Quick-response dynamic wireless power supply system and method suitable for mobile equipment

Abstract

The invention discloses a quick-response dynamic wireless power supply system and method suitable for mobile equipment. The method comprises the following steps: S1, acquiring running state data of the mobile equipment in real time; s2, preprocessing the operation state data; s3, predicting a power supply position track and a power demand of the mobile device based on the improved self-attention echo state network; s4, constructing a dynamic power supply control instruction set; s5, an array type wireless energy emission control module is driven, and energy directional transmission is carried out; s6, collecting parameters of a receiving end in real time at the receiving end of the mobile equipment; s7, correcting the control instruction set through the reinforcement learning adjustment model; and S8, updating network parameters and control strategies through incremental learning. According to the method, the improved self-attention echo state network and reinforcement learning are combined, intelligent prediction and closed-loop control of dynamic wireless power supply of the mobile equipment are achieved, and the method has the advantages of being fast in response, high in stability and high in self-optimization capacity.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless power supply, and in particular, to a fast-response dynamic wireless power supply system and method applicable to mobile devices. Background Art

[0002] With the wide application of mobile devices such as smart wearable devices, service robots, drones, and mobile transportation terminals in industries, military, medical, and consumer electronics fields, how to achieve continuous power supply for these mobile devices has become one of the key problems restricting their long-term stable operation. Traditional mobile devices mainly rely on internal batteries for power supply. Although they are convenient for deployment and movement, their battery life is limited, the charging cycle is frequent, and the degree of manual intervention is high, making it difficult to meet the requirements of long-time continuous operation.

[0003] To solve the limitations of battery power supply, various wireless power supply technologies have emerged in recent years, including electromagnetic induction, magnetic resonance coupling, radio frequency long-distance transmission, etc. These technologies can achieve non-contact energy transmission for devices within a certain distance and degree of freedom, providing a new idea for continuous power supply for mobile devices. However, most traditional wireless power supply systems are designed based on the premise of a fixed position or orientation. The transmitter usually needs to be aligned with the receiver to achieve efficient energy transmission. Therefore, they cannot adapt to the rapidly changing positions and postures of mobile devices in a dynamic space. In addition, most existing wireless power supply control systems adopt static or regularized adjustment methods and cannot achieve adaptive optimization according to the state changes of the devices, easily resulting in problems such as power supply deviation, power waste, or insufficient power supply.

[0004] Some studies have tried to introduce a power supply model based on control algorithms into the wireless energy transmission system. For example, by obtaining the current position and movement trajectory of the device and combining PID control or model predictive control to adjust the transmission power and direction. However, such methods generally rely on accurate prior modeling, are difficult to adapt to complex uncertain factors in the environment, and have a slow response to nonlinear dynamic systems. In addition, limited by the existing sensing and feedback mechanisms, the system often has problems such as lag, slow convergence, and coarse adjustment granularity, making it difficult to provide an efficient response in fast-moving or high-frequency changing power supply scenarios.

[0005] In recent years, artificial intelligence, especially deep learning and reinforcement learning, has demonstrated powerful capabilities in fields such as predictive control, sequence modeling, and system adaptive optimization. Some studies have introduced it into the scheduling and control of wireless power supply systems and achieved certain results. For example, using recurrent neural networks to predict mobile trajectories, or adopting deep reinforcement learning for parameter adjustment and power optimization. However, these methods still face multiple technical challenges in practical engineering applications. First, the traditional recurrent neural network structure has the problem of long-term dependence and is difficult to accurately capture the dynamic characteristics of multiple time scales in complex nonlinear power supply systems. Second, the policy convergence speed of reinforcement learning is relatively slow, and model parameters are not easily transferable, resulting in the inability to quickly adapt to changes in device states during actual deployment. Moreover, most existing systems lack high-frequency real-time feedback mechanisms and self-learning capabilities and cannot accumulate experience and self-evolve the model based on historical power supply behaviors.

[0006] Based on this, the existing technologies still have the following prominent problems: First, in a complex environment with variable mobile states, it is difficult for existing wireless power supply systems to achieve continuous, directional, and high-efficiency power supply to target devices. Second, the control strategy lacks real-time performance and adaptive capabilities and cannot be dynamically adjusted in a timely manner according to feedback information such as changes in the attitude and power consumption of mobile devices. Third, there is a lack of a system architecture that integrates data collection, prediction, feedback, self-learning, and control closed-loop, resulting in a low overall intelligence level of the system and the inability to support rapid response and continuous optimization in complex power supply scenarios.

[0007] Therefore, there is an urgent need for an intelligent wireless power supply solution that can combine temporal modeling capabilities with real-time control adjustment mechanisms and has self-learning capabilities to solve the problems of unstable power supply, lagging adjustment, and lack of intelligent control in the existing technologies for mobile devices, and achieve efficient energy following, closed-loop control, and continuous evolution for dynamic targets, providing stable, reliable, and efficient wireless power supply guarantees for intelligent devices. Summary of the Invention

[0008] An object of the present invention is to propose a fast-response dynamic wireless power supply method applicable to mobile devices. The present invention combines an improved self-attention echo state network and a reinforcement learning model to construct a high-precision temporal prediction and adaptive control mechanism, realizing dynamic prediction and real-time regulation of the wireless power supply trajectory and power demand of mobile devices during operation, and having the advantages of fast response speed, strong power supply stability, strong self-learning ability, and adaptability to complex scenarios.

[0009] A fast-response dynamic wireless power supply method applicable to mobile devices according to an embodiment of the present invention includes the following steps:

[0010] S1. Real-time collect the operation state data of the mobile device;

[0011] S2. Preprocess the operation status data to construct a power supply status sequence;

[0012] S3. Input the power supply status sequence into a time series prediction network composed of an improved self-attention echo state network to predict the power supply position trajectory and power demand of the mobile device at future time steps;

[0013] S4. Construct a dynamic power supply control instruction set according to the prediction results;

[0014] S5. Based on the control instruction set, drive the array-type wireless energy emission control module to perform magnetic resonance or radio frequency energy directional transmission according to the power supply position trajectory;

[0015] S6. Real-time collect the received power, voltage, current and mobile device attitude parameters at the mobile device receiving end;

[0016] S7. Construct a state vector based on the collected actual power, voltage, current and mobile device attitude parameters, input it into the reinforcement learning adjustment model, output a corrected control instruction set, and dynamically adjust the emission parameters of the array-type wireless energy emission control module;

[0017] S8. Store the prediction data, feedback error, control parameters and power supply performance indicators generated in each round of power supply tasks in the self-learning optimization module, and the self-learning optimization module updates the parameters of the self-attention echo state network and the control strategy of the reinforcement learning adjustment model through incremental learning.

