Wireless charging power adaptive adjustment method and system based on machine learning

Abstract

The invention relates to the technical field of wireless charging, and discloses a wireless charging power adaptive adjustment method and system based on machine learning, and the method comprises the steps: collecting and preprocessing multi-dimensional parameter data of a charging environment; learning the data by using a recurrent neural network, and establishing a correlation model of environmental factors and transmission loss; modeling a charging environment based on a graph structure, and extracting spatial relationship characteristics by using a graph neural network; a neighborhood sampling and dynamic pruning technology optimization model is adopted, and a hierarchical power decision system is constructed; adjusting the charging power by adopting feed-forward control according to the output of the decision-making system; according to the invention, by introducing the deep learning technology, accurate prediction and dynamic adjustment of the charging power are realized, and the charging efficiency and stability are improved; collaborative optimization of multiple charging points is realized by adopting graph structure modeling; and the calculation complexity is reduced through an optimization technology, so that the system is suitable for a large-scale application scene.

Description

Technical Field

[0001] The present invention relates to the field of wireless charging technology, and more specifically, to a method and system for adaptively adjusting wireless charging power based on machine learning. Background Art

[0002] With the widespread adoption of smart mobile devices and the rapid development of the Internet of Things (IoT), wireless charging technology has garnered widespread attention as a convenient power supply method. Wireless charging uses electromagnetic fields to transfer power from a transmitter to a receiver, eliminating the need for physical connections and significantly improving the user experience. Currently, wireless charging technology is widely used in consumer electronics such as smartphones and wearables, and is gradually expanding into areas such as smart homes and industrial manufacturing.

[0003] However, the power transmission efficiency of wireless charging is affected by a variety of environmental factors, such as temperature, humidity, and electromagnetic interference. These factors can cause dynamic changes in the transmission channel characteristics, affecting charging efficiency and stability. Traditional wireless charging systems typically use fixed power settings or simple feedback control mechanisms, which have difficulty accurately sensing and adapting to environmental changes, resulting in energy waste and low charging efficiency.

[0004] Furthermore, in multi-device charging scenarios, complex interactions exist between multiple charging points. Traditional charging systems often employ independent control methods, lacking effective coordination mechanisms, making it difficult to achieve globally optimal power configuration. Furthermore, with the rapid growth in the number of IoT devices and the ever-expanding scale of charging networks, higher demands are being placed on the system's real-time responsiveness and computational efficiency. Therefore, a wireless charging method is needed that can adaptively adjust power, achieve multi-point coordination, and support large-scale applications. Summary of the Invention

[0005] The present invention provides a wireless charging power adaptive adjustment method and system based on machine learning, which solves the technical problem of accurate prediction and dynamic adjustment of charging power during wireless charging in the related art, thereby improving charging efficiency and stability, and solving the technical problem of unbalanced power distribution in multi-device charging scenarios.

[0006] The present invention provides a method for adaptively adjusting wireless charging power based on machine learning, comprising the following steps: Collect and preprocess multi-dimensional parameter data of the charging environment to obtain a preprocessed environmental parameter data set; A recurrent neural network is used to learn the preprocessed environmental parameter data set, establish a correlation model between environmental factors and transmission loss, and output a transmission channel model; Based on the preprocessed environmental parameter dataset, multiple charging points and receiving devices in the charging environment are modeled as a graph structure. A graph neural network is used to extract spatial relationship features and output a spatial relationship model. Neighborhood sampling and dynamic pruning techniques are used to reduce computational complexity in the transmission channel model and spatial relationship model, and a hierarchical power decision system is constructed based on the optimized model output. According to the power regulation strategy output by the decision system, the charging power is adjusted in advance using a feedforward control method.

[0007] In a preferred embodiment, the step of collecting and preprocessing multi-dimensional parameter data of the charging environment includes: Collect environmental parameter data, equipment parameter data, charging effect data and time information; Clean, normalize and extract features from the collected data; Dynamically adjust the data collection frequency according to the rate of environmental change.

[0008] In a preferred embodiment, the step of using a recurrent neural network to learn the preprocessed environmental parameter data set includes: Build a recurrent neural network model based on the long short-term memory network, including the input layer, LSTM layer, attention layer and fully connected layer; Extract and fuse features at different time scales, including short-term features, medium-term features, and long-term features; The back-propagation algorithm is used to train the model to minimize the error between the predicted transmission loss value and the actual measured value.

[0009] In a preferred embodiment, the step of modeling multiple charging points and receiving devices in the charging environment as a graph structure includes: Construct a graph structure containing charging point nodes and device nodes, and the edges represent the influence relationship between nodes; Assigning feature vectors to nodes, including location coordinates, power parameters, charging status, and other information; Assign weights to edges to indicate the strength of influence.

[0010] In a preferred embodiment, the step of reducing computational complexity of the transmission channel model and the spatial relationship model by using neighborhood sampling and dynamic pruning technology includes: Design a two-level sampling strategy to control the number of neighbor samples for each node; Dynamically adjust the pruning threshold based on system load and temporarily ignore low-information connections; Adopt hybrid format storage strategy and dedicated sparse matrix calculation library to accelerate the operation process; Through model quantization and knowledge distillation technology, the knowledge of complex models is transferred to simple models, reducing the demand for computing resources.

[0011] In a preferred embodiment, the step of constructing a hierarchical power decision system based on the optimized model output includes: Construct a global-local two-layer decision-making architecture: the global layer coordinates power allocation among multiple charging points, and the local layer fine-tunes the power parameters of individual charging points; Construct a comprehensive optimization function to balance the goals of charging efficiency, power utilization, heat loss, and interference control; Use deep reinforcement learning methods to achieve power optimization decisions; A dynamic adaptation mechanism for continuous optimization, including incremental learning strategies, exploration-exploitation balance, and anomaly detection and recovery.

[0012] In a preferred embodiment, the step of adjusting the charging power in advance by adopting a feedforward control method includes: Build a predictive control model to predict future environmental parameter changes and calculate the optimal control sequence; Design control signal chains to convert decisions into specific hardware operations; Deploy a multi-level monitoring and feedback system to continuously evaluate control effectiveness; A closed-loop optimization system with continuous improvement is achieved to continuously improve the control effect.

[0013] In a preferred embodiment, the power optimization decision is realized by using a deep reinforcement learning method, specifically including: Using the Actor-Critic architecture, the Actor network outputs power regulation actions, and the Critic network evaluates the value of state-action pairs; The PPO algorithm is used, and the truncated objective function is used to limit the strategy update amplitude; Improve learning efficiency through parallel environment training and prioritized experience replay.

[0014] In a preferred embodiment, the neighborhood sampling strategy is based on the idea of ​​the GraphSAGE algorithm, which allocates sampling probabilities through the importance sampling algorithm and dynamically adjusts the number of samples for different nodes; Dynamic pruning technology evaluates the importance of connections based on attention weights, sets an adaptive threshold mechanism, introduces a time decay factor, and designs a periodic re-evaluation mechanism to prevent important connections from being permanently ignored.

