A digital twin electromagnetic wave propagation prediction regulation system and method based on data fusion

Abstract

The application relates to the technical field of electric wave propagation prediction, and discloses a digital twin electric wave propagation prediction regulation and control system and method based on algorithm fusion, wherein the system comprises an environment perception layer, a twin modeling layer, a trajectory prediction layer, a dynamic regulation layer and a closed-loop optimization layer which are coupled in a top-down step-by-step linkage mode; the environment perception layer is configured to construct a three-dimensional environment model with electromagnetic characteristic parameters of ground objects; the twin modeling layer is configured to construct an electric wave propagation digital twin; the trajectory prediction layer is configured to predict a future moving trajectory of a user; the dynamic regulation layer is configured to cooperatively optimize phase shift parameters of a spatial intelligent electromagnetic time sequence reconstruction metasurface array and base station power and beam parameters; and the closed-loop optimization layer is configured to complete error traceability analysis and incremental model correction, thereby forming a full-link regulation and control closed loop. The application can improve electric wave prediction accuracy in a complex scene, realize pre-compensation for weak communication coverage, and effectively improve communication coverage quality and spectrum utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of radio wave propagation prediction technology, and in particular to a digital twin radio wave propagation prediction and control system and method that integrates induction and computation. Background Technology

[0002] With the large-scale deployment of 5G mobile communication technology and the in-depth iterative research and development of 6G mobile communication technology, high-frequency communication bands such as millimeter waves and terahertz, with their advantages of ultra-large bandwidth and ultra-high transmission rate, have become the core carriers supporting high-end applications such as massive terminal access, ultra-low latency service transmission, and immersive communication. However, high-frequency electromagnetic waves have shorter wavelengths and weaker diffraction and penetration capabilities, making them highly susceptible to obstruction by complex scenarios such as buildings, vegetation, and terrain undulations. This results in severe path propagation loss and signal fading, creating numerous weak communication coverage blind spots. This can easily lead to problems such as link jitter, momentary disconnection, and communication discontinuity, greatly restricting the service stability of high-frequency communication in complex terrain scenarios.

[0003] Meanwhile, the widespread adoption of connected vehicles, the Internet of Things, industrial internet, and immersive mobile services has led to an exponential increase in the number of wireless terminals accessing the entire network. This exacerbates the contradiction between limited spectrum resources and explosive communication demands, placing higher demands on the accuracy of radio wave propagation prediction, the real-time performance of network control, and the utilization rate of spectrum resources. Therefore, constructing a high-precision, low-latency, dynamically iterative, and predictably controllable intelligent radio wave propagation prediction system has become a core challenge that needs to be overcome in next-generation mobile communication technologies.

[0004] Existing technologies for predicting and controlling radio wave propagation still have many limitations, which are explained in detail below: In environmental modeling, the modeling level is low and the dynamic adaptability is poor. Traditional radio wave modeling mostly relies on empirical statistical models or deterministic ray tracing mechanism models. Scene data depends on manual surveying and static low-precision geographic information, which cannot achieve refined and automated batch annotation of the electromagnetic characteristics of ground objects, resulting in a large deviation between the simulation scene and the real electromagnetic environment. At the same time, the static modeling mode cannot adapt to dynamic environments such as urban building renewal, seasonal changes in vegetation, and changes in scene occlusion, and the model's generalization ability is weak.

[0005] In terms of real-time simulation of global electromagnetic fields, there is an inherent technical contradiction between modeling accuracy and real-time simulation. Traditional pure physical mechanism models offer high solution accuracy, but rely on large-scale iterative numerical calculations, resulting in high computational load and long simulation time, making them unable to support real-time dynamic control. Pure data-driven models offer fast computation speed and good real-time performance, but lack constraints from physical propagation rules, leading to low fitting reliability and overfitting in complex scenarios, failing to meet the dual requirements of high accuracy and low latency in global electromagnetic field simulations.

[0006] In terms of control mode, existing systems generally adopt a passive response mechanism, which only triggers resource adjustment after detecting a decline or interruption in the quality of the communication link. This results in a significant control delay, which cannot meet the communication continuity requirements in high-speed mobile scenarios and lacks the ability to predict user mobile behavior.

[0007] In terms of system architecture, communication functions and environmental perception functions have long been separated. Perception data is only used for environmental monitoring and is not deeply integrated with communication dynamic control, resulting in a waste of perception resources. At the same time, it is impossible to use the echo information of communication signals to further improve the accuracy of environmental perception.

[0008] In terms of spectrum resource scheduling, most schemes adopt static allocation rules or simple cognitive radio technology, lacking the ability to make dynamic optimization decisions based on the real-time electromagnetic environment, and making it difficult to achieve efficient utilization of spectrum resources according to the spatiotemporal variation characteristics of radio wave propagation.

[0009] In summary, existing technologies cannot construct a closed-loop system that integrates sensing, prediction, and control, and are ill-suited to meet the future mobile communication system's demands for high-precision, high-dynamic, and highly intelligent radio wave propagation management. Summary of the Invention

[0010] To address the aforementioned issues, this invention proposes a digital twin radio wave propagation prediction and control system and method that integrates induction and computation. By constructing a digital twin of radio wave propagation, a closed-loop linkage between environment, propagation, and control can be achieved, thereby improving communication coverage quality and spectrum utilization efficiency in complex scenarios.

[0011] The technical solution adopted in this invention is as follows: A digital twin radio wave propagation prediction and control system integrating sensing and computation includes a top-down, hierarchically coupled environment perception layer, twin modeling layer, trajectory prediction layer, dynamic control layer, and closed-loop optimization layer. The environmental perception layer is configured to acquire multi-source remote sensing and terrain elevation data of the target area, construct a three-dimensional environmental model with electromagnetic characteristic parameters of ground objects, and output it to the twin modeling layer. The twin modeling layer is configured to be constrained by the physical laws of electromagnetic propagation, integrate the prior knowledge model of radio wave propagation with the enhanced sparse measured data, construct a digital twin of radio wave propagation, generate a full-area electromagnetic background field and output it to the trajectory prediction layer. The trajectory prediction layer is configured to predict the user's future movement trajectory, identify potential weak communication coverage areas by combining electromagnetic background field, and generate a pre-prediction command to output to the dynamic control layer. The dynamic control layer is configured to collaboratively optimize the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array, base station power, and beam parameters according to the pre-judgment instruction, thereby enhancing user coverage. The closed-loop optimization layer is configured to compare the measured communication data on site with the twin prediction values, complete the error source analysis and incremental model correction, and form a full-link control closed loop.

[0012] Furthermore, in the environmental perception layer, multi-source remote sensing and terrain elevation data of the target area are acquired, and a three-dimensional environmental model with attached electromagnetic characteristic parameters of ground objects is constructed, including: Acquire multi-source remote sensing data and terrain elevation data of the target area, preprocess the raw data and register it to a unified coordinate system and resolution; Based on the processed data, land features are classified to distinguish between building, vegetation, water, and bare land scene types; Based on the material structure and electromagnetic properties of various land features, the corresponding electromagnetic loss, reflection and scattering parameters are labeled for the 3D scene.

[0013] Furthermore, in the twin modeling layer, constrained by the physical laws of electromagnetic propagation, a priori knowledge model of radio wave propagation is integrated with enhanced sparse measured data to construct a digital twin of radio wave propagation, generating a full-area electromagnetic background field, including: The physical laws of electromagnetic propagation are used as the physical mechanism for network training, and the enhanced sparse measured data are used as the supervision signal. A lightweight digital twin of radio wave propagation is obtained through collaborative training of the physical constraints and the sparse measured data. The lightweight radio wave propagation digital twin is used to deduce the global electromagnetic characteristics and generate the global electromagnetic background field.

