Control algorithm for comprehensive energy storage device of elevator

Abstract

The invention provides a control algorithm for a comprehensive energy storage device of an elevator, and belongs to the technical field of elevator energy storage control, and the method comprises the following steps: 1, collecting local data, and extracting federal features; 2, training a local meta-strategy, and uploading gradient homomorphic encryption; 3, global model aggregation and edge side model issuing; 4, digital twinborn body prediction and adversarial disturbance enhancement are carried out; and step 5, Shapley value distribution and cooperative control of a plurality of stored energy. The comprehensive energy consumption is reduced by 22%-30%, compared with traditional group control and single energy storage control, the equivalent cycle index of the lithium battery is reduced by 40% or above, multi-elevator cluster strategy sharing is achieved through a federated learning framework, the energy storage distribution strategy is optimized in real time in combination with the digital twinborn, and the cross-scene adaptability is improved while data privacy is guaranteed. A multi-agent game mechanism is innovatively introduced, the problem of energy storage resource competition is solved, and the method is suitable for complex elevator network systems such as super high-rise buildings and subway hubs.

Description

Technical Field

[0001] The present invention relates to the technical field of elevator energy storage control, and in particular to a control algorithm for an elevator integrated energy storage device. Background Art

[0002] An elevator is an unbalanced system consisting of a car and a counterweight. This system is characterized by potential energy loads, and the direction of the load torque varies with the car load. When the elevator is unloaded (car weight is less than the counterweight) ascending, fully loaded (car weight is greater than the counterweight) descending, or decelerating before leveling, the drive motor is in a regenerative braking state, converting mechanical energy into electrical energy. In traditional elevators, this process often uses high-power resistors to dissipate this energy. This not only wastes energy but also generates significant heat, causing the machine room to overheat. To ensure the proper functioning of electronic components, the machine room is often equipped with cooling equipment such as fans and air conditioners, resulting in secondary energy waste. With the continuous improvement of national energy conservation and emission reduction policies and the strengthening of the legal system, elevators with energy regeneration functions have quietly become a standard feature of major elevator manufacturers.

[0003] In existing elevators equipped with energy feedback, when the elevator is fully loaded and descending, or unloaded and ascending, and the drive motor is in the generating state, the variable frequency control system will feed the energy generated by the elevator system back to the power grid in real time. However, the output power of existing elevators equipped with energy feedback fluctuates greatly during acceleration and deceleration, which has a significant impact on the power grid. Furthermore, the energy feedback technology in existing elevators is real-time feedback, which simultaneously inputs harmonics into the power grid, causing grid pollution. Furthermore, most existing electricity meters use one-way billing, and the energy fed back from elevator operation does not reverse the user's electricity consumption, resulting in no actual reduction in electricity bills. For users, the energy fed back by the elevator does not generate any real economic value. How to address the impact of elevator energy feedback on the power grid, avoid polluting the grid, and simultaneously reduce the power consumption of users' electricity meters and save on electricity bills has become a technical issue that urgently needs to be addressed in the development of elevators equipped with energy feedback.

[0004] Therefore, it is necessary to design an elevator comprehensive energy storage device that can store energy in real time, save electricity, and not transmit electricity to the power grid to pollute the power grid. The capacity configuration of traditional elevator energy storage devices (such as supercapacitors and lithium batteries) usually relies on empirical formulas or static optimization models, and has the following problems: (1) Poor scene adaptability: fixed capacity is difficult to cope with dynamically changing passenger flow, floor load and operating time requirements. (2) Insufficient multi-objective trade-offs: It is necessary to optimize energy consumption, cost, response speed and other goals at the same time, and traditional methods are prone to fall into single-objective local optimality. (3) Inefficient data utilization: Historical operation data and real-time interactive information are not fully exploited, and the adaptive ability is weak. Therefore, it is necessary to design a control algorithm for elevator comprehensive energy storage devices. Summary of the Invention

[0005] The purpose of the present invention is to provide a control algorithm for an elevator integrated energy storage device to solve the technical problems existing in the background technology.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] The control algorithm of the elevator integrated energy storage device comprises the following steps:

[0008] Step 1: Local data collection and federated feature extraction;

[0009] Step 2: Local meta-strategy training and gradient homomorphic encryption upload;

[0010] Step 3: Global model aggregation and edge model delivery;

[0011] Step 4: Digital twin prediction, enhanced against disturbances;

[0012] Step 5: Shapley value allocation and coordinated control of several energy storages.