[0018] Optionally, the operation status data includes the mobile device position coordinates, speed vector, attitude angle and instantaneous power demand.

[0019] Optionally, the preprocessing includes normalization processing, standardization processing, time series reconstruction and feature vector fusion.

[0020] Optionally, S3 specifically includes:

[0021] S31. Construct a time series prediction network, which is composed of an improved self-attention echo state network. The improved self-attention echo state network includes an input encoding layer, a self-attention layer, a gating mechanism, a spatial perception state update layer and an output prediction layer, and the input is the power supply status sequence where F i ={f1,f2,…,f m} represents the fusion feature vector at the i-th time step, m represents the feature dimension, p represents the number of time steps of the power supply status sequence, and the improved self-attention echo state network introduces a gating mechanism and a spatial perception weight matrix;

[0022] S32. In the input encoding layer, for each fusion feature vector Fi Perform a linear mapping to generate an intermediate feature representation;

[0023] S33. In the self-attention layer, calculate query vectors, key vectors, and value vectors based on the intermediate feature representation, and perform attention weighting calculations for all time steps to obtain an attention-weighted vector:

[0024]

[0025] where, Z i represents the attention-weighted vector, d k represents the key vector dimension, p represents the number of time steps of the power supply state sequence, softmax represents the normalization operation, Q i represents the query vector, K i represents the key vector, V j represents the value vector, and T represents the transpose operation of the vector;

[0026] S34. Input the attention-weighted vector into the self-attention echo state network with a gated mechanism to calculate the gating parameters for the current time step:

[0027] z i = σ(W z ·Z i + U z ·H i-1 ), r i = σ(W r ·Z i + U r ·H i-1 );

[0028] where, z i represents the update gate, r i represents the reset gate, σ represents the Sigmoid activation function, W z and W r represent the gating input weights, U z and U r represent the gating recurrent weights, and H i-1 represents the hidden state at time step i - 1;

[0029] Construct a spatial perception recurrent connection matrix:

[0030] W = α · exp(-β · ||pos i - pos j || 2 );

[0031] where, W represents the connection weight, α represents the connection strength factor, β represents the spatial decay coefficient, exp represents the natural exponential function, pos i and pos jDenote the corresponding position vector, ||·|| 2 Denote the Euclidean distance;

[0032] Update the hidden state:

[0033] H i = z i ⊙ tanh(W in · Z i + W · (r i ⊙ H i-1 ));

[0034] Wherein, H i Denotes the hidden state at time step i, z i Denotes the update gate, ⊙ denotes the Hadamard product, W in Denotes the input connection matrix, tanh denotes the hyperbolic tangent activation function, pos j Denotes Z i Denotes the attention weighted vector, W denotes the connection weight, r i Denotes the reset gate, H i-1 Denotes the hidden state at time step i - 1;

[0035] S35. Input the final hidden state sequence {H1, H2, …, H p} into the output prediction layer, and calculate the power supply position trajectory and power demand at future time steps through a multi-layer perceptron:

[0036]

[0037] Wherein, Denotes the power supply position trajectory, Denotes the power demand, (x k , y k , z k ) denotes the spatial power supply coordinates at the k-th prediction time step, P k Denotes the power demand at the corresponding time step.

[0038] Optionally, the control instruction set includes the transmission power adjustment value, frequency offset, beam direction angle, and phase compensation amount of the wireless power supply system.

[0039] Optionally, the S4 specifically includes:

[0040] S41. Based on the prediction result sequence Construct a dynamic power supply control instruction set, and generate a corresponding power supply parameter set for each time step j Where A j Denotes the transmission power adjustment value, f j Denotes the frequency offset, θ j Denotes the beam direction angle, represents the phase compensation amount, (x j , y j , z j ) represents the spatial power supply coordinates at the j-th predicted time step;

[0041] S42. Calculate the transmission power adjustment value based on the predicted power demand P j and the preset system loss parameter;

[0042] S43. Calculate the frequency offset based on the predicted position (x j , y j , z j ):

[0043] f j = f0 + γ·Δd j ;

[0044] where f j represents the frequency offset, f0 represents the system reference transmission frequency, γ represents the frequency offset sensitivity coefficient caused by position change, and Δd j = ||(x j , y j , z j ) - (x j-1 , y j-1 , z j-1 )|| represents the Euclidean distance of position change;

[0045] S44. Calculate the beam direction adjustment angle based on the predicted position and the transmission array reference coordinate system:

[0046] θ j = arctan2(y j - y0, x j - x0);

[0047] where θ j represents the beam direction adjustment angle, arctan represents the arctangent function, and (x0, y0) represents the projection coordinates of the center position of the power supply array on the horizontal plane;

[0048] S45. Introduce a phase correction mechanism and calculate the phase compensation amount based on the distance change:

[0049]

[0050] where represents the phase compensation amount, f j represents the frequency offset, d j represents the spatial distance between the center of the transmission array and the receiving target, and c represents the propagation speed of electromagnetic waves in the air;

[0051] S46. Integrate the power supply parameter sets generated at all time steps in a time series to form a dynamic power supply control instruction set C = {C1, C2, …, C k}, and drive the array-type wireless energy transmission module to perform dynamic power supply regulation.

[0052] Optionally, the S5 specifically includes:

[0053] S51. Based on the dynamic power supply control instruction set C = {C1, C2, …, C k}, sequentially extract the control parameter sets at each time step and transmit them to the wireless energy transmission control module for real-time configuration;

[0054] S52. Receive the power supply position trajectory and synchronize it with the current clock, control the array-type wireless energy transmission control module to switch to the dynamic power supply state, and perform position alignment and energy modulation operations at the set time step;

[0055] S53. Adjust the angle θ j of the beam direction according to the control parameter set, adjust the spatial transmission direction of the phased array or coil group, achieve precise spatial steering to align with the current or predicted position of the target device, and form a directional power supply channel;

[0056] S54. Apply differential phase control to each transmitting unit in the array according to the phase compensation amount in the control parameter set, perform transmission phase synchronization and interference construction, so that the output waveform forms a focused interference energy beam in the target direction;

[0057] S55. According to the transmission power adjustment value A j in the control parameter set, adjust the output power level of the power amplifier module or voltage regulation module, achieve dynamic matching of the current power demand of the device, and control the system output power to change in real time with the target power consumption;

[0058] S56. According to the frequency offset f j in the control parameter set, control the operating frequency of the radio frequency signal source or magnetic resonance source, and complete the real-time switching of frequency synthesis and transmitted signals;

[0059] S57. Input the control parameters in consecutive time steps into the array-type wireless energy transmission control module in the order of time series, drive the system to complete parameter refresh, direction adjustment, and transmission control at each time step, ensure continuous coverage of energy in space and time, and achieve dynamic directional and stable power supply to the mobile target device.

