An elevator operation energy consumption optimal speed planning method and system

Abstract

The application discloses an elevator operation energy consumption optimal speed planning method and system, and relates to the technical field of intelligent device control.The application collects multi-dimensional operation parameters and passenger flow data through edge perception, combines with inherent constants of the device to construct a dynamic speed trajectory candidate engine, generates and screens compliant candidate trajectories, builds a forward-looking planning model, realizes passenger flow and working condition prediction by using an LSTM network and standardizes the output results, constructs a multi-objective cost function by using an MPC algorithm, and rolls to solve the comprehensive optimal speed trajectory.The elevator operates based on the optimal trajectory and real-time returns operation data, relies on a double duel deep Q network to build a dynamic energy consumption evaluation model, combines with smoothness scoring to construct a reward function, completes parameter iteration through experience replay and gradient back propagation, adaptively corrects the weight of the forward-looking planning model, and loads the optimized model parameters before the next operation.The application is suitable for energy-saving, comfortable and efficient collaborative control of building elevators under multiple working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent device control technology, specifically to an optimal speed planning method and system for elevator operation with optimal energy consumption. Background Technology

[0002] With the acceleration of urbanization and the widespread adoption of high-rise buildings, elevators, as core equipment in vertical transportation, are experiencing a continuous increase in their energy consumption as a proportion of total building energy consumption. Elevator energy conservation and intelligent speed planning have become key technological directions for building energy efficiency management. Currently, IEEE and domestic research institutions have verified the significant effectiveness of intelligent algorithms in elevator energy-saving control. Elevator control technology has gradually evolved from traditional fixed-curve regulation to a dynamic optimization stage that integrates model prediction and machine learning. It is widely used in various building scenarios such as residential buildings, commercial complexes, and industrial plants, effectively reducing elevator operating energy consumption and improving passenger comfort. In this context, industry professionals continue to explore and optimize control strategies, striving to further balance energy consumption, efficiency, and passenger experience, and address the problem of declining control performance under complex operating conditions.

[0003] Chinese patent (publication number CN110902514A) discloses an elevator operation control method. This method combines car load, transmission efficiency and balance coefficient to complete parameter calibration, adjust the elevator operation status and realize energy consumption control. However, this method is based on static parameters to complete the control, does not take into account the changes in operating conditions caused by passenger flow fluctuations, has insufficient control strategy flexibility, and does not perform overall optimization design of the operating speed curve.

[0004] Chinese patent (publication number CN119263009B) discloses an elevator energy consumption optimization control method. This method distinguishes the operating mode according to the stopping distance and adjusts the speed and acceleration parameters accordingly to balance energy consumption and transportation efficiency. However, this method relies on a fixed threshold to divide the operating range, and the trajectory adjustment method is singular. It cannot generate multiple alternative operating schemes according to real-time operating conditions, and the range of operating condition adaptability is limited.

[0005] Chinese patent (publication number CN111753468A) discloses an elevator system control method based on deep reinforcement learning. This method uses reinforcement learning algorithms to iteratively update the control strategy to achieve energy-saving operation. However, this method only uses a basic reinforcement learning framework to design the control logic, lacks the ability to predict the operating state in advance, has relatively simple evaluation dimensions, and the accuracy and stability of strategy iteration need to be improved.

[0006] In addition, existing technologies generally suffer from insufficient dimensions of operational status perception, rigid control strategies, and a lack of ability to predict operating conditions in advance, making it difficult to simultaneously achieve the comprehensive control objectives of low energy consumption, high efficiency, and high comfort in complex and ever-changing operating environments.

[0007] In summary, current elevator energy consumption and speed control technologies still have many shortcomings and are difficult to adapt to dynamically changing on-site conditions. The industry urgently needs a more comprehensive elevator speed planning and energy consumption optimization control solution. Summary of the Invention

[0008] Based on the above-mentioned technical problems, this application discloses an optimal speed planning method and system for elevator operation energy consumption, wherein the method is specifically as follows:

[0009] By collecting real-time elevator operating parameters at the millisecond level through the edge perception module and combining them with the building's time-series passenger flow data, multi-dimensional time-varying operating parameters are obtained.

[0010] Based on the collected multi-dimensional time-varying operating parameters, a dynamic speed trajectory candidate engine coupled with elevator electromechanical loss, mechanical resistance, and load loss is established. The engine generates multiple sets of candidate speed trajectories that meet the constraints of acceleration, jerk, comfort, and running efficiency according to real-time working condition boundary conditions.

[0011] A forward-looking planning model is constructed. The forward-looking planning model uses the LSTM time series prediction algorithm to predict the floor call trend and passenger flow changes within a preset time period. Combined with the predictive control MPC algorithm, it performs global optimization on multiple candidate speed trajectories to generate a forward-looking optimal speed trajectory.

[0012] The elevator operates based on the optimal speed trajectory, and real-time elevator operation parameters are collected when the elevator actually executes the optimal trajectory.

[0013] A dynamic energy consumption assessment model is constructed using a dual-duel deep Q-network reinforcement learning architecture. The dynamic energy consumption assessment model periodically performs sample playback and parameter iteration on the look-ahead planning model based on real-time elevator operating parameters, and adaptively corrects the energy consumption weights under different time-varying operating conditions to obtain an optimized look-ahead planning model.

[0014] Before the next elevator operation, the elevator speed is planned using an optimized look-ahead planning model.

[0015] Preferably, the dynamic speed trajectory candidate engine specifically involves: reading real-time operating condition boundary parameters and current multi-dimensional time-varying operating parameters, and dividing the trajectory into acceleration, uniform acceleration, uniform speed, deceleration, and acceleration / deceleration segments based on the S-shaped speed trajectory segmentation principle;

[0016] Within the boundary constraints, the acceleration and peak velocity are discretely sampled, and the parameters and durations corresponding to each segment that meet the adjustment are dynamically generated, and differentiated candidate velocity trajectories are generated in batches.

[0017] The total predicted energy consumption for each candidate trajectory is calculated one by one. The energy consumption index, comfort index and timeliness index are combined for screening. Illegal trajectories that exceed the threshold are eliminated, and the set of compliant candidate trajectories with lower energy consumption and meeting the comfort and timeliness constraints are retained.

[0018] Preferably, the calculation of the predicted total energy consumption for each group of candidate trajectories specifically involves: for each group of candidate velocity trajectories, extracting segmented velocity and acceleration sequences within the complete trajectory runtime sequence; combining real-time collected multi-dimensional time-varying operating parameters and inherent equipment constants; calculating elevator electromechanical losses, mechanical resistance, and load losses according to the trajectory segmented time sequence; and accumulating and coupling the three types of losses in real time to obtain the complete predicted total energy consumption for a single group of candidate trajectories, as shown in the formula:

[0019]

[0020] in, This represents the total energy consumption for a single elevator trip. The electrical losses of the motor are calculated using the following formula:

[0021]

[0022] in, This refers to the stator current of the motor. For the stator copper resistor of the motor, This represents the real-time iron loss power of the motor. In the formula, This refers to the real-time speed of the motor. For real-time motor temperature, The reference temperature is room temperature. These are the fitting coefficients for iron loss;

[0023] The formula for calculating the switching losses of the frequency converter is as follows:

[0024]

[0025] in, This is the switching loss coefficient (inherent constant) of the frequency converter. The real-time switching frequency is dynamically calculated based on the operating conditions.