[0013] Furthermore, the specific process of step 1 is:

[0014] Vibration accelerometers are installed at the car floor and guide rail connections, a gyroscope is installed on the car roof, a current sensor is installed on the main motor power supply line, a voltage sensor is installed at the output end of the energy storage device, and temperature sensors are installed on the lithium battery pack and supercapacitor module. A crowd counting camera is installed at the elevator hall entrance to collect data in real time. Kalman filtering is used to jointly calibrate the vibration accelerometer and gyroscope data to eliminate installation deviations. During normal operation, 10Hz sampling is used. When abnormal vibration is detected, that is, when the amplitude exceeds the threshold, the sampling is switched to 500Hz to reduce storage overhead.

[0015] Data preprocessing and feature extraction: vibration signals are denoised using wavelet thresholding. Image data is segmented into pedestrian areas in real time using YOLOv5 to remove background noise. Time series features are generated, including statistical features, frequency domain features, and nonlinear features. The mean, variance, and peak-to-peak value within the sliding window are statistical features. The FFT spectrum peak position and wavelet packet energy entropy are frequency domain features. Nonlinear features include approximate entropy and sample entropy. Spatiotemporal feature alignment and time alignment are performed. Based on the dynamic time warping algorithm, the vibration signal time series of different elevators are synchronized, spatial alignment is performed, elevator topology relationships are constructed through graph convolutional networks, and multi-node features are integrated.

[0016] Privacy protection and data desensitization: embed a privacy protection mechanism in the local data collection stage and use local sensitive hashing to map high-dimensional features to low-dimensional space, ensuring that the same data hashes collide and different data hashes are separated.

[0017] Furthermore, the specific process of step 2 is as follows: each elevator node is treated as an independent task. Task characteristics include floor distribution, passenger flow pattern, and energy storage type. An actor-critic architecture is adopted. The actor outputs the energy storage allocation action, and the critic evaluates the action value. The morning and evening peak scenarios are prioritized, and the training weight of the night maintenance scenario is gradually increased. The FGSM attack is introduced to generate adversarial samples to enhance the robustness of the strategy.

[0018] Homomorphic encryption uses CKKS encryption to convert floating-point gradients into integers. The error compensation mechanism ensures accuracy. The server performs additive homomorphic operations on the encrypted gradients. Threshold encryption is introduced in the gradient aggregation stage to prevent single-point key leakage. The ciphertext is uploaded in blocks and in parallel, supporting breakpoint resumption.

[0019] Furthermore, the specific process of step 3 is:

[0020] Adaptive weighted aggregation dynamically adjusts the aggregation weight based on the amount of elevator node data and contribution:

[0021]

[0022] Shapleyk is the marginal contribution of the kth node to the global model. Model compression and communication, gradient sparsification, uploading the first 5% non-zero gradient parameters, compression rate > 90%, compressing 32-bit floating-point parameters to 8-bit integers, removing redundant parameters with gradient absolute value < ε, ε = 0.01, dividing the model into blocks and transmitting them in parallel via the QUIC protocol, supporting breakpoint resumption, using threshold encryption in the gradient aggregation stage to prevent single-point key leakage, and adding Gaussian noise when updating the global model;

[0023] Model adaptation and lightweighting: knowledge of the global model is transferred to a lightweight student model, reducing the number of student model parameters by 70% and increasing inference speed by 2 times. Low-precision matrix multiplication replaces floating-point operations. A dynamic hot update mechanism and differential update only send the changed model parameters, achieving a compression rate of >80%. Double-buffered loading: first run the old model, then download the new model, verify its integrity, and then switch to double buffering.

[0024] Edge resource scheduling, heterogeneous computing optimization, high computing tasks are offloaded to the edge GPU, real-time control tasks are executed in hard real-time on the CPU side, and the Zstd algorithm is used to compress model weights, reducing memory usage by 50%.