[0060] Optionally, the S7 specifically includes:

[0061] S71. Obtain the actual power, voltage, current, and mobile device attitude parameters feedback by the receiver at the current time step, and splice and construct the state vector S by combining the transmission control instruction set and the target coordinates at the current time step j ;

[0062] S72. Construct a reward function in the state-action space based on the received power error, attitude deviation, and electrical parameter deviation:

[0063]

[0064] Among them, r j represents the reward function, describing the immediate reward at the current time step, L j represents the loss function, w1, w2, w3, and w4 respectively represent the weighting coefficients of each error, ΔP j represents the difference between the transmitted power and the received power, ΔΨ j represents the attitude deviation vector, ||·|| 2 represents the two-norm operation, ΔV j represents the voltage deviation, ΔI j represents the current deviation;

[0065] S73. Use the state vector S j and the reward function, and input them into the reinforcement learning adjustment model for training. The reinforcement learning adjustment model includes a policy network π θ (a j |S j ) and a value network V ω (S j ), where π θ represents the policy network characterized by the parameter θ, a j represents the control action, V ω represents the value network characterized by the parameter ω;

[0066] S74. Update the policy network parameters using a policy gradient-based reinforcement learning algorithm:

[0067]

[0068] Among them, J(θ) represents the policy objective function, which is the expectation of the cumulative reward of the policy network within a time window of length p, represents the gradient of the policy objective function with respect to the policy parameter θ, E represents the mathematical expectation, π θ (a j |S j ) represents the policy network, r j represents the reward function, represents the gradient operation;

[0069] S75. The execution policy network outputs an action As the corrected control instruction set Replace the original control parameters and transmit them to the array - type wireless energy emission control module for control execution, where Represents the corrected transmission power, Represents the corrected frequency, Represents the corrected direction angle, Represents the corrected phase compensation.

[0070] Optionally, the S8 specifically includes:

[0071] S81. After each round of power supply task execution is completed, collect the prediction data generated during the task, and at the same time collect the control parameter sequence used during the power supply process, the receiver feedback data, and the power supply performance indicators recorded by the system;

[0072] S82. Statistically generate task evaluation indicators for the power supply performance indicators collected during the task. The task evaluation indicators include the average power error, the mean square deviation of attitude change, and the average control energy consumption, reflecting the control accuracy and execution efficiency of the power supply process;

[0073] S83. Combine the control instructions, received feedback, prediction results, and power supply performance indicators in each time step to construct training samples, form a time - series sample set based on state - action - feedback, and input it into the self - learning optimization module;

[0074] S84. Adopt an incremental learning strategy to update the parameters of the self - attention echo state network, and perform online backpropagation training on the self - attention echo state network using the new sample sequence, minimizing the mean square error loss function between the predicted trajectory and the actual feedback position;

[0075] S85. Based on the state information, control actions, and reward feedback in the time - series sample set, use the experience replay mechanism to train the reinforcement learning adjustment model and update the policy network parameters;

[0076] S86. Replace the corresponding modules in the original control framework with the parameters of the updated self - attention echo state network and the reinforcement learning adjustment model, and apply them to the next round of power supply tasks.

[0077] A fast - response dynamic wireless power supply system applicable to mobile devices according to an embodiment of the present invention includes:

[0078] An operating state acquisition module for real - time acquisition of the operating state data of the mobile device;

[0079] A data pre - processing module for normalizing, standardizing, time - series reconstructing, and feature - vector fusing the operating state data to construct a power supply state sequence;

[0080] A timing prediction module, which is used to input a power supply state sequence into a timing prediction network composed of an improved self-attention echo state network to predict the power supply position trajectory and power demand of a mobile device at future time steps;

[0081] A control instruction generation module, which is used to construct a dynamic power supply control instruction set according to the prediction result;

[0082] An energy emission control module, which is used to drive an array-type wireless energy emission control module based on the control instruction set to perform magnetic resonance or radio frequency energy directional transmission according to the predicted power supply position trajectory;

[0083] A receiving state acquisition module, which is used to collect the received power, voltage, current and mobile device attitude parameters in real time at the receiving end of the mobile device and generate feedback state data;

[0084] A reinforcement learning adjustment module, which is used to construct a state vector based on the feedback state data and input it into a reinforcement learning model composed of a policy network and a value network, output a corrected control instruction set, and realize dynamic adjustment of the transmission power, frequency, direction and phase to construct a closed-loop control system;

[0085] A self-learning optimization module, which is used to store the prediction data, feedback error, control parameters and power supply performance indicators in each round of power supply tasks, and update the parameters of the self-attention echo state network and the control strategy of the reinforcement learning adjustment model based on the incremental learning strategy.

[0086] The beneficial effects of the present invention are as follows:

[0087] First of all, the present invention proposes a fast-response dynamic wireless power supply method and system suitable for mobile devices. Aiming at the problems of lagging response, rigid control and insufficient adaptability of existing wireless power supply technologies in dynamic environments, by constructing a timing prediction model composed of an improved self-attention echo state network, the prediction accuracy of the future operating state of the mobile device is significantly improved. The network introduces a gating mechanism and spatially aware recurrent connection weights in its structure, which can effectively model the multi-time scale dependencies in the power supply state sequence and enhance the sensitivity to spatial position changes, so as to achieve high-accuracy prediction of the power supply position trajectory and power demand.

[0088] Secondly, the present invention further combines the prediction results to construct a dynamic power supply control instruction set. The instruction set not only covers key control parameters such as transmit power adjustment, frequency offset, beam direction adjustment, and phase compensation, but also can be continuously updated with the change of time steps to realize the real-time drive of the array wireless energy transmission module. The system can perform directional power supply according to the position and attitude of the mobile device at different time points, and complete direction alignment and energy focusing in advance through the predicted trajectory, effectively reducing the spatial energy loss and power supply delay, and ensuring the energy supply continuity and stability during the continuous movement of the device.

[0089] In addition, in order to overcome the limitations of the traditional control system such as regulation lag and unidirectional output, the present invention constructs a reinforcement learning adjustment model. By real-time collecting received power, voltage, current, and attitude feedback parameters, dynamically constructing a state vector and inputting it into the policy network and value network, and combining the optimization method based on policy gradient to output a corrected control instruction set, a fast-response closed-loop control mechanism is formed. This mechanism not only has the ability to sense errors in real time and self-adjust control instructions, but also has the ability to continuously optimize strategies, enabling the system to maintain efficient and stable energy following ability when the motion state of the mobile device changes rapidly.

