An elevator real-time scheduling optimization system based on edge computing

Abstract

The application discloses an elevator real-time scheduling optimization system based on edge computing, comprising: a multi-source scheduling data acquisition and preprocessing module, which generates a standardized scheduling input vector set; an event-aware gate module, which receives an input vector and historical indicators, calculates a skip probability and outputs a variable step length value; a variable step length control module, which controls hidden state updating and generates a hidden state cache sequence; a skip attention pooling module, which generates an abstract vector; an improved SkipRNN hidden state updating module, which updates the hidden state; a linear readout layer module, which generates a target floor instruction sequence; a group control scheduling controller interaction module, which generates a multi-objective reward value; and a meta-reinforcement learning head module, which updates event-aware gate and step length mapping parameters and completes closed-loop optimization. The application realizes intelligent elevator scheduling, rapid response and energy consumption optimization, and improves operation efficiency and passenger satisfaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of elevator intelligent scheduling, and in particular to an elevator real-time scheduling optimization system based on edge computing. BACKGROUND

[0002] The existing elevator scheduling method mainly makes decisions based on fixed rules or centralized control strategies, and often adopts first-come-first-served, shortest distance first or time window first scheduling optimization. In actual application, some systems introduce machine learning algorithms based on cloud deployment to try to improve scheduling efficiency through historical data mining. Such methods rely on centralized servers for state calculation and scheduling decisions, and have problems of large response delay and poor edge adaptability, which cannot meet the real-time requirements under high-frequency passenger flow changes and multi-objective scheduling demands. At the same time, some edge-deployed lightweight models lack the ability to dynamically perceive the boarding events, and the scheduling strategy is difficult to jointly optimize according to the elevator state, passenger flow density and energy consumption level, resulting in an imbalance between scheduling actions and operating energy consumption.

[0003] Although the existing recurrent neural network structure has sequence modeling capability, its running efficiency is limited under edge computing conditions. The SkipRNN (recurrent neural network with skip state update) structure introduces a skip mechanism to reduce the amount of calculation, but the skip decision is based on fixed gating logic, which cannot dynamically perceive changes in the boarding environment and cannot optimize the skip strategy according to performance feedback, which is likely to cause key state updates to be missed, affecting scheduling accuracy. At the same time, the original SkipRNN lacks a compensation mechanism for input information during skipping, and does not have the ability to model the content of the skipped period, which is likely to cause information fragmentation and affect the model's ability to capture long-term dependencies. The existing scheduling strategy lacks a closed-loop adaptive optimization mechanism, making it difficult to achieve dynamic feedback correction between the scheduling strategy and system operation effect, and cannot meet the intelligent, multi-objective and low-latency group control elevator scheduling requirements.

[0004] Therefore, how to provide an elevator real-time scheduling optimization system based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an elevator real-time scheduling optimization system based on edge computing. The present application fully integrates edge computing and event perception mechanisms, and describes in detail the scheduling process of skip control, state caching and closed-loop reinforcement learning strategy, which has the advantages of low computational overhead, fast response speed and high scheduling accuracy.

[0006] According to the elevator real-time scheduling optimization system based on edge computing of the present application, it comprises:

[0007] The multi-source scheduling data acquisition and preprocessing module is configured to acquire elevator multi-source scheduling data and perform preprocessing to generate a standardized scheduling input vector set.

[0008] The event-aware gate module is configured to receive the standardized scheduling input vector set, the historical average waiting time indicator and the historical average energy consumption indicator, input the event-aware gate, calculate the skip probability and output the variable step value.

[0009] The variable step control module is configured to input the variable step value into the improved SkipRNN and control the variable step control unit to keep the previous hidden state when the variable step value k is greater than zero, and perform hidden state update when the variable step value k is equal to zero, while caching k input vectors to generate a hidden state cache sequence.

[0010] The skip attention pooling module is configured to perform attention pooling operation on the hidden state cache sequence to generate an abstract vector.

[0011] The improved SkipRNN hidden state update module is configured to combine the abstract vector with the next time step standardized scheduling input vector to update the improved SkipRNN hidden state.

[0012] The linear readout layer module is configured to input the updated hidden state into the linear readout layer to generate a target floor instruction sequence.

[0013] The group control scheduling controller interaction module is configured to write the target floor instruction sequence into the group control scheduling controller, receive the scheduling controller feedback execution result, calculate the real-time average waiting time indicator and the real-time average energy consumption indicator, and generate a multi-objective reward value.

[0014] The meta-reinforcement learning head module is configured to input the multi-objective reward value into the policy optimization structure to update the event-aware gate parameters and the variable step mapping parameters, and return to the skip probability calculation step to complete the closed-loop scheduling optimization process.

[0015] Optionally, the elevator multi-source scheduling data in the multi-source scheduling data acquisition and preprocessing module is composed of elevator state data, passenger request data and floor passenger flow density data; the elevator state data records the current floor, running direction, door state and load value in sequence; the passenger request data records the call floor, target floor and request timestamp in sequence; the floor passenger flow density data is the real-time passenger count of the waiting area of each floor; the preprocessing includes, in sequence, alignment operation based on a unified time reference, missing data filling operation based on linear interpolation, noise filtering operation based on a sliding window mean and numerical normalization operation based on minimum and maximum specification, and the output standardized scheduling input vector is spliced according to the order of the elevator state data, the passenger request data and the floor passenger flow density data.

[0016] Optionally, the step of calculating the skip step probability and outputting the variable step length value in the event-aware gate module specifically comprises:

[0017] reading the current time step input vector in the standardized scheduling input vector set in chronological order;

[0018] sequentially splicing the current time step input vector and the historical average waiting time indicator to generate a first spliced vector;

[0019] continuously splicing the first spliced vector and the historical average energy consumption indicator to generate an event-aware input vector;

[0020] inputting the event-aware input vector into the linear transformation unit in the event-aware gate, and the linear transformation unit performs matrix multiplication and bias addition operations on the event-aware input vector to output an intermediate calculation result vector;

[0021] inputting the intermediate calculation result vector into the activation function unit, and the activation function unit performs Sigmoid function transformation on the intermediate calculation result vector to obtain a normalized skip step probability value;

[0022] inputting the skip step probability value into the variable step length mapping unit, and mapping the skip step probability value to an integer variable step length value k through a preset interval mapping function, wherein the variable step length value k is limited between 0 and K, and K is a preset maximum skip step length.