[0025] Furthermore, the specific process of step 4 is as follows: constructing a dynamic digital twin to simulate the physical behavior and degradation process of the elevator energy storage system in real time, improving the model's resistance to noise and attacks through adversarial disturbance training to prevent strategy failure, and dynamically adjusting the control strategy based on the prediction results of the digital twin to achieve virtual-physical interactive closed-loop optimization;

[0026] Dynamic twin modeling, physics engine, using COMSOL Multiphysics to simulate the electrochemical degradation process of energy storage units, Q is the remaining capacity, η is the Coulomb efficiency, and k(T) is the temperature dependence coefficient. Based on OpenFOAM, the system simulates elevator mechanical vibration and structural fatigue. An LSTM-Transformer hybrid network is used to predict the load for the next 10 seconds. Online prediction and correction are performed. When the prediction error exceeds a threshold, online fine-tuning is triggered. This method combines the output of the physical engine with data-driven predictions through Bayesian weighted fusion.

[0027] Adversarial perturbation generation: Generate perturbations that maximize model loss, inject the vibration spectrum of the digital twin into real sensor data, test the robustness of the strategy, generate an adversarial loss function, gradually increase the perturbation intensity to avoid premature model collapse, and finally perform robustness evaluation indicators and adversarial sample detection.

[0028] Furthermore, the specific process of step 5 is as follows: quantifying the marginal contributions of supercapacitors, lithium batteries, and flywheels through Shapley values ​​to solve the tragedy of the commons problem, dynamically adjusting the energy storage coordination strategy based on real-time scenarios, balancing energy recovery efficiency, life loss, and cost, and calculating Shapley values ​​in real time to update control instructions and adapt to the instantaneous load changes of elevators;

[0029] Characteristic function definition, benefit indicator, R(S) = ω1·ΔEregen + ω2·ηlife - ω3·ΔPgrid, where ΔEregen is the energy recovery increment, ηlife is the lifetime efficiency, and ΔPgrid is the grid fluctuation penalty. Multi-time scale calculation: in the short term of 1 second, supercapacitors dominate the instantaneous power response; in the medium term of 1 minute, lithium batteries bear the base load; in the long term of >5 minutes, the flywheel covers the continuous power shortage scenario;

[0030] Incremental benefit comparison: When energy storage unit i is added to subset S, the difference between the energy recovery increment and the life loss increment is calculated; ΔRi = ΔEregen(S∪{i})-ΔEregen(S)-λ·(ΔQloss(S∪{i})-ΔQloss(S)). Scenario-adaptive weighting: When multiple energy storage units compete for the same load, power is allocated according to the Shapley value ratio. When the lithium battery overtemperature is detected, the weight is automatically reduced and the load is transferred to the supercapacitor, providing real-time contribution evaluation and feedback.

[0031] The present invention has the following beneficial effects due to the adoption of the above technical solution:

[0032] This system reduces overall energy consumption by 22%-30%. Compared to traditional group control with single energy storage, it reduces lithium battery equivalent cycles by over 40%. It leverages a federated learning framework to enable multi-elevator cluster strategy sharing and integrates digital twins to optimize energy storage allocation strategies in real time, enhancing cross-scenario adaptability while protecting data privacy. It also innovatively introduces a multi-agent game mechanism to address competition for energy storage resources, making it suitable for complex elevator network systems such as super-high-rise buildings and subway hubs. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0034] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.

[0035] like Figure 1 As shown, the control algorithm of the elevator integrated energy storage device includes the following steps:

[0036] Step 1: Local data collection and federated feature extraction. Vibration accelerometers are installed on the car floor and guide rail connections, a gyroscope is installed on the car roof, a current sensor is installed on the main motor power supply line, a voltage sensor is installed at the output end of the energy storage device, and a temperature sensor is installed on the lithium battery pack and supercapacitor module. A crowd counting camera is installed at the elevator hall entrance to collect data in real time. Kalman filtering is used to jointly calibrate the vibration accelerometer and gyroscope data to eliminate installation deviations. During normal operation, 10Hz sampling is used. When abnormal vibration is detected, that is, when the amplitude exceeds the threshold, the sampling is switched to 500Hz to reduce storage overhead.

[0037] Data preprocessing and feature extraction: vibration signals are denoised using wavelet thresholding. Image data is segmented into pedestrian areas in real time using YOLOv5 to remove background noise. Time series features are generated, including statistical features, frequency domain features, and nonlinear features. The mean, variance, and peak-to-peak value within the sliding window are statistical features. The FFT spectrum peak position and wavelet packet energy entropy are frequency domain features. Nonlinear features include approximate entropy and sample entropy. Spatiotemporal feature alignment and time alignment are performed. Based on the dynamic time warping algorithm, the vibration signal time series of different elevators are synchronized, spatial alignment is performed, elevator topology relationships are constructed through graph convolutional networks, and multi-node features are integrated.