[0015] In a preferred embodiment, a wireless charging power adaptive adjustment system based on machine learning is used to perform a wireless charging power adaptive adjustment method based on machine learning, including: A data acquisition module is used to collect and preprocess multi-dimensional parameter data of the charging environment to obtain a preprocessed environmental parameter data set; The transmission channel modeling module is used to use a recurrent neural network to learn the preprocessed environmental parameter data set, establish a correlation model between environmental factors and transmission loss, and output a transmission channel model; The spatial relationship modeling module is used to model multiple charging points and receiving devices in the charging environment as a graph structure based on the preprocessed environmental parameter dataset, extract spatial relationship features using a graph neural network, and output a spatial relationship model; An optimization decision module is used to reduce computational complexity of the transmission channel model and spatial relationship model using neighborhood sampling and dynamic pruning techniques, and to build a hierarchical power decision system based on the optimized model outputs; The control execution module is used to adjust the charging power in advance using a feedforward control method according to the power regulation strategy output by the decision system.

[0016] The beneficial effects of the present invention are: The machine learning-based wireless charging power adaptive adjustment method provided by the present invention establishes a correlation model between environmental factors and transmission loss by introducing recurrent neural networks and graph neural networks, realizes accurate prediction and dynamic adjustment of charging power, and significantly improves charging efficiency and stability.

[0017] By constructing a hierarchical power decision-making system and a feedforward control mechanism, the present invention can predict environmental changes in advance and make corresponding power adjustments, significantly reducing energy loss during the charging process and improving the environmental adaptability and robustness of the system.

[0018] In the multi-device charging scenario, the present invention adopts graph structure modeling and spatial relationship feature extraction technology to achieve collaborative optimization among multiple charging points, solve the problem of unbalanced power distribution in traditional methods, and enable the overall charging network to reach the optimal working state.

[0019] To address the challenges of large-scale application scenarios, this invention innovatively adopts optimization techniques such as neighborhood sampling and dynamic pruning, which significantly reduces the computational complexity. At the same time, it ensures model performance through methods such as knowledge distillation, enabling the system to achieve rapid response in large-scale scenarios such as smart cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a method for adaptively adjusting wireless charging power based on machine learning of the present invention; Figure 2 is a bar graph comparing the charging efficiencies of the present invention; Figure 3 It is a line graph showing the environmental adaptability of the present invention; Figure 4 It is a bubble chart showing the relationship between computational complexity and accuracy of the present invention. DETAILED DESCRIPTION

[0021] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0022] At least one embodiment of the present invention discloses a method for adaptively adjusting wireless charging power based on machine learning, such as Figure 1 As shown, the following steps are included: Step 1: Collect and preprocess multi-dimensional parameter data of the charging environment to obtain a preprocessed environmental parameter data set; It includes the following sub-steps: Step 1.1, data collection and preprocessing; Collect multi-dimensional parameter data of the charging environment, including but not limited to: Environmental parameters: temperature, humidity, electromagnetic field strength, environmental noise, etc.; Device parameters: device type, battery capacity, battery status, charging reception efficiency, etc.; Charging effect data: actual charging power, charging efficiency, transmission loss, heat generation, etc.; Time information: timestamp, charging duration, historical charging records, etc.

[0023] The collected data is cleaned, normalized and feature extracted, and features of different scales are converted into the same numerical range to facilitate subsequent modeling and analysis.

[0024] In some implementations, an adaptive sampling frequency strategy can be employed to dynamically adjust the data collection frequency based on the rate of environmental change. For example, when the environment is relatively stable (e.g., at night), the sampling frequency can be reduced to once per minute; while when the environment is rapidly changing (e.g., during periods of high activity), the sampling frequency can be increased to several times per second to capture transient changes.

[0025] For data with high noise levels, the Kalman filter algorithm can be applied for preprocessing, effectively filtering out random fluctuations while preserving the true trend. Kalman filtering involves both prediction and update steps, enabling optimal estimation of the true state based on historical data and current observations, improving data quality.

[0026] Step 1.2, recurrent neural network model construction; According to this application, it is necessary to build a recurrent neural network model based on the Long Short-Term Memory (LSTM) network to capture the temporal relationship between environmental parameters and transmission loss. The key structure of this model includes: Input layer: receives multi-dimensional environmental parameters and device parameters; LSTM layer: processes time series features and captures the pattern of parameter changes over time; Attention layer: assigns different importance weights to features at different time points; Fully connected layer: integrates features and maps them to transmission loss prediction values; The model processes sequence data through LSTM units to capture long-term and short-term dependencies, then focuses on the features of key time points through the attention mechanism, and finally outputs the prediction results through the fully connected layer.

[0027] In long sequence processing, the model can retain early important information and filter out irrelevant information, greatly improving prediction accuracy.

[0028] In some implementations, a Gated Recurrent Unit (GRU) can be used instead of an LSTM as the basic unit of a recurrent neural network. The GRU has a simpler structure, with only two gating mechanisms: an update gate and a reset gate. This reduces computational complexity and can even outperform the LSTM on certain datasets. The GRU uses these two gating mechanisms to control information flow, enabling the model to both retain long-term information and adapt to short-term changes.

[0029] Step 1.3, multi-scale feature fusion; In order to capture both short-term and long-term environmental change patterns, this application uses multi-scale feature fusion technology to effectively combine features at different time scales: Short-term features: Use sliding windows to extract high-frequency change features in the most recent time period (e.g., within 5 minutes); Medium-term features: extract the fluctuation trend features within a medium time period (such as within 1 hour); Long-term features: extract periodic change features over a long period of time (such as 1 day or 1 week); Multi-scale features are combined through weighted fusion, with the system automatically learning the importance weights of features at each time scale to achieve optimal fusion. This approach can simultaneously capture both instantaneous changes and long-term trends, significantly improving the comprehensiveness and accuracy of forecasts.

[0030] In some implementations, adaptive fusion methods can be used to dynamically adjust the weights of features at different time scales. For example, a gating mechanism can be used to adjust weights based on the current rate of environmental change, giving higher weight to long-term features when the environment is stable and paying more attention to short-term features when the environment is changing dramatically, thus achieving intelligent feature importance assessment.

[0031] Step 1.4, transmission loss prediction model optimization; According to this application, the system uses a backpropagation algorithm to train the model to minimize the error between the predicted transmission loss value and the actual measured value. The training process uses mean squared error as the loss function and gradually adjusts the model parameters through gradient descent. Regularization terms are also introduced to prevent overfitting and improve the model's generalization ability.

[0032] The model optimization process in this application focuses on practicality and robustness, rather than simply pursuing low error on the training set. Techniques such as cross-validation are used to evaluate the model's performance on unseen data to ensure that the model can adapt to various changes in the actual application environment.

[0033] In some embodiments, a more complex loss function can be used to optimize the model. For example, a weighted sum of mean squared error and mean absolute error can be combined as a loss function to balance the sensitivity to large errors and small errors, thereby improving the adaptability of the model to different error ranges.

[0034] To further improve model performance, ensemble learning methods can be used to combine the prediction results of multiple basic models. By training multiple LSTM models with different structures or initialization parameters and then combining their prediction results using weighted averaging or voting, the prediction bias of individual models can be effectively reduced, improving the overall prediction stability and accuracy.