[0023] Optionally, the step of inputting the variable step length value into the improved SkipRNN and controlling through the variable step length control unit in the variable step length control module specifically comprises:

[0024] reading the current time step variable step length value k, and inputting the variable step length value k into the variable step length control unit in the improved SkipRNN;

[0025] in the case that the variable step length value k is greater than zero, the variable step length control unit disables the hidden state update operation of the current time step, directly passes the hidden state of the previous time step as the hidden state of the current time step, writes the standardized scheduling input vector of the current time step into the skip input cache area, and then appends the standardized scheduling input vector of each time step to the skip input cache area in turn to form a skip input cache sequence with a length of k in the next k time steps;

[0026] when the variable step length value k is equal to zero, the variable step length control unit triggers the hidden state update operation, inputs the standardized scheduling input vector of the current time step and the hidden state of the previous time step into the hidden state update unit together, performs recursive calculation, and generates the updated hidden state of the current time step;

[0027] After the skip input buffer sequence length reaches the variable step length value k, the skip input buffer sequence is written into the skip attention summary unit, and the skip attention summary unit performs weight normalization operation based on the additive attention mechanism on each time step input vector in the skip input buffer sequence to generate a summary vector;

[0028] The summary vector is spliced with the normalized scheduling input vector corresponding to the time step of the summary vector as the composite input of the next effective hidden state update operation, used to compensate for the missing input information during the skip, and update the hidden state of the improved SkipRNN;

[0029] After the update is completed, the skip input buffer area is emptied, and the next round of skip control operation is waited.

[0030] Optionally, the hidden state update operation in the improved SkipRNN hidden state update module specifically includes:

[0031] When the variable step length control unit receives a signal that the variable step length value k is equal to zero, the normalized scheduling input vector of the current time step is read , and the hidden state of the previous time step is read ;

[0032] The normalized scheduling input vector and the hidden state are input into the state update unit of the improved SkipRNN, and the state update unit performs weighted summation, nonlinear activation and state transformation operations to generate the initial updated hidden state of the current time step ;

[0033] It is judged whether the generated summary vector Δs exists in the skip input buffer area, and when the summary vector Δs exists, the summary vector Δs is spliced with the current time step input vector to form a fusion input vector ;

[0034] The fusion input vector and the hidden state of the previous time step are input into the state update unit again to repeatedly perform the weighted summation, nonlinear activation and state transformation operations to overwrite the initial update result and generate the final fusion updated hidden state ;

[0035] The final hidden state is written into the hidden state register, and the skip input buffer area and the summary vector buffer area are emptied to prepare for the next time step skip control;

[0036] In the case where the variable step length value k is greater than zero, the hidden state update operation is skipped, and only the hidden state of the previous time step Directly pass to the current time step, as the current hidden state.

[0037] Optionally, the generation of the target floor instruction sequence in the linear readout layer module specifically includes:

[0038] Read the updated hidden state of the current time step output by the improved SkipRNN, and input the hidden state into the full connection transformation unit of the linear readout layer;

[0039] In the linear readout layer, perform a linear mapping operation on the hidden state vector, generate an unnormalized instruction vector by matrix multiplication with a weight matrix and adding a bias vector;

[0040] Perform activation function transformation on the unnormalized instruction vector to generate a normalized scheduling instruction vector;

[0041] Perform a maximum value selection operation on the normalized scheduling instruction vector to output the target floor number with the maximum activation value as the target scheduling floor of the current time step;

[0042] On the basis of the target scheduling floor, combine the current elevator running direction, door state and load state information to generate a target floor instruction sequence;

[0043] Write the target floor instruction sequence into the task queue of the group control scheduling controller.

[0044] Optionally, the generation of the multi-target reward value in the group control scheduling controller interaction module specifically includes:

[0045] Write the target floor instruction sequence into the task execution queue of the group control scheduling controller to trigger the controller to execute the elevator scheduling action;

[0046] After each scheduling action is completed, receive the actual execution result data returned by the group control scheduling controller;

[0047] According to the call time and elevator arrival time of each passenger request in the actual execution result data, calculate the corresponding waiting time, and average the waiting time of all passenger requests in the set sliding time window to obtain the real-time average waiting time index of the current time window;

[0048] According to the motor running power and running time length in the actual execution result data, calculate the cumulative energy consumption value in the current time window, and divide by the window duration to obtain the real-time average energy consumption index;

[0049] Calculate the difference between the real-time average waiting time index and the real-time average energy consumption index and the preset performance reference value to generate the waiting time deviation term and the energy consumption deviation term;

[0050] Perform a weighted sum operation on the latency bias term and the energy consumption bias term to generate a multi-objective reward value for the current time step.

[0051] Optionally, the closed-loop scheduling optimization process in the meta-reinforcement learning head module specifically includes:

[0052] Read the multi-objective reward value of the current time step, and form a state-action-reward triple with the event-aware input vector and the skip step probability value of the corresponding time step;

[0053] Input the state-action-reward triple into the policy update unit in the meta-reinforcement learning head, and calculate the gradient of the event-aware gate parameter according to the policy gradient optimization method;

[0054] Update the linear transformation weight matrix and bias vector in the event-aware gate using the gradient of the event-aware gate parameter to generate new event-aware gate parameters;

[0055] After the event-aware gate parameter is updated, the difference between the skip step probability value and the corresponding reward value is taken as the advantage signal and input to the variable step length mapping parameter adjustment unit;

[0056] The variable step length mapping parameter adjustment unit corrects the mapping function of the skip step probability to the step length interval according to the advantage signal, and adjusts the skip step length corresponding to each interval boundary value;

[0057] Write the updated event-aware gate parameter and variable step length mapping parameter into the parameter storage area, and return the event-aware gate to the step of calculating the skip step probability for skip step control calculation in the next time step;

[0058] After every preset number of time steps, perform a round of parameter synchronization operation to replace the parameters in the current event-aware gate and variable step length mapping module to complete a policy evolution period in the closed-loop scheduling optimization process.