[0014] Furthermore, in the twin modeling layer, the physical mechanism soft constraint penalty term is used to replace the numerical iterative solution process of electromagnetic propagation; the sparse measured data is used to drive network convergence; and the lightweight radio wave propagation digital twin is used to deduce global electromagnetic feature data and integrate it to generate the global electromagnetic background field.

[0015] Furthermore, in the trajectory prediction layer, predicting the user's future movement trajectory includes: Collect historical location information of user terminals and extract motion features; collect the user's current location and correct positioning errors. Based on the historical location information and the current location, the user's future movement trajectory is inferred to generate a predicted location sequence; The predicted location sequence is spatially superimposed and matched with the electromagnetic background field of the entire region to establish a spatiotemporal coupling relationship between the user's movement path and the state of the electromagnetic field.

[0016] Furthermore, in the trajectory prediction layer, potential weak coverage areas for communication are identified by combining the electromagnetic background field, and advance prediction instructions are generated, including: Based on the spatiotemporal coupling correlation, the estimated received signal strength is calculated for each predicted location, and areas with signal strength below the communication threshold are selected as potential weak coverage areas. Assess the probability of communication interruption risk in each potential weak coverage area and classify them into obstructed weak areas, edge weak areas, or overlapping interference weak areas; For different risk levels and weak area types, generate advance prediction instructions carrying location information, risk level information and appropriate control strategy information.

[0017] Furthermore, in the dynamic control layer, the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array are collaboratively optimized in relation to the base station power and beam parameters, including: Analyze the pre-judgment instructions and divert the control tasks according to the weak area type and risk level in the pre-judgment instructions; For different weak region types, beamforming optimization, power equalization optimization, or metasurface phase compensation optimization are performed respectively to obtain the base station parameters and the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array; the spatial intelligent electromagnetic timing reconstruction metasurface array includes multiple tunable electromagnetic units, and the multiple tunable electromagnetic units realize spatial electromagnetic wave reconstruction by timing control of the phase shift state of each unit. By integrating the base station parameters with the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array, a joint control scheme is generated.

[0018] Furthermore, the dynamic control layer also includes: According to the joint control scheme, before the user arrives in a potentially weak coverage area, the phase shift calibration of the spatial intelligent electromagnetic timing reconstruction metasurface array and the pre-adjustment of the base station's transmit power and beam pointing are completed. When multiple sets of the aforementioned spatial intelligent electromagnetic timing reconstruction metasurface arrays are deployed in the scene, multi-array collaborative linkage control is initiated. The real-time environmental perception data and coverage status data generated by the control operation are transmitted back to the environmental perception layer to assist in updating the three-dimensional environmental model.

[0019] Furthermore, in the closed-loop optimization layer, the measured communication data from the field is compared with the twin prediction values ​​to complete error source analysis and incremental model correction, including: Collect measured communication data and network link performance data from user terminals, compare the measured communication data with the simulated predicted values ​​at the corresponding positions of the digital twin, and quantify the error; when the error exceeds a preset threshold, trace the source to environmental adaptation error, modeling mechanism error, or equipment parameter error; Based on the source tracing results, incremental parameter correction is performed using the newly added measured communication data to update the digital twin; the corrected model parameters are then synchronized to the twin modeling layer.

[0020] A digital twin radio wave propagation prediction and control method integrating inductive and computational fusion includes: Acquire multi-source remote sensing and terrain elevation data of the target area, and construct a three-dimensional environmental model with attached electromagnetic characteristic parameters of ground objects; Constrained by the physical laws of electromagnetic propagation, and integrating the prior knowledge model of radio wave propagation with enhanced sparse measured data, a digital twin of radio wave propagation is constructed to generate the electromagnetic background field of the entire region. Predict the user's future movement trajectory, identify potential weak communication coverage areas based on the electromagnetic background field, and generate advance prediction instructions; Based on the aforementioned pre-judgment instructions, the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array are collaboratively optimized with the base station power and beam parameters to achieve enhanced user coverage; By comparing the actual on-site communication data with the twin prediction values, error source analysis and incremental model correction are completed, forming a closed loop for full-link control.

[0021] The beneficial effects of this invention are as follows: 1. This invention constructs a complete system consisting of five inseparable, progressively interconnected layers: an environmental perception layer, a twin modeling layer, a trajectory prediction layer, a dynamic control layer, and a closed-loop optimization layer. This achieves integrated operation across the entire communication chain, from perception and modeling to prediction, control, and optimization. This architecture breaks away from the fragmented nature of traditional communication systems, directly using environmental perception results to construct a digital twin of radio wave propagation. Based on the electromagnetic simulation results of the twin, coverage risks are predicted in advance, and the control module is driven to perform forward-looking operations in real time. Simultaneously, the twin is continuously corrected using real-time measured data, forming a self-evolving closed-loop system. Compared to traditional passive response systems, this invention can adjust resources before communication link quality deteriorates, significantly shortening system response time, effectively solving the link interruption problem caused by high-frequency communication blockage, and greatly improving the continuity and stability of communication coverage in high-speed mobile scenarios.

[0022] 2. This invention achieves deep integration and coordinated scheduling of communication, sensing, and computing resources. The environmental sensing layer acquires refined geographic and electromagnetic information of the target area through multi-source heterogeneous data fusion technology, providing precise environmental input for communication control. Computing resources provide computing power support for the construction of lightweight digital twins, prediction of user trajectories, and optimization of control parameters. Meanwhile, the measured data generated during communication is used to improve sensing accuracy and the fidelity of the digital twin. This integrated design of communication, sensing, and computing avoids data silos and resource redundancy between different functional modules, enabling the system to dynamically allocate various resources according to actual business needs and environmental conditions. In the context of increasingly scarce spectrum resources, this significantly improves the system's spectrum utilization and overall capacity.

[0023] 3. This invention constructs a global modeling system driven by both soft constraints of physical mechanisms and lightweight sparse measured data, balancing the completeness of global electromagnetic characterization with online real-time inference capabilities. The proposed lightweight digital twin construction method for electromagnetic wave propagation, driven by both knowledge and data, balances the constraints of physical laws with the adaptability of measured data. By integrating the fundamental physical laws of electromagnetic propagation into the training process of the neural network, it avoids the technical contradictions of high computational load and slow inference in pure physical mechanism modeling, and insufficient physical reliability due to the reliance on massive dense sampling in pure data-driven models. On the one hand, it relies on a priori knowledge base of electromagnetic wave propagation to solidify electromagnetic physical rules, ensuring that the spatial distribution of the global electromagnetic field conforms to the real propagation mechanism and improving the generalization and adaptation capabilities for complex terrain scenarios. On the other hand, it relies only on a small number of sparse sensing nodes to collect data, combined with global data augmentation technology to fill spatial data gaps, eliminating the need for large-scale dense deployment of acquisition equipment and reducing engineering deployment hardware costs. Meanwhile, a physical mechanism soft constraint penalty term is introduced, abandoning the traditional high-iteration numerical solution method. On the basis of completing the complete global electromagnetic characterization of the entire region, the computational load is greatly reduced and the overall deduction latency is compressed, realizing the compatibility and unity of global complete modeling and online real-time deduction, and solving the technical problem of balancing the integrity and real-time performance of global electromagnetic modeling.

[0024] 4. The predictive coverage enhancement technology employed in this invention achieves precise spatiotemporal coverage through accurate prediction of user movement trajectories and proactive identification of weak coverage areas. The dynamic control layer can specifically optimize the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface, the base station's transmit power, and beam pointing according to the different causes of weak coverage areas. It also supports the coordinated control of multiple sets of spatial intelligent electromagnetic timing reconstruction metasurfaces to form superimposed coverage enhancement beams. This multi-parameter joint optimization control method not only effectively overcomes signal attenuation caused by obstacles but also reduces the overall energy consumption of the base station while ensuring user communication quality, improving the system's energy efficiency and coverage range. This better meets the needs of future high-density, high-mobility wireless communication scenarios.