[0035] Step 2: Use a recurrent neural network to learn the preprocessed environmental parameter data set, establish a correlation model between environmental factors and transmission loss, and output a transmission channel model; It includes the following sub-steps: Step 2.1, graph structure construction; According to this application, based on the actual layout of the charging environment, a graph structure G = (V, E) is constructed: Node set V: includes charging point nodes (transmitter) and device nodes (receiver); Edge set E: represents the influence relationship between nodes, including the charging relationship from charging point to device, the interference relationship between charging points, the competition relationship between devices, etc. Each node has a feature vector, including: Charging point node characteristics: location coordinates, maximum output power, current working status, etc.; Device node characteristics: location coordinates, battery capacity, current power level, urgency of charging needs, etc.

[0036] Each edge has a weight, which represents the strength of the impact. The initial weight can be determined based on factors such as physical distance and power size, and then dynamically adjusted through learning.

[0037] In a specific implementation of a smart office, nodes can be represented by fixed charging points in different areas such as conference rooms, work areas, and rest areas, as well as employees' mobile devices such as phones, laptops, and tablets. Edge weights can be initialized based on the distance between the device and the charging point, for example, using a value inversely proportional to the square of the distance.

[0038] In some implementations, a heterogeneous graph network can be used to represent the charging environment, with different node and edge types having different feature representations. For example, a graph structure can be defined as a four-tuple consisting of a node set, an edge set, a node type mapping function, and an edge type mapping function. Node types can be categorized as charging points, mobile devices, fixed devices, etc.; edge types can be categorized as charging relationships, interference relationships, competitive relationships, etc.

[0039] In order to handle the dynamic changes of nodes (such as the joining and leaving of mobile devices), a dynamic graph update mechanism can be used.

[0040] When a new device is detected, the system creates a new node, initializes its feature vector, and calculates the relationship between the new node and existing nodes, performing a partial graph update rather than rebuilding the entire graph. This incremental graph update approach significantly reduces computational overhead and enables the system to adapt to highly dynamic scenarios.

[0041] Step 2.2, graph convolutional network layer construction; According to this application, the graph convolutional network architecture consists of three main components: a graph convolution layer, a batch normalization layer, and an activation function layer. The core of the graph convolution operation is to perform neighborhood aggregation on node features and update the node representation using the node's first-order neighbor information.

[0042] In the implementation of this application, the graph convolutional network specifically includes: Two layers of graph convolutional layers, the first layer outputs 64-dimensional features, and the second layer outputs 32-dimensional features; Each layer is followed by a batch normalization layer to help stabilize the training process - ReLU is used as the activation function to provide nonlinear transformation capabilities; Add residual connections to alleviate the gradient disappearance problem.

[0043] It’s important to note that in smart city scenarios, graph convolutional networks can effectively capture the spatial relationships between charging points in different areas. For example, when a large-scale event is held in a certain area of ​​the city, charging demand in that area will increase. Charging points in surrounding areas can obtain this information through graph convolutional networks and adjust their power strategies in advance.

[0044] In some embodiments, a spatial-temporal graph convolutional network (ST-GCN) can be used instead of a regular graph convolutional network to capture both spatial relationships and temporal evolution.

[0045] The key to ST-GCN is the introduction of convolution operations in the time dimension, which treats the features of nodes at different time points as time series for processing. This enables the model to simultaneously process the temporal evolution of spatial topology and node features, such as capturing the cyclical change pattern of charging demand.

[0046] A graph convolutional network with adaptive topology can be used to dynamically learn the connectivity between nodes. The core idea is to use the adjacency matrix as a learnable parameter, dynamically constructing the graph structure based on the similarity or correlation of node features. This approach can discover implicit node relationships in the data, beyond predefined physical distances, and is particularly suitable for charging environments with complex interference or nonlinear effects.

[0047] Step 2.3, graph attention network layer construction; According to this application, the graph attention network consists of a multi-head attention mechanism, where each attention head independently calculates attention weights and then merges the results. The core idea of ​​graph attention is to assign different importance weights to different neighboring nodes, allowing the network to focus on the most relevant node information.

[0048] In the implementation of this application, the graph attention network specifically includes: 8 attention heads, each head calculates different aspects of node relationships; Each attention head outputs 16-dimensional features, which are combined to form 128-dimensional features; Use LeakyReLU as the activation function in attention weight calculation; Add skip connections to help transfer original feature information.

[0049] In application examples in industrial production environments, there are often functional dependencies between multiple production equipment in a factory, and graph attention networks can automatically learn the strength of these dependencies.

[0050] When a key device is running out of power, the graph attention network will assign a higher attention weight to the device, giving priority to its charging needs and avoiding production line stalls due to insufficient power.

[0051] Step 2.4, node embedding and relationship representation; This application uses a multi-layer graph neural network to learn node representations and generate embedding vectors that contain topological structure and node feature information. Specific implementations include: First, the local structural features of the nodes are extracted through a two-layer graph convolutional network; Then, the long-distance dependencies between nodes are captured through the graph attention network layer; Finally, the features are mapped to an embedding space of fixed dimension through a fully connected layer; In smart home scenarios, the system can learn family members' device usage habits and location preferences. For example, if it detects that user A typically places her phone at the bedroom charging station at 10 p.m., the system will adjust the power parameters of that charging station in advance to achieve optimal charging results.

[0052] At the same time, the system can identify sources of charging interference (such as interference with nearby charging points when a microwave oven is operating) and mitigate the interference by adjusting the operating frequency of the affected charging points.

[0053] In some embodiments, a contrastive learning approach may be employed to train node embeddings by maximizing the mutual information between related node representations while minimizing the mutual information between unrelated nodes.

[0054] Contrastive learning trains charging points in the same area as positive sample pairs and charging points in different areas as negative sample pairs, so that nodes with similar functions or relationships are closer in the embedding space, which is conducive to subsequent collaborative optimization.

[0055] To handle node heterogeneity, a meta-path guided node embedding method can be used.

[0056] First, several types of meta-paths are defined (such as "charging point-device-charging point" or "device-charging point-device"). Node embeddings are then calculated for each meta-path. Finally, the embeddings from different meta-paths are fused using an attention mechanism. This approach captures the complex semantic relationships between different node types and enhances the expressiveness of charging network modeling.

[0057] Step 3: Based on the preprocessed environmental parameter dataset, multiple charging points and receiving devices in the charging environment are modeled as a graph structure, and a graph neural network is used to extract spatial relationship features and output a spatial relationship model. It includes the following sub-steps: Step 3.1, neighborhood sampling technique; The neighborhood sampling strategy in this application is based on the GraphSAGE algorithm, which reduces computational complexity by controlling the number of neighbors sampled for each node. Specific implementations include: Design a two-level sampling strategy: a maximum of 15 nodes are sampled from the first-level neighbors and a maximum of 10 nodes are sampled from the second-level neighbors; Using the importance sampling algorithm, the sampling probability is allocated according to the importance of the node (such as the urgency of the device power, the power size of the charging point, etc.); Through batch processing technology, a fixed number of node subsets are processed each time instead of the entire graph structure; According to this application, in smart city application scenarios, when there are thousands of charging points distributed in the city, traditional methods need to process millions of node relationships, and the computational complexity is unacceptable.

[0058] In some implementations, an adaptive neighborhood sampling strategy can be employed to dynamically adjust the number of samples based on node characteristics and network structure. For example, for a key charging point at the center of the network, the number of neighbor samples can be increased to 25 to 30 to obtain more comprehensive environmental information; while for low-power charging points at the edge of the network, the number of neighbor samples can be reduced to 5 to 8 to conserve computing resources.