[0026] The formula for calculating mechanical friction and wind resistance loss is as follows:

[0027]

[0028] in, For environmental wind pressure resistance, For the real-time operating speed of the elevator, This is the coefficient of friction of the track (an inherent constant of the equipment). The total mass of the car. This is the acceleration due to gravity.

[0029] Preferably, the forward-looking planning model is a two-layer intelligent energy consumption planning model that integrates LSTM time-series passenger flow prediction and MPC rolling time-domain optimization. The model is divided into an upper-layer time-series prediction and a lower-layer rolling optimization planning. The upper-layer time-series prediction is responsible for predicting the floor call trends and passenger load changes in the short term based on historical passenger flow time-series data, providing prior information on the operating conditions for speed planning. The lower-layer rolling optimization planning is responsible for performing rolling time-domain multi-objective optimization on a batch of candidate speed trajectories based on the predicted operating conditions and the current real-time operating conditions, and selecting the forward-looking speed trajectory that has the best overall performance in terms of energy consumption, timeliness, and comfort.

[0030] Preferably, the upper-level time series prediction specifically involves using a time series feature sequence composed of the number of elevator calls, elevator start and stop frequency, average load, and time period labels for each floor within a fixed historical period as input to the LSTM network.

[0031] Temporal feature memorization and iterative updates are achieved through an LSTM gating mechanism, based on the hidden layer temporal features output by the LSTM. The probability of floor calls, passenger flow load trends, and start / stop times within a preset time period are predicted by regression analysis of the fully connected layer.

[0032] The prediction results are standardized by using the Sigmoid activation function of the output layer to obtain standardized prior working condition constraint information, which is then sent down to the lower layer.

[0033] Preferably, the lower-level rolling optimization planning adopts the Model Predictive Control (MPC) algorithm. Based on the predicted operating conditions output by the upper-level LSTM, it performs rolling time-domain global optimization on the screened compliant candidate trajectories. The specific process of global optimization is as follows:

[0034] Set the time domain length T for rolling optimization, and discretize the future running process into N control steps;

[0035] The multi-objective minimization objective function for MPC is established as follows:

[0036]

[0037] in, For the comprehensive cost function, with Minimize as the objective. For the first Energy consumption per control step This is due to runtime deviation. For the first The comfort fluctuation value of each control step. These are the weighting coefficients for energy consumption, runtime deviation, and comfort fluctuation values, respectively.

[0038] In each rolling time domain, traverse all compliant candidate velocity trajectories and solve for the optimal trajectory sequence corresponding to the minimum objective function value;

[0039] Execute the optimal speed command for the current moment, update the real-time operating conditions and LSTM predicted operating conditions in the next moment, and re-roll and iterate to finally output a forward-looking optimal speed trajectory that meets the constraints of energy consumption, timeliness and comfort.

[0040] Preferably, the dynamic energy consumption assessment model is an adaptive iterative optimization assessment model based on a dual-duel deep Q-network;

[0041] The model is based on the full-dimensional coupling of elevator energy consumption, operating comfort, and operating efficiency. Through a reinforcement learning iterative mechanism, it evaluates the performance of different operating conditions and different speed trajectory strategies in real time, and realizes dynamic correction and adaptive optimization of the weight parameters of the forward planning model.

[0042] Preferably, the dual-duel depth Q-network specifically adopts a dual-head duel network structure with a shared feature extraction layer and independent output of two branches, decoupling the total action value Q-value into two independent parts: state value and action advantage;

[0043] Among them, the value branch outputs a status value based on the current global operating condition, which represents the basic quality of the current overall elevator operating environment, load, and equipment status.

[0044] The advantageous branch combines the operating conditions and different speed control actions to output the relative advantage benefits of each candidate action, representing the differences in energy consumption and comfort benefits of different acceleration, peak speed, and deceleration slope adjustment strategies.

[0045] By fusing the outputs of two branches through branch aggregation operations and combining the expected value of the advantage to eliminate the coupling ambiguity between state value and action advantage, a precise total action value is obtained, enabling quantitative evaluation of different velocity trajectory strategies. The formula is as follows:

[0046]

[0047] in, For total value, , These are the state-value function and action-advantage function, respectively, obtained through network training and fitting. This is the expected value of the mean advantage of all possible speed actions in the current state.

[0048] Preferably, the reinforcement learning iteration specifically involves constructing a reward function based on the logic that the lower the energy consumption, the higher the smoothness, the smaller the duration deviation, and the higher the single-trip reward value, thereby driving the model to continuously converge to the optimal policy.

[0049] Build an experience replay pool to store the status, actions, rewards, and next status for each run;

[0050] In each iteration, a preset number of historical samples are randomly selected in batches to avoid overfitting caused by temporal correlation.

[0051] Calculate the loss function between the target Q value and the predicted Q value, update the network parameters through gradient backpropagation, and correct the energy consumption weight coefficient under different loads, temperatures, and travel conditions.

[0052] After the iteration is completed, the updated weight parameters are output, thus completing the optimization and upgrade of the forward planning model.

[0053] The system includes an edge perception acquisition module, a dynamic speed trajectory candidate engine module, a forward planning model module, an energy consumption assessment and iterative optimization module, and an elevator operation execution module, specifically:

[0054] The edge perception acquisition module collects real-time elevator operating parameters and combines them with building time-series passenger flow data to obtain multi-dimensional time-varying operating parameters;

[0055] The dynamic speed trajectory candidate engine module constructs a coupling mechanism of elevator electromechanical loss, mechanical resistance, and load loss based on the collected multi-dimensional time-varying operating parameters. According to the real-time working condition boundary conditions, it generates multiple sets of candidate speed trajectories that meet the constraints of acceleration, jerk, comfort, and running efficiency.

[0056] The forward planning model module incorporates the LSTM time series prediction algorithm and the MPC model prediction control algorithm to predict the floor call trends and passenger flow changes within a preset time period, and to perform global optimization on multiple candidate speed trajectories to generate a forward-looking optimal speed trajectory.

[0057] The energy consumption assessment and iterative optimization module has a built-in dual-duel deep Q-network reinforcement learning architecture. It periodically completes sample playback and parameter iteration based on real-time elevator operating parameters, adaptively corrects the energy consumption weights under different time-varying operating conditions, and completes the iterative optimization of the forward planning model.