[0025] 5. This invention constructs an incremental closed-loop correction mechanism, enabling long-term autonomous evolution of the digital twin model and extending its effective lifespan. Most existing electromagnetic simulation models are static open-loop structures with fixed parameters after model finalization. However, with urban building renovations, seasonal environmental changes, and continuous electromagnetic environment drift, model operating errors accumulate, resulting in a short effective lifespan and high maintenance costs. This invention employs an incremental local parameter correction strategy, updating weights only in areas where errors exceed limits, eliminating the need for global retraining of the entire model and significantly reducing computational power consumption. Simultaneously, it accurately distinguishes error sources such as environmental adaptation errors, mechanism modeling errors, and equipment parameter drift errors, continuously feeding measured operating data back to the twin modeling layer to dynamically update the global electromagnetic background field. This allows the system to adapt to dynamically changing real electromagnetic environments for extended periods, extending the stable operating cycle of the digital twin system and reducing long-term maintenance and iteration costs. Attached Figure Description

[0026] Figure 1 This is a diagram of the architecture of a digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods, according to Embodiment 2 of the present invention.

[0027] Figure 2 This is a flowchart of a digital twin radio wave propagation prediction and control method based on the fusion of inductive and computational methods, according to Embodiment 3 of the present invention.

[0028] Figure 3 This is a flowchart illustrating the dynamic control stage of Embodiment 3 of the present invention. Detailed Implementation

[0029] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0030] Example 1 This embodiment provides a digital twin electromagnetic wave propagation prediction and control system that integrates induction and computation. The system adopts a five-layer hierarchical coupled and linked architecture, including, from top to bottom, an environmental perception layer, a twin modeling layer, a trajectory prediction layer, a dynamic control layer, and a closed-loop optimization layer. Each layer is functionally independent but has deeply interconnected data flow, jointly constructing an integrated operation system for global electromagnetic modeling, risk prediction, collaborative control, and closed-loop iteration. The environment perception layer is configured to acquire multi-source remote sensing and terrain elevation data of the target area, construct a three-dimensional environment model with attached electromagnetic characteristic parameters of ground objects, and output it to the twin modeling layer.

[0031] The twin modeling layer is configured to use the physical laws of electromagnetic propagation as constraints, integrate the prior knowledge model of radio wave propagation with the enhanced sparse measured data, construct a digital twin of radio wave propagation, generate a full-area electromagnetic background field, and output it to the trajectory prediction layer.

[0032] The trajectory prediction layer is configured to predict the user's future movement trajectory, identify potential weak communication coverage areas by combining electromagnetic background field, and generate a pre-prediction command to be output to the dynamic control layer.

[0033] The dynamic control layer is configured to collaboratively optimize the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array, base station power, and beam parameters according to the pre-judgment instruction, thereby enhancing user coverage.

[0034] The closed-loop optimization layer is configured to compare the measured communication data on site with the twin prediction values, complete the error source analysis and incremental model correction, and form a full-link control closed loop.

[0035] Preferably, the environmental perception layer acquires multi-source remote sensing data and terrain elevation data of the target area, preprocesses the acquired data and registers them to a unified coordinate system and resolution; performs intelligent classification of ground features based on the processed data and outputs the classification results; constructs a three-dimensional environmental model based on the classification results and terrain elevation data; and labels the electromagnetic characteristic parameters of the corresponding ground features on the three-dimensional environmental model and outputs them to the twin modeling layer.

[0036] Specifically, the environmental perception layer acquires optical images, synthetic aperture radar images, digital elevation models, and laser point cloud data of the target area through multi-source remote sensing. The acquired raw data undergoes geometric distortion correction, radiometric normalization, and clutter denoising in sequence, and all data is registered to a unified geographic coordinate system. In this embodiment, in the preferred configuration example for a typical urban scene, the grid resolution is dynamically set according to the scene. Typical urban scenes use a 10m level (compared to traditional decimeter-level fine modeling, reducing the data volume to 1%~5%), significantly simplifying the data volume while retaining electromagnetic boundary constraints. A lightweight multi-scale convolutional neural network is used to automatically classify land features, identifying different land feature types such as buildings, vegetation, water bodies, roads, and bare land. For each grid cell, fixed electromagnetic parameters (with dimensionless relative permittivity) are assigned according to different land feature types: buildings have a relative permittivity of 6.0 and conductivity of 0.05 S / m; dry vegetation has a relative permittivity of 3.5 and conductivity of 0.01 S / m; water bodies have a relative permittivity of 80 and conductivity of 0.5 S / m; roads have a relative permittivity of 4.0 and conductivity of 0.01 S / m; and bare land has a relative permittivity of 4.5 and conductivity of 0.01 S / m. Based on the classification results and elevation data, a three-dimensional mesh model is constructed, and the above electromagnetic characteristic parameters such as permittivity, conductivity, and surface roughness are labeled on the corresponding land feature grids in the model. Finally, a high-precision three-dimensional environment model is generated and output to the twin modeling layer.

[0037] It should be noted that by fusing multi-source heterogeneous data and semantic-level ground feature classification, the physical environmental characteristics of the target area can be accurately restored, providing precise physical boundary constraints for subsequent radio wave propagation modeling and effectively reducing the deviation between environmental data and actual scenes in traditional modeling methods. At the same time, this layer abandons the redundant modeling method of massive fine geometric meshes, and only performs fine mesh subdivision for ground feature areas that affect radio wave propagation. While retaining complete electromagnetic boundary constraints, it simplifies the data volume and reduces the burden on backend real-time inference. Traditional modeling directly performs full-scale fine mesh subdivision, resulting in a huge amount of data that cannot meet the requirements of second-level online inference.

[0038] Preferably, the twin modeling layer constructs a knowledge-sparse measured data dual-driven architecture, and uses physical mechanism soft constraint penalty terms to replace the traditional large-scale iterative solution of ray tracing, so as to realize lightweight global electromagnetic rapid deduction.

[0039] Specifically, the twin modeling layer pre-imports and integrates various prior knowledge models of radio wave propagation, such as the ITU-recommended model, ray tracing model, and empirical attenuation formula, to form a structured prior knowledge base, which provides the basic propagation constraints of Maxwell's equations. A small number of distributed sensing nodes are deployed in the target area to continuously collect sparse measured data of radio wave propagation, such as received signal strength, delay spread, and multipath information. A generative lightweight adversarial network is used to augment the collected sparse measured data, and simulated propagation data covering the entire area is generated by learning the distribution characteristics of the measured data to fill the gaps in spatial sampling. The theoretical propagation data output from the prior knowledge base is spatiotemporally aligned with the augmented measured data and used together as input data for subsequent fusion training. For example, based on empirical values ​​of relevant distances in urban environments and spatial sampling theorems, the initial sampling interval is set. After verification of sampling density sensitivity, the same whole-area prediction accuracy can be achieved with 20% to 30% of the monitoring points of traditional road testing schemes. Based on a 10km × 10km urban area, this scheme only needs to deploy 20 collection points (traditional gridded road testing schemes usually require about 70 fixed monitoring points supplemented by manual road testing under the same coverage accuracy). The actual received power and delay spread of each point are collected as measured samples. A lightweight adversarial network is used to perform spatial interpolation enhancement on the sparse point data to complete the grid-level equivalent propagation samples in the entire area.

[0040] This embodiment does not employ the traditional PINN (Physical Information Neural Network) hard-boundary forced solution method. Instead, it uses the residuals of electric field curl and magnetic field divergence as soft constraints in the loss function to construct a hierarchical transfer training architecture: the first layer uses ray tracing to generate 100,000 sets of virtual samples in batch simulation to complete the network's basic intervention training; the second layer only uses sparse measured point samples for fine-tuning, adding a soft constraint term of path loss exponential decay to the loss function to drive the network towards minimizing the residuals of the wave equation. The entire training process does not require large-scale numerical iteration calculations, ultimately resulting in a lightweight radio twin network.