[0038] Privacy protection and data desensitization: embed a privacy protection mechanism in the local data collection stage and use local sensitive hashing to map high-dimensional features to low-dimensional space, ensuring that the same data hashes collide and different data hashes are separated.

[0039] Step 2: Local meta-policy training and gradient homomorphic encryption upload. Each elevator node is treated as an independent task. Task characteristics include floor distribution, passenger flow patterns, and energy storage type. An actor-critic architecture is used. The actor outputs energy storage allocation actions, and the critic evaluates the value of the actions. Morning and evening peak scenarios are prioritized, and the training weight of nighttime maintenance scenarios is gradually increased. FGSM attacks are introduced to generate adversarial samples to enhance the robustness of the strategy.

[0040] Homomorphic encryption uses CKKS encryption to convert floating-point gradients into integers. The error compensation mechanism ensures accuracy. The server performs additive homomorphic operations on the encrypted gradients. Threshold encryption is introduced in the gradient aggregation stage to prevent single-point key leakage. The ciphertext is uploaded in blocks and in parallel, supporting breakpoint resumption.

[0041] Step 3: Global model aggregation, edge-side model delivery. Adaptive weight aggregation, dynamically adjusts the aggregation weight according to the amount of elevator node data and contribution:

[0042]

[0043] Shapleyk is the marginal contribution of the kth node to the global model. Model compression and communication, gradient sparsification, uploading the first 5% non-zero gradient parameters, compression rate > 90%, compressing 32-bit floating-point parameters to 8-bit integers, removing redundant parameters with gradient absolute value < ε, ε = 0.01, dividing the model into blocks and transmitting them in parallel via the QUIC protocol, supporting breakpoint resumption, using threshold encryption in the gradient aggregation stage to prevent single-point key leakage, and adding Gaussian noise when updating the global model;

[0044] Model adaptation and lightweighting: knowledge of the global model is transferred to a lightweight student model, reducing the number of student model parameters by 70% and increasing inference speed by 2 times. Low-precision matrix multiplication replaces floating-point operations. A dynamic hot update mechanism and differential update only send the changed model parameters, achieving a compression rate of >80%. Double-buffered loading: first run the old model, then download the new model, verify its integrity, and then switch to double buffering.

[0045] Edge resource scheduling, heterogeneous computing optimization, high computing tasks are offloaded to the edge GPU, real-time control tasks are executed in hard real-time on the CPU side, and the Zstd algorithm is used to compress model weights, reducing memory usage by 50%.

[0046] Step 4: Digital twin prediction and enhanced anti-disturbance. A dynamic digital twin is constructed to simulate the physical behavior and degradation process of the elevator energy storage system in real time. Through anti-disturbance training, the model's resistance to noise and attacks is improved to prevent strategy failure. Based on the digital twin's prediction results, the control strategy is dynamically adjusted to achieve virtual-physical interactive closed-loop optimization.

[0047] Dynamic twin modeling, physics engine, using COMSOL Multiphysics to simulate the electrochemical degradation process of energy storage units, Q is the remaining capacity, η is the Coulomb efficiency, and k(T) is the temperature dependence coefficient. Based on OpenFOAM, the system simulates elevator mechanical vibration and structural fatigue. An LSTM-Transformer hybrid network is used to predict the load for the next 10 seconds. Online prediction and correction are performed. When the prediction error exceeds a threshold, online fine-tuning is triggered. This method combines the output of the physical engine with data-driven predictions through Bayesian weighted fusion.

[0048] Adversarial perturbation generation: Generate perturbations that maximize model loss, inject the vibration spectrum of the digital twin into real sensor data, test the robustness of the strategy, generate an adversarial loss function, gradually increase the perturbation intensity to avoid premature model collapse, and finally perform robustness evaluation indicators and adversarial sample detection.

[0049] Step 5: Shapley value allocation and coordinated control of multiple energy storage systems. Shapley values ​​are used to quantify the marginal contributions of supercapacitors, lithium batteries, and flywheels, solving the tragedy of the commons problem. The energy storage coordination strategy is dynamically adjusted based on real-time scenarios to balance energy recovery efficiency, lifespan loss, and cost. Shapley values ​​are calculated in real time and control instructions are updated to adapt to sudden changes in elevator load.