[0058] The elevator operation execution module controls the elevator operation according to the optimal speed trajectory, and collects elevator operation parameters in real time during the execution of the elevator trajectory and transmits them back to the system.

[0059] Compared with the prior art, the technical solution of this application has the following technical effects:

[0060] This invention designs a dynamic velocity trajectory candidate generation mechanism, which generates multiple sets of candidate trajectories in batches based on the S-shaped operation principle and boundary constraints, and completes energy consumption calculation and trajectory screening according to multiple loss coupling models. It can output diverse compliant trajectories under different loads and travel conditions, thus broadening the adaptability of control strategies.

[0061] This invention adopts a two-layer forward planning architecture that combines time-series prediction with rolling optimization. Based on the time-series algorithm, it predicts the trend of passenger flow changes and outputs the operating condition constraints in advance. Through a multi-objective optimization algorithm, it performs global optimization on candidate trajectories, which effectively improves the foresight of speed planning and simultaneously optimizes elevator energy consumption, running efficiency and ride smoothness.

[0062] This invention builds a dynamic evaluation module through a deep reinforcement learning architecture. Based on a multi-dimensional reward function and experience replay mechanism, it continuously iterates and optimizes the control parameters. It adopts an offline iteration and online activation operation mode, which can adaptively adjust the control weights according to equipment aging, environmental fluctuations, and changes in passenger flow, so as to ensure the stable energy consumption optimization effect during the long-term operation of the elevator.

[0063] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0064] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0066] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0067] Figure 1 A flowchart of an optimal speed planning method for elevator operation with optimal energy consumption;

[0068] Figure 2 This is an architecture diagram of the forward-looking planning model in this application;

[0069] Figure 3 This is a diagram of the architecture of the dynamic energy consumption assessment model in this application;

[0070] Figure 4 This is an architecture diagram of an elevator operation energy-optimal speed planning system.

[0071] Figure 5 This is a comparison chart of the average energy consumption data per trip per day in the tests of each method in the embodiments of this application;

[0072] Figure 6 This is a comparison chart of the average daily runtime data for each method tested in the embodiments of this application;

[0073] Figure 7 This is a thermographic comparison chart of the daily accelerometer data for each method tested in the embodiments of this application;

[0074] Figure 8 This is a probability density map of waiting time during peak periods in the testing of each method in the embodiments of this application. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0076] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0077] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0078] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0079] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0080] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0081] Example 1 describes a method for optimal speed planning for elevator operation with optimal energy consumption. Figure 1 As shown, it specifically includes:

[0082] By collecting real-time elevator operating parameters at the millisecond level through the edge perception module and combining them with the building's time-series passenger flow data, multi-dimensional time-varying operating parameters are obtained.

[0083] Based on the collected multi-dimensional time-varying operating parameters, a dynamic speed trajectory candidate engine coupled with elevator electromechanical loss, mechanical resistance, and load loss is established. The engine generates multiple sets of candidate speed trajectories that meet the constraints of acceleration, jerk, comfort, and running efficiency according to real-time working condition boundary conditions.

[0084] A forward-looking planning model is constructed. The forward-looking planning model uses the LSTM time series prediction algorithm to predict the floor call trend and passenger flow changes within a preset time period. Combined with the predictive control MPC algorithm, it performs global optimization on multiple candidate speed trajectories to generate a forward-looking optimal speed trajectory.

[0085] The elevator operates based on the optimal speed trajectory, and real-time elevator operation parameters are collected when the elevator actually executes the optimal trajectory.

[0086] A dynamic energy consumption assessment model is constructed using a dual-duel deep Q-network reinforcement learning architecture. The dynamic energy consumption assessment model periodically performs sample playback and parameter iteration on the look-ahead planning model based on real-time elevator operating parameters, and adaptively corrects the energy consumption weights under different time-varying operating conditions to obtain an optimized look-ahead planning model.

[0087] Before the next elevator operation, the elevator speed is planned using an optimized look-ahead planning model.

[0088] Furthermore, the multi-dimensional time-varying operating parameters include the real-time load mass of the car, the real-time travel distance of the elevator, the stator current of the traction motor, the rotor speed, the motor temperature, the grid voltage fluctuation value, the car motion drag coefficient, the wire rope friction loss coefficient, the building time-series passenger flow density, and the floor call sequence data.

[0089] Furthermore, the inherent constants of the equipment do not need to be collected in real time, including the stator copper resistance of the motor, the coefficient of friction of the track, the weight of the car, and the reference parameters of the motor iron loss.

[0090] Furthermore, the dynamic velocity trajectory candidate engine specifically reads real-time operating condition boundary parameters and current multi-dimensional time-varying operating parameters, and divides the velocity trajectory into acceleration, uniform acceleration, uniform speed, deceleration, and acceleration / deceleration segments based on the S-shaped velocity trajectory segmentation principle.

[0091] Within the boundary constraints, the acceleration and peak velocity are discretely sampled, and the parameters and durations corresponding to each segment that meet the adjustment are dynamically generated, and differentiated candidate velocity trajectories are generated in batches.

[0092] The total predicted energy consumption for each candidate trajectory is calculated one by one. The energy consumption index, comfort index and timeliness index are combined for screening. Illegal trajectories that exceed the threshold are eliminated, and the set of compliant candidate trajectories with lower energy consumption and meeting the comfort and timeliness constraints are retained.

[0093] Furthermore, the dynamic generation of operating parameters and duration for each segment that meets the constraints is specifically as follows: the acceleration segment uses maximum jerk. To constrain the time taken, dynamic calculation is used to accelerate the setup time. ,in To achieve the optimal target acceleration under the current operating conditions, the velocity in this stage smoothly increases from 0 to... ;

[0094] If the speed limit threshold is not reached during the uniform acceleration phase, maintain constant acceleration. Run, dynamically match duration Until the speed approaches the motor's efficient speed range or the preset peak speed;

[0095] Whether to maintain the constant speed segment is dynamically determined based on the remaining travel distance. If the long-stroke condition is met, the optimal steady-state speed is maintained. The running time is solved by the closed-loop calculation of the remaining displacement difference; the short-stroke working condition is directly set to zero for the uniform speed time, and the uniform speed interval is canceled.

[0096] The deceleration and acceleration / deceleration phases calculate the deceleration start point in advance based on the target stopping position, and dynamically allocate the deceleration slope and acceleration / deceleration duration with a comfort threshold as a constraint, ensuring that the speed returns to zero upon arrival at the station without speed overshoot or braking shock.