[0041] It should be noted that traditional pure ray tracing single-scene simulation takes tens of minutes to several hours, which is completely unacceptable for real-time prediction by mobile users; ordinary pure data neural networks do not consider the physical laws of electromagnetic waves at all, and will show abnormal field strength jumps in obstructed areas. This embodiment relies on soft constraints to balance physical realism and real-time performance, and only requires a small number of sparse measurement points to complete training. Under large-scale deployment in the field, the overall cost of hardware procurement, installation and operation and maintenance is significantly reduced compared with traditional solutions.

[0042] Preferably, the twin modeling layer uses the physical laws of electromagnetic propagation as model constraints to construct a fusion training network, and trains a digital twin of electromagnetic wave propagation through a lightweight dual-drive physical soft constraint modeling method of the twin modeling layer; the trained lightweight real-time digital twin of electromagnetic wave propagation outputs full-area electromagnetic feature data, and integrates the full-area electromagnetic feature data into an electromagnetic background field.

[0043] Specifically, the electromagnetic propagation physical laws corresponding to Maxwell's equations are used as the physical mechanism soft constraint penalty term for network training, replacing the traditional high-iteration numerical solution, thus compressing the model inference latency while ensuring physical reliability. The finite sparse measured data enhanced by dual-drive knowledge data is used as the supervision signal, driving the network to converge quickly with a very small number of samples, avoiding the time overhead caused by large-scale label data collection. Through collaborative training supervised by physical constraints and sparse measured data, a lightweight digital twin of electromagnetic wave propagation adapted to complex terrain scenarios is obtained. Based on this lightweight twin, the rapid inference of global electromagnetic features is completed, generating a continuous and high-precision global electromagnetic background field, which is then output to the trajectory prediction layer.

[0044] It should be noted that by employing a dual-drive collaborative training approach combining soft constraints based on physical mechanisms and lightweight training based on sparse measured data, a lightweight digital twin of radio wave propagation is obtained that is adaptable to complex terrain scenarios and balances accuracy and real-time performance. This approach avoids predictions that contradict physical realities from purely data-driven models, improving the twin's generalization ability and prediction accuracy. Furthermore, traditional pure-mechanism radio wave modeling relies on large-scale, high-iteration numerical solutions, offering high physical deduction accuracy but incurring enormous computational overhead and significant real-time latency, failing to meet the real-time response requirements of dynamic prediction and control. The resulting full-area electromagnetic background field provides comprehensive environmental electromagnetic information support for subsequent coverage prediction.

[0045] Preferably, the trajectory prediction layer collects historical location information of the user terminal and extracts user motion features to build a spatially aware temporal prediction network. Based on the user's historical motion features and current location, it predicts the user's future movement trajectory and generates a predicted location sequence, realizing grid-by-grid spatiotemporal binding and matching between the user's movement trajectory and the global electromagnetic field.

[0046] Specifically, the trajectory prediction layer continuously collects historical location information reported by user terminals, extracting motion features such as user speed, direction of movement, dwell time, and turning patterns from historical trajectory sequences; it obtains the user's current location information in real time through base station positioning or satellite positioning, and uses an adaptive filtering algorithm to correct positioning errors; it employs a long-term feature inference network that integrates spatiotemporal correlation constraints to perform deep learning on the user's historical motion features, combines prior information from three-dimensional geographic grid space to mine the user's movement behavior patterns, adaptively captures short-term motion details and long-term movement trends, and thus accurately predicts the user's continuous movement trajectory over a future period, generating a high-density predicted location sequence containing timestamps; it then performs spatial coordinate matching between the dynamically predicted location sequence and the global electromagnetic background field output by the twin modeling layer, realizing the spatiotemporal coupling correlation between the user's movement path and the time-varying electromagnetic environment, and completing the dynamic binding of trajectory and electromagnetic environment.

[0047] It should be noted that by using deep learning of users' historical movement characteristics, it is possible to accurately infer users' movement intentions and predict their future movement trajectories, providing a sufficient time window for advance coverage control and solving the response delay problem of traditional passive control.

[0048] Preferably, the trajectory prediction layer calculates the estimated received signal strength at each predicted location, identifies areas with signal strength below the communication threshold as potential weak coverage areas, assesses the interruption risk probability and weak area type of potential weak coverage areas, and generates differentiated prediction instructions containing target information and control requirements for the identified potential weak coverage areas and outputs them to the dynamic control layer.

[0049] Specifically, the trajectory prediction layer relies on the spatially biased long-time-series trajectory extrapolation engine to output continuous position extrapolation results. It performs grid-by-grid spatiotemporal coupling matching between the refined predicted position sequence and the electromagnetic background field output by the twin modeling layer to accurately calculate the estimated received signal strength corresponding to each predicted spatiotemporal node. It compares the real-time field strength prediction value with the preset communication threshold to identify areas with signal strength lower than the communication threshold as potential weak coverage areas. For each potential weak coverage area, it assesses the probability of communication interruption, uses a lightweight gradient boosting tree model, and determines the type of weak area as occlusion type, far-field type, multipath null type, or insufficient power type based on electromagnetic characteristics, and simultaneously provides an interruption risk confidence level between 0 and 1.

[0050] In this embodiment, for each valid weak region, a pre-judgment instruction in structured TLV (Type-Length-Value) format is generated: - Type segment: 01 Occlusion type, 02 Far-field type, 03 Multipath zero-dimple type; - Length field: Total length of subsequent parameters in bytes; - Value range: Enter the latitude and longitude of the weak area center, risk level 1 to 5, beam weight adjustment vector of each antenna port, and power boost step size in sequence.

[0051] Then, the generated structured TLV format pre-prediction instruction queue is output to the dynamic control layer.

[0052] It should be noted that by using trajectory-electromagnetic field bidirectional spatiotemporal correlation analysis under geospatial grid constraints, the limitations of traditional static field strength screening can be overcome. This allows for the early detection of dynamic coverage defects on user movement paths, enabling early warning of communication quality risks and providing precise timing basis and data support for predictive and differentiated joint compensation control in dynamic control layers.

[0053] Preferably, the dynamic control layer receives and parses the structured instruction information in the prediction instruction queue, and completes the intelligent diversion of control tasks according to the weak coverage cause type identified in the prediction instruction through the weak differentiation and deep enhancement joint optimization solution engine; for weak coverage scenarios with different mechanisms, it adaptively matches exclusive optimization solution logic, outputs accurate control parameters, and integrates multi-dimensional parameter structures to construct a global joint control scheme.

[0054] Specifically, the dynamic control layer receives the structured TLV format pre-prediction instruction queue transmitted by the trajectory prediction layer, and parses the key information in each instruction, such as the target's future spatiotemporal coordinates, estimated arrival time, required signal gain, and weak area cause type. Based on the weak area type identified in the prediction instruction, the intelligent control task is distributed, and different types of weak areas are sent to the corresponding optimization processing modules.

[0055] Specifically, for occlusion-type, far-field-type, and multipath null-type weak regions, a joint deep reinforcement learning algorithm architecture is adopted to iteratively optimize the phase shift matrix of each sub-unit of the spatial intelligent electromagnetic timing reconstruction metasurface array and the three-dimensional beam pointing parameters of the base station. For power-insufficient weak regions, a constraint-constraint analytical optimization solution mechanism is used to adaptively solve for the optimal power allocation coefficient of the base station. The various optimal control parameters obtained from the solution are integrated to form a joint control scheme for each potential weak region.

[0056] It should be noted that by performing adaptive flow control based on the type of weak area, the most suitable optimization algorithm can be adopted for coverage problems of different causes, which improves the real-time performance and accuracy of the control parameter solution and ensures that the control effect can accurately match the needs of coverage enhancement.