[0090] Finally, the present invention introduces a self-learning optimization module, which converts the prediction data, feedback error, control parameters, and power supply performance indicators in each round of power supply task into training samples, and updates the parameters of the self-attention echo state network and the reinforcement learning model through incremental learning, opening up a full-process intelligent closed-loop from prediction modeling to feedback regulation and then to model evolution. Compared with the traditional wireless power supply method, the present invention can continuously accumulate historical power supply experience, improve the generalization ability of the model and the adaptability of control strategies, and maintain the performance stability and intelligence during long-term operation in multi-task, multi-path or complex dynamic scenarios. Description of the Drawings

[0091] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0092] Figure 1 is a flowchart of a fast-response dynamic wireless power supply method for mobile devices proposed by the present invention;

[0093] Figure 2 is a schematic structural diagram of an improved self-attention echo state network of a fast-response dynamic wireless power supply method for mobile devices proposed by the present invention;

[0094] Figure 3 is a schematic diagram of a self-learning optimization module of a fast-response dynamic wireless power supply method for mobile devices proposed by the present invention. Detailed Embodiments

[0095] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0096] Reference Figures 1 - 3 , a fast-response dynamic wireless power supply method applicable to mobile devices, comprising the following steps:

[0097] S1. Real-time collect the operation status data of the mobile device;

[0098] S2. Preprocess the operation status data to construct a power supply status sequence;

[0099] S3. Input the power supply status sequence into a time series prediction network composed of an improved self-attention echo state network to predict the power supply position trajectory and power demand of the mobile device in future time steps;

[0100] S4. Construct a dynamic power supply control instruction set according to the prediction results;

[0101] S5. Based on the control instruction set, drive the array-type wireless energy emission control module to perform magnetic resonance or radio frequency energy directional transmission according to the power supply position trajectory;

[0102] S6. Real-time collect the received power, voltage, current and mobile device attitude parameters at the receiving end of the mobile device;

[0103] S7. Construct a state vector based on the collected actual power, voltage, current and mobile device attitude parameters, input it into a reinforcement learning adjustment model, output a corrected control instruction set, and dynamically adjust the emission parameters of the array-type wireless energy emission control module;

[0104] S8. Store the prediction data, feedback error, control parameters and power supply performance indicators generated in each round of power supply tasks into a self-learning optimization module, and the self-learning optimization module updates the parameters of the self-attention echo state network and the control strategy of the reinforcement learning adjustment model through an incremental learning method.

[0105] This method realizes the accurate prediction and dynamic control of the power supply demand of mobile devices by introducing a self-attention echo state network and a reinforcement learning model, solves problems such as response lag, energy waste and inaccurate spatial positioning in traditional wireless power supply, and improves the real-time performance, reliability and intelligent level of wireless power supply.

[0106] In this embodiment, the operation status data includes the mobile device position coordinates, velocity vector, attitude angle and instantaneous power demand.

[0107] The present invention collects multi-dimensional operating state data including position coordinates, velocity vectors, attitude angles, and instantaneous power demands, providing richer input features for constructing the power supply state sequence, which helps improve the accuracy of the time series prediction network and the environmental adaptability of the control system.

[0108] In this embodiment, it is characterized in that the preprocessing includes normalization processing, standardization processing, time series reconstruction, and feature vector fusion.

[0109] Through preprocessing steps such as normalization, standardization, time series reconstruction, and feature vector fusion, the present invention can effectively reduce the noise interference and dimensional inconsistency problems of the original data, and improve the training stability and generalization ability of the prediction model.

[0110] In this embodiment, S3 specifically includes:

[0111] S31. Construct a time series prediction network, which is composed of an improved self-attention echo state network. The improved self-attention echo state network includes an input encoding layer, a self-attention layer, a gating mechanism, a spatial perception state update layer, and an output prediction layer. The input is the power supply state sequence where F i ={f1, f2, …, f m} represents the fused feature vector at the i-th time step, m represents the feature dimension, p represents the number of time steps of the power supply state sequence, and the improved self-attention echo state network introduces a gating mechanism and a spatial perception weight matrix;

[0112] S32. In the input encoding layer, perform a linear mapping on each fused feature vector F i to generate an intermediate feature representation;

[0113] S33. In the self-attention layer, calculate the query vector, key vector, and value vector based on the intermediate feature representation, and perform attention weighting calculation for all time steps to obtain the attention weighted vector:

[0114]

[0115] where, Z i represents the attention weighted vector, d k represents the key vector dimension, p represents the number of time steps of the power supply state sequence, softmax represents the normalization operation, Q i represents the query vector, K i represents the key vector, V j represents the value vector, and T represents the transpose operation of the vector;

[0116] S34. Input the attention weighted vector into the self-attention echo state network with a gating mechanism to calculate the gating parameter of the current time step:

[0117] z i = σ(W z ·Z i + U z ·H i-1 ), r i = σ(W r ·Z i + U r ·H i-1 );

[0118] Among them, z i represents the update gate, r i represents the reset gate, σ represents the Sigmoid activation function, W z and W r represent the gated input weights, U z and U r represent the gated recurrent weights, H i-1 represents the hidden state at time step i - 1;

[0119] Construct the spatial perception recurrent connection matrix:

[0120] W = α·exp(-β·||pos i - pos j || 2 );

[0121] Among them, W represents the connection weight, α represents the connection strength factor, β represents the spatial decay coefficient, exp represents the natural exponential function, pos i and pos j represent the corresponding position vectors, ||·|| 2 represents the Euclidean distance;

[0122] Update the hidden state:

[0123] H i = z i ⊙ tanh(W in ·Z i + W·(r i ⊙ H i-1 ));

[0124] Among them, H i represents the hidden state at time step i, z i represents the update gate, ⊙ represents the Hadamard product, W in represents the input connection matrix, tanh represents the hyperbolic tangent activation function, pos j represents Z i represents the attention weighted vector, W represents the connection weight, r i represents the reset gate, H i-1Represents the hidden state at time step i-1;

[0125] S35. Input the final hidden state sequence {H1, H2, …, H p} into the output prediction layer, and calculate the power supply position trajectory and power demand at future time steps through a multi-layer perceptron:

[0126]

[0127] where, represents the power supply position trajectory, represents the power demand, (x k , y k , z k ) represents the spatial power supply coordinates at the k-th prediction time step, and P k represents the power demand at the corresponding time step.

[0128] The present invention adopts an improved self-attention echo state network, integrating a gating mechanism and a spatial perception state update layer, which can effectively enhance the model's ability to model complex movement trajectories and power demand changes, and improve the accuracy and timeliness of power supply trajectory and demand prediction.

[0129] In this embodiment, the control instruction set includes the transmission power adjustment value, frequency offset, beam direction angle, and phase compensation amount of the wireless power supply system.