[0097] Furthermore, the calculation of the predicted total energy consumption for each candidate trajectory group is as follows: For each candidate velocity trajectory group, the segmented velocity and acceleration sequences within the complete trajectory running sequence are extracted. Combined with real-time collected multi-dimensional time-varying operating parameters and inherent constants of the equipment, the elevator electromechanical losses, mechanical resistance, and load losses are calculated separately according to the trajectory segmented time sequence. The three types of losses are accumulated and coupled in real time to obtain the complete predicted total energy consumption for a single candidate trajectory group. The formula is:

[0098]

[0099] in, This represents the total energy consumption for a single elevator trip. The electrical losses of the motor are calculated using the following formula:

[0100]

[0101] in, This refers to the stator current of the motor. For the stator copper resistor of the motor, This represents the real-time iron loss power of the motor. In the formula, This refers to the real-time speed of the motor. For real-time motor temperature, The reference temperature is room temperature. These are the fitting coefficients for iron loss;

[0102] The formula for calculating the switching losses of the frequency converter is as follows:

[0103]

[0104] in, This is the switching loss coefficient (inherent constant) of the frequency converter. The real-time switching frequency is dynamically calculated based on the operating conditions.

[0105] The formula for calculating mechanical friction and wind resistance loss is as follows:

[0106]

[0107] in, For environmental wind pressure resistance, For the real-time operating speed of the elevator, This is the coefficient of friction of the track (an inherent constant of the equipment). The total mass of the car. This is the acceleration due to gravity.

[0108] Furthermore, such as Figure 2 The diagram shows the architecture of the forward-looking planning model. Specifically, the forward-looking planning model is a two-layer intelligent energy consumption planning model that integrates LSTM time-series passenger flow prediction and MPC rolling time-domain optimization. The model is divided into an upper-layer time-series prediction and a lower-layer rolling optimization planning. The upper-layer time-series prediction is responsible for predicting the floor call trends and passenger load changes in the short term based on historical passenger flow time-series data, providing prior information on the operating conditions for speed planning. The lower-layer rolling optimization planning is responsible for performing rolling time-domain multi-objective optimization on a batch of candidate speed trajectories based on the predicted operating conditions and the current real-time operating conditions, and selecting the forward-looking speed trajectory that comprehensively optimizes energy consumption, timeliness, and comfort.

[0109] Furthermore, the upper-level time series prediction specifically involves using the number of elevator calls, elevator start and stop frequency, average load, and time period labels within a fixed historical period to form a time series feature sequence, which is then used as the input to the LSTM network.

[0110] Temporal feature memorization and iterative updates are achieved through an LSTM gating mechanism, based on the hidden layer temporal features output by the LSTM. The probability of floor calls, passenger flow load trends, and start / stop times within a preset time period are predicted by regression analysis of the fully connected layer.

[0111] The prediction results are standardized by using the Sigmoid activation function of the output layer to obtain standardized prior working condition constraint information, which is then sent down to the lower layer.

[0112] Furthermore, this LSTM time series prediction algorithm adopts a two-layer stacked LSTM structure, with 64 neurons in each layer. The input dimension matches the temporal feature dimension with 8, the hidden layer dimension is set to 64, and the iteration time step is set to 10. A dropout layer is added between the two LSTM layers with a dropout rate of 0.2 to suppress overfitting during temporal training. A fully connected layer is connected at the end of the network to map the hidden layer feature dimension to the predicted output dimension. The output layer is configured with a Sigmoid activation function to complete data standardization.

[0113] The lower-level rolling optimization planning specifically involves using the Model Predictive Control (MPC) algorithm. Based on the predicted operating conditions output by the upper-level LSTM, it performs rolling time-domain global optimization on the selected compliant candidate trajectories. The specific process of global optimization is as follows:

[0114] Set the time domain length T for rolling optimization, and discretize the future running process into N control steps;

[0115] The multi-objective minimization objective function for MPC is established as follows:

[0116]

[0117] in, For the comprehensive cost function, with Minimize as the objective. For the first Energy consumption per control step This is due to runtime deviation. For the first The comfort fluctuation value of each control step. These are the weighting coefficients for energy consumption, runtime deviation, and comfort fluctuation values, respectively.

[0118] In each rolling time domain, traverse all compliant candidate velocity trajectories and solve for the optimal trajectory sequence corresponding to the minimum objective function value;

[0119] Execute the optimal speed command for the current moment, update the real-time operating conditions and LSTM predicted operating conditions in the next moment, and re-roll and iterate to finally output a forward-looking optimal speed trajectory that meets the constraints of energy consumption, timeliness and comfort.

[0120] Furthermore, the MPC algorithm adopts a single-layer rolling optimization and time-series layered execution architecture, which is divided into three progressive logics: prediction layer, optimization layer, and execution update layer, with each layer nested and iterated in a closed loop.

[0121] The optimal parameters for the MPC algorithm are set as follows: preset rolling optimization time domain length 1s, discrete control step size 0.05s, number of discrete control steps N=20 in a single time domain; energy consumption weighting coefficient. Time-weighted coefficient Comfort weighted coefficient To enable dynamic adaptive adjustment of parameters, the initial base values ​​are set to 0.6, 0.2, and 0.2 respectively.

[0122] The constraint boundaries of the MPC algorithm inherit the standardized operating condition thresholds of the upper-level LSTM and the elevator hardware limit parameters, including speed, acceleration, jerk upper limit constraints and runtime constraints.

[0123] Furthermore, such as Figure 3 The diagram shows the architecture of the dynamic energy consumption assessment model, which is specifically an adaptive iterative optimization assessment model built on a dual-duel deep Q-network.

[0124] The model is based on the full-dimensional coupling of elevator energy consumption, operating comfort, and operating efficiency. Through a reinforcement learning iterative mechanism, it evaluates the performance of different operating conditions and different speed trajectory strategies in real time, and realizes dynamic correction and adaptive optimization of the weight parameters of the forward planning model.

[0125] Furthermore, the dual-duel deep Q-network is specifically designed as follows: it adopts a dual-headed duel network structure with a shared feature extraction layer and independent output of two branches, decoupling the total action value Q-value into two independent parts: state value and action advantage.

[0126] Among them, the value branch outputs a status value based on the current global operating condition, which represents the basic quality of the current overall elevator operating environment, load, and equipment status.

[0127] The advantageous branch combines the operating conditions and different speed control actions to output the relative advantage benefits of each candidate action, representing the differences in energy consumption and comfort benefits of different acceleration, peak speed, and deceleration slope adjustment strategies.

[0128] By fusing the outputs of two branches through branch aggregation operations and combining the expected value of the advantage to eliminate the coupling ambiguity between state value and action advantage, a precise total action value is obtained, enabling quantitative evaluation of different velocity trajectory strategies. The formula is as follows:

[0129]

[0130] in, For total value, , These are the state-value function and action-advantage function, respectively, obtained through network training and fitting. This is the expected value of the mean advantage of all possible speed actions in the current state.