[0050] Characteristic function definition, benefit indicator, R(S) = ω1·ΔEregen + ω2·ηlife - ω3·ΔPgrid, where ΔEregen is the energy recovery increment, ηlife is the lifetime efficiency, and ΔPgrid is the grid fluctuation penalty. Multi-time scale calculation: in the short term of 1 second, supercapacitors dominate the instantaneous power response; in the medium term of 1 minute, lithium batteries bear the base load; in the long term of >5 minutes, the flywheel covers the continuous power shortage scenario;

[0051] Incremental benefit comparison: When energy storage unit i is added to subset S, the difference between the energy recovery increment and the life loss increment is calculated; ΔRi = ΔEregen(S∪{i})-ΔEregen(S)-λ·(ΔQloss(S∪{i})-ΔQloss(S)). Scenario-adaptive weighting: When multiple energy storage units compete for the same load, power is allocated according to the Shapley value ratio. When the lithium battery overtemperature is detected, the weight is automatically reduced and the load is transferred to the supercapacitor, providing real-time contribution evaluation and feedback.

[0052] The Shapley value contribution distribution model is:

[0053]

[0054] Where N is the set of energy storage units (supercapacitors, lithium batteries, flywheels, etc.), R(S) is the synergistic benefit of subset S, and φi is the Shapley value of the i-th energy storage unit.

[0055] Matters not covered by the present invention are known technologies.

[0056] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. The control algorithm of the elevator integrated energy storage device is characterized by: The method comprises the following steps: Step 1: Local data collection and federated feature extraction; Step 2: Local meta-strategy training and gradient homomorphic encryption upload; Step 3: Global model aggregation and edge model delivery; Step 4: Digital twin prediction, enhanced against disturbances; Step 5: Shapley value allocation and coordinated control of several energy storages.

2. The control algorithm of the elevator integrated energy storage device according to claim 1 is characterized in that: The specific process of step 1 is: Vibration accelerometers are installed at the car floor and guide rail connections, a gyroscope is installed on the car roof, a current sensor is installed on the main motor power supply line, a voltage sensor is installed at the output end of the energy storage device, and temperature sensors are installed on the lithium battery pack and supercapacitor module. A crowd counting camera is installed at the elevator hall entrance to collect data in real time. Kalman filtering is used to jointly calibrate the vibration accelerometer and gyroscope data to eliminate installation deviations. During normal operation, 10Hz sampling is used. When abnormal vibration is detected, that is, when the amplitude exceeds the threshold, the sampling is switched to 500Hz to reduce storage overhead. Data preprocessing and feature extraction: vibration signals are denoised using wavelet thresholding. Image data is segmented into pedestrian areas in real time using YOLOv5 to remove background noise. Time series features are generated, including statistical features, frequency domain features, and nonlinear features. The mean, variance, and peak-to-peak value within the sliding window are statistical features. The FFT spectrum peak position and wavelet packet energy entropy are frequency domain features. Nonlinear features include approximate entropy and sample entropy. Spatiotemporal feature alignment and time alignment are performed. Based on the dynamic time warping algorithm, the vibration signal time series of different elevators are synchronized, spatial alignment is performed, elevator topology relationships are constructed through graph convolutional networks, and multi-node features are integrated. Privacy protection and data desensitization: embed a privacy protection mechanism in the local data collection stage and use local sensitive hashing to map high-dimensional features to low-dimensional space, ensuring that the same data hashes collide and different data hashes are separated.

3. The control algorithm of the elevator integrated energy storage device according to claim 1 is characterized in that: The specific process of step 2 is as follows: each elevator node is treated as an independent task. Task characteristics include floor distribution, passenger flow patterns, and energy storage type. An actor-critic architecture is adopted. The actor outputs energy storage allocation actions, and the critic evaluates the value of the actions. Morning and evening peak scenarios are prioritized for learning, and the training weight of nighttime maintenance scenarios is gradually increased. FGSM attacks are introduced to generate adversarial samples to enhance the robustness of the strategy. Homomorphic encryption uses CKKS encryption to convert floating-point gradients into integers. The error compensation mechanism ensures accuracy. The server performs additive homomorphic operations on the encrypted gradients. Threshold encryption is introduced in the gradient aggregation stage to prevent single-point key leakage. The ciphertext is uploaded in blocks and in parallel, supporting breakpoint resumption.