[0130] The present invention realizes the joint multi-parameter regulation of wireless energy transmission by constructing a control instruction set including transmission power, frequency offset, beam direction angle, and phase compensation, improves the spatial focusing ability and energy transmission efficiency, and meets the power supply requirements of different power consumption scenarios.

[0131] In this embodiment, the specific steps of S4 are as follows:

[0132] S41. Based on the prediction result sequence construct a dynamic power supply control instruction set, and generate a corresponding power supply parameter set for each time step j where A j represents the transmission power adjustment value, f j represents the frequency offset, θ j represents the beam direction angle, represents the phase compensation amount, (x j , y j , z j ) represents the spatial power supply coordinates at the j-th prediction time step;

[0133] S42. Calculate the transmission power adjustment value according to the predicted power demand P j and the preset system loss parameters;

[0134] S43. Calculate the frequency offset based on the predicted position (x j , y j , z j ):

[0135] f j = f0 + γ·Δd j ;

[0136] where f j represents the frequency offset, f0 represents the system reference emission frequency, γ represents the frequency offset sensitivity coefficient caused by position change, and Δd j = ||(x j , y j , z j ) - (x j-1 , y j-1 , z j-1 )|| represents the Euclidean distance of the position change;

[0137] S44. Calculate the beam direction adjustment angle based on the predicted position and the emission array reference coordinate system:

[0138] θ j = arctan2(y j - y0, x j - x0);

[0139] where θ j represents the beam direction adjustment angle, arctan represents the arctangent function, and (x0, y0) represents the projection coordinates of the center position of the power supply array on the horizontal plane;

[0140] S45. Introduce a phase correction mechanism and calculate the phase compensation amount based on the distance change:

[0141]

[0142] where represents the phase compensation amount, f j represents the frequency offset, d j represents the spatial distance between the center of the emission array and the receiving target, and c represents the propagation speed of electromagnetic waves in the air;

[0143] S46. Integrate the power supply parameter sets generated at all time steps in time series to form a dynamic power supply control instruction set C = {C1, C2,..., C k}, and drive the array - type wireless energy emission module to execute dynamic power supply regulation.

[0144] The present invention realizes the physical interpretability and accurate calculation of control instructions by gradually decomposing the prediction results to calculate various power supply control parameters and deriving the frequency offset and phase compensation values based on the physical modeling method, improving the robustness and adaptive ability of the power supply system regulation and control.

[0145] In this embodiment, step S5 specifically includes:

[0146] S51. Based on the dynamic power supply control instruction set C = {C1, C2, …, C k}, successively extract the control parameter sets at each time step and transmit them to the wireless energy emission control module for real-time configuration;

[0147] S52. Receive the power supply position trajectory and synchronize it with the current clock, control the array-type wireless energy emission control module to switch to the dynamic power supply state, and perform position alignment and energy modulation operations at the set time step;

[0148] S53. According to the beam direction adjustment angle θ in the control parameter set j , adjust the spatial emission direction of the phased array or coil group to achieve accurate spatial steering alignment with the current or predicted position of the target device, and form a directional power supply channel;

[0149] S54. According to the phase compensation amount in the control parameter set , apply differential phase control to each emission unit in the array, perform emission phase synchronization and interference construction, so that the output waveform forms a focused interference energy beam in the target direction;

[0150] S55. According to the transmission power adjustment value A in the control parameter set j , adjust the output power level of the power amplifier module or voltage regulation module to achieve dynamic matching of the current power demand of the device, and control the system output power to change in real time with the target power consumption;

[0151] S56. According to the frequency offset f in the control parameter set j , control the operating frequency of the radio frequency signal source or magnetic resonance source to complete the real-time switching of frequency synthesis and transmitted signals;

[0152] S57. Input the control parameters in consecutive time steps into the array-type wireless energy emission control module in the order of time series, drive the system to complete parameter refreshing, direction adjustment and emission control at each time step, ensure continuous coverage of energy in space and time, and achieve dynamic directional stable power supply to the mobile target device.

[0153] By applying the control parameters of each time step to the array - type wireless energy transmission control module and combining the direction adjustment and transmission phase control of the phased array, the present invention realizes the space - time joint energy directional focusing, significantly improving the power supply continuity and stability for high - speed moving devices.

[0154] In this embodiment, the S7 specifically includes:

[0155] S71. Obtain the actual power, voltage, current, and mobile device attitude parameters fed back by the receiver at the current time step, and splice and construct the state vector S by combining the transmission control instruction set and the target coordinates at the current time step j ;

[0156] S72. Construct a reward function in the state - action space based on the received power error, attitude deviation amount, and electrical parameter deviation:

[0157]

[0158] where r j represents the reward function, describing the immediate reward at the current time step, L j represents the loss function, w1, w2, w3, and w4 respectively represent the weighting coefficients of each error, ΔP j represents the difference between the transmitted power and the received power, ΔΨ j represents the attitude deviation vector, ||·|| 2 represents the two - norm operation, ΔV j represents the voltage deviation, ΔI j represents the current deviation;

[0159] S73. Use the state vector S j and the reward function, and input them into the reinforcement learning adjustment model for training. The reinforcement learning adjustment model includes a policy network π θ (a j |S j ) and a value network V ω (S j ), where π θ represents the policy network characterized by the parameter θ, a j represents the control action, and V ω represents the value network characterized by the parameter ω;

[0160] S74. Update the policy network parameters using a reinforcement learning algorithm based on policy gradients:

[0161]

[0162] where J(θ) represents the policy objective function, which is the expectation of the cumulative reward of the policy network within a time window of length p denotes the gradient of the policy objective function with respect to the policy parameter θ, E denotes the mathematical expectation, and π θ (a j |S j ) represents the policy network, r j represents the reward function, denotes the gradient operation;

[0163] S75. Execute the action output by the policy network as the corrected control instruction set to replace the original control parameters and transmit them to the array-type wireless energy emission control module for control execution, where denotes the corrected transmission power, denotes the corrected frequency, denotes the corrected direction angle, denotes the corrected phase compensation.

[0164] Through the introduction of a reinforcement learning model and the design of a reward function based on power error, attitude deviation, and electrical parameter differences, the present invention realizes a dynamic self-optimization control mechanism for feedback regulation, significantly improves the response ability of the system to the attitude changes of mobile devices and environmental disturbances, and constructs an intelligent closed-loop control system.

