment, and the results are shown in Table 2.

[0066] Ablation experiments show that the fusion of current and light curtain features contributes the most (RMSE increases by about 40% and energy consumption increases by about 11% after removal); asymmetric loss, rolling prediction, and online adaptation all contribute significantly to performance.

[0067] Unexpected results verified: Door operator health check: During testing, the health assessment module detected that the wear index of the bearing in elevator door operator No. 3 gradually increased from 0.32 to 0.78 (on day 25), issuing a warning. Maintenance personnel inspected and confirmed the bearing wear, replaced it, and the system returned to normal. This warning was issued approximately 45 days earlier than the usual maintenance cycle.

[0068] Weighing sensor drift compensation: In the renovation test of an old elevator (12 years of use), the zero-point drift of the weighing sensor reached ±8%. After adopting the recursive identification algorithm of this invention, the calibrated weighing error was reduced to ±2.1%, and no hardware was replaced.

[0069] Working principle The working principle of this invention is based on the following logical loop: 1. Physical Sensing Layer: The door operator current signal and the light curtain obstruction signal describe the passenger influx process during door opening from two dimensions: mechanical and geometric timing, respectively. The current integral is positively correlated with the total impulse generated by passengers pushing the door, and the light curtain switching frequency is positively correlated with the activity level of the number of people entering and exiting. The two complement each other and can distinguish different congestion patterns.

[0070] 2. Temporal Modeling Layer: The LSTM network, through its gating mechanism, can capture the temporal dependencies between current and light curtain signals. The input layer organizes features by time step, the hidden state conveys information from past moments, and the output layer regresses the final load. The asymmetric loss function causes the model to tend to overestimate the load during training to match the asymmetry of actual energy consumption costs.

[0071] 3. Feedforward Control Layer: The group control scheduler obtains the load forecast value before the doors close and adjusts the stop allocation in advance accordingly. During heavy loads, other elevators' pre-stopping is canceled; during light loads, skipping floors is allowed; and in case of anomalies, the standby elevator is activated. This is equivalent to transforming traditional feedback control (adjusting subsequent elevators after doors close) into feedforward control (adjusting the current elevator and other elevators during door closing), gaining a lead time of 2-7 seconds.

[0072] 4. Adaptive Evolutionary Layer: Online incremental updates enable the model to continuously adapt to changes in passenger flow distribution; recursive identification of gantry parameters allows the model to compensate for mechanical aging; LSTM state vector transfer is used for health diagnosis; and the identification results can also be used to calibrate the weighing sensor. These mechanisms work together to maintain the long-term performance of the system.

[0073] Industrial application verification In addition to the above-mentioned commercial complex tests, this invention has also been verified in the following three typical scenarios: Scenario A: A super high-rise office building (35 floors, 6 elevators). Data from 6 months of operation shows that the average waiting time for elevators during the morning peak hours (8:00-10:00) has decreased from 34.2 seconds to 22.5 seconds, and the number of invalid elevator dispatches during peak hours has decreased by 72%, with the system saving an average of about 820 kWh of electricity per month.

[0074] Scenario B: A tertiary hospital (outpatient building, 5 elevators). To address emergency response needs, this system reduced the average response time for emergency elevator calls from 18.6 seconds to 9.3 seconds. The door operator health diagnosis module successfully issued warnings for two door operator malfunctions, preventing elevator outages.

[0075] Scenario C: An old residential community (18 stories, 2 elevators, 15 years old). Before the renovation, the weighing sensors suffered from severe drift, frequently causing false overload alarms. After adopting this system, online identification and automatic calibration of weighing errors reduced the false overload alarm rate by 86%. Simultaneously, predictive scheduling reduced ineffective starts and stops, decreasing the average monthly number of elevator mechanical components (brakes, door operators) by approximately 24%, and is expected to extend component replacement cycles by approximately 30%.

[0076] The above verification results show that the present invention has repeatable energy-saving effects and improved reliability in both new elevator construction and old elevator renovation scenarios.

[0077] Other implementation methods Those skilled in the art will understand that the parameters in the above embodiments (such as sampling frequency, time window, number of LSTM hidden units, loss function weights, thresholds, etc.) can be adjusted according to specific application scenarios. For example, in high-rise buildings, the prediction window can be extended to 5 seconds; in hospital scenarios, the asymmetric weights corresponding to emergency needs can be further increased. These adjustments do not depart from the scope of protection of this invention.