[0059] Historical information can be combined to guide the current sampling process, retaining important historical neighbor information through a memory mechanism. Specifically, the system maintains a neighbor importance cache that records the influence of each neighbor over multiple time steps in the past, and prioritizes neighbor nodes with greater historical influence during sampling.

[0060] This approach is particularly suitable for scenarios where stable relationships exist in the charging network, such as long-term impact relationships between fixed devices.

[0061] Step 3.2, attention-guided dynamic pruning; According to this application, the dynamic pruning algorithm uses attention weights to evaluate the importance of connections, automatically identifying and temporarily ignoring low-information connections. Specific implementations include: Set up an adaptive threshold mechanism to dynamically adjust the pruning threshold based on system load; Introducing a time decay factor to gradually reduce the weight of connections that have not been used for a long time; Design a periodic re-evaluation mechanism to prevent important connections from being permanently ignored; The core of attention-guided pruning is to set a dynamic threshold. When the attention weight of a connection falls below the threshold, it is temporarily ignored, thereby reducing the amount of computation. It should be noted that the system will periodically re-evaluate all connections to ensure that important information is not missed.

[0062] It should be noted that in the public service area application example, the charging system in the airport terminal can set a lower threshold (such as 0.05) during off-peak hours to retain more connections and improve prediction accuracy; During peak hours, the threshold is increased (e.g., 0.15), ignoring more low-importance connections and prioritizing computing efficiency and charging needs of key equipment.

[0063] An uncertainty-based pruning strategy can be introduced to consider not only the size of the attention weight but also the uncertainty of its estimation.

[0064] The system estimates the variance of each attention weight through Monte Carlo sampling or ensemble learning methods, and combines the mean and variance to make decisions. This method can retain connections that are not the most important but are relatively stable, improving the robustness of the system.

[0065] Step 3.3, sparse matrix calculation optimization; This application uses sparse matrix representation and computational techniques to accelerate the computation of graph neural networks. Specific implementations include: Adopt a mixed format storage strategy: Use CSR (Compressed Sparse Row) format to store node feature matrices; Use CSC (Compressed Sparse Column) format to store the adjacency matrix; Dynamically evaluate and select the storage format that best suits the current computing task; Leverage dedicated sparse matrix computation libraries for acceleration: Use multi-threaded parallel computing on the CPU; Leverage the GPU's sparse tensor acceleration cores on supported hardware; Use block matrix multiplication algorithm to improve cache hit rate; According to this application, in the smart home scenario application example, even on a resource-constrained home smart central control device (such as a smart home control center with a low-power processor), through sparse matrix optimization, the system can still complete all calculations of the charging network containing 30 device nodes within 200 milliseconds, meeting real-time response requirements.

[0066] In some implementations, a mixed-precision computing strategy can be employed to further improve computational efficiency while maintaining accuracy. The system converts most computations to half-precision floating-point (FP16) format, while using single-precision floating-point (FP32) format only for critical computational steps (such as the final loss calculation). This mixed-precision strategy can increase computational speed by approximately 1.5-2 times while maintaining algorithmic performance.

[0067] An automatic format selection algorithm can be introduced to dynamically select the optimal sparse representation format based on the sparsity pattern of the current matrix and the characteristics of the computing task.

[0068] The system maintains a performance evaluation model to determine the best format based on the following factors: Matrix dimensions and proportion of non-zero elements; The type of computation (e.g., matrix-vector multiplication, matrix-matrix multiplication); Hardware characteristics (e.g., cache size, SIMD instruction set); The expected access pattern (row-major or column-major).

[0069] Step 3.4, model quantization and knowledge distillation; According to this application, through model quantization and knowledge distillation technology, computing resource requirements are further reduced, enabling efficient deployment on edge devices. Specific implementations include: Model quantization technology: Adopt a mixed precision strategy and use different quantization precisions for different layers; Key layers (such as the attention calculation layer in the figure) maintain FP16 precision, and non-key layers use INT8 precision; Use quantization-aware training to simulate quantization errors during training; Knowledge distillation implementation: Teacher model: 3 layers of graph convolution + 2 layers of graph attention + 2 layers of full connection, with approximately 500K parameters; Student model: 2 layers of graph convolution + 1 layer of graph attention + 1 layer of full connection, with about 100K parameters; The dual-temperature distillation method was used, and the soft target temperature parameter T was set to 3; The weight ratio of hard labels and soft labels in the distillation loss function is 0.3:0.7; The core idea of ​​knowledge distillation is to let the simple model (student model) learn the behavior of the complex model (teacher model) rather than just the correct answer.

[0070] In this way, the student model can maintain the performance of the teacher model as much as possible while significantly reducing the number of parameters.

[0071] In an application example in an industrial production environment, each local control unit in a factory workshop only needs to deploy a lightweight student model, while a complete teacher model is deployed in the central control room.

[0072] The student model is responsible for handling daily charging scheduling tasks. When encountering complex abnormal situations, it will request the teacher model in the central control room to assist in decision-making.

[0073] In some implementations, a progressive knowledge distillation approach can be employed to gradually compress the model through a multi-stage distillation process. First, the model is distilled from the original complex model (e.g., 5-layer graph convolution + 3-layer graph attention) to a medium-complexity model (e.g., 3-layer graph convolution + 2-layer graph attention), and then from the medium-complexity model to a lightweight model (e.g., 2-layer graph convolution + 1-layer graph attention).

[0074] This multi-stage distillation method can alleviate the information loss from directly going from complex to simple, achieving higher compression rates while maintaining model performance.

[0075] A method combining feature distillation and relational distillation can be introduced to learn not only the output distribution of the teacher model, but also the intermediate layer feature representations and inter-node relationship knowledge. This multi-objective distillation method can more comprehensively transfer the knowledge of the teacher model, allowing the lightweight student model to better simulate the behavior and performance of the teacher model while maintaining low computational complexity.

[0076] Step 4: Neighborhood sampling and dynamic pruning techniques are used to reduce computational complexity in the transmission channel model and spatial relationship model, and a hierarchical power decision system is constructed based on the optimized model output. It includes the following sub-steps: Step 4.1: Construction of global-local two-layer decision-making architecture; The hierarchical decision-making architecture in this application includes two levels: global and local, to achieve collaborative optimization at different scales. Specific implementations include: Global layer decision module: The core structure is a multi-layer fully connected neural network, and the input nodes are graph embedding vectors and system state vectors; Use attention pooling operation to compress graph structure information into a fixed-dimensional vector; Output the power allocation strategy and priority score of each charging point; Local layer decision module: Each charging point is independently deployed and responsible for fine-tuning local power parameters; Adjust specific power parameters based on the prediction results of the recurrent neural network; Receive the constraints of the global layer and optimize within the allowed range; The two-layer decision modules collaborate through a two-way communication mechanism: The global layer passes constraints and priority information to the local layer; The local layer feeds back the execution results and local environment status to the global layer; Adopt an iterative update strategy to gradually optimize the global configuration.