[0059] The beneficial effects of the present application are:

[0060] (1) Improve the efficiency of elevator scheduling, dynamically adjust the model skip frequency by introducing an improved SkipRNN combined with an event-aware gate mechanism, reduce redundant calculations, improve the running efficiency and response speed of the model on edge devices, and meet the high timeliness requirements of real-time elevator scheduling control;

[0061] (2) Reduce energy consumption and user waiting time, simultaneously optimize the average energy consumption and average waiting time indicators using a multi-objective reward mechanism, dynamically balance the energy efficiency of elevator operation and passenger service quality, and have the dual effects of energy saving and emission reduction and user experience improvement in actual deployment;

[0062] (3) Realize the adaptive update of the scheduling strategy, continuously optimize the event perception gate parameters and the variable step length mapping parameters through the meta-reinforcement learning mechanism, realize the self-learning and adaptation of the scheduling strategy when facing different passenger flow modes and use scenarios, and have good generalization ability and application expansion;

[0063] (4) The system design emphasizes the low delay and high responsiveness requirements of scheduling decisions, deploys a lightweight standardized scheduling input vector generation module, a step control unit and a linear readout layer module on the elevator edge device, so that most of the decision-making processes can be completed on the local terminal, reduces the dependence on the central server, meets the principle of on-site processing close to the data source side of edge computing, and the group control scheduling controller interaction module can realize real-time two-way communication and feedback collection with the edge group control system, forming a typical edge computing intelligent closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0064] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application, and do not constitute a limitation of the application. In the drawings:

[0065] Fig. 1 A module structure diagram of an elevator real-time scheduling optimization system based on edge computing is provided for the application;

[0066] Fig. 2 An improved SkipRNN step mechanism structure diagram of an elevator real-time scheduling optimization system based on edge computing is provided for the application;

[0067] Fig. 3 A multi-objective reward optimization closed loop structure diagram of an elevator real-time scheduling optimization system based on edge computing is provided for the application. DETAILED DESCRIPTION

[0068] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0069] Reference Figs. 1-3 An elevator real-time scheduling optimization system based on edge computing, comprising:

[0070] A multi-source scheduling data acquisition and preprocessing module for acquiring elevator multi-source scheduling data and performing preprocessing to generate a standardized scheduling input vector set;

[0071] An event perception gate module for receiving the standardized scheduling input vector set, historical average waiting time indicators and historical average energy consumption indicators, inputting the event perception gate, calculating the step probability and outputting the variable step length value;

[0072] a variable step length control module for inputting a variable step length value into the improved SkipRNN, controlled by a variable step length control unit, keeping the previous hidden state when the variable step length value k is greater than zero, performing hidden state update when the variable step length value k is equal to zero, and simultaneously caching k input vectors to generate a hidden state cache sequence;

[0073] a skip attention pooling module for performing an attention pooling operation on the hidden state cache sequence to generate an abstract vector;

[0074] an improved SkipRNN hidden state update module for merging the abstract vector with a next time step standardized scheduling input vector to update the improved SkipRNN hidden state;

[0075] a linear readout layer module for inputting the updated hidden state into a linear readout layer to generate a target floor instruction sequence;

[0076] a group control scheduling controller interaction module for writing the target floor instruction sequence into a group control scheduling controller, receiving scheduling controller feedback execution results, calculating real-time average waiting time indicators and real-time average energy consumption indicators, and generating a multi-objective reward value;

[0077] a meta-reinforcement learning head module for inputting the multi-objective reward value into a policy optimization structure to update event perception gate parameters and variable step length mapping parameters, returning to the skip probability calculation step, and completing a closed-loop scheduling optimization process.

[0078] In this embodiment, the modules are implemented through the following methods:

[0079] Collecting elevator multi-source scheduling data, performing preprocessing to generate a standardized scheduling input vector set;

[0080] Concatenating the standardized scheduling input vector set with historical average waiting time indicators and historical average energy consumption indicators, inputting into an event perception gate, calculating a skip probability and outputting a variable step length value;

[0081] a variable step length control module for inputting a variable step length value into the improved SkipRNN, controlled by a variable step length control unit, keeping the previous hidden state when the variable step length value k is greater than zero, performing hidden state update when the variable step length value k is equal to zero, and simultaneously caching k input vectors to generate a hidden state cache sequence;

[0082] an improved SkipRNN hidden state update module for merging the abstract vector with a next time step standardized scheduling input vector to update the improved SkipRNN hidden state;

[0083] In this embodiment, the attention pooling operation is used to extract the summary of the skip input cache sequence in the improved SkipRNN to realize the weighted fusion compensation of the input information of the skip time step. The specific process is as follows: each standardized scheduling input vector in the skip input cache sequence is combined with the context query vector of the current time step, and then input into a scoring function to generate a corresponding attention score; all scores are normalized by a Softmax function to obtain the attention weight of each time step input vector; all skip input vectors are weighted and summed according to their attention weights to generate a summary vector.

[0084] The updated hidden state is input into a linear readout layer to generate a target floor instruction sequence.

[0085] The target floor instruction sequence is written into a group control scheduler, the execution result of the scheduler is received, the real-time average waiting time index and the real-time average energy consumption index are calculated, and a multi-objective reward value is generated.

[0086] The multi-objective reward value is input into the meta-reinforcement learning head to update the event perception gate parameters and the variable step length mapping parameters, and the updated event perception gate is returned to the step of calculating the skip probability to complete the closed-loop scheduling optimization process.

