Virtual space driven integrated sensing and computing method for improving grid observability

Abstract

The present application relates to a kind of virtual space driven integrated power grid observability promotion method of sensing algorithm, belong to electric communication technical field.Face to power grid dynamic perception and data transmission demand, using time division multiple access transmission form and data compression mechanism, solve the communication resource competition problem of measurement data and service data transmission, optimize data transmission efficiency.For edge side data processing and delay control, through digital twinning, the high-precision real-time mirror of physical state is realized, through world model, the multi-step forward-looking prediction of future state is completed, the cognitive inconsistency quantization double-model cognitive bias is introduced, and the cognitive confidence is defined, the adaptive dynamic trigger of measurement perception is realized.The peak information age accurately depicts the information aging degree of the whole link from compression, queuing, transmission, retransmission to successful decoding, based on virtual space joint cognitive driving sensing algorithm global resource collaborative scheduling, the observability of power grid state and service operation stability are improved.

Description

Technical Field

[0001] This invention relates to a virtual space-driven integrated method for improving the observability of power grids through communication and sensing, belonging to the field of electrical communication technology. Background Technology

[0002] With the large-scale grid connection of distributed renewable energy generators, flexible loads, and energy storage batteries, the operating characteristics of smart grids are becoming increasingly dynamic and complex. The difficulty of comprehensively, accurately, and in real-time sensing of grid status has significantly increased, leading to a substantial decrease in grid observability. Time-sensitive power services such as real-time monitoring, renewable energy dispatching, and fault control heavily rely on reliable grid status observation capabilities, imposing stringent requirements on data freshness and timely decision-making. Taking distributed photovoltaic dispatching as an example, insufficient observability directly causes distorted state perception and an unclear overall situation. Decisions based on outdated or erroneous state information can trigger supply-demand imbalances or even large-scale power backflow, seriously jeopardizing the safe and economical operation of the grid. Integrated sensing, computing, and communication technologies can achieve collaborative optimization of heterogeneous resources, effectively mitigating the timeliness losses accumulated throughout the entire lifecycle of data sensing, transmission, and computing. This has become a core technological path for improving grid observability and ensuring the stable and efficient operation of the grid.

[0003] Information uncertainty profoundly impacts the operational efficiency of integrated communication and computing networks. It distorts state perception, leads to improper network resource allocation, and consequently causes power business information to become outdated and suffer from severe timeliness loss, affecting system observability. Deep reinforcement learning (DRL), as a mainstream artificial intelligence technology for sequential decision-making in dynamic environments, can obtain optimal strategies under conditions of incomplete and uncertain information through trial-and-error iteration, end-to-end optimization, and environment-adaptive decision-making. Typical deep reinforcement learning algorithms include Deep Q-network (DQN), Deterministic actor-critic (DAC), and Deep Deterministic Policy Gradient (DDPG), which have been widely applied in the management practices of integrated communication and computing networks. However, existing research on the integration of sensing and computing focuses on the coordinated scheduling of physical domain resources. Traditional deep reinforcement learning relies on single-step rewards, has a limited decision-making perspective, lacks high-fidelity virtual mapping of the power grid's operating state and long-term prediction capabilities, and is difficult to accurately characterize the state coupling and evolution characteristics across time slots. This results in insufficient generalization ability in high-dimensional complex networks, which restricts the effective improvement of power grid observability.

[0004] Digital twins (DTs) connect physical entities to a virtual space through real-time measurement and sensing, intelligent learning, and bidirectional interaction. World models (WMs), on the other hand, learn the inherent state evolution patterns from historical datasets, enabling multi-step predictions without relying on real-time observations. They compensate for the response lag of digital twins through forward-looking state extrapolation, thus supporting proactive resource scheduling. However, inherent discrepancies exist between the perceptions of the physical system by digital twins and world models, making it difficult to form a unified and self-consistent virtual space. This inconsistency directly weakens the reliability of state characterization and forward-looking prediction, hindering the global optimization of integrated sensing, perception, and computation. To address these issues, this invention proposes a virtual space-driven method for improving the observability of a power grid through integrated sensing, perception, and computation. By constructing a self-consistent virtual space that coordinates digital twins and world models, and integrating state characterization based on confidence perception with long-term forward-looking extrapolation, this method specifically addresses key challenges such as inconsistent virtual space perception and limited observability improvement.

[0005] In view of the above-mentioned shortcomings, the present invention aims to create a virtual space-driven integrated sensing and computing method for improving the observability of power grids, making it more valuable for industrial applications. Summary of the Invention

[0006] To address the aforementioned technical problems, the purpose of this invention is to provide a virtual space-driven integrated method for improving the observability of a power grid.