[0077] Application example in public service areas: In large public spaces such as airport terminals, the global decision-making module monitors the grid load and device distribution across the entire area. If a sudden increase in foot traffic in area A is detected, the power quota for charging points in that area will be increased in advance. At the same time, each local charging point will make fine adjustments within the power quota allocated at the global level based on the actual situation of the surrounding equipment to ensure the charging efficiency of user equipment.

[0078] In some embodiments, a three-tier decision-making architecture may be employed, with an intermediate tier added to handle coordination optimization issues within a region.

[0079] The three-tier architecture includes: Global layer: responsible for resource allocation and load balancing of the entire charging network; Regional layer: responsible for the coordination between multiple charging points in a specific area; Local layer: responsible for adjusting the specific parameters of a single charging point.

[0080] This three-layer architecture is particularly suitable for large-scale scenarios, such as commercial complexes or industrial parks, where there are clear regional divisions. The regional layer can reduce the computational burden of the global layer while providing a more macro optimization perspective than the local layer.

[0081] An adaptive hierarchical structure can be introduced to dynamically adjust the decision-making level according to the network size and complexity. In small-scale networks (such as home environments), it can be simplified to a single-layer or two-layer structure; In large-scale networks (such as smart cities), it can be expanded to a multi-layer structure.

[0082] The system uses the following indicators to determine whether the hierarchy needs to be adjusted: Number of network nodes; Inter-region communication delay; The complexity of the optimization problem; Availability of computing resources.

[0083] Through the adaptive hierarchical structure, the system can maintain efficient operation in scenarios of different scales, avoiding structural redundancy in small scenarios while meeting the coordination needs in large scenarios.

[0084] Step 4.2, multi-objective optimization function construction; This application designs an optimization function that integrates multiple objectives and balances different performance indicators. The specific implementation includes: Main optimization goals: Charging efficiency target: ; in, Indicates the overall charging efficiency; Represents the collection of all devices receiving charging; An index representing a single device; Representation device Received charging power; Representation device Environmental parameters; Representation device The charging efficiency function is related to the received power and environmental conditions.

[0085] Power utilization target: ; in, Indicates the overall power utilization; Represents the set of all charging points; An index representing a single charging point; Indicates charging point Current output charging power; Indicates charging point The formula calculates the ratio of the actual output power of all charging points to their maximum power capacity, reflecting the resource utilization efficiency of the charging system.

[0086] Heat loss target: ; in, Indicates overall heat loss; Represents the set of all charging points; An index representing a single charging point; Indicates charging point Heat loss function; Indicates charging point Current output charging power; Indicates charging point Environmental parameters; Represents the collection of all devices receiving charging; An index representing a single device; Representation device Heat loss function; Representation device Received charging power; This formula calculates the total heat loss of the charging point and the receiving device, including charging point losses (related to output power and environment) and device losses (related to received power).

[0087] Interference control objectives: ; in, Indicates the overall interference intensity; Represents the set of all charging points; 、 Respectively represent 、 An index of different charging points; Indicates charging point and Interference function between Indicates charging point Current output charging power; Indicates charging point Current output charging power; Indicates charging point and The formula calculates the sum of interference between all charging point pairs, and the interference intensity is related to the output power and distance of the charging point.

[0088] Comprehensive optimization function: ; in, represents the comprehensive optimization objective function; Represents the weight coefficient of the charging efficiency target; The weight coefficient representing the power utilization target; Indicates the weight coefficient of the heat loss target; represents the weight coefficient of the interference control target; Represents the charging efficiency objective function value; represents the power utilization objective function value; Represents the heat loss objective function value; Represents the value of the disturbance control objective function.

[0089] Weight adaptive adjustment mechanism: Dynamically adjust weight parameters according to scene characteristics; Prioritize charging efficiency in office environments; Prioritize power utilization in public spaces; Increase the weight of heat control in temperature-sensitive environments; In smart home application examples: the system can automatically adjust the optimization weight according to the preferences of family members. For example, for users who pay attention to the health of the device, the system will increase the weight of heat control ( to 0.4), reducing the charging speed but extending the battery life; and for users who often need fast charging, the system will increase the weight of charging efficiency ( Increased to 0.6), giving priority to ensuring charging speed.

[0090] In some embodiments, a Pareto multi-objective optimization method may be used. Instead of combining multiple objectives using a weighted sum, multiple objective functions are optimized simultaneously to find a Pareto optimal solution set.

[0091] Specifically, the system uses a multi-objective evolutionary algorithm (such as NSGA-II or MOEA / D) to generate a series of non-dominated solutions and then selects the most appropriate solution based on the current system state and user preferences.

[0092] The multi-objective optimization problem is expressed as: ; in, The objective function vector representing the multi-objective optimization problem; 、 、 Respectively represent 、 、 An optimization goal, such as charging efficiency; Indicates the number of optimization objectives; represents the decision variables, including the power configuration parameters of all charging points; Represents the feasible solution space of the decision variables, which contains all power configuration schemes that meet the system constraints; Indicates that all objective functions are minimized simultaneously.

[0093] Step 4.3, optimization algorithm implementation; This application uses deep reinforcement learning methods to achieve power optimization decisions, specifically including: Actor-Critic Architecture: Actor network: a 3-layer fully connected network with the output layer using the Tanh activation function to limit the range of action; Critic network: a 4-layer fully connected network that outputs value estimates for state-action pairs; Shared feature extraction layers to enhance feature utilization efficiency.

[0094] PPO (Proximal Policy Optimization) algorithm implementation: Use a truncated objective function to limit the magnitude of policy updates: ; in, Represents the truncated objective function of the PPO algorithm; Represents the parameters of the policy network; represents the empirical expectation at time step t; represents the probability ratio of the new and old strategies; represents the advantage function estimate; Represents a truncation function that limits the input value to a specified range; Represents the cutoff range parameter, usually set to 0.2.

[0095] Use a multi-step return calculation method to balance bias and variance; Add entropy regularization term to encourage policy exploration.

[0096] Parallel environment training: Simulate multiple charging environments for training at the same time; Use experience replay buffer to improve sample utilization efficiency; Use priority experience replay technology to focus on learning key samples; Application example in an industrial production environment: On an automated production line in a factory, the charging requirements of various devices have obvious differences in priority.

[0097] Step 4.4, dynamic adaptation and online learning; This application implements a dynamic adaptation mechanism for continuous optimization, specifically including: Incremental learning strategy: Keep the core model parameters and only update the output layer weights; Use a sliding window dataset to prioritize the latest data; Set up a learning rate decay mechanism to reduce the update step size over time; Explore and exploit balance: Use the ε-greedy strategy to try non-optimal actions with small probability; As the system stability improves, the exploration probability is gradually reduced; Reset exploration rate regularly to cope with sudden changes in the environment; Anomaly detection and recovery: Monitor decision-making performance indicators and detect abnormal performance degradation; When an exception occurs, fall back to the safe configuration and collect exception data; Use abnormal data for targeted training to enhance robustness; In a smart office scenario, the system can adapt to daily changes and special events in the office environment. For example, if it detects a pattern of employees arriving in large numbers on Monday mornings, it will increase the power allocation to charging points in the office area in advance. When a large meeting is scheduled, the charging points in the conference room will be pre-adjusted to high-capacity mode.