[0087] In this embodiment, the multi-source scheduling data acquisition and preprocessing module includes elevator multi-source scheduling data composed of elevator state data, passenger request data, and floor passenger flow density data; the elevator state data records the current floor, the running direction, the door state, and the load value in sequence; the passenger request data records the call floor, the target floor, and the request timestamp in sequence; the floor passenger flow density data is the real-time passenger count of the waiting area of each floor; the preprocessing includes sequentially performing alignment operation based on unified time reference, missing data filling operation based on linear interpolation, noise filtering operation based on sliding window mean, and numerical normalization operation based on minimum maximum specification, and the output standardized scheduling input vector is spliced according to the order of elevator state data, passenger request data, and floor passenger flow density data.

[0088] In this embodiment, the step of calculating the skip probability and outputting the variable step length value in the event perception gate module includes:

[0089] The current time step input vector in the set of standardized scheduling input vectors is read in time sequence;

[0090] The current time step input vector is sequentially spliced with the historical average waiting time index to generate a first spliced vector; the historical average waiting time index refers to the statistical result of the average user waiting time of all passenger requests in the elevator system within a set time sliding window, which is used to reflect the service response capability of the current elevator scheduling strategy within a period of time;

[0091] The first splicing vector is spliced with a historical average energy consumption index to generate an event-aware input vector; the historical average energy consumption index refers to the statistical result of the average user waiting time of the elevator system for all elevator requests within a set time sliding window, and is used to reflect the service response capability of the current elevator dispatching strategy within a period of time;

[0092] The event-aware input vector is input into a linear transformation unit in the event-aware gate, and the linear transformation unit performs matrix multiplication and bias addition operations on the event-aware input vector to output an intermediate calculation result vector;

[0093] The intermediate calculation result vector is input into an activation function unit, and the activation function unit performs Sigmoid function transformation on the intermediate calculation result vector to obtain a normalized skip probability value;

[0094] The skip probability value is input into a variable step length mapping unit, and the skip probability value is mapped to an integer variable step length value k through a preset interval mapping function, and the variable step length value k is limited between 0 and K, where K is a preset maximum skip step length.

[0095] In the embodiment, the step of inputting the variable step length value into the improved SkipRNN and controlling it through the variable step length control unit in the variable step length control module includes:

[0096] The current time step variable step length value k is read and input into the variable step length control unit in the improved SkipRNN;

[0097] When the variable step length value k is greater than zero, the variable step length control unit disables the hidden state update operation of the current time step, directly passes the hidden state of the previous time step as the hidden state of the current time step, writes the standardized dispatching input vector of the current time step into the skip input cache area, and then appends the standardized dispatching input vector of each time step to the skip input cache area in turn to form a skip input cache sequence with a length of k in the next k time steps;

[0098] When the variable step length value k is equal to zero, the variable step length control unit triggers the hidden state update operation, inputs the standardized dispatching input vector of the current time step and the hidden state of the previous time step into the hidden state update unit together, performs recursive calculation to generate the updated hidden state of the current time step;

[0099] After the length of the skip input cache sequence reaches the variable step length value k, the skip input cache sequence is written into the skip attention summary unit, and the skip attention summary unit performs weight normalization operation based on the additive attention mechanism on the input vectors of each time step in the skip input cache sequence to generate a summary vector;

[0100] In this embodiment, to solve the problem of missing input information during skipping, an additive attention mechanism-based skipping input summary method is introduced to extract key feature information from the skipping input cache sequence and generate a summary vector to compensate for the context information break caused by skipping hidden state update. The additive attention mechanism uses learnable parameters to weight and score the input vectors at each time step during skipping. Specifically, first, each input vector in the cache sequence and the query vector at the current time step are input into the additive attention scoring function together to calculate the attention score of each input vector. Then, all scores are normalized to generate normalized attention weights. Finally, the input vectors in the skipping cache sequence are weighted and summed according to the corresponding weights to generate a summary vector. The summary vector is a compact representation of the skipping interval information and is concatenated with the input vector at the current time step to update the hidden state.

[0101] Through the additive attention mechanism described above, the present application can realize dynamic extraction and fusion of input information under the skipping update strategy, retain the computational efficiency advantage brought by the skipping strategy, improve the expression ability of the model for key input fragments, and enhance the accuracy of the scheduling strategy and the robustness of the system.

[0102] The summary vector and the normalized scheduling input vector corresponding to the summary vector are concatenated as the composite input for the next effective hidden state update operation to compensate for the missing input information during skipping and update the improved SkipRNN hidden state.

[0103] After the update is completed, the skipping input cache area is emptied, and the next round of skipping control operation is waited.

[0104] In this embodiment, the improved SkipRNN described in the present application introduces three structural innovations of variable step length control, skipping input caching, and attention summary compensation in the skipping mechanism compared to the original SkipRNN. The original SkipRNN only determines whether to skip the current update based on the current input and hidden state through fixed gating logic, while the present application generates a dynamically variable skipping step length k through the event-aware gate, caches the continuous input sequence during skipping, and performs summary fusion on the cache sequence through the attention mechanism at the effective time step. This reduces the computational load while maintaining the ability to extract key information, improving the adaptability and decision-making accuracy of the model to complex dynamic scheduling scenarios.

[0105] In this embodiment, the hidden state update operation in the improved SkipRNN hidden state update module specifically includes:

[0106] When the variable step length control unit receives a signal that the variable step length value k is equal to zero, the normalized scheduling input vector at the current time step is read and the hidden state of the previous time step is also read ;

[0107] The standardized scheduling input vector and the hidden state The state update unit of the improved SkipRNN performs weighted summation, nonlinear activation and state transformation operations to generate the initial updated hidden state of the current time step ;

[0108] Determine whether the generated summary vector Δs exists in the skip input buffer, and when the summary vector Δs exists, concatenate the summary vector Δs and the current time step input vector to form a fusion input vector ;

[0109] The fusion input vector and the hidden state of the previous time step are input into the state update unit again, and the weighted summation, nonlinear activation and state transformation operations are repeatedly performed to overwrite the initial update result and generate the final updated hidden state of the fusion ;