[0057] Preferably, the dynamic control layer, based on the joint control scheme, realizes advanced timing control, and adjusts the power and beam parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array and the base station in advance; for multi-array deployment scenarios, it initiates a multi-node collaborative control mechanism and initiates a collaborative control scheme; and feeds back the real-time sensing data generated during the control execution process to the environmental sensing layer to update the three-dimensional environmental model.

[0058] Specifically, the dynamic control layer pre-sets the execution sequence of control commands based on the spatiotemporal timing information built into the prediction commands. Before the user arrives at a potential weak zone, it sends control commands to the spatial intelligent electromagnetic timing reconstruction metasurface array controller and the base station. The metasurface array controller adjusts the phase shift parameters of each subarray according to the commands, so that the reflected beam is aligned with the location where the user is about to arrive. The base station simultaneously completes the preset adjustment of the transmission power and beam pointing angle. When multiple sets of spatial intelligent electromagnetic timing reconstruction metasurface arrays are deployed in the system, the multi-array surface collaborative control algorithm is activated to jointly optimize the phase configuration parameters of multiple arrays. The beam vector superposition forms a full-domain enhanced coverage main lobe, eliminating the problems of limited single-array control gain and local coverage collapse. During the control execution process, real-time sensing data such as environmental echoes and channel state changes are continuously collected and fed back to the environmental sensing layer for iterative updates of the three-dimensional environmental model, ensuring the real-time performance and scene adaptability of environmental modeling.

[0059] It should be noted that by using time-series pre-judgment and advanced parameter pre-setting control logic, precise spatiotemporal coverage is achieved, with the signal arriving before the person arrives. This completely changes the traditional passive control mode of interruption followed by adjustment. The coverage enhancement capability in complex environments is further improved through the collaborative control of multiple metasurface arrays, while real-time data feedback ensures the timeliness of the environmental model. Combined with a five-layer architecture bidirectional data feedback mechanism, closed-loop self-iterative optimization of twin modeling, trajectory prediction, and dynamic control is realized.

[0060] Preferably, the closed-loop optimization layer collects the measured communication data reported by the user terminal and the link performance data on the network side in real time, compares the measured data with the predicted values ​​of the corresponding positions of the radio wave propagation digital twin, and calculates the error. When the calculated error exceeds a preset threshold, error source localization analysis is initiated. Based on the newly collected measured data, incremental training and local parameter correction are performed on the radio wave propagation digital twin. The corrected parameters are fed back to the twin modeling layer to update the radio wave propagation digital twin, thus completing the full-link closed-loop optimization.

[0061] Specifically, the closed-loop optimization layer collects real-time communication data reported by user terminals, such as measured received signal strength, signal-to-interference-plus-noise ratio, and actual location, while simultaneously collecting network-side performance data such as link quality and handover success rate. It compares the collected measured data with the predicted values ​​of the radio wave propagation digital twin at the corresponding locations, calculating error statistics such as mean absolute error and root mean square error. When the calculated error exceeds a preset threshold, error source localization analysis is initiated to determine whether the error originates from environmental changes, model bias, or hardware failure. If the error originates from environmental changes or model bias, the radio wave propagation digital twin is incrementally and lightweightly trained using newly collected measured data, with minor adjustments to the neural network weight parameters and iterative updates to the empirical parameters in the prior knowledge base. Only local corrections are made to the parameters corresponding to the biased scenarios, eliminating the need for global retraining and ensuring model iteration efficiency and real-time performance. The closed-loop optimization layer sends the corrected parameter set to the twin modeling layer through a feedback channel to update the radio wave propagation digital twin, completing one closed-loop self-optimization process.

[0062] It should be noted that, through real-time error feedback and the autonomous evolution mechanism of online incremental local correction of the model, the digital twin of radio wave propagation can continuously evolve in continuous interaction with the real environment, dynamically adapt to time-varying scenarios, and always maintain an accurate mapping of the real electromagnetic environment, ensuring that the system can maintain stable prediction and control performance when the environment changes dynamically.

[0063] Accordingly, this embodiment provides a digital twin radio wave propagation prediction and control method that integrates inductive and computational methods, including: Acquire multi-source remote sensing and terrain elevation data of the target area, and construct a three-dimensional environmental model with attached electromagnetic characteristic parameters of ground objects; Constrained by the physical laws of electromagnetic propagation, and integrating the prior knowledge model of radio wave propagation with enhanced sparse measured data, a digital twin of radio wave propagation is constructed to generate the electromagnetic background field of the entire region. Predict the user's future movement trajectory, identify potential weak communication coverage areas based on the electromagnetic background field, and generate advance prediction instructions; Based on the aforementioned pre-judgment instructions, the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array are collaboratively optimized with the base station power and beam parameters to achieve enhanced user coverage; By comparing the actual on-site communication data with the twin prediction values, error source analysis and incremental model correction are completed, forming a closed loop for full-link control.

[0064] Specifically, firstly, a 3D environmental model labeled with electromagnetic characteristic parameters is constructed by collecting, preprocessing, and classifying multi-source remote sensing and terrain elevation data, and then using ground feature classification. Secondly, a digital twin of electromagnetic wave propagation is obtained by integrating a priori knowledge model of electromagnetic wave propagation with sparse measured enhanced data samples and training it with the physical laws of electromagnetic propagation as constraints, generating a dynamically updated electromagnetic background field for the entire region. Subsequently, future movement trajectories are predicted based on users' historical and real-time location data. A high-precision long-term trajectory is output through a spatially biased long-term trajectory inference engine, which is spatiotemporally coupled with the electromagnetic background field grid by grid to identify potential weak coverage areas in advance and generate time-series prediction instructions. Further, based on the weak area type, an adaptive matching optimization solution strategy is used to jointly optimize the phase shift matrix of the metasurface array for spatial intelligent electromagnetic time-series reconstruction, as well as the base station power and beam parameters, to achieve pre-compensation for the continuity of mobile user coverage. Finally, measured user communication data is collected and compared with the twin prediction values. When the error exceeds the limit, the digital twin is locally corrected, forming a continuously optimized full-link closed loop of perception-modeling-prediction-control-optimization.

[0065] It should be noted that this method deeply integrates communication, sensing, computing technologies with digital twin technology to construct a new technical system that includes physical mechanism constraint modeling, spatial bias long-term prediction, weak-distinction hierarchical intelligent control, and incremental closed-loop self-correction. It effectively adapts to the characteristics of high-frequency communication that is sensitive to blockage and has drastic channel time-varying features, and achieves accurate prediction and active dynamic control of radio wave propagation. It effectively solves the problems of high-frequency communication being susceptible to blockage and having poor coverage continuity, while improving the utilization efficiency of spectrum resources and the overall robustness of the system.

[0066] Example 2 See Figure 1 This embodiment provides a digital twin radio wave propagation prediction and control system that integrates inductive and computational computing, comprising five execution stages coupled from top to bottom: The first layer, the environmental perception layer, goes beyond simple data collection. It constructs a multi-scale, multi-modal electromagnetic environment cognition foundation, primarily achieving deep fusion of multi-modal heterogeneous data and environmental understanding. Specifically, the environmental perception layer includes a data access module (including various types of parsers), a data processing module, and a modeling output module, outputting a high-precision 3D environmental model to the second layer, the twin modeling layer.