[0098] Step 5: Based on the power regulation strategy output by the decision system, the charging power is adjusted in advance using a feedforward control method; It includes the following sub-steps: Step 5.1, predictive control model construction; According to this application, the predictive control model integrates time series prediction and control theory, and can adjust power in advance according to environmental trends. Specific implementations include: Future state prediction: Use the recurrent neural network model in step 1 to predict the changes in environmental parameters in the next 5 to 30 minutes; The prediction accuracy decreases as the prediction time increases; Output the predicted environmental parameter vector for subsequent control decisions.

[0099] Control strategy planning: Calculate the optimal control sequence based on the current state and the predicted environment; Use model predictive control to plan control actions for multiple time steps in the future; Only the first step of the control sequence is actually executed, and subsequent steps are adjusted as the prediction is updated.

[0100] Confidence Assessment: Calculate the confidence interval of the prediction results and evaluate the reliability of the prediction; When the confidence level is lower than the threshold, it automatically switches to the traditional feedback control mode; The confidence assessment is based on the ratio of the prediction variance to the maximum acceptable variance.

[0101] In the smart home scenario, the system can predict the interference to the nearby charging environment when family members use high-power appliances such as microwave ovens, and adjust the power and frequency parameters of the relevant charging points 5 to 10 seconds in advance.

[0102] When the system detects that the user is about to return home (such as a smart door lock unlocking signal or location information approaching home), it will start the charging points of the user's commonly used devices in advance to ensure that the user can use the devices immediately after returning home.

[0103] In some embodiments, an integrated prediction method can be used to integrate the results of multiple prediction models to improve prediction accuracy and robustness.

[0104] The system runs multiple forecasting models simultaneously, including: Time series prediction model based on recurrent neural network; State estimation model based on Kalman filter; Trend analysis models based on statistical methods; Rule-based expert knowledge model.

[0105] The system integrates the predictions of each model through a dynamic weighting mechanism, where the weights are determined by the model's recent performance. This integrated prediction approach maintains stable prediction accuracy across a wide range of environmental conditions, making it particularly well-suited for complex and changing charging environments.

[0106] Optionally, an adaptive prediction time domain mechanism can be introduced to dynamically adjust the prediction time range according to the rate of environmental change and prediction reliability.

[0107] When the environment is stable, the system can extend the prediction time domain to 30 to 60 minutes, providing a longer-term planning perspective; In cases where the environment changes rapidly or uncertainty is high, the system will shorten the prediction time domain to 5 to 10 minutes to ensure the accuracy of short-term predictions.

[0108] The adjustment of the prediction time domain takes into account factors such as current prediction confidence and environmental stability to achieve intelligent time domain control.

[0109] Step 5.2, control signal generation and execution; According to this application, a complete control signal chain is designed to convert the decision into specific hardware operations. The specific implementation includes: Control signal type: Power regulation signal: controls the output power, divided into 16 discrete gears or continuous adjustment; Frequency adjustment signal: fine-tune the resonant frequency and optimize transmission efficiency; Mode switching signal: switch between high-efficiency mode, low-power mode, and safe mode; Working status signal: control sleep, wake-up, emergency shutdown and other states; Signal adaptation and conversion: Adapt the signal according to the hardware characteristics of different charging points; Convert the decision value into a control command that can be recognized by the hardware; Add checksum to ensure signal transmission accuracy; Communication and execution mechanism: Prioritize low-latency communication protocols (such as the dedicated MQTT protocol); Set up hierarchical priority queues to give priority to emergency control signals; Implement automatic retransmission and confirmation mechanism to ensure control reliability.

[0110] It should be noted that in an industrial production environment, the system customizes control signal formats and communication protocols for different production equipment.

[0111] For critical control equipment, redundant communication links and real-time response protocols are used to ensure that control signals are delivered within 5 milliseconds; For general sensor devices, low-power communication protocols and batch control signals are used to reduce system overhead.

[0112] At the same time, all control signals are encrypted to prevent unauthorized access or malicious interference.

[0113] In some embodiments, a progressive control signal generation strategy may be adopted to avoid sudden changes in the control signal causing impact on the system.

[0114] When the system needs to transition from the current power state to the target state, instead of sending the target control signal directly, a series of transition signals are generated according to a smooth function, where the smooth function can be a simple linear function or a more complex S-shaped curve.

[0115] This gradual control is particularly suitable for equipment or environments that are sensitive to power fluctuations, and can reduce electromagnetic interference and temperature fluctuations caused by sudden power changes.

[0116] Step 5.3, real-time monitoring and feedback; Based on this application, a multi-level monitoring and feedback system was built to continuously evaluate the control effect. The specific implementation includes: Sensor network deployment: Deploy multiple types of sensors in the charging environment; Temperature sensor: monitors charging point and device temperature to prevent overheating; Electromagnetic field sensors: monitor power transmission efficiency and interference; Current / voltage sensor: accurately measures actual charging power; Data collection and processing: Use distributed data collection architecture to reduce network transmission pressure; Edge nodes perform preliminary data filtering and anomaly detection; The central node performs data fusion and global analysis; Effect evaluation and deviation analysis: Calculate the deviation between the actual effect and the expected target; Deviation analysis is based on the relative error between actual and predicted values; When the deviation exceeds the threshold, the adjustment mechanism is triggered; In the public service area, a hierarchical monitoring network is deployed for the charging system in the airport terminal. Each charging point is equipped with basic sensors, a local monitoring node is set up in each charging area, and a central monitoring center is set up for the entire terminal.

[0117] When the system detects an abnormally high temperature at a charging point in area A, it will immediately reduce the power of the charging point and issue a maintenance alarm; At the same time, analyze whether the abnormality is caused by environmental factors (such as direct sunlight) or equipment failure, and take corresponding measures.

[0118] In some implementations, a federated edge computing architecture can be employed for data processing, reducing data transmission and improving system responsiveness. Each edge node not only collects data but also runs a lightweight neural network for preliminary analysis and decision-making, reporting only aggregated results and anomalies to the central node. Edge nodes regularly synchronize model parameters with the central node, integrating distributed learning with global optimization.

[0119] Step 5.4, closed-loop optimization mechanism; According to this application, a continuously improved closed-loop optimization system is realized, which continuously improves the control effect. The specific implementation includes: Historical data management: Establish a hierarchical data storage structure and determine the retention period based on the value of the data; Key scenario data (such as abnormal events and new equipment) is permanently stored; Routine data is stored in a rolling manner and updated regularly; Periodic model updates: Retrain the model on a predetermined schedule (e.g., weekly); Incremental training method, which preserves historical knowledge while adapting to new changes; Model version management, supporting rollback when performance degrades; System performance analysis: Generate performance reports regularly to identify system bottlenecks; Key indicators include: prediction accuracy, control response time, and energy utilization; Adjust optimization strategies and algorithm parameters based on analysis results.

[0120] In smart city applications, the control system of the city-level charging network achieves comprehensive closed-loop optimization. Every morning, the system automatically analyzes the operating data of the past 24 hours, identifies peak and low charging periods, and adjusts resource allocation strategies accordingly. A comprehensive model update is performed every weekend to integrate newly collected data to optimize the prediction model; a system performance report is generated once a month to analyze the charging efficiency and energy utilization of each region and provide decision support for the city's energy management department.