[0165] In this embodiment, the S8 specifically includes:

[0166] S81. After each round of power supply task execution is completed, collect the prediction data generated during the task, and at the same time collect the control parameter sequence used during the power supply process, the receiver feedback data, and the power supply performance indicators recorded by the system;

[0167] S82. Statistically generate task evaluation indicators for the power supply performance indicators collected during the task. The task evaluation indicators include the average power error, the mean square deviation of attitude changes, and the average control energy consumption, which reflect the control accuracy and execution efficiency of the power supply process;

[0168] S83. Combine the control instructions, received feedback, prediction results, and power supply performance indicators in each time step to construct training samples, form a time series sample set based on state-action-feedback, and input it into the self-learning optimization module;

[0169] S84. Update the self-attention echo state network parameters using an incremental learning strategy, and perform online backpropagation training on the self-attention echo state network using the new sample sequence to minimize the mean square error loss function between the predicted trajectory and the actual feedback position;

[0170] S85. Based on the state information, control actions, and reward feedback in the time series sample set, use the experience replay mechanism to train the reinforcement learning adjustment model and update the policy network parameters;

[0171] S86. Replace the corresponding modules in the original control framework with the parameters of the updated self-attention echo state network and the reinforcement learning adjustment model, and apply them to the next round of power supply tasks.

[0172] The present invention adopts an incremental learning mechanism to continuously optimize the model online, realizes the adaptive update of the power supply system in multiple rounds of tasks, solves the problems of model aging and insufficient adaptability, and effectively improves the control performance and system intelligence level in long-term usage scenarios.

[0173] A fast-response dynamic wireless power supply system applicable to mobile devices, comprising:

[0174] An operating state acquisition module for real-time acquisition of the operating state data of the mobile device;

[0175] A data preprocessing module for normalizing, standardizing, time series reconstruction and feature vector fusion of the operating state data to construct a power supply state sequence;

[0176] A time series prediction module for inputting the power supply state sequence into a time series prediction network composed of an improved self-attention echo state network to predict the power supply position trajectory and power demand of the mobile device at future time steps;

[0177] A control instruction generation module for constructing a dynamic power supply control instruction set according to the prediction results;

[0178] An energy emission control module for driving an array-type wireless energy emission control module based on the control instruction set to perform magnetic resonance or radio frequency energy directional transmission according to the predicted power supply position trajectory;

[0179] A reception state acquisition module for real-time acquisition of received power, voltage, current and mobile device attitude parameters at the mobile device receiving end and generating feedback state data;

[0180] A reinforcement learning adjustment module for constructing a state vector based on the feedback state data and inputting it into a reinforcement learning model composed of a policy network and a value network, outputting a corrected control instruction set, realizing dynamic adjustment of the transmission power, frequency, direction and phase, and constructing a closed-loop control system;

[0181] A self-learning optimization module for storing prediction data, feedback errors, control parameters and power supply performance indicators in each round of power supply tasks, and updating the parameters of the self-attention echo state network and the control strategy of the reinforcement learning adjustment model based on an incremental learning strategy.

[0182] This system integrates functions of data acquisition, time series prediction, dynamic control, feedback regulation, and self-learning optimization through a modular structure, forming an intelligent wireless power supply platform with prediction and closed-loop self-optimization capabilities, significantly improving the adaptability, intelligence, and practicality of the wireless power supply system for complex scenarios.

[0183] Example 1:

[0184] To verify the feasibility of the present invention in implementation, the present invention is applied to the dynamic wireless power supply scenario of an internal logistics robot system in a smart factory. The factory is located in the Industrial Park of Suzhou, Jiangsu Province, mainly producing electronic components and precision machining parts. The internal logistics system consists of multiple automatic guided vehicles (AGVs), which undertake tasks such as material distribution, in-station replenishment, and finished product transfer.

[0185] In the traditional power supply mode, AGV equipment needs to rely on fixed charging piles to complete periodic charging. According to the factory scheduling data, each AGV enters the charging state about 3 times a day on average, and each time it takes an average of 15 to 20 minutes, which not only significantly increases the equipment downtime, but also limits the scheduling ability of the flexible production line. In addition, the number of fixed charging piles deployed in the factory is limited and cannot cover all operating areas. Some AGVs need to detour a long way to the charging area when the battery is low, further increasing the ineffective driving time and affecting the operation efficiency.

[0186] In the embodiment of the present invention, an array-type wireless energy emission control module is installed in the airspace area above the north-south main channel of the factory, with a total coverage area of about 2,500 square meters. The wireless power supply system predicts the power supply trajectory based on the historical running trajectory, attitude angle, speed change, and power demand of the AGV through an improved self-attention echo state network, and generates a real-time control instruction set. The system adopts the magnetic resonance directional transmission method, and actively performs beam alignment, power adjustment, and frequency matching according to the predicted position of the AGV to achieve power supply while running.

[0187] After deploying the system of the present invention, it runs continuously for 31 days, collects operation data all-weather, and compares it with the similar data under the traditional charging mode.

[0188] Table 1 Comparison table of experimental data

[0189]

[0190]

[0191] From the perspective of power supply continuity and task coherence, in the traditional method, the AGV needs to stop for charging about 3 times a day, resulting in a relatively long non-operating time. In the present invention, through trajectory prediction and dynamic energy transmission technology, the device realizes continuous power supply, and the number of charging stops is reduced to 0, completely eliminating the task interruption problem during the charging process, and enabling the device to always be in a schedulable state.

[0192] In terms of working hours, the average daily working hours of the device in the traditional method is 20.8 hours, while under the support of the present invention, this value is increased to 22.4 hours, with an increase of 7.7%. At the same time, the daily reactive waiting time is reduced from 1.7 hours to 0.5 hours, a reduction of more than 70%, indicating that the system significantly improves the operation efficiency and time utilization rate.

[0193] In terms of power supply accuracy, the present invention effectively improves the energy transmission accuracy by introducing an improved self-attention echo state network and a reinforcement learning closed-loop control model. Experimental data shows that the average power supply error of the traditional method is 14.6%, while the present invention reduces it to 3.7%, and the error reduction amplitude is as high as 74.7%, with more accurate power supply and stronger stability. In addition, due to the large amplitude of attitude angle changes during the operation of the AGV device, it is difficult for the traditional wireless power supply system to cope with, so the mean square deviation of the fluctuations caused by attitude changes reaches 0.048 in the traditional system; while in the present invention, this value is reduced to 0.022, and the stability is improved by 54.2%, ensuring the energy alignment ability of the system under dynamic postures.

[0194] In terms of operation efficiency, the AGV supported by the traditional system completes an average of 161 tasks per day, while under the support of the system of the present invention, this value rises to 180 tasks, with an increase of 11.8%, indicating that the stability of the power supply system has a direct promoting effect on the overall logistics efficiency. In terms of response speed, the response time of the system of the present invention is 120 milliseconds, a reduction of 76% compared with 500 milliseconds of the traditional system, enabling it to quickly respond to changes in the moving state of the device and adapt to high-speed operation scenarios.