[0131] Furthermore, the reinforcement learning iteration specifically involves constructing a reward function based on the logic that lower energy consumption, higher smoothness, smaller duration deviation, and higher single-trip reward value, driving the model to continuously converge to the optimal policy. The formula is:

[0132]

[0133] in, For the reward function, These are weighted by energy consumption, comfort, and timeliness. Total energy consumption, This is the deviation between the actual running time and the standard running time. Rate the smoothness. In the formula, As the baseline for perfect smoothness constant, As a comfort penalty factor, The elevator is given real-time acceleration, and the integral term is the absolute cumulative fluctuation of acceleration throughout the entire process, which is used to quantify the degree of impact and vibration during the entire operation.

[0134] Build an experience replay pool to store the status, actions, rewards, and next status for each run;

[0135] In each iteration, a preset number of historical samples are randomly selected in batches to avoid overfitting caused by temporal correlation.

[0136] Calculate the loss function between the target Q value and the predicted Q value, update the network parameters through gradient backpropagation, and correct the energy consumption weight coefficient under different loads, temperatures, and travel conditions.

[0137] After the iteration is completed, the updated weight parameters are output, thus completing the optimization and upgrade of the forward planning model.

[0138] Furthermore, the iterative activation of the optimized forward planning model is as follows: the dynamic energy consumption assessment model adopts an offline caching and online replacement mechanism. After a single run is completed, sample iteration is performed, and the current elevator operation control is not affected during the iteration. After the iteration converges and the weight parameters are updated, the forward planning model parameters are globally refreshed before the next elevator power-on call task is started, so that the new round of speed planning is adapted to the latest equipment operating conditions and environmental parameters, and the strategy is continuously self-optimized.

[0139] This embodiment details a method for optimal speed planning for elevator operation with optimal energy consumption. It collects multi-dimensional operating parameters and passenger flow data through edge sensing, and constructs a dynamic speed trajectory candidate engine by combining inherent equipment constants to generate and filter compliant candidate trajectories. A forward-looking planning model is built, using an LSTM network to predict passenger flow and operating conditions and standardize the output results. A multi-objective cost function is then constructed using the MPC algorithm to continuously solve for the comprehensive optimal speed trajectory. The elevator operates based on the optimal trajectory and transmits operating data back in real time. A dynamic energy consumption evaluation model is built using a dual-duel deep Q-network, and a reward function is constructed by combining smoothness scores. Parameter iteration is completed through experience playback and gradient backpropagation, adaptively adjusting the weights of the forward-looking planning model. The optimized model parameters are loaded before the next operation.

[0140] Example 2, based on Example 1, describes in detail an elevator operation energy-optimal speed planning system, including an edge perception acquisition module, a dynamic speed trajectory candidate engine module, a look-ahead planning model module, an energy consumption assessment iterative optimization module, and an elevator operation execution module, specifically:

[0141] The edge perception acquisition module collects real-time elevator operating parameters and combines them with building time-series passenger flow data to obtain multi-dimensional time-varying operating parameters;

[0142] The dynamic speed trajectory candidate engine module constructs a coupling mechanism of elevator electromechanical loss, mechanical resistance, and load loss based on the collected multi-dimensional time-varying operating parameters. According to the real-time working condition boundary conditions, it generates multiple sets of candidate speed trajectories that meet the constraints of acceleration, jerk, comfort, and running efficiency.

[0143] The forward planning model module incorporates the LSTM time series prediction algorithm and the MPC model predictive control algorithm to predict the floor call trends and passenger flow changes within a preset time period, and performs global optimization on multiple candidate speed trajectories to generate the forward-looking optimal speed trajectory.

[0144] The energy consumption assessment and iterative optimization module has a built-in dual-duel deep Q-network reinforcement learning architecture. It regularly performs sample playback and parameter iteration based on real-time elevator operating parameters, adaptively corrects the energy consumption weights under different time-varying operating conditions, and completes the iterative optimization of the forward planning model.

[0145] The elevator operation execution module controls the elevator operation based on the optimal speed trajectory, and collects elevator operation parameters in real time during the execution of the elevator trajectory and sends them back to the system.

[0146] Furthermore, the dynamic velocity trajectory candidate engine module includes a working condition boundary reading subunit, an S-shaped trajectory segmentation generation subunit, a trajectory parameter dynamic rotor matching unit, and a trajectory filtering subunit;

[0147] The working condition boundary reading subunit reads real-time working condition boundary parameters and multi-dimensional time-varying operating parameters;

[0148] The S-shaped trajectory segmentation generation subunit is based on the S-shaped velocity trajectory segmentation principle, which divides the elevator running trajectory into acceleration segment, uniform acceleration segment, uniform speed segment, deceleration segment, and acceleration / deceleration segment.

[0149] The trajectory parameter dynamic matching rotor unit performs discrete sampling of acceleration and peak velocity within the boundary constraint range, dynamically generates the corresponding operating parameters and duration of each segment, and generates differentiated candidate velocity trajectories in batches;

[0150] The trajectory screening and filtering subunit calculates the predicted total energy consumption for each group of candidate trajectories, and combines energy consumption indicators, comfort indicators and timeliness indicators to jointly screen and eliminate illegal trajectories that exceed the threshold, and retain the set of compliant candidate trajectories that meet the constraints.

[0151] Furthermore, the dynamic velocity trajectory candidate engine module has a built-in energy consumption coupling calculation subunit. For each group of candidate velocity trajectories, the energy consumption coupling calculation subunit extracts the time-series segmented velocity and acceleration operation sequences, combines real-time operation parameters and inherent constants of the equipment, calculates segmented electrical losses, mechanical resistance, and load losses, and accumulates and couples them in real time to obtain the predicted total energy consumption of a single group of candidate trajectories.

[0152] Furthermore, the forward-looking planning model module is a two-layer nested intelligent energy consumption planning architecture, including an upper-layer time-series prediction sub-module and a lower-layer rolling optimization planning sub-module;

[0153] The upper-level time-series prediction submodule predicts the trend of elevator calls and changes in passenger load in the floor in the near future based on historical passenger flow time-series data, and outputs standardized prior operating condition constraint information to provide prior operating condition boundaries for speed planning.

[0154] The lower-level rolling optimization planning submodule performs rolling time-domain multi-objective optimization on a batch of candidate speed trajectories based on the upper-level predicted operating conditions and the current real-time operating conditions, and selects the forward-looking speed trajectory that has the best overall performance in terms of energy consumption, timeliness, and comfort.

[0155] Furthermore, the energy consumption assessment and iterative optimization module incorporates a dynamic energy consumption assessment model, which is built on a dual-duel deep Q-network architecture. The dynamic energy consumption assessment model relies on a multi-dimensional evaluation system of elevator energy consumption, comfort, and timeliness. Through a reinforcement learning iterative mechanism, it evaluates the performance of different operating conditions and different speed trajectory strategies in real time, dynamically corrects the energy consumption weight parameters of the forward planning model, and achieves adaptive optimization and upgrading of the speed planning strategy.