[0121] In some embodiments, a multi-time scale optimization strategy may be used to perform corresponding optimization adjustments based on changes in different time scales: Short-term optimization (minute level): fine-tuning parameters based on real-time status changes; Mid-term optimization (daily): adjust control strategies based on intraday pattern changes; Long-term optimization (monthly): Analyze long-term trends and optimize algorithms and model architectures.

[0122] Different optimization methods and objective functions are used for different timescales. For example, short-term optimization uses gradient descent to directly adjust parameters, medium-term optimization uses Bayesian optimization to adjust hyperparameters, and long-term optimization uses evolutionary algorithms or reinforcement learning to adjust model structure. This multi-timescale optimization strategy can simultaneously address short-term fluctuations and long-term trend changes, maintaining the continuous and efficient operation of the system.

[0123] Active learning mechanisms can be introduced to collect the most valuable data samples for model updates. The system evaluates the uncertainty or expected information gain of data samples and actively selects the most valuable scenarios for in-depth observation and learning.

[0124] The system also implements automated A / B testing to continuously evaluate the performance differences between new and existing strategies. While ensuring safety, the system deploys new strategies in a small area and compares them with a control group. Performance data is then collected to determine whether to promote them. This gradual optimization and deployment approach reduces the risk of policy updates while ensuring continuous improvement in system performance.

[0125] Application examples of this implementation: In order to verify the actual effect of the wireless charging power adaptive adjustment method based on machine learning proposed in this application, a real application example in a smart office scenario is given below.

[0126] Application scenarios: This application scenario is a smart office area located in a modern office building with a total area of ​​approximately 2,000 square meters, including multiple functional spaces: an open office area, 4 meeting rooms, 2 rest areas and 1 reception hall.

[0127] A total of 48 wireless charging points have been deployed throughout the office, located on desks, walls, and public facilities in various functional areas. Approximately 120 employees in the office regularly use over 200 wireless charging devices, including smartphones, tablets, laptops, and smart wearables.

[0128] The charging environment in this office area has the following characteristics: High device density: During peak hours (such as meetings), more than 20 devices may need to be charged simultaneously in a single area; Multiple interference sources: changes in the electromagnetic environment caused by office equipment (such as monitors and printers), communication equipment, and personnel movement; Variable usage patterns: Charging requirements vary greatly in different scenarios, such as meetings, collaboration, and personal work; Priority requirements are clear: key equipment such as conference equipment and presentation equipment need to be given priority for charging.

[0129] Implementation process example: System deployment and initialization: The following hardware facilities are deployed in this smart office area: 48 wireless charging points with adjustable power, ranging from 5W to 30W; 24 environmental sensors to monitor parameters such as temperature, humidity, and electromagnetic field strength; 8 edge computing nodes, distributed in various functional areas; 1 central server, deployed in the equipment room.

[0130] System initialization includes the following steps: Establish digital twin mapping of charging points and devices to form an initial graph structure; Collect two weeks of basic operating data, including equipment usage patterns and environmental parameter changes; Use the collected data to train initial models, including a multi-scale transmission channel model and a graph-structured spatial relationship model; Deploy the optimized model to edge computing nodes and central servers.

[0131] Multi-scale transmission channel modeling example: In actual deployment, the system first conducts detailed modeling of the charging environment in each area.

[0132] Taking the open office area as an example, the system collected the following data: Temperature data: The office temperature typically fluctuates between 23-26°C during working hours; Population density data: There are three peaks of population density at 9:00 in the morning, 13:00 after lunch, and 16:00 during tea time; Electromagnetic environment data: Printers and copiers will generate short-term electromagnetic interference when working.

[0133] Based on this data, the system constructed an LSTM network model with an input dimension of 12 (including various environmental parameters), a hidden layer dimension of 64, and an output dimension of 4 (transmission loss-related parameters). After two weeks of training, the model was able to accurately predict changes in transmission efficiency at different times and locations.

[0134] For example, the system observed that transmission losses in Area G (near the printing area) increased by an average of 18% between 3:00 PM and 4:00 PM due to concentrated office equipment use and frequent personnel activity. Based on this prediction, the system automatically increased the output power of the charging points in this area by 15% at 2:50 PM to compensate for potential losses and maintain stable charging efficiency.

[0135] Example of graph structure spatial relationship modeling: In the conference room scenario, the system builds a heterogeneous graph network consisting of charging points and devices. Taking Conference Room 3 as an example, it is equipped with 6 charging points, and a typical meeting usually has 8 to 12 devices that need to be charged.

[0136] The system models this spatial relationship through graph convolutional networks and graph attention networks and finds that: Charging point No. 2 in the center of the conference table has the strongest connection with devices at the four seating positions; When a projector is used in a meeting room, a charging point close to the projector device may cause slight interference; The device at the meeting host's position typically has the highest charging priority.

[0137] Based on these relationship models, the system automatically adjusts the charging points in the conference room to their optimal configuration 30 minutes in advance when it detects a large conference is scheduled. For example, during an important client meeting with 12 attendees, the system automatically increases the power of the charging point near the speaker to its maximum (30W) while simultaneously adjusting any potentially interfering charging points to a low-interference mode (reducing power by 20% but optimizing frequency parameters). This ensures stable operation of presentation equipment while meeting the charging needs of other attendees.

[0138] Sparse computing optimization example: The computational challenges faced by the system when processing the entire office area's charging network were significant. Before optimization, processing the entire graph network, consisting of over 200 nodes (charging points and devices), took over two seconds to compute, failing to meet real-time response requirements.

[0139] After applying the sparse compute optimization, the system performs the following optimizations: For each charging point, only the 15 most influential neighbor nodes are retained for calculation; Use dynamic pruning technology to temporarily ignore connections with attention weights below 0.08; The sparse matrix is ​​stored in CSR format and multi-threaded parallel computing is used.

[0140] After optimization, the computation time for a network of the same size was reduced to 120 milliseconds, reducing computing resource usage by 78%, while reducing decision accuracy by only 3%. This enables the system to process the entire office area's charging network in real time on ordinary office-grade servers, without relying on high-performance computing clusters.

[0141] When the office area held an open day event and the number of visitors surged to over 200, the system successfully controlled the calculation time to within 200 milliseconds by further increasing the pruning threshold (from 0.08 to 0.12) and reducing the number of neighbors retained by each node (from 15 to 10), ensuring the system continued to run smoothly.

[0142] Hierarchical power optimization decision example: In daily operation, the system uses a two-tier decision-making structure to manage charging resources in the entire office area. In a typical workday example: Global level decision making: The system detects that two important meetings are taking place simultaneously at 10:30 am, and a surge in charging demand is expected in the conference room; The global decision-making layer allocates 35% of the day's charging power quota to the conference room area; Reduce the power quota of rest areas with few people to 10%; Reserve 50% of the power quota for open office areas to meet daily work needs.

[0143] Local level decision making: In conference room 3, the local decision module identified that the battery level of the demonstration tablet was less than 30%; Immediately raise the priority of the demonstration device to the highest level and allocate the maximum charging power; At the same time, it detects that the battery levels of three participants' mobile phones are sufficient (>70%), and automatically reduces the charging power of these devices; Dynamically adjust the frequency parameters of each charging point to minimize mutual interference.