[0195] In addition, by statistically analyzing the annual operation efficiency, it is estimated that the present invention can save about 438 hours of charging waiting time for a single device every year, greatly improving the availability of the device and the overall scheduling efficiency of the system. At the same time, the annual operation cycle availability rate of the device is increased from 86.7% to 93.3%, an increase of 7.6 percentage points, which is conducive to deployment and application in scenarios with extremely high requirements for beat control such as manufacturing and warehousing.

[0196] It is worth mentioning that in the present invention, a self-learning mechanism based on incremental learning is also introduced, and the average daily adaptive control parameter adjustment frequency reaches 112 times, which is much higher than the traditional static power supply strategy, indicating that the system has strong online optimization and dynamic adaptation capabilities, and constructs a truly intelligent power supply closed-loop control system.

[0197] As described above, it is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A fast-response dynamic wireless power supply method applicable to mobile devices, characterized in that, It includes the following steps: S1. Real-time collect the operation status data of the mobile device; S2. Preprocess the operation status data and construct a power supply status sequence; S3. Input the power supply status sequence into a time series prediction network composed of an improved self-attention echo state network to predict the power supply position trajectory and power demand of the mobile device in future time steps; S4. Construct a dynamic power supply control instruction set according to the prediction results; S5. Based on the control instruction set, drive the array-type wireless energy emission control module to perform magnetic resonance or radio frequency energy directional transmission according to the power supply position trajectory; S6. Real-time collect the received power, voltage, current and mobile device attitude parameters at the mobile device receiving end; S7. Construct a state vector based on the collected actual power, voltage, current and mobile device attitude parameters, input it into the reinforcement learning adjustment model, output a corrected control instruction set, and dynamically adjust the emission parameters of the array-type wireless energy emission control module; S8. Store the prediction data, feedback error, control parameters and power supply performance indicators generated in each round of power supply task into the self-learning optimization module, and the self-learning optimization module updates the parameters of the self-attention echo state network and the control strategy of the reinforcement learning adjustment model through incremental learning.

2. The fast response dynamic wireless power supply method applicable to mobile devices according to claim 1, wherein The operation status data includes the mobile device position coordinates, velocity vector, attitude angle and instantaneous power demand.

3. The fast response dynamic wireless power supply method applicable to a mobile device according to claim 1, wherein The preprocessing includes normalization processing, standardization processing, time series reconstruction and feature vector fusion.

4. A fast response dynamic wireless power supply method applicable to mobile devices according to claim 1, characterized in that, The specific content of S3 includes: S31. Construct a time series prediction network, which is composed of an improved self-attention echo state network. The improved self-attention echo state network includes an input encoding layer, a self-attention layer, a gating mechanism, a spatial perception state update layer, and an output prediction layer. The input is a power supply state sequence where F i ={f1, f2, …, f m} represents the fusion feature vector at the i-th time step, m represents the feature dimension, p represents the number of time steps of the power supply state sequence, and the improved self-attention echo state network introduces a gating mechanism and a spatial perception weight matrix; S32. In the input encoding layer, for each fused feature vector F i perform a linear mapping to generate an intermediate feature representation; S33. In the self-attention layer, calculate the query vector, key vector and value vector based on the intermediate feature representation, and perform attention weighting calculation for all time steps to obtain an attention weighted vector: Among them, Z i represents the attention weighted vector, d k represents the key vector dimension, p represents the time steps of the power supply state sequence, softmax represents the normalization operation, Q i represents the query vector, K i represents the key vector, V j represents the value vector, T represents the transpose operation of the vector; S34. Input the attention weighted vector into the self-attention echo state network with a gating mechanism to calculate the gating parameter of the current time step: z i = σ(W z ·Z i + U z ·H i-1 ), r i = σ(W r ·Z i + U r ·H i-1 ); where z i represents the update gate, r i represents the reset gate, σ represents the Sigmoid activation function, W z and W r represent the gated input weights, U z and U r represent the gated recurrent weights, H i-1 represents the hidden state at time step i-1; Construct a spatial perception recursive connection matrix: W = α·exp(-β·||pos i - pos j || 2 ); where W represents the connection weight, α represents the connection strength factor, β represents the spatial decay coefficient, exp represents the natural exponential function, pos i and pos j represent the corresponding position vectors, and ||·|| 2 represents the Euclidean distance; Update the hidden state: H i = z i ⊙tanh(W in ·Z i + W·(r i ⊙H i-1 )); Among them, H i represents the hidden state at time step i, z i represents the update gate, ⊙ represents the Hadamard product, W in represents the input connection matrix, tanh represents the hyperbolic tangent activation function, pos j represents Z i represents the attention weighted vector, W represents the connection weight, r i represents the reset gate, H i-1 represents the hidden state at time step i - 1; S35. Input the final hidden state sequence {H1, H2, …, H p} into the output prediction layer, and calculate the power supply position trajectory and power demand for future time steps through a multi-layer perceptron: Among them, represents the power supply position trajectory, represents the power demand, (x k , y k , z k ) represents the spatial power supply coordinates at the k-th predicted time step, and P k represents the power demand at the corresponding time step.

5. A fast-response dynamic wireless power supply method applicable to mobile devices according to claim 1, characterized in that, The control instruction set includes the emission power adjustment value, frequency offset, beam direction angle and phase compensation amount of the wireless power supply system.

6. The fast response dynamic wireless power supply method applicable to a mobile device according to claim 1, characterized in that, The specific content of S4 includes: S41. Based on the prediction result sequence Construct a dynamic power supply control instruction set to generate a corresponding power supply parameter set for each time step j where A j represents the transmit power adjustment value, f j represents the frequency offset, θ j represents the beam direction angle, represents the phase compensation amount, (x j , y j , z j ) represents the spatial power supply coordinates at the j-th predicted time step; S42. Calculate the transmission power adjustment value based on the predicted power demand P j and the preset system loss parameter; S43. Calculate the frequency offset based on the predicted position (x j , y j , z j ): f j = f0 + γ·Δd j ; Among them, f j represents the frequency offset, f0 represents the system reference emission frequency, γ represents the frequency offset sensitivity coefficient caused by position change, and Δd j = ||(x j , y j , z j ) - (x j-1 , y j-1 , z j-1 )|| represents the Euclidean distance of the position change; S44. Calculate the beam direction adjustment angle based on the predicted position and the reference coordinate system of the emission array; θ j = arctan2(y j - y0, x j - x0); where θ j represents the beam direction adjustment angle, arctan represents the arctangent function, and (x0, y0) represents the projection coordinates of the center position of the power supply array on the horizontal plane; S45. Introduce a phase correction mechanism and calculate the phase compensation amount based on the distance change; Among them, represents the phase compensation amount, and f j represents the frequency offset, and d j represents the spatial distance between the center of the transmitting array and the receiving target, and c represents the propagation speed of electromagnetic waves in the air; S46. Integrate the power supply parameter sets generated at all time steps in time series to form a dynamic power supply control instruction set C = {C1, C2, …, C k}, and drive the array - type wireless energy emission module to execute dynamic power supply regulation.