[0110] The final hidden state is written into the hidden state register, and the skip input buffer and the summary vector buffer are emptied to prepare for the next time step skip control;

[0111] In the embodiment, the hidden state register is used to store the hidden state generated by the improved SkipRNN at each time step, and is a key intermediate structure for implementing skip control and recursive state preservation. The register is used to write the latest generated hidden state after each hidden state update operation for the next time step to call, and keeps the hidden state of the previous time step unchanged during the skip, ensuring that the hidden state has sustainable tracking ability under skip control;

[0112] Specifically, when the variable step value k output by the event-aware gate is equal to zero, the current time step hidden state generated by the state update unit will be immediately written into the hidden state register as the starting state for the next time step state calculation after triggering the hidden state update operation. When the variable step value k is greater than zero, the state update operation is skipped, and the hidden state register does not occur at this time. The system directly reads the hidden state saved in the register of the previous time step for transmission, avoiding unnecessary calculation. The register can be mapped to a high-priority buffer in the memory address space in the edge computing device, with low read-write delay and high access frequency, which can support the high-frequency state calling demand in the fast inference scenario;

[0113] By setting the hidden state register, the application realizes the breakpoint reservation and lossless continuation of the recurrent neural network state under the skip mechanism, can support the cooperative execution of skip and update operations, and enhances the engineering deployment feasibility of the model structure and the stability of the scheduling response.

[0114] In the case where the variable step size k is greater than zero, the hidden state update operation is skipped, and only the last time step hidden state is directly passed to the current time step as the current hidden state.

[0115] In the embodiment, the generation of the target floor instruction sequence in the linear readout layer module specifically includes:

[0116] The current time step updated hidden state output by the improved SkipRNN is read, and the hidden state is input into the full connection variable unit of the linear readout layer.

[0117] In the linear readout layer, a linear mapping operation is performed on the hidden state vector, a weight matrix is multiplied by matrix multiplication, and a bias vector is added to generate an unnormalized instruction vector; the weight matrix is a trainable two-dimensional real matrix, denoted as , with dimensions , where L is the total number of floors that the elevator control system can reach, and H is the dimension of the hidden state vector. In the linear readout process, the hidden state vector is input, first multiplied by the weight matrix , and then added to the bias vector to form an unnormalized instruction vector. Each element in the unnormalized instruction vector corresponds to a candidate target floor scheduling score. After passing through an activation function such as the Softmax function, a normalized scheduling probability distribution is formed, and finally the scheduling target floor at the current time step is obtained through the maximum activation value index.

[0118] An activation function transformation is performed on the unnormalized instruction vector to generate a normalized scheduling instruction vector.

[0119] A maximum value selection operation is performed on the normalized scheduling instruction vector to output the target floor number with the maximum activation value as the target scheduling floor at the current time step.

[0120] On the basis of the target scheduling floor, the current elevator running direction, door state and load state information are combined to generate a target floor instruction sequence.

[0121] The target floor instruction sequence is written into the task queue of the group control scheduling controller.

[0122] Specifically, in a 20-story office building, the elevator dispatch system is deployed with the improved SkipRNN model and linear readout layer of the present application. The system sets the hidden state vector dimension to 64, the total number of floors that the elevator can reach is 20, and the weight matrix in the linear readout layer is dimension , the bias vector dimension .

[0123] When the model runs to a certain time step t, the current hidden state output by SkipRNN is a real number vector with a dimension of 64 ;

[0124] The system inputs to the linear readout layer and performs the following operations:

[0125] Calculate the score vector: ; where , represents a 20-dimensional real number vector, representing the dispatch score for each target floor.

[0126] Input into the Softmax activation function to generate a normalized dispatch probability distribution;

[0127] Assuming the 12th floor has the highest score, i.e.: ;

[0128] Accordingly, the system determines that under the current dispatch state, the elevator should be dispatched to the 12th floor first, and then generates a complete dispatch instruction sequence as follows by combining the current running direction of the elevator, which is "up", the door state, which is "closed", and the load state, which is "non-full load":

[0129] Target floor: 12th floor;

[0130] Running direction: up;

[0131] Door operation: keep closed;

[0132] Task queue action: insert priority dispatch bit;

[0133] The instruction sequence is immediately written to the group control dispatch controller and triggers the elevator control logic to perform the corresponding actions. The entire reasoning process is completed in the edge node, with an average delay of less than 150ms, meeting the real-time requirements.

[0134] In this embodiment, the generation of the multi-target reward value in the group control dispatch controller interaction module specifically includes:

[0135] write the target floor instruction sequence into the task execution queue of the group control scheduling controller, trigger the controller to execute the elevator scheduling action;

[0136] After each scheduling action is completed, the actual execution result data returned by the group control scheduling controller is received, and the result data includes the execution start time, arrival time, number of passengers taking the elevator, and corresponding floor record of the scheduling instruction, motor running time length and power record;

[0137] According to the call time and elevator arrival time of each passenger request in the actual execution result data, the corresponding waiting time is calculated, and the waiting times of all passenger requests are averaged in the set sliding time window to obtain the real-time average waiting time index of the current time window;

[0138] According to the motor running power and running time length in the actual execution result data, the cumulative energy consumption value in the current time window is calculated, and then divided by the window time length to obtain the real-time average energy consumption index;

[0139] The real-time average waiting time index and the real-time average energy consumption index are respectively subtracted from the preset performance reference value to generate a waiting time deviation term and an energy consumption deviation term;

[0140] The waiting time deviation term and the energy consumption deviation term are subjected to weighted summation operation to generate a multi-objective reward value of the current time step.

[0141] In this embodiment, in order to realize the dynamic trade-off between service efficiency and energy consumption level of the elevator scheduling strategy, a multi-objective reward value generated based on real-time performance feedback is set, the multi-objective reward value is used as an input signal of the meta-reinforcement learning module to guide the update of the event perception gate and the variable step length mapping parameter, and the multi-objective reward value is composed of two parts: a waiting time deviation term and an energy consumption deviation term, which respectively measure the relative performance of the scheduling response performance and the energy efficiency level in the current time window.