[0144] Through this hierarchical decision-making process, the system achieves optimal global resource allocation while meeting the needs of key devices. During a month-long test, the number of low-battery device incidents during meetings decreased by 92%, while overall charging efficiency increased by 24%.

[0145] Feedforward control execution example: The system's feedforward control capability excels in responding to rapidly changing environments. Here's a specific example: Time: Wednesday 15:20; Environmental changes: The system calendar detects that a general meeting (approximately 80 people) will be held in the conference room at 16:00; Prediction results: During the pre-meeting preparation period (15:40-16:00), the demand for charging in large conference rooms will surge, which may lead to competition and reduced efficiency. System feedforward control action: Starting at 15:25, the power of the six charging points in the conference room will be gradually increased by 30% to prepare for the upcoming load peak; At 15:30, a reminder will be pushed to the devices of relevant personnel, suggesting that the devices be placed on the charging point in advance; 15:35 Pre-optimize the charging points in the large conference room and adjust the frequency parameters of each charging point to minimize interference; At 15:40, peak mode is started, dynamic load balancing is activated, and the power distribution of each charging point is adjusted according to the actual number of connected devices.

[0146] Through this feedforward control, the system successfully avoided competition for charging resources at the start of the meeting. All participating devices were charged to an average of over 85% before the meeting began, ensuring stable operation during the meeting. Compared to the previous passive response mode, the charging satisfaction rate increased by 45% and the waiting time for device charging was reduced by 68%.

[0147] Technical effect verification: like Figures 2 to 4 Shown are the comparison of charging efficiency; environmental adaptability performance; and the relationship between computational complexity and accuracy.

[0148] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A wireless charging power adaptive adjustment method based on machine learning, characterized in that: The following steps are involved: Collect and preprocess multi-dimensional parameter data of the charging environment to obtain a preprocessed environmental parameter data set; A recurrent neural network is used to learn the preprocessed environmental parameter data set, establish a correlation model between environmental factors and transmission loss, and output a transmission channel model; Based on the preprocessed environmental parameter dataset, multiple charging points and receiving devices in the charging environment are modeled as a graph structure. A graph neural network is used to extract spatial relationship features and output a spatial relationship model. Neighborhood sampling and dynamic pruning techniques are used to reduce computational complexity in the transmission channel model and spatial relationship model, and a hierarchical power decision system is constructed based on the optimized model output. According to the power regulation strategy output by the decision system, the charging power is adjusted in advance using a feedforward control method.

2. The method for adaptively adjusting wireless charging power based on machine learning according to claim 1, characterized in that: The step of collecting and preprocessing multi-dimensional parameter data of the charging environment includes: Collect environmental parameter data, equipment parameter data, charging effect data and time information; Clean, normalize and extract features from the collected data; Dynamically adjust the data collection frequency according to the rate of environmental change.

3. The method for adaptively adjusting wireless charging power based on machine learning according to claim 1, characterized in that: The step of using a recurrent neural network to learn the pre-processed environmental parameter data set includes: Build a recurrent neural network model based on the long short-term memory network, including the input layer, LSTM layer, attention layer and fully connected layer; Extract and fuse features at different time scales, including short-term features, medium-term features, and long-term features; The back-propagation algorithm is used to train the model to minimize the error between the predicted transmission loss value and the actual measured value.

4. The method for adaptively adjusting wireless charging power based on machine learning according to claim 1, wherein: The step of modeling multiple charging points and receiving devices in the charging environment into a graph structure includes: Construct a graph structure containing charging point nodes and device nodes, and the edges represent the influence relationship between nodes; Assigning feature vectors to nodes, including location coordinates, power parameters, charging status, and other information; Assign weights to edges to indicate the strength of influence.

5. The method for adaptively adjusting wireless charging power based on machine learning according to claim 1, characterized in that: The step of reducing computational complexity of the transmission channel model and the spatial relationship model by using neighborhood sampling and dynamic pruning technology includes: Design a two-level sampling strategy to control the number of neighbor samples for each node; Dynamically adjust the pruning threshold based on system load and temporarily ignore low-information connections; Adopt hybrid format storage strategy and dedicated sparse matrix calculation library to accelerate the operation process; Through model quantization and knowledge distillation technology, the knowledge of complex models is transferred to simple models, reducing the demand for computing resources.

6. The method for adaptively adjusting wireless charging power based on machine learning according to claim 1, characterized in that: The step of constructing a hierarchical power decision system based on the optimized model output includes: A global-local two-layer decision-making architecture is constructed. The global layer coordinates the power distribution among multiple charging points, and the local layer fine-tunes the power parameters of individual charging points. Construct a comprehensive optimization function to balance the goals of charging efficiency, power utilization, heat loss, and interference control; Use deep reinforcement learning methods to achieve power optimization decisions; A dynamic adaptation mechanism for continuous optimization, including incremental learning strategies, exploration-exploitation balance, and anomaly detection and recovery.

7. The method for adaptively adjusting wireless charging power based on machine learning according to claim 1, characterized in that: The step of adjusting the charging power in advance by adopting the feedforward control method includes: Build a predictive control model to predict future environmental parameter changes and calculate the optimal control sequence; Design control signal chains to convert decisions into specific hardware operations; Deploy a multi-level monitoring and feedback system to continuously evaluate control effectiveness; A closed-loop optimization system with continuous improvement is achieved to continuously improve the control effect.

8. The method for adaptively adjusting wireless charging power based on machine learning according to claim 6, characterized in that: The power optimization decision-making is realized by using a deep reinforcement learning method, specifically including: Using the Actor-Critic architecture, the Actor network outputs power regulation actions, and the Critic network evaluates the value of state-action pairs; The PPO algorithm is used, and the truncated objective function is used to limit the strategy update amplitude; Improve learning efficiency through parallel environment training and prioritized experience replay.

9. The method for adaptively adjusting wireless charging power based on machine learning according to claim 5, characterized in that: The neighborhood sampling strategy is based on the GraphSAGE algorithm, which allocates sampling probabilities through the importance sampling algorithm and dynamically adjusts the number of samples for different nodes. Dynamic pruning technology evaluates the importance of connections based on attention weights, sets an adaptive threshold mechanism, introduces a time decay factor, and designs a periodic re-evaluation mechanism to prevent important connections from being permanently ignored.

10. A wireless charging power adaptive adjustment system based on machine learning, used to execute the wireless charging power adaptive adjustment method based on machine learning according to any one of claims 1 to 9, characterized in that: include: A data acquisition module is used to collect and preprocess multi-dimensional parameter data of the charging environment to obtain a preprocessed environmental parameter data set; The transmission channel modeling module is used to use a recurrent neural network to learn the preprocessed environmental parameter data set, establish a correlation model between environmental factors and transmission loss, and output a transmission channel model; The spatial relationship modeling module is used to model multiple charging points and receiving devices in the charging environment as a graph structure based on the preprocessed environmental parameter dataset, extract spatial relationship features using a graph neural network, and output a spatial relationship model; An optimization decision module is used to reduce computational complexity of the transmission channel model and spatial relationship model using neighborhood sampling and dynamic pruning techniques, and to build a hierarchical power decision system based on the optimized model output; The control execution module is used to adjust the charging power in advance using a feedforward control method according to the power regulation strategy output by the decision system.

Images