[0156] This embodiment details an elevator operation energy-optimized speed planning system, including an edge perception acquisition module, a dynamic speed trajectory candidate engine module, a look-ahead planning model module, an energy consumption assessment and iterative optimization module, and an elevator operation execution module. The edge perception acquisition module collects multi-dimensional parameters of elevator operation and passenger flow. The dynamic speed trajectory candidate engine module generates and filters compliant candidate trajectories in batches based on the S-shaped trajectory principle. The look-ahead planning model module uses an LSTM network to predict passenger flow conditions and standardize the output. It then constructs a multi-objective cost function using the MPC algorithm and performs rolling optimization to obtain the optimal speed trajectory. The elevator operation execution module controls the equipment to operate according to the trajectory and transmits data back. The energy consumption assessment and iterative optimization module relies on a dual-duel deep Q-network, combining a reward function and an experience replay mechanism to complete model iteration, updating weight parameters offline and refreshing parameters before the next task starts, thus achieving adaptive optimization of the planning strategy.

[0157] Example 3, based on Example 1 or 2, details the implementation and verification of this method on a permanent magnet synchronous traction elevator with a load capacity of 1000kg and a rated speed of 1.75m / s in a 28-story commercial office building, as follows:

[0158] The elevator operates an average of 1,200 trips per day, with dense passenger flow during the morning peak (7:30-9:00) and evening peak (17:30-19:00), and sparse passenger flow during off-peak hours. The experimental period is 30 consecutive days, with basic data collected and model initialization completed in the first 10 days, and comparative testing conducted in the following 20 days.

[0159] During implementation, the edge sensing acquisition frequency of this method is 10ms / time; the LSTM temporal prediction network is a two-layer stacked network with 64 neurons per layer, an input dimension of 8, a time step of 10, and a dropout rate of 0.2; the MPC rolling optimization has a temporal length of 1s, a control step of 0.05s, a single temporal step count of 20, and initial weights... The dual-duel deep Q-network has a shared layer of 128→64 neurons, with 32 neurons in each branch, a learning rate of 0.001, a replay pool capacity of 10000, a batch size of 64, and a target network update interval of 50 rounds; the smoothness score is based on 100, and the comfort penalty coefficient is... .

[0160] This method uses an edge perception module to collect elevator operating parameters and building passenger flow data at the millisecond level, fusing them to obtain multi-dimensional time-varying operating parameters. Based on these multi-dimensional time-varying operating parameters, a coupling mechanism for electromechanical losses, mechanical resistance, and load losses is constructed. Multiple sets of candidate speed trajectories that meet the constraints are generated in batches according to real-time operating condition boundary conditions. A two-layer look-ahead planning architecture is adopted, using a time-series prediction algorithm to predict the floor call trends and passenger flow changes within a preset time period. Combined with a model predictive control algorithm, the candidate speed trajectories are globally optimized to generate a look-ahead optimal speed trajectory. The elevator operation is controlled according to the optimal speed trajectory, and actual operating parameters are collected synchronously. Sample playback and parameter iteration are performed periodically based on actual operating parameters, adaptively correcting the energy consumption weights under different time-varying operating conditions. The optimized model parameters are loaded before the next elevator operation task is started, realizing continuous optimization of the speed planning strategy.

[0161] This elevator lobby in the building has a total of 4 elevators. The other three elevators are used and controlled by the existing mainstream Event Triggered Model Predictive Control (ET-MPC), Imitation Learning-Reinforcement Learning Hybrid Scheduling Method (IL-RL), and Absolute Residual Distance Adaptive Speed ​​Control Method (ARDP).

[0162] Among them, the Event Triggered Model Predictive Control (ET-MPC) method establishes the elevator system state model and measurement model, and designs multi-objective performance indicators; it uses a neural network learning model to initialize the control law and obtain an approximate optimal control solution; it constructs an event triggering mechanism, which initiates online optimization only when the system state deviation exceeds a preset threshold; in the optimization stage, it uses a shortened prediction time domain to calculate the optimal speed command sequence, executes the command at the current moment and updates the system state, thereby realizing dynamic tracking and adjustment of elevator speed;

[0163] The Imitation Learning-Reinforcement Learning Hybrid Scheduling Method (IL-RL) collects historical operational data of expert scheduling strategies, completes imitation learning pre-training through a behavior cloning algorithm, and obtains an initial control strategy. The pre-trained strategy is used as the initial network for reinforcement learning, and a multi-dimensional reward function including energy consumption, waiting time, and runtime is constructed. A direct effect update interval mechanism is introduced to optimize the training efficiency of reinforcement learning. By interacting with the environment, the network parameters are continuously updated iteratively to achieve adaptive optimization of the elevator operation strategy.

[0164] The Absolute Remaining Distance Adaptive Speed ​​Control (ARDP) method uses position sensors to acquire real-time information about the elevator's current position and target floor; calculates the absolute remaining distance between the current position and the target position; dynamically adjusts the running speed, acceleration, and jerk parameters based on the absolute remaining distance and the elevator's dynamic characteristics; and gradually reduces the speed as it approaches the target floor to achieve smooth stopping and precise leveling of the elevator, thereby improving operating efficiency and passenger comfort.

[0165] The average energy consumption per trip per day during the 20-day test for each method was calculated, and the data is shown in Table 1 below:

[0166] Table 1. Average energy consumption per trip per day during the testing of each method.

[0167] Test days Invention Solution ET-MPC IL-RL ARDP 1 0.243 0.248 0.241 0.265 5 0.229 0.242 0.236 0.258 10 0.215 0.239 0.232 0.255 15 0.211 0.238 0.230 0.254 20 0.209 0.236 0.229 0.253 20-day average 0.219 0.241 0.234 0.256

[0168] According to Table 1 and Figure 5 The comparison chart of average energy consumption data per trip per day for each method shows that, within the 20-day testing period, the power consumption data of each method has been optimized to some extent as the database has been enriched and the model has been iterated. Among them, the method of this method, which uses a dynamic energy consumption evaluation model built with a dual-duel deep Q-network reinforcement learning architecture to iterate the look-ahead planning model, can continuously learn the optimal control strategy under different operating conditions, and achieve more efficient model optimization. It has the largest and fastest power consumption data optimization within 20 days, which verifies the long-term operating advantages and stability of this method.

[0169] The average daily runtime of each method during the 20-day test was calculated, and the data is shown in Table 2 below:

[0170] Table 2. Average daily runtime during each method test.

[0171] Test days Invention Solution ET-MPC IL-RL ARDP 1 23.2 23.5 23.3 22.9 5 22.6 23.0 22.9 22.6 10 22.2 22.8 22.5 22.5 15 22.1 22.7 22.4 22.4 20 22.0 22.6 22.3 22.3 20-day average 22.3 22.9 22.7 22.5

[0172] According to Table 2 and Figure 6As shown in the comparison chart of the average daily runtime data of each method during the test, this method achieves long-term incremental optimization through continuous iterative optimization mechanism during the test period. After multiple rounds of adaptive parameter adjustment, the runtime steadily decreases and gradually converges. Finally, the average runtime over 20 days is better than all control groups, proving that this scheme can continuously explore the space for runtime efficiency optimization during the iteration process and has obvious advantages in efficiency optimization potential and overall runtime performance.