[0142] Specifically, the system receives the execution result returned by the group control scheduling controller at each time step, calculates the average waiting time of all valid elevator requests in the current sliding time window , and subtracts the set target waiting time to obtain the waiting time deviation term ;

[0143] According to the motor running power and running time length in the scheduling execution process, the average energy consumption per unit time is calculated , and the difference between the target energy consumption level is obtained to obtain the energy consumption deviation term: ;

[0144] ​The two deviation terms will be multiplied by adjustable weight coefficients With Finally, the multi-objective reward value of the current time step is formed The calculation formula is ; wherein With The weights of the service response priority and the energy saving target are denoted, and satisfy The dynamic configuration can be performed according to the building dispatching strategy requirement.

[0145] The multi-objective reward value can reflect the contribution of the dispatching behavior to the user experience and the energy consumption optimization, has real-time and adjustability, and is an important quantitative index for supporting the adaptive strategy evolution of the closed-loop dispatching optimization system of the application. Through the introduction of the reward mechanism, the system can automatically strengthen the beneficial strategy and suppress the inefficient behavior according to the current dispatching effect, and improve the intelligence and stability of the overall dispatching system.

[0146] In the embodiment, the closed-loop dispatching optimization process in the meta-reinforcement learning head module specifically includes:

[0147] The multi-objective reward value of the current time step is read, and the state-action-reward triple is composed of the multi-objective reward value, the event perception input vector of the corresponding time step and the jump probability value.

[0148] The state-action-reward triple is input into the policy updating unit in the meta-reinforcement learning head, and the gradient of the event perception gate parameter is calculated according to the policy gradient optimization method.

[0149] The linear transformation weight matrix and the bias vector in the event perception gate are updated using the gradient of the event perception gate parameter, and a new event perception gate parameter is generated.

[0150] After the event perception gate parameter is updated, the difference between the jump probability value and the corresponding reward value is taken as the advantage signal, and is input into the variable step length mapping parameter adjustment unit.

[0151] The variable step length mapping parameter adjustment unit corrects the mapping function of the jump probability to the step length interval according to the advantage signal, and adjusts the jump length corresponding to each interval boundary value.

[0152] In the embodiment, in order to realize the dynamic matching of the jump length and the current dispatching performance, a variable step length mapping function adjustment mechanism driven by the advantage signal is designed, which acts on the probability-step length mapping process in the jump control process, and the core is a group of learnable interval boundary parameters, which is used to map the normalized jump probability value output by the event perception gate Into the discrete integer type jump step length value , wherein is the maximum jump step length.

[0153] Initially, the interval boundaries are defined as a set of monotonically increasing real numbers:

[0154] ;

[0155] The mapping function is defined as: if the step probability falls into the interval , the corresponding step length ;

[0156] In each scheduling feedback period, the system obtains the multi-objective reward value obtained at the current time step, and the actual step length corresponding to the step probability , and the difference between the policy expected reward and the actual step length ;

[0157] ;

[0158] wherein represents the historical average reward corresponding to the step length value ;

[0159] When the advantage signal , it means that the performance brought by the current step length is better than the average level, and the system will increase the mapping probability range of the interval through the fine-tuning mechanism; otherwise, if , the interval range is reduced, and the policy adjusts the step length by applying a perturbation to the interval boundary :

[0160] ;

[0161] wherein is the learning rate, used to control the boundary adjustment amplitude;

[0162] The above interval boundary is adjusted and normalized to ensure that all boundaries meet the conditions of monotonic increase and fixed start and end range, maintaining the integrity and effectiveness of the probability mapping function within the definition domain; this mechanism can realize the adaptive evolution of the step length distribution to the scheduling environment, automatically adjusting the step strategy in different scenarios, and improving the dynamic response ability and long-term policy optimization level of the system performance;

[0163] The updated event-aware gate parameters and variable step length mapping parameters are written into the parameter storage area, and the event-aware gate is returned to the step of calculating the step probability for the next time step of step control calculation;

[0164] After a predetermined number of time steps, a round of parameter synchronization operation is performed to replace the parameters in the current event-aware gate and variable step length mapping module, to complete the policy evolution period in a closed-loop scheduling optimization process.

[0165] Example 1: To verify the feasibility of the application in practice, the application is applied to the elevator group control system of a certain high-rise building for scheduling optimization experiment. The high-rise building has five elevators covering thirty floors. The system collects real-time multi-source scheduling data of multiple elevators through edge collection devices, including elevator running state, floor button request, up and down passenger number change, door switch state, etc. To ensure data quality, the collected data is first aligned with time stamp, missing data is completed, and normalized processing is performed to construct a standardized scheduling input vector set, which is input into the event-aware gate module.

[0166] In actual operation, the event-aware gate receives the input vector and combines the historical average waiting time index and the historical average energy consumption index to dynamically calculate the skip probability and output the variable step value. The skip value is used to control the state update rhythm of the improved SkipRNN, thereby skipping redundant state transitions and improving computational efficiency.

[0167] Experiments show that when the variable step length dynamically adjusts between 1~4, the elevator state prediction accuracy improves by about 7.3%, and the redundant calculation decreases by 34.8%. In the skip state, the system performs attention pooling operation on the cached multi-segment input vector to compensate for the input information during the skip period and enhance the hidden state expression ability.

[0168] The updated hidden state generates floor instructions through the linear readout layer and is sent to the group control scheduler. The system quantitatively analyzes the feedback information of the controller and the scheduling effect at the current time (real-time average waiting time and energy consumption), outputs a multi-objective reward value, and the reward value enters the meta-reinforcement learning head module to drive the optimization update of the event-aware gate parameters and the variable step length mapping parameters, realizing closed-loop adaptive scheduling optimization.