[0067] The second layer, the twin modeling layer, breaks through the inherent limitations of traditional single-mechanism models and ordinary data-driven models. It constructs a dual-driven collaborative training architecture based on soft constraints of physical mechanisms and lightweight sparse measured data. Relying on a priori knowledge base of electromagnetic propagation to provide underlying physical rules of electromagnetic propagation, it completes model correction using only a small number of sparse measured communication samples, eliminating the need for large-scale global sampling. The physical priors of Maxwell's equations are used as soft constraint penalty terms in the neural network, replacing the traditional high-iteration numerical solution process. Simultaneously, a lightweight adversarial network is used to densify and repair sparse measured points. This reduces computational overhead while ensuring the derivation results conform to the laws of electromagnetic propagation, balancing the credibility and real-time performance of global electromagnetic derivation. It constructs a self-evolving digital twin of electromagnetic propagation, achieving accurate mapping of electromagnetic characteristics across the entire region, and realizing the goal of a self-evolving digital twin engine driven by soft constraints of physical mechanisms and lightweight sparse measured data.

[0068] Specifically, the twin modeling layer includes a knowledge base module (ITU module, ray tracing engine, empirical formula library, etc.), a data augmentation module (data cleaner, sparse interpolator, etc.), a fusion training module (physical information neural network, dual-drive fusion engine, twin manager), and an output module (coverage prediction map generator, multipath delay spectrum calculator, channel impulse response generator, twin data service interface). The twin modeling layer primarily constructs a digital twin of radio wave propagation and outputs the global electromagnetic background field (coverage map, delay spectrum, etc.) to the third-layer trajectory prediction layer. Here, ITU refers to the International Telecommunication Union (ITU), and the ITU module is a functional module integrating a series of standards and algorithms developed by the ITU for radio propagation prediction and communication system design.

[0069] The third layer, the trajectory prediction layer, introduces a spatiotemporal correlation attention mechanism to perform deep pattern mining on the user's long-term motion characteristics, completing the deduction of mobile behavior and the reasoning of operational intent. It performs joint probabilistic calculations by combining the predicted trajectory with the global electromagnetic background field output by the twin, achieving advanced early warning and probabilistic positioning of future weakly covered areas, and realizing the purpose of user mobile intent reasoning and blind spot prediction based on the spatiotemporal attention mechanism.

[0070] Specifically, the trajectory prediction layer includes a data acquisition module (historical trajectory storage, real-time positioning receiver, motion feature extractor, etc.), a geospatial bias long-term time-series trajectory extrapolation engine module, a weak area identification module (radio wave propagation twin interface, coverage weak area detector, risk probability estimator, etc.), and an instruction generation module (predictive instruction encoder, target location calculator, arrival time estimator, required gain calculator, instruction queue manager, etc.). The trajectory prediction layer mainly generates a standardized TLV time-series predictive instruction queue (including target location, estimated arrival time, link compensation gain, weak area type, and control priority), which is then sent to the fourth-layer dynamic control layer.

[0071] The fourth layer is the dynamic control layer, which overturns the traditional passive mode of interruption followed by adjustment and constructs a predictive beam timing scheduling strategy based on future spatiotemporal location. Relying on a weak-distinction hierarchical deep enhancement joint solution engine, it conducts forward-looking control parameter planning, driving the spatial temporal metasurface array collaborative hardware management module to perform phase preset control on the spatial intelligent electromagnetic timing reconstruction metasurface array. This allows the electromagnetic beam to target the area the user is about to arrive in advance, achieving precise spatiotemporal coverage where the signal arrives before the user, thus realizing the goal of integrated predictive beam tracking and energy scheduling encompassing sensing, communication, and control.

[0072] Specifically, the dynamic control layer includes a weakly differentiated, hierarchical, and deep-enhancement joint optimization engine, a spatial-temporal metasurface array collaborative hardware management module, a collaborative scheduling module, and an instruction output module (phase shift parameter generator, power allocation coefficient generator, beam pointing calculator, execution timing orchestrator, etc.). The dynamic control layer issues spatial-temporal metasurface array control instructions to enhance the radio wave propagation coverage in the physical world.

[0073] The fifth layer is the closed-loop optimization layer, which constructs a real-time error feedback channel from the physical world to the digital world. The deviation between the actual measured data and the twin's predicted values ​​is used as an iterative incentive signal to trigger the model's online learning. Simultaneously, an incremental lightweight training strategy is used to fine-tune the model parameters. Through this cognitive closed-loop mechanism of scenario testing-deviation feedback-parameter correction, the digital twin continuously evolves through ongoing interaction with the real environment, achieving a leapfrog improvement in system performance and realizing the goal of real-time error feedback, online twin correction, and knowledge evolution.

[0074] Specifically, the closed-loop optimization layer includes a data feedback module (terminal measurement collector, network performance monitor, target location collector, etc.), an analysis and diagnosis module, a model calibration module (online learner, weight fine-tuning engine, knowledge base updater), and a feedback output module (sibling model update parameters, classification model fine-tuning parameters, propagation model correction coefficients, etc.), thereby outputting the sibling model update parameters, classification model fine-tuning parameters, propagation model correction coefficients, etc., to the second sibling modeling layer to update the sibling model.

[0075] It should be noted that the above five-layer architecture is not a simple stacking of functions. No single layer can be effective or generate substantial regulatory benefits on its own. Only when the five layers work together in a coordinated manner with their interdependent relationships can a closed-loop intelligent decision-making chain be formed, from real-time environmental awareness to the issuance of optimal control commands. This achieves an overall performance leap that cannot be achieved by a single layer or fragmented solution.

[0076] Example 3 See Figure 2 This embodiment provides a digital twin radio wave propagation prediction and control method that integrates induction and computation. Based on a five-layer linkage architecture, it sequentially completes environmental perception, twin modeling, trajectory prediction, dynamic control, and closed-loop optimization, achieving high-precision electromagnetic modeling, user movement trajectory prediction, weak coverage risk warning, predictive beam control, and continuous model self-evolution. The specific steps are as follows: Phase 1: Environmental Perception.

[0077] The first step involves acquiring basic data for the target area using multi-source remote sensing methods, including optical remote sensing imagery, synthetic aperture radar (SAR) imagery, digital elevation model (DEM), and LiDAR point cloud data. The second step involves unified preprocessing of the aforementioned basic data, including geometric correction, radiometric calibration, and denoising, and registering data from different sources to a unified coordinate system and resolution. The third step uses a multi-scale convolutional neural network (CNN) model to intelligently classify the preprocessed imagery, identifying land cover types such as buildings, vegetation, water bodies, roads, and bare land, and outputting the classification label and confidence score for each pixel. The fourth step involves constructing a 3D mesh model (resolution ≤ 1 meter) based on the classification results and elevation data, and labeling the electromagnetic property parameters (dielectric constant, conductivity, surface roughness) of each land cover on the model, ultimately generating a high-precision 3D environment model, which is then passed as output to the next stage.

[0078] Phase Two: Twin Modeling

[0079] The first step involves importing and integrating various prior knowledge models of radio wave propagation, including ITU-R recommended models (such as P.526 and P.833), ray tracing models, and empirical attenuation formulas such as the Okumura-Hata model, to form a prior knowledge base. The second step involves deploying a small number of sensing nodes within the target area to collect sparse measured data, primarily including Received Signal Strength Reduction (RSRP), delay spread, and multipath information. The third step utilizes a generative lightweight adversarial network to enhance the sparse measured data, generating simulated propagation data covering the entire area and filling spatial sampling gaps. The fourth step constructs a dual-driven collaborative training architecture combining soft constraints of physical mechanisms and lightweight sparse measured data, employing a dual-loss coupling training mechanism of mechanistic constraints and measured fitting. This approach reduces computational overhead while ensuring the simulation results conform to the laws of electromagnetic propagation, balancing the credibility of the global electromagnetic simulation with real-time performance, and constructing a self-evolving digital twin of radio wave propagation. The fifth step involves using the trained twin to output full-area electromagnetic characteristic data, including the Received Signal Strength Distribution Map (RSRP Map), coverage probability map, multipath delay spectrum (PDP), and channel impulse response (CIR). These data together constitute the full-area electromagnetic background field and are then transmitted to the third stage.