[0173] The jerk data from the 20-day test for each method were statistically analyzed, and the data is shown in Table 3 below:

[0174] Table 3. Daily accelerometer data for each method test.

[0175] Test days Invention Solution ET-MPC IL-RL ARDP 1 0.86 0.93 0.90 1.04 5 0.79 0.90 0.86 0.99 10 0.74 0.88 0.83 0.97 15 0.72 0.87 0.82 0.96 20 0.70 0.86 0.81 0.95 20-day average 0.75 0.88 0.84 0.98

[0176] According to Table 3 and Figure 7 The heatmap showing the daily acceleration data comparison for each method test reveals that this method, based on a forward-looking planning model, achieved smoother acceleration performance in the early stages. With continuous iterations, the average acceleration was the lowest in the group, resulting in less impact during gear changes. This led to a final smoothness score of 89.6, significantly higher than ET-MPC (82.3), IL-RL (84.1), and ARDP (78.5), and an 8.87% improvement in smoothness compared to the best control group. Thanks to continuous iteration and optimization of gear control parameters throughout the test, this solution continuously optimized the acceleration and deceleration curves, effectively reducing running bumps and impacts, ultimately achieving a significant advantage in ride smoothness.

[0177] The average waiting time during peak hours for each method was calculated, and the data is shown in Table 4 below:

[0178] Table 4 shows the average waiting time during peak periods in the tests for each method.

[0179] Test days Invention Solution ET-MPC IL-RL ARDP 1 21.2 23.3 20.5 25.5 5 19.8 22.4 19.6 24.8 10 18.9 22.3 19.0 24.5 15 18.7 22.2 18.6 24.3 20 18.6 22.0 18.4 24.2 20-day average 19.3 22.4 19.1 24.7

[0180] According to Table 4 and Figure 8 As shown in the peak-hour waiting time probability density maps of each method test, based on the time-series passenger flow prediction and closed-loop iterative optimization mechanism, this solution continuously reduces the waiting time throughout the entire test cycle, with a longer optimization cycle. In contrast, the three control groups only underwent minor optimization in the early stages of the test before entering a plateau period, demonstrating the outstanding adaptive optimization capability of this method and its greater potential for long-term performance improvement.

[0181] This embodiment details the implementation and verification of the proposed method on a 1000kg load, 1.75m / s rated speed permanent magnet synchronous traction elevator in a 28-story commercial office building. The invention was compared with three existing mainstream elevator speed control schemes: ET-MPC, IL-RL, and ARDP. A 20-day tracking and statistical analysis was conducted using four quantitative indicators: energy consumption, runtime, acceleration, and peak waiting time. The results show that the proposed method, utilizing a multi-level forward-looking speed control architecture based on millisecond-level edge data acquisition, LSTM passenger flow time-series prediction, MPC trajectory optimization, and dual-duel deep Q-network closed-loop iterative optimization, continuously iterates and optimizes all performance indicators throughout the entire cycle. The energy consumption reduction and optimization rate are superior to all control methods. The average runtime is the best in the group, the acceleration value is the lowest, and the ride smoothness is significantly improved. Although the peak waiting time is second best overall, the optimization range is the largest, and the long-term adaptive optimization potential is outstanding. This fully demonstrates that the proposed method can balance energy saving, operational efficiency, and ride comfort.

[0182] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for optimal speed planning of elevator operation energy consumption, characterized in that, include: By collecting real-time elevator operating parameters at the millisecond level through the edge perception module and combining them with the building's time-series passenger flow data, multi-dimensional time-varying operating parameters are obtained. Based on the collected multi-dimensional time-varying operating parameters, a dynamic speed trajectory candidate engine coupled with elevator electromechanical loss, mechanical resistance, and load loss is established. The engine generates multiple sets of candidate speed trajectories that meet the constraints of acceleration, jerk, comfort, and running efficiency according to real-time working condition boundary conditions. A forward-looking planning model is constructed. The forward-looking planning model uses the LSTM time series prediction algorithm to predict the floor call trend and passenger flow changes within a preset time period. Combined with the predictive control MPC algorithm, it performs global optimization on multiple candidate speed trajectories to generate a forward-looking optimal speed trajectory. The elevator operates based on the optimal speed trajectory, and real-time elevator operation parameters are collected when the elevator actually executes the optimal trajectory. A dynamic energy consumption assessment model is constructed using a dual-duel deep Q-network reinforcement learning architecture. The dynamic energy consumption assessment model periodically performs sample playback and parameter iteration on the look-ahead planning model based on real-time elevator operating parameters, and adaptively corrects the energy consumption weights under different time-varying operating conditions to obtain an optimized look-ahead planning model. Before the next elevator operation, the elevator speed is planned using an optimized look-ahead planning model.

2. The elevator operation energy-optimal speed planning method according to claim 1, characterized in that, The dynamic speed trajectory candidate engine specifically reads real-time operating condition boundary parameters and current multi-dimensional time-varying operating parameters, and divides the trajectory into acceleration, uniform acceleration, uniform speed, deceleration, and acceleration / deceleration segments based on the S-shaped speed trajectory segmentation principle. Within the boundary constraints, the acceleration and peak velocity are discretely sampled, and the parameters and durations corresponding to each segment that meet the adjustment are dynamically generated, and differentiated candidate velocity trajectories are generated in batches. The total predicted energy consumption for each candidate trajectory is calculated one by one. The energy consumption index, comfort index and timeliness index are combined for screening. Illegal trajectories that exceed the threshold are eliminated, and the set of compliant candidate trajectories with lower energy consumption and meeting the comfort and timeliness constraints are retained.

3. The elevator operation energy-optimal speed planning method according to claim 2, characterized in that, The calculation of the predicted total energy consumption for each candidate trajectory group specifically involves: for each candidate velocity trajectory group, extracting segmented velocity and acceleration sequences within the complete trajectory runtime sequence; combining real-time collected multi-dimensional time-varying operating parameters and inherent equipment constants; calculating elevator electromechanical losses, mechanical resistance, and load losses according to the trajectory segmented time sequence; and accumulating and coupling these three types of losses in real time to obtain the complete predicted total energy consumption for a single candidate trajectory group, as shown in the formula: in, This represents the total energy consumption for a single elevator trip. The electrical losses of the motor are calculated using the following formula: in, This refers to the stator current of the motor. For the stator copper resistor of the motor, This represents the real-time iron loss power of the motor. In the formula, This refers to the real-time speed of the motor. For real-time motor temperature, The reference temperature is room temperature. These are the fitting coefficients for iron loss; The formula for calculating the switching losses of the frequency converter is as follows: in, This is the switching loss coefficient (inherent constant) of the frequency converter. The real-time switching frequency is dynamically calculated based on the operating conditions. The formula for calculating mechanical friction and wind resistance loss is as follows: in, For environmental wind pressure resistance, For the real-time operating speed of the elevator, This is the coefficient of friction of the track (an inherent constant of the equipment). The total mass of the car. This is the acceleration due to gravity.