[0169] The experimental scenario of the application particularly investigates the performance comparison between the morning and evening peak periods (8:00-9:00 and 5:00-6:00) and the ordinary periods (10:00-11:00 and 14:00-15:00). During the morning and evening peak periods, the traditional fixed step length RNN scheduling system has frequent state updates and response lags, resulting in an average waiting time of 56.8 seconds and an increase in passenger complaint rate. After using the application, the average waiting time decreases to 39.2 seconds, the energy saving rate increases to 17.6%, and the overall operation stability and user satisfaction are significantly improved.

[0170] The following table shows the performance comparison data of the traditional method and the application method in different time periods, including average waiting time, average energy consumption, passenger satisfaction score and other dimensions.

[0171] Table 1: Comparison of performance evaluation table of elevator scheduling optimization system

[0172]

[0173] The above Table 1 shows the comparison between the traditional fixed step RNN scheduling method and the elevator scheduling optimization system based on the improved SkipRNN and meta-reinforcement learning mechanism proposed by the present application in terms of multiple key performance indicators, including average waiting time, average single energy consumption, step rate, and passenger satisfaction score, within different time periods.

[0174] During the early peak period of 8:00, the average waiting time of the traditional method is 56.8 seconds, the energy consumption is 14.7 kJ, the step rate is 0, and the satisfaction score is only 2.8, showing that the traditional method has a response lag and insufficient energy utilization efficiency under high load conditions. In contrast, the method of the present application reduces the average waiting time to 39.2 seconds, the energy consumption to 12.1 kJ, while achieving a step rate of 48.5%, and the satisfaction score is improved to 4.3, indicating that the system effectively alleviates congestion during peak periods, improves operating efficiency and user experience.

[0175] During the relatively stable midday period of 11:00, the method of the present application also exhibits superior performance; the average waiting time is reduced from 34.7 seconds of the traditional method to 24.6 seconds, the energy consumption is reduced from 11.3 kJ to 9.8 kJ, the step rate is increased to 52.1%, and the satisfaction score is increased to 4.6, indicating that even during off-peak periods, the system can reduce redundant calculations through the step mechanism, achieving energy efficiency optimization and service improvement.

[0176] During the late peak period of 17:30, the traditional method again exhibits a long waiting time (58.3 seconds) and high energy consumption (15.0 kJ), while the present application reduces the waiting time to 40.5 seconds and the energy consumption to 12.4 kJ, while maintaining a step rate of 46.9% and a satisfaction score of 4.2, indicating that the system has stable scheduling performance and strong resource control capability under high load scenarios.

[0177] In summary, the step mechanism and reinforcement learning parameter adaptive optimization method proposed by the present application exhibit significantly better scheduling performance than the traditional method in multiple time periods, allowing flexible adjustment of the calculation strategy under different load conditions, improving scheduling response efficiency and reducing system energy consumption, enhancing user experience and system economy.

[0178] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes within the technical scope disclosed by the present application according to the technical solutions and inventive concepts of the present application, which should be covered within the protection scope of the present application.

Claims

1. A real-time elevator scheduling and optimization system based on edge computing, characterized in that, include: The multi-source scheduling data acquisition and preprocessing module is used to acquire multi-source elevator scheduling data and perform preprocessing to generate a standardized set of scheduling input vectors. The event-aware gate module is used to receive a set of standardized scheduling input vectors, historical average waiting time indicators, and historical average energy consumption indicators, input them into the event-aware gate, calculate the skipping probability, and output a variable step size value. The variable step size control module is used to input the variable step size value into the improved SkipRNN, and control it through the variable step size control unit to generate the hidden state buffer sequence. The skip-step attention pooling module is used to perform attention pooling operations on the hidden state cache sequence to generate a summary vector; An improved SkipRNN hidden state update module is used to merge the summary vector with the normalized scheduling input vector of the next time step and update the hidden state of the improved SkipRNN. The linear readout layer module is used to input the updated hidden state into the linear readout layer to generate the target floor instruction sequence; The group control and dispatch controller interaction module is used to write the target floor instruction sequence into the group control and dispatch controller, receive the execution results fed back by the dispatch controller, calculate the real-time average waiting time index and the real-time average energy consumption index, and generate multi-objective reward values. The meta-reinforcement learning head module is used to input multi-objective reward values ​​into the policy optimization structure, update the event-aware gate parameters and variable step size mapping parameters, return to the skip step probability calculation step, and complete the closed-loop scheduling optimization process.

2. The elevator real-time scheduling and optimization system based on edge computing according to claim 1, characterized in that, The multi-source scheduling data acquisition and preprocessing module consists of elevator status data, passenger request data, and floor passenger density data. The elevator status data sequentially records the current floor, direction of travel, door status, and load value. The passenger request data sequentially records the calling floor, target floor, and request timestamp. The floor passenger density data is the real-time passenger count in the waiting area of ​​each floor. The preprocessing includes sequentially performing alignment operations based on a unified time reference, missing data imputation operations based on linear interpolation, noise filtering operations based on sliding window mean, and numerical normalization operations based on min-maximum norm. The output standardized scheduling input vector is concatenated according to the order of elevator status data, passenger request data, and floor passenger density data.

3. The elevator real-time scheduling and optimization system based on edge computing according to claim 2, characterized in that, The specific steps in the event-aware gate module for calculating the skip probability and outputting the variable step size value include: Read the current time step input vector from the standardized scheduling input vector set step by step in chronological order; The current time step input vector is sequentially concatenated with the historical average waiting time index to generate the first concatenated vector. The first concatenated vector is further concatenated with the historical average energy consumption index to generate an event-aware input vector. The event-aware input vector is input into the linear transformation unit in the event-aware gate. The linear transformation unit performs matrix multiplication and bias addition operations on the event-aware input vector and outputs the intermediate calculation result vector. The intermediate calculation result vector is input into the activation function unit, and the activation function unit performs a Sigmoid function transformation on the intermediate calculation result vector to obtain the normalized skip probability value. The skip probability value is input into the variable step size mapping unit, and the skip probability value is mapped to an integer variable step size value k through a preset interval mapping function. The variable step size value k is limited to between 0 and K, where K is the preset maximum skip step size.