4. The control algorithm of the elevator integrated energy storage device according to claim 1 is characterized in that: The specific process of step 3 is: Adaptive weighted aggregation dynamically adjusts the aggregation weight based on the amount of elevator node data and contribution: Shapleyk is the marginal contribution of the kth node to the global model. Model compression and communication, gradient sparsification, uploading the first 5% non-zero gradient parameters, compression rate > 90%, compressing 32-bit floating-point parameters to 8-bit integers, removing redundant parameters with gradient absolute value < ε, ε = 0.01, dividing the model into blocks and transmitting them in parallel via the QUIC protocol, supporting breakpoint resumption, using threshold encryption in the gradient aggregation stage to prevent single-point key leakage, and adding Gaussian noise when updating the global model; Model adaptation and lightweighting: knowledge of the global model is transferred to a lightweight student model, reducing the number of student model parameters by 70% and increasing inference speed by 2 times. Low-precision matrix multiplication replaces floating-point operations. A dynamic hot update mechanism and differential update only send the changed model parameters, achieving a compression rate of >80%. Double-buffered loading: first run the old model, then download the new model, verify its integrity, and then switch to double buffering. Edge resource scheduling, heterogeneous computing optimization, high computing tasks are offloaded to the edge GPU, real-time control tasks are executed in hard real-time on the CPU side, and the Zstd algorithm is used to compress model weights, reducing memory usage by 50%.

5. The control algorithm of the elevator integrated energy storage device according to claim 1 is characterized in that: The specific process of step 4 is as follows: construct a dynamic digital twin to simulate the physical behavior and degradation process of the elevator energy storage system in real time. Through adversarial disturbance training, the model's resistance to noise and attacks is improved to prevent strategy failure. The control strategy is dynamically adjusted based on the prediction results of the digital twin to achieve virtual-physical interactive closed-loop optimization. Dynamic twin modeling, physics engine, using COMSOL Multiphysics to simulate the electrochemical degradation process of energy storage units, Q is the remaining capacity, η is the Coulomb efficiency, and k(T) is the temperature dependence coefficient. Based on OpenFOAM, the system simulates elevator mechanical vibration and structural fatigue. An LSTM-Transformer hybrid network is used to predict the load for the next 10 seconds. Online prediction and correction are performed. When the prediction error exceeds a threshold, online fine-tuning is triggered. This method combines the output of the physical engine with data-driven predictions through Bayesian weighted fusion. Adversarial perturbation generation: Generate perturbations that maximize model loss, inject the vibration spectrum of the digital twin into real sensor data, test the robustness of the strategy, generate an adversarial loss function, gradually increase the perturbation intensity to avoid premature model collapse, and finally perform robustness evaluation indicators and adversarial sample detection.

6. The control algorithm of the elevator integrated energy storage device according to claim 1 is characterized in that: The specific process of step 5 is as follows: quantifying the marginal contributions of supercapacitors, lithium batteries, and flywheels through Shapley values ​​to solve the tragedy of the commons problem, dynamically adjusting the energy storage coordination strategy based on real-time scenarios, balancing energy recovery efficiency, life loss, and cost, and calculating Shapley values ​​in real time to update control instructions and adapt to the instantaneous load changes of the elevator. Characteristic function definition, benefit indicator, R(S) = ω1·ΔEregen + ω2·ηlife - ω3·ΔPgrid, where ΔEregen is the energy recovery increment, ηlife is the lifetime efficiency, and ΔPgrid is the grid fluctuation penalty. Multi-time scale calculation: in the short term of 1 second, supercapacitors dominate the instantaneous power response; in the medium term of 1 minute, lithium batteries bear the base load; in the long term of >5 minutes, the flywheel covers the continuous power shortage scenario; Incremental benefit comparison: When energy storage unit i is added to subset S, the difference between the energy recovery increment and the life loss increment is calculated; ΔRi = ΔEregen(S∪{i})-ΔEregen(S)-λ·(ΔQloss(S∪{i})-ΔQloss(S)). Scenario-adaptive weighting: When multiple energy storage units compete for the same load, power is allocated according to the Shapley value ratio. When the lithium battery overtemperature is detected, the weight is automatically reduced and the load is transferred to the supercapacitor, providing real-time contribution evaluation and feedback.

Images