4. The elevator operation energy-optimal speed planning method according to claim 1, characterized in that, The forward planning model is specifically a two-layer intelligent energy consumption planning model that integrates LSTM time-series passenger flow prediction and MPC rolling time-domain optimization. The model is divided into an upper-layer time-series prediction and a lower-layer rolling optimization planning. The upper-layer time-series prediction is responsible for predicting the floor call trend and passenger load changes in the short term based on historical passenger flow time-series data, providing prior information on the working conditions for speed planning. The lower-level rolling optimization planning is responsible for performing rolling time-domain multi-objective optimization on a batch of candidate speed trajectories based on the predicted operating conditions and the current real-time operating conditions, and selecting the forward-looking speed trajectory that has the best overall performance in terms of energy consumption, timeliness, and comfort.

5. The elevator operation energy-optimal speed planning method according to claim 4, characterized in that, The upper-level time series prediction specifically involves using a time series feature sequence composed of the number of elevator calls, elevator start and stop frequency, average load, and time period label within a fixed historical period as the input to the LSTM network. Temporal feature memorization and iterative updates are achieved through an LSTM gating mechanism, based on the hidden layer temporal features output by the LSTM. The probability of floor calls, passenger flow load trends, and start / stop times within a preset time period are predicted by regression analysis of the fully connected layer. The prediction results are standardized by using the Sigmoid activation function of the output layer to obtain standardized prior working condition constraint information, which is then sent down to the lower layer.

6. The elevator operation energy-optimal speed planning method according to claim 5, characterized in that, The lower-level rolling optimization planning adopts the Model Predictive Control (MPC) algorithm. Based on the predicted operating conditions output by the upper-level LSTM, it performs rolling time-domain global optimization on the screened compliant candidate trajectories. The specific process of global optimization is as follows: Set the time domain length T for rolling optimization, and discretize the future running process into N control steps; The multi-objective minimization objective function for MPC is established as follows: in, For the comprehensive cost function, with Minimize as the objective. For the first Energy consumption per control step This is due to runtime deviation. For the first The comfort fluctuation value of each control step. These are the weighting coefficients for energy consumption, runtime deviation, and comfort fluctuation values, respectively. In each rolling time domain, traverse all compliant candidate velocity trajectories and solve for the optimal trajectory sequence corresponding to the minimum objective function value; Execute the optimal speed command for the current moment, update the real-time operating conditions and LSTM predicted operating conditions in the next moment, and re-roll and iterate to finally output a forward-looking optimal speed trajectory that meets the constraints of energy consumption, timeliness and comfort.

7. The elevator operation energy-optimal speed planning method according to claim 1, characterized in that, The dynamic energy consumption assessment model is specifically an adaptive iterative optimization assessment model based on a dual-duel deep Q-network. The model is based on the full-dimensional coupling of elevator energy consumption, operating comfort, and operating efficiency. Through a reinforcement learning iterative mechanism, it evaluates the performance of different operating conditions and different speed trajectory strategies in real time, and realizes dynamic correction and adaptive optimization of the weight parameters of the forward planning model.

8. The elevator operation energy-optimal speed planning method according to claim 7, characterized in that, The dual-duel depth Q-network is specifically designed as a dual-headed duel network structure with a shared feature extraction layer and independent output of two branches, which decouples the total action value Q-value into two independent parts: state value and action advantage. Among them, the value branch outputs a status value based on the current global operating condition, which represents the basic quality of the current overall elevator operating environment, load, and equipment status. The advantageous branch combines the operating conditions and different speed control actions to output the relative advantage benefits of each candidate action, representing the differences in energy consumption and comfort benefits of different acceleration, peak speed, and deceleration slope adjustment strategies. By fusing the outputs of two branches through branch aggregation operations and combining the expected value of the advantage to eliminate the coupling ambiguity between state value and action advantage, a precise total action value is obtained, enabling quantitative evaluation of different velocity trajectory strategies. The formula is as follows: in, For total value, , These are the state-value function and action-advantage function, respectively, obtained through network training and fitting. This is the expected value of the mean advantage of all possible speed actions in the current state.

9. The elevator operation energy-optimal speed planning method according to claim 8, characterized in that, The reinforcement learning iteration specifically involves constructing a reward function based on the logic that the lower the energy consumption, the higher the smoothness, the smaller the duration deviation, and the higher the single-trip reward value, thereby driving the model to continuously converge to the optimal policy. Build an experience replay pool to store the status, actions, rewards, and next status for each run; In each iteration, a preset number of historical samples are randomly selected in batches to avoid overfitting caused by temporal correlation. Calculate the loss function between the target Q value and the predicted Q value, update the network parameters through gradient backpropagation, and correct the energy consumption weight coefficient under different loads, temperatures, and travel conditions. After the iteration is completed, the updated weight parameters are output, thus completing the optimization and upgrade of the forward planning model.

10. An elevator operation energy-efficient speed planning system, used to implement the elevator operation energy-efficient speed planning method according to any one of claims 1-9, comprising an edge perception acquisition module, a dynamic speed trajectory candidate engine module, a look-ahead planning model module, an energy consumption assessment iterative optimization module, and an elevator operation execution module, as detailed below: The edge perception acquisition module collects real-time elevator operating parameters and combines them with building time-series passenger flow data to obtain multi-dimensional time-varying operating parameters; The dynamic speed trajectory candidate engine module constructs a coupling mechanism of elevator electromechanical loss, mechanical resistance, and load loss based on the collected multi-dimensional time-varying operating parameters. According to the real-time working condition boundary conditions, it generates multiple sets of candidate speed trajectories that meet the constraints of acceleration, jerk, comfort, and running efficiency. The forward planning model module incorporates the LSTM time series prediction algorithm and the MPC model prediction control algorithm to predict the floor call trends and passenger flow changes within a preset time period, and to perform global optimization on multiple candidate speed trajectories to generate a forward-looking optimal speed trajectory. The energy consumption assessment and iterative optimization module has a built-in dual-duel deep Q-network reinforcement learning architecture. It periodically completes sample playback and parameter iteration based on real-time elevator operating parameters, adaptively corrects the energy consumption weights under different time-varying operating conditions, and completes the iterative optimization of the forward planning model. The elevator operation execution module controls the elevator operation according to the optimal speed trajectory, and collects elevator operation parameters in real time during the execution of the elevator trajectory and transmits them back to the system.

Images