[0007] The present invention provides a virtual space-driven, integrated sensing and computing method for improving the observability of a power grid, comprising:

[0008] S1. Construct a virtual space-driven integrated power grid observation enhancement system model, which includes a virtual space measurement data perception model, a power business data transmission model, an edge computing model, an edge-side virtual space model, and a power business data timeliness loss model. S2. Based on the system model, establish an integrated optimization problem of sensing and computing, with the weighted sum of minimizing the timeliness loss of power business data and the cognitive confidence of the virtual space model as the optimization objective. The long-term average constraint is transformed into an instantaneous optimization penalty term, and decomposed into a joint sensing-communication optimization sub-problem and a computing resource allocation optimization sub-problem. S3. Execute the virtual space-driven integrated optimization algorithm for sensing and computing, and use the perception-communication optimization algorithm driven by cognition and guided by world model trajectory to solve the joint optimization sub-problem of perception and communication. Use the edge computing resource allocation algorithm oriented towards minimizing the timeliness loss of power business to solve the optimization sub-problem of computing resource allocation, so as to realize the collaborative scheduling of sensing and computing resources and improve the observability of the power grid.

[0009] Furthermore, the construction of the edge-side virtual space model includes: Based on static data, virtual space measurement data, and learnable parameters, a digital twin model is constructed to achieve a high-precision real-time mirror of the physical power grid state. The digital twin error is minimized through gradient iteration and soft update strategies. A world model is established based on historical operational data and system dynamics. Multi-step forward-looking state simulations are performed under conditions without real-time observation. The parameters of the world model are updated using a low-frequency fine-tuning mechanism. The cognitive inconsistency between the digital twin and the world model is calculated, and a cognitive confidence level is defined as an adaptive trigger for measuring data perception.

[0010] Furthermore, the cognitive inconsistency refers to the degree of discrepancy between the digital twin and the world model in their descriptions of the physical system's state, serving as the trigger for measurement data perception. The cognitive confidence decreases as cognitive inconsistency increases, and is used to dynamically adjust the frequency of measurement data uploads and the allocation of communication resources.

[0011] Furthermore, the power business data timeliness loss model adopts peak information age (PAoI) quantification, where PAoI is used to characterize the maximum information aging degree of business data throughout the entire process from compression, queuing, transmission, retransmission to decoding.

[0012] Furthermore, the cognitive-driven and world model trajectory-guided perception-communication optimization algorithm includes: The world model's forward inference step size is dynamically adjusted based on cognitive confidence, and an enhanced state space is constructed by integrating cognitive trajectory sequences and cognitive confidence. Using the virtual space measurement and sensing variables and the compression ratio of business data transmission as the action space, an Actor-Critic architecture is adopted to generate sensing decisions. Multi-step sequence sampling is performed based on trajectory similarity, and cumulative rewards are used instead of single-step rewards to overcome the short-sightedness problem in decision-making in deep reinforcement learning, thereby achieving joint optimization of perception strategy and compression ratio.

[0013] Furthermore, the dynamic look-ahead step size adjustment mechanism is as follows: When the cognitive confidence level is higher than the high threshold, the minimum look-ahead step size is used. When the cognitive confidence level is below a low threshold, the maximum look-forward step size is used. When cognitive confidence is in the middle range, use a linear growth step size.

[0014] Furthermore, the multi-step sequence sampling determines the sampling probability by calculating the cosine similarity between the current cognitive trajectory and the historical sample trajectory. When the cognitive confidence is low, similar trajectories are sampled first, and when the cognitive confidence is high, uniform sampling is performed.

[0015] Furthermore, the edge computing resource allocation algorithm includes: The evolution of PAoI is smoothly and continuously approximated by the hyperbolic tangent function, transforming the non-convex optimization problem into a convex optimization problem; Edge computing resources are allocated based on KKT conditions to prioritize the data decoding and processing of latency-sensitive power services, thereby reducing the risk of information aging and transmission congestion.

[0016] A virtual space-driven, integrated sensing and computing power grid observation enhancement system, used to execute the aforementioned virtual space-driven, integrated sensing and computing power grid observation enhancement method, includes: The equipment layer is used to collect power business data and virtual space measurement data; The network layer, consisting of 6G base stations and edge servers, is used to provide communication and computing services; The virtual space layer, deployed on edge servers, is used to build a self-consistent virtual space that coordinates digital twins and world models, enabling real-time mirroring of physical states and multi-step forward-looking simulations. The control layer deploys deep reinforcement learning agents to perform integrated synesthesia-computation joint optimization decisions. The cloud-based business layer is used to support distributed photovoltaic dispatching, load forecasting, and energy storage charging and discharging dispatching services based on optimized grid observation capabilities.