[0080] Phase 3: Trajectory Prediction.

[0081] The first step involves collecting historical location information from user terminals to construct long-term trajectory samples, recording motion characteristics such as movement speed, direction, dwell time, and turning patterns. The second step involves acquiring the user's current location in real time via GPS or base station positioning, and applying a filtering and correction algorithm to eliminate random positioning noise and location jitter, outputting high-precision real-time location information. The third step utilizes a spatial offset long-term trajectory extrapolation mechanism to deeply learn the user's long-term motion patterns and behavior patterns, predicting the user's movement trajectory within the next T seconds, and outputting the predicted location sequence and its confidence interval. The fourth step involves spatiotemporally coupling and matching the predicted future location with the global electromagnetic background field output in the second stage, calculating the estimated received signal strength at each predicted location, identifying areas with signal strength below the communication threshold, marking them as potential weak coverage areas, and assessing the probability of outage and the type of weak area (obstruction, far field, multipath nulls, or insufficient power). The fifth step is to generate standardized timing prediction instructions for each identified weak region. The instructions include the target user ID, future location coordinates, estimated arrival time, required gain value, weak region type identifier, and priority level, forming a prediction instruction queue, which is then output to the fourth stage.

[0082] Phase 4: Dynamic Regulation.

[0083] The first step is to receive the prediction instruction queue from the third stage and parse out information such as the target location, arrival time, and required gain in each instruction.

[0084] The second step involves traffic splitting based on the type of weak area: If the weak area is due to obstruction, far-field, or multipath null, it enters the spatial array beam coordination optimization module. This module constructs a state space containing the user's long-term location, channel state, and the working state of the spatial intelligent electromagnetic timing reconstruction metasurface unit. It defines the action space consisting of the spatial intelligent electromagnetic timing reconstruction metasurface phase shift parameters and the base station beam pointing, and uses the communication coverage improvement effect and system energy efficiency as optimization objectives. The metasurface phase matrix and base station beam pointing parameters are obtained through strategy iteration. If the weak area is caused by limited transmit power, it enters the intelligent power allocation module, which solves for the optimal power allocation coefficient for each user under the constraints of the base station's total power limit and the user's minimum signal-to-interference-plus-noise ratio.

[0085] The third step is to combine the metasurface phase shift matrix optimized by the spatial array beam coordination optimization module with the power coefficient obtained by intelligent power allocation, that is, to adjust the phase parameters of each sub-unit of the spatial intelligent electromagnetic timing reconstruction metasurface array in advance, so that the reflected beam is pre-pointed to the spatial location where the user is about to arrive.

[0086] The fourth step is to determine the number of airspace intelligent electromagnetic timing reconfiguration metasurface arrays deployed by the system. When multiple arrays are deployed, a collaborative control strategy is initiated to jointly optimize the phase configuration of each array group to form a collaboratively enhanced beam. When only a single array is deployed, the single array control scheme is executed directly.

[0087] The fifth step involves generating the final control instruction set, which includes the metasurface phase shift parameter matrix, power allocation coefficients, beam pointing angle, and timing execution information. These instructions are then sent to the array controller and base station via the control channel, allowing the physical devices to execute the coverage enhancement operation. Simultaneously, real-time sensing data generated after the control execution (such as echoes and environmental changes) is fed back to the preprocessing step in the first stage to update the 3D environment model.

[0088] Phase 5: Closed-loop optimization.

[0089] The first step involves real-time collection of measurement data reported by user terminals, including measured RSRP, SINR (signal-to-interference-plus-noise ratio), actual location information, and network-side performance data such as link quality and handover success rate. The second step compares the measured data with the predicted spatial location of the second-stage radio wave propagation twin, calculating statistics such as mean absolute error (MAE) and root mean square error (RMSE) to determine if the deviation exceeds a preset threshold (e.g., 2dB). If the deviation is within the limit, the current twin model is maintained; if the deviation exceeds the limit, error source localization is initiated to analyze whether the error originates from dynamic environmental changes, inherent model bias, or hardware failure. The third step triggers a model correction and update mechanism, using newly collected measured data to incrementally train the twin model, fine-tuning the weights of the deep neural network, correcting empirical parameters in the prior knowledge base, and generating a correction parameter set containing twin model weight update parameters, classification model fine-tuning parameters, and propagation model correction coefficients. The fourth step involves sending the set of calibration parameters through a feedback channel to the dual-driven collaborative training module of physical mechanism soft constraints and sparse measured data lightweighting within the twin modeling layer of the second stage. This updates the digital twin of radio wave propagation, completing the model's self-evolution. Subsequently, based on the updated twin, further predictions and adjustments are performed, forming a continuously optimized closed loop.

[0090] See Figure 3 The dynamic control phase can be achieved through the following steps: Step 1: Receive and parse the prediction command Receive the prediction instruction queue from the third stage (trajectory prediction). Parse each instruction to extract the following key information: the target user's future location coordinates, estimated arrival time, required signal gain, weak area type (occlusion, far field, multipath null, or insufficient power), and priority level, completing the pre-classification and parameter analysis of the control task.

[0091] Step 2: Determine the type of weakness and redirect traffic to the corresponding optimization module. Based on the weak region type obtained from the analysis, the following branch decision is made: If the weak region type is occlusion, far field, or multipath null, then proceed to the spatial array beam co-optimization process module (step 3). If the weak zone type is insufficient power, then proceed to the global power resource optimization and allocation process module (step four).

[0092] Note: In a real system, this process can be executed in parallel; this is a description of logical branches.

[0093] Step 3: Deep reinforcement learning optimization (for occlusion / multipath problems) 3.1 Constructing a multi-dimensional state space: including the current user location, dynamic channel state information (CSI), and the state of each unit of the spatial intelligent electromagnetic timing reconstruction metasurface array.

[0094] 3.2 Define the control action space: including the phase shift configuration parameters of each sub-unit of the array and the three-dimensional beam pointing angle of the base station.

[0095] 3.3 Design a multi-objective optimization reward function: comprehensively weigh the signal coverage enhancement effect, communication link quality, and system energy consumption efficiency to achieve a balanced and optimal multi-dimensional index.

[0096] 3.4 Determine if online iterative optimization is needed: Supports dual-mode parameter solving: In the online running state, the control strategy parameters are continuously iterated and updated to adapt to dynamic electromagnetic scenarios; in the offline prediction scenario, a pre-optimized mature strategy model is loaded to directly infer and output the optimal control parameters.

[0097] 3.5 Final output array optimal phase shift matrix and base station beam pointing parameters.

[0098] Step 4: Optimize and allocate power resources across the entire domain (addressing power shortage issues) 4.1 Set constraints: upper limit of total base station transmit power, minimum signal-to-interference-plus-noise ratio (SINR) threshold for each user.

[0099] 4.2 Two types of adaptive solution branches are provided to adapt to scenarios with different levels of complexity: Branch 1: Constrained analytical optimization solution of the branch, which directly and analytically solves the global optimal power allocation coefficient based on strict boundary constraints, and is suitable for simple steady-state scenarios; Branch 2: Iterative adaptive optimization solution branch, through a progressive iterative optimization and allocation strategy, quickly converges to obtain an approximate optimal solution, adapting to dynamic, complex, and time-varying scenarios.

[0100] 4.3 Output results: Optimal power allocation coefficients and global power balance configurations for each user.

[0101] Step 5: Predictive Coverage Enhancement (Joint Optimization Results) The array phase shift matrix output from the third step is fused with the power allocation optimization result output from the fourth step. Based on the pre-configured phase shift parameters of each sub-unit of the metasurface array using spatial intelligent electromagnetic timing reconstruction, the reflected beam is pre-aligned with the spatiotemporal region where the user will arrive in the future, achieving predictive pre-coverage enhancement.