4. The elevator real-time scheduling and optimization system based on edge computing according to claim 3, characterized in that, The steps in the variable step size control module that input the variable step size value into the improved SkipRNN and control it through the variable step size control unit specifically include: Read the variable step size value k of the current time step and input the variable step size value k into the variable step size control unit in the improved SkipRNN; When the variable step size value k is greater than zero, the variable step size control unit disables the hidden state update operation of the current time step, directly passes the hidden state of the previous time step as the hidden state of the current time step, writes the normalized scheduling input vector of the current time step into the skip input buffer, and then appends the normalized scheduling input vector of each time step to the skip input buffer in the next k time steps, forming a skip input buffer sequence of length k. When the variable step size value k equals zero, the variable step size control unit triggers the hidden state update operation, inputs the normalized scheduling input vector of the current time step and the hidden state of the previous time step into the hidden state update unit, performs recursive calculation, and generates the updated hidden state of the current time step. After the length of the skip input buffer sequence reaches the variable step size value k, the skip input buffer sequence is written into the skip attention summarization unit. The skip attention summarization unit performs a weight normalization operation based on the additive attention mechanism on the input vector of each time step in the skip input buffer sequence to generate a summary vector. The digest vector is concatenated with the normalized scheduling input vector of the time step corresponding to the digest vector, and used as the composite input for the next effective hidden state update operation. This is used to compensate for the lack of input information during the skip step and update the hidden state of the improved SkipRNN. After the update is complete, clear the step input buffer and wait for the next round of step control operation.

5. The elevator real-time scheduling and optimization system based on edge computing according to claim 4, characterized in that, The hidden state update operation in the improved SkipRNN hidden state update module specifically includes: When the variable step size control unit receives a signal that the variable step size value k is equal to zero, it reads the normalized scheduling input vector of the current time step. Simultaneously read the hidden state from the previous time step. ; Standardize the scheduling input vector With hidden state Input the state update unit of the improved SkipRNN. The state update unit performs weighted summation, nonlinear activation, and state transition operations to generate the initial updated hidden state for the current time step. ; Determine if a generated summary vector Δs exists in the jump step input buffer. If the summary vector Δs exists, combine it with the current time step input vector. The vectors are concatenated to form a fused input vector. ; fuse input vectors Hidden state from the previous time step Input the state update unit again, repeat the weighted summation, nonlinear activation, and state transition operations, overwrite the initial update result, and generate the final hidden state after fusion update. ; The final hidden state Write to the hidden state register, clear the jump input buffer and the digest vector buffer, and prepare for the jump control of the next time step; If the variable step size k is greater than zero, skip the hidden state update operation and only update the hidden state of the previous time step. It is passed directly to the current time step and used as the current hidden state.

6. The elevator real-time scheduling and optimization system based on edge computing according to claim 5, characterized in that, The generation of the target floor instruction sequence in the linear readout layer module specifically includes: Read the hidden state updated at the current time step by the output of the improved SkipRNN, and input the hidden state into the fully connected transformation unit of the linear readout layer; In the linear readout layer, a linear mapping operation is performed on the hidden state vector. By performing matrix multiplication with the weight matrix and adding the bias vector, an unnormalized instruction vector is generated. The activation function transformation is performed on the unnormalized instruction vector to generate a normalized scheduling instruction vector; Perform a maximum value selection operation on the normalized scheduling instruction vector and output the target floor number with the maximum activation value as the target scheduling floor for the current time step. Based on the target floor, and combined with the current elevator running direction, door status and load status information, a target floor instruction sequence is generated; Write the target floor instruction sequence into the task queue of the group control and scheduling controller.

7. The elevator real-time scheduling and optimization system based on edge computing according to claim 6, characterized in that, The generation of multi-objective reward values ​​in the group control scheduling controller interaction module specifically includes: Write the target floor instruction sequence into the task execution queue of the group control and scheduling controller, triggering the controller to execute elevator scheduling actions; After each scheduling action is completed, receive the actual execution result data returned by the group control scheduling controller; Based on the actual execution results data, the corresponding waiting time is calculated for each passenger's requested elevator call time and elevator arrival time. The waiting time of all passengers' requests is averaged within the set sliding time window to obtain the real-time average waiting time index for the current time window. Based on the motor operating power and operating time in the actual execution results data, calculate the cumulative energy consumption value within the current time window, divide it by the window duration, and obtain the real-time average energy consumption index. The difference between the real-time average waiting time index and the real-time average energy consumption index and the preset performance reference value is calculated to generate waiting time deviation item and energy consumption deviation item. Perform a weighted summation operation on the waiting time deviation term and the energy consumption deviation term to generate the multi-objective reward value for the current time step.

8. The elevator real-time scheduling and optimization system based on edge computing according to claim 7, characterized in that, The closed-loop scheduling optimization process in the meta-reinforcement learning head module specifically includes: Read the multi-objective reward value of the current time step, and combine the multi-objective reward value with the event-aware input vector and the step-skip probability value of the corresponding time step to form a state-action-reward triplet; The state-action-reward triplet is input into the policy update unit in the meta-reinforcement learning head, and the gradient of the event-aware gate parameters is calculated according to the policy gradient optimization method. The gradients of the event-aware gate parameters are used to update the linear transformation weight matrix and bias vector in the event-aware gate, generating new event-aware gate parameters. After the event-aware gate parameters are updated, the difference between the skip probability value and the corresponding reward value is used as an advantage signal and input to the variable step size mapping parameter adjustment unit. The variable step size mapping parameter adjustment unit corrects the mapping function from the step probability to the step size interval based on the dominant signal, and adjusts the step length corresponding to the boundary value of each interval. Write the updated event-aware gate parameters and variable step size mapping parameters into the parameter storage area, and return the event-aware gate to the step of calculating the skip probability for use in the skip control calculation of the next time step; After each preset number of time steps, a round of parameter synchronization operation is performed to replace the parameters in the current event-aware gate and variable step size mapping module, so as to complete the strategy evolution cycle in a closed-loop scheduling optimization process.

Images