7. A rapid response dynamic wireless power supply method applicable to mobile devices according to claim 1, characterized in that The specific content of S5 includes: S51. Based on the dynamic power supply control instruction set C = {C1, C2, …, C k}, sequentially extract the control parameter sets at each time step and transmit them to the wireless energy transmission control module for real-time configuration; S52. Receive the power supply position trajectory and synchronize it with the current clock, control the array-type wireless energy emission control module to switch to the dynamic power supply state, and perform position alignment and energy modulation operations at the set time steps; S53. Adjust the angle θ of the beam direction according to the control parameter set j , and adjust the spatial emission direction of the phased array or coil group to achieve precise alignment of the spatial steering with the current or predicted position of the target device, thereby forming a directional power supply channel S54. According to the phase compensation amount in the control parameter set Apply differential phase control to each transmitting unit in the array, perform transmitting phase synchronization and interference construction, so that the output waveform forms a focused interference energy beam in the target direction; S55. Adjust the output power level of the power amplifier module or the voltage regulation module according to the transmission power adjustment value A in the control parameter set, so as to achieve dynamic matching of the current power demand of the device, and make the output power of the control system change in real time with the change of the target power consumption; j ​ S56. According to the frequency offset f in the set of control parameters j , control the operating frequency of the radio frequency signal source or the magnetic resonance source to complete the real-time switching of frequency synthesis and transmitted signals; S57. Input the control parameters in consecutive time steps into the array-type wireless energy emission control module in the order of time series, drive the system to complete parameter refreshing, direction adjustment and emission control at each time step, ensure continuous coverage of energy in space and time, and achieve dynamic directional stable power supply for the mobile target device.

8. A fast-response dynamic wireless power supply method applicable to mobile devices according to claim 1, characterized in that, The specific content of S7 includes: S71. Obtain the actual power, voltage, current, and mobile device attitude parameters feedback by the receiver at the current time step, and splice and construct the state vector S by combining the transmission control instruction set and the target coordinates at the current time step j ; S72. Construct a reward function in the state-action space based on the received power error, attitude deviation amount and electrical parameter deviation; Among them, r j represents the reward function, describing the immediate reward at the current time step, and L j represents the loss function, where w1, w2, w3, and w4 respectively represent the weighting coefficients of each error, and ΔP j represents the difference between the transmit power and the receive power, and ΔΨ j represents the attitude deviation vector, ||·|| 2 represents the two-norm operation, and ΔV j represents the voltage deviation, and ΔI j represents the current deviation; S73. Using the state vector S j and the reward function, input them into the reinforcement learning adjustment model for training. The reinforcement learning adjustment model includes a policy network π θ (a j |S j ) and a value network V ω (S j ), where π θ represents the policy network characterized by the parameter θ, a j represents the control action, and V ω represents the value network characterized by the parameter ω; S74. Update the policy network parameters using a reinforcement learning algorithm based on policy gradient; Among them, \(J(\theta)\) represents the policy objective function, which is the expectation of the cumulative reward of the policy network within a time window of length \(p\). represents the gradient of the policy objective function with respect to the policy parameter \(\theta\), \(E\) represents the mathematical expectation, and \(\pi\) θ (a j |S j ) represents the policy network, \(r\) j represents the reward function, represents the gradient operation; S75. Execute the action output by the policy network As the corrected control instruction set Replace the original control parameters and transmit them to the array wireless energy emission control module for control execution, where Represents the corrected transmission power Represents the corrected frequency Represents the corrected direction angle Represents the corrected phase compensation 9. A fast-response dynamic wireless power supply method applicable to mobile devices according to claim 1, characterized in that, The specific content of S8 includes: S81. After each round of power supply task execution is completed, collect the prediction data generated during the task, and at the same time collect the control parameter sequence used during the power supply process, the receiver feedback data, and the power supply performance indicators recorded by the system; S82. Statistically generate task evaluation indicators for the power supply performance indicators collected during the task. The task evaluation indicators include average power error, mean square deviation of attitude change, and average control energy consumption, reflecting the control accuracy and execution efficiency of the power supply process; S83. Combine the control instructions, received feedback, prediction results, and power supply performance indicators in each time step to construct training samples, form a time series sample set based on state-action-feedback, and input it into the self-learning optimization module; S84. Update the parameters of the self-attention echo state network using the incremental learning strategy, and perform online backpropagation training on the self-attention echo state network using the new sample sequence to minimize the mean square error loss function between the predicted trajectory and the actual feedback position; S85. Based on the state information, control actions, and reward feedback in the time series sample set, use the experience replay mechanism to train the reinforcement learning adjustment model and update the policy network parameters; S86. Replace the corresponding modules in the original control framework with the parameters of the updated self-attention echo state network and the reinforcement learning adjustment model, and apply them to the next round of power supply tasks.

10. A fast-response dynamic wireless power supply system applicable to mobile devices, which executes the fast-response dynamic wireless power supply method applicable to mobile devices according to any one of claims 1 to 9, characterized in that, It includes: An operating state acquisition module for real-time acquisition of the operating state data of the mobile device; A data preprocessing module for normalizing, standardizing, reconstructing the time series, and fusing feature vectors of the operating state data to construct a power supply state sequence; A time series prediction module for inputting the power supply state sequence into a time series prediction network composed of an improved self-attention echo state network to predict the power supply position trajectory and power demand of the mobile device at future time steps; A control instruction generation module for constructing a dynamic power supply control instruction set according to the prediction results; An energy emission control module for driving the array-type wireless energy emission control module based on the control instruction set to perform magnetic resonance or radio frequency energy directional transmission according to the predicted power supply position trajectory; A receiving state acquisition module for real-time acquisition of received power, voltage, current, and mobile device attitude parameters at the mobile device receiver and generating feedback state data; A reinforcement learning adjustment module for constructing a state vector based on the feedback state data and inputting it into a reinforcement learning model composed of a policy network and a value network to output a corrected control instruction set to achieve dynamic adjustment of the transmission power, frequency, direction, and phase, and construct a closed-loop control system; A self-learning optimization module for storing the prediction data, feedback error, control parameters, and power supply performance indicators in each round of power supply tasks, and updating the parameters of the self-attention echo state network and the control strategy of the reinforcement learning adjustment model based on the incremental learning strategy.

Images