[0102] Step 6: Determine if multi-array collaboration exists. In multi-array deployment scenarios, a multi-unit collaborative control strategy is initiated, and the phase configuration of multiple metasurface arrays is jointly and iteratively optimized. The main lobe is enhanced and covered by beam vector superposition. In single-array deployment scenarios, an independent array control strategy is directly executed.

[0103] Step 7: Generate the final control instruction set Based on the above optimization results, a complete set of control instructions is generated, including: Array phase shift parameter matrix (phase value of each cell); Base station power allocation coefficient; Beam pointing angle; Execution sequence (i.e., when to start the adjustment and the duration).

[0104] Step 8: Send commands to physical devices The spatial temporal metasurface array collaborative hardware management module sends control commands to the metasurface array hardware and base station execution unit, and the physical equipment performs predictive coverage enhancement actions.

[0105] Step 9: Enter the closed-loop optimization phase After the command is issued, it waits for the user terminal to report the measured signal data (RSRP, SINR, position, etc.) and uses this measured data as feedback input to the fifth stage (closed-loop optimization) for error analysis and incremental model correction.

[0106] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A digital twin radio wave propagation prediction and control system integrating inductive and computational methods, characterized in that, It includes a top-down, hierarchically coupled environment perception layer, twin modeling layer, trajectory prediction layer, dynamic control layer, and closed-loop optimization layer: The environmental perception layer is configured to acquire multi-source remote sensing and terrain elevation data of the target area, construct a three-dimensional environmental model with electromagnetic characteristic parameters of ground objects, and output it to the twin modeling layer. The twin modeling layer is configured to be constrained by the physical laws of electromagnetic propagation, integrate the prior knowledge model of radio wave propagation with the enhanced sparse measured data, construct a digital twin of radio wave propagation, generate a full-area electromagnetic background field and output it to the trajectory prediction layer. The trajectory prediction layer is configured to predict the user's future movement trajectory, identify potential weak communication coverage areas by combining electromagnetic background field, and generate a pre-prediction command to output to the dynamic control layer. The dynamic control layer is configured to collaboratively optimize the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array, base station power, and beam parameters according to the pre-judgment instruction, thereby enhancing user coverage. The closed-loop optimization layer is configured to compare the measured communication data on site with the twin prediction values, complete the error source analysis and incremental model correction, and form a full-link control closed loop.

2. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 1, characterized in that, The environmental perception layer acquires multi-source remote sensing and terrain elevation data of the target area, and constructs a three-dimensional environmental model with attached electromagnetic characteristic parameters of ground features, including: Acquire multi-source remote sensing data and terrain elevation data of the target area, preprocess the raw data and register it to a unified coordinate system and resolution; Based on the processed data, land features are classified to distinguish between building, vegetation, water, and bare land scene types; Based on the material structure and electromagnetic properties of various land features, the corresponding electromagnetic loss, reflection and scattering parameters are labeled for the 3D scene.

3. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 1, characterized in that, In the twin modeling layer, constrained by the physical laws of electromagnetic propagation, a priori knowledge model of radio wave propagation is integrated with enhanced sparse measured data to construct a digital twin of radio wave propagation, generating a full-area electromagnetic background field, including: The physical laws of electromagnetic propagation are used as the physical mechanism for network training, and the enhanced sparse measured data are used as the supervision signal. A lightweight digital twin of radio wave propagation is obtained through collaborative training of the physical constraints and the sparse measured data. The lightweight radio wave propagation digital twin is used to deduce the global electromagnetic characteristics and generate the global electromagnetic background field.

4. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 3, characterized in that, In the twin modeling layer, the physical mechanism soft constraint penalty term is used to replace the electromagnetic propagation numerical iterative solution process; The sparse measured data is used to drive network convergence; the lightweight radio wave propagation digital twin is used to deduce global electromagnetic characteristic data and integrate it to generate the global electromagnetic background field.

5. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 1, characterized in that, In the trajectory prediction layer, predicting the user's future movement trajectory includes: Collect historical location information of user terminals and extract motion features; collect the user's current location and correct positioning errors. Based on the historical location information and the current location, the user's future movement trajectory is inferred to generate a predicted location sequence; The predicted location sequence is spatially superimposed and matched with the electromagnetic background field of the entire region to establish a spatiotemporal coupling relationship between the user's movement path and the state of the electromagnetic field.

6. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 5, characterized in that, In the trajectory prediction layer, potential weak coverage areas for communication are identified by combining the electromagnetic background field, and advance prediction instructions are generated, including: Based on the spatiotemporal coupling correlation, the estimated received signal strength is calculated for each predicted location, and areas with signal strength below the communication threshold are selected as potential weak coverage areas. Assess the probability of communication interruption risk in each potential weak coverage area and classify them into obstructed weak areas, edge weak areas, or overlapping interference weak areas; For different risk levels and weak area types, generate advance prediction instructions carrying location information, risk level information and appropriate control strategy information.

7. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 1, characterized in that, In the dynamic control layer, the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array are collaboratively optimized in relation to the base station power and beam parameters, including: Analyze the pre-judgment instructions and divert the control tasks according to the weak area type and risk level in the pre-judgment instructions; For different weak region types, beamforming optimization, power equalization optimization, or metasurface phase compensation optimization are performed respectively to obtain the base station parameters and the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array; the spatial intelligent electromagnetic timing reconstruction metasurface array includes multiple tunable electromagnetic units, and the multiple tunable electromagnetic units realize spatial electromagnetic wave reconstruction by timing control of the phase shift state of each unit. By integrating the base station parameters with the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array, a joint control scheme is generated.

8. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 7, characterized in that, The dynamic control layer also includes: According to the joint control scheme, before the user arrives in a potentially weak coverage area, the phase shift calibration of the spatial intelligent electromagnetic timing reconstruction metasurface array and the pre-adjustment of the base station's transmit power and beam pointing are completed. When multiple sets of the aforementioned spatial intelligent electromagnetic timing reconstruction metasurface arrays are deployed in the scene, multi-array collaborative linkage control is initiated. The real-time environmental perception data and coverage status data generated by the control operation are transmitted back to the environmental perception layer to assist in updating the three-dimensional environmental model.

9. The digital twin radio wave propagation prediction and control system based on the fusion of inductive and computational methods according to claim 1, characterized in that, In the closed-loop optimization layer, the measured communication data from the field is compared with the twin prediction values ​​to complete error source analysis and incremental model correction, including: Collect measured communication data and network link performance data from user terminals, compare the measured communication data with the simulated predicted values ​​at the corresponding positions of the digital twin, and quantify the error; when the error exceeds a preset threshold, trace the source to environmental adaptation error, modeling mechanism error, or equipment parameter error; Based on the source tracing results, incremental parameter correction is performed using the newly added measured communication data to update the digital twin; the corrected model parameters are then synchronized to the twin modeling layer.

10. A digital twin radio wave propagation prediction and control method integrating inductive and computational fusion, characterized in that, include: Acquire multi-source remote sensing and terrain elevation data of the target area, and construct a three-dimensional environmental model with attached electromagnetic characteristic parameters of ground objects; Constrained by the physical laws of electromagnetic propagation, and integrating the prior knowledge model of radio wave propagation with enhanced sparse measured data, a digital twin of radio wave propagation is constructed to generate the electromagnetic background field of the entire region. Predict the user's future movement trajectory, identify potential weak communication coverage areas based on the electromagnetic background field, and generate advance prediction instructions; Based on the aforementioned pre-judgment instructions, the phase shift parameters of the spatial intelligent electromagnetic timing reconstruction metasurface array are collaboratively optimized with the base station power and beam parameters to achieve enhanced user coverage; By comparing the actual on-site communication data with the twin prediction values, error source analysis and incremental model correction are completed, forming a closed loop for full-link control.

Images