[0078] Furthermore, the temporal neural network model in this invention is not limited to LSTM; it can also employ architectures such as Gated Recurrent Units (GRUs) or Transformers, as long as it can process temporal fusion features and output regression values. The asymmetric loss function can also be replaced with weighted quantile loss or other forms of asymmetric cost function. The number of rolling prediction steps and the correction threshold can be dynamically set according to the door opening time.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An elevator group control energy-saving scheduling method based on load prediction, characterized in that, Includes the following steps: Step A: During the elevator door opening process, collect the drive current signal of the door operator motor and the infrared blocking signal of the door zone light curtain in real time; Step B: Extract the current time-domain features from the driving current signal and extract the occlusion timing event features from the infrared occlusion signal; Step C: Input the current time-domain features and occlusion timing event features into the pre-trained time-series neural network model, and output the final load prediction value of this door opening process; wherein, the time-series neural network model is a Long Short-Term Memory (LSTM) network, and its loss function during the training phase is an asymmetric loss function. The asymmetric loss function applies a higher penalty weight to under-predicted samples whose predicted values ​​are lower than the true values ​​than to over-predicted samples whose predicted values ​​are higher than the true values. The penalty weight of the asymmetric loss function is dynamically adjusted according to the current direction of elevator operation and the degree of congestion on the floor: during peak hours of upward movement or on congested floors, the penalty weight for under-prediction is increased to 2 to 5 times that for over-prediction; during off-peak hours of downward movement, the penalty weight for under-prediction is reduced to 1 to 1.5 times that for over-prediction. Step D: Before the door closing action is completed, send the final load prediction value to the group control scheduler; Step E: The group control scheduler dynamically adjusts the stop allocation strategy for this elevator and other related elevators based on the final load prediction value.

2. The method according to claim 1, characterized in that, The current time-domain characteristics include the current peak value, current integral value, and current rise slope within the first N seconds after the door is opened; the occlusion timing event characteristics include the occlusion frequency, the longest continuous occlusion time, and the number of occlusion-non-occlusion switching, where N is 3 to 5 seconds.

3. The method according to claim 1, characterized in that, The docking allocation strategy in step E includes: If the final load prediction value exceeds the first threshold, a "This elevator will be full" signal is sent to the group control scheduler, and the pre-stop instructions of other elevators to this floor are cancelled. If the final load prediction value is lower than the second threshold, the elevator is allowed to pass through at least two pre-stopped floors in this trip. If the final load prediction value is in the middle range and the deviation from the historical pattern exceeds the set range, the backup elevator will be activated in advance to standby on the adjacent floor.

4. The method according to claim 1, characterized in that, Steps A to C are repeated multiple times within the same door opening cycle: intermediate load prediction values ​​are output at the 1st, 2nd, and 3rd seconds respectively, with the intermediate load prediction value output at the 3rd second serving as the final load prediction value; when the difference between the intermediate load prediction value output in the later time exceeds the correction threshold, the group control scheduler performs incremental correction on the issued docking allocation instructions.

5. The method according to claim 1, characterized in that, The temporal neural network model automatically distinguishes the following two congestion modes by jointly learning the current time-domain features and the occlusion time-series event features: High current integration and low obstruction switching frequency: determined to be a multi-person parallel congestion influx mode; Low current integration and high obstruction switching frequency: determined to be a single-person frequent entry and exit mode; It also outputs the corresponding load prediction confidence level according to different modes.

6. The method according to claim 1, characterized in that, It also includes step F: extracting the internal state vector of the time-series neural network model, inputting the internal state vector into the gantry crane health assessment module, outputting the gantry crane guide rail lubrication or bearing wear index, and issuing a maintenance warning when the wear index exceeds the threshold.

7. The method according to claim 1, characterized in that, It also includes step G: during the actual operation of the elevator, the real load value obtained by the weighing sensor after the door is closed is used to perform online incremental updates on the time-series neural network model; at the same time, the actual friction coefficient and response delay of the door operator system are recursively identified, and the identification results are used as one of the input features of the model, so that the model can adapt to the mechanical parameter drift of the old elevator.

8. An elevator group control energy-saving scheduling system based on load prediction, characterized in that, include: The current acquisition module is used to acquire the drive current signal of the door motor in real time during the opening of the elevator door; The light curtain signal acquisition module is used to acquire the infrared occlusion signal of the light curtain in the door area in real time during the elevator door opening period; The edge computing unit, with a built-in temporal neural network model, is used to output the final load prediction value of this door opening process based on the driving current signal and the infrared blocking signal. The group control scheduler receives the final load prediction value and dynamically adjusts the stop allocation strategy of this elevator and other related elevators based on the prediction value before the door closing action is completed. The edge computing unit is connected to the group control scheduler via a real-time communication bus. The temporal neural network model is trained using an asymmetric loss function, and the penalty weight of the asymmetric loss function is dynamically adjusted according to the current direction of elevator operation and the degree of floor congestion: during peak upward travel periods or on congested floors, the penalty weight for under-prediction is increased to 2 to 5 times that for over-prediction; during off-peak downward travel periods, the penalty weight for under-prediction is reduced to 1 to 1.5 times that for over-prediction.

Images