[0017] A virtual space-driven, integrated sensing and computing power grid observation enhancement device, used to implement the aforementioned virtual space-driven, integrated sensing and computing power grid observation enhancement method, includes: The model building unit is used to build a virtual space-driven integrated power grid observation enhancement system model, which includes a virtual space measurement data perception model, a power business data transmission model, an edge computing model, an edge-side virtual space model, and a power business data timeliness loss model. The optimization modeling unit is used to establish an integrated optimization problem of sensing and computing based on the system model. The optimization objective is to minimize the weighted sum of the timeliness loss of power business data and the cognitive confidence of the virtual space model. The long-term average constraint is transformed into an instantaneous optimization penalty term and decomposed into a joint optimization sub-problem of sensing and communication and a sub-problem of computing resource allocation. The algorithm execution unit is used to execute the virtual space-driven integrated optimization algorithm for sensing and computing. It uses a perception and communication optimization algorithm driven by cognition and guided by world model trajectory to solve the joint optimization sub-problem of perception and communication, and uses an edge computing resource allocation algorithm oriented towards minimizing the timeliness loss of power business to solve the optimization sub-problem of computing resource allocation, so as to realize the collaborative scheduling of sensing and computing resources and improve the observability of the power grid.

[0018] By means of the above-described solution, the present invention has at least the following advantages: (1) This invention proposes an integrated optimization framework for sensing and computing driven by digital twins and world models. First, to meet the needs of dynamic sensing and data transmission in the power grid, a virtual space measurement data sensing model and a power business data transmission model are constructed. Time-division multiple access transmission and data compression mechanisms are adopted to solve the communication resource competition problem between measurement data and business data transmission and optimize data transmission efficiency. Second, for edge-side data processing and latency control, an edge computing model and an edge-side virtual space model based on measurement data are built. A high-precision real-time mirror of the physical state is achieved through digital twins, and multi-step forward prediction of future states is completed through world models. Cognitive inconsistency is introduced to quantify the cognitive bias of the two models and define cognitive confidence to achieve adaptive dynamic triggering of measurement sensing. Finally, a timeliness loss model for power business data is established. The peak information age accurately depicts the information aging degree of data from compression, queuing, transmission, retransmission to successful decoding. Based on virtual space joint cognition, the global resource collaborative scheduling of sensing and computing is driven to improve the observability of power grid status and the stability of business operation.

[0019] (2) This invention proposes a co-optimization algorithm for perception and communication driven by cognition and guided by world model trajectory, which overcomes the limitations of short-sighted decision-making and insufficient generalization ability in traditional deep reinforcement learning. Based on the adaptive adjustment of the step size of the world model look-ahead inference based on cognitive confidence, the cognitive trajectory sequence in virtual space is fused with cognitive confidence to construct an enhanced state space, thereby improving the state perception capability in complex dynamic environments. At the same time, based on the multi-step sequence sampling and cumulative reward mechanism of trajectory similarity, the short-sighted decision-making problem caused by the single-step reward of traditional DRL is overcome, and the collaborative optimization of the compression ratio of virtual space measurement perception and business data transmission is achieved. This not only alleviates the competition for communication resources but also improves transmission efficiency, effectively reduces the timeliness loss of power business data, and improves the observability of the power grid.

[0020] (3) This invention proposes an edge computing resource allocation method aimed at minimizing the timeliness loss of power business operations, achieving a synergistic reduction in the PAoI of business data and decoding latency. By smoothly and continuously approximating the PAoI evolution using a hyperbolic tangent function, the non-convex optimization problem is transformed into a convex optimization problem that can be solved efficiently. Edge computing resources are adaptively allocated based on KKT conditions. Under the premise of meeting the upper limit constraint of edge server computing resources, priority is given to ensuring the data decoding and processing of latency-sensitive businesses, significantly reducing the degree of information aging and the risk of transmission congestion. While improving the utilization rate of computing resources, the timeliness and observability of the entire power grid interconnection computing link are comprehensively enhanced. Compared with existing resource allocation strategies based solely on queue backlog or fixed priority, this method can accurately respond to the timeliness requirements of power business data, significantly reduce the degree of information aging, provide high-quality data support for latency-sensitive businesses such as distributed photovoltaic scheduling and load forecasting, and further improve the observability of the system.

[0021] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show a certain embodiment of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a diagram of the virtual space-driven integrated sensing and computing power grid observation enhancement system of the present invention; Figure 2 This is a flowchart of the virtual space-driven integrated optimization method for sensory computation of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] This invention comprises two parts: a virtual space-driven integrated sensing and computing power grid observation enhancement system and method, which are described in detail below.

[0026] This invention proposes a virtual space-driven, integrated sensing and computing power grid observation enhancement system, such as... Figure 1 As shown, it includes the device layer, network layer, virtual space layer, control layer, and cloud service layer. The details of each layer are as follows.

[0027] The device layer consists of IoT terminals deployed on power equipment such as distributed photovoltaics and charging piles. It is responsible for sensing two types of key data: one is power business data for the cloud business layer, whose latency determines the service quality; the other is measurement data used to build and update the virtual space, whose accuracy and completeness directly affect the cognitive confidence of the virtual space and the observability of the system.

[0028] The network layer consists of 6G base stations and edge servers, providing communication and computing services for IoT devices within the coverage area, supporting efficient data transmission and near-end processing.

[0029] The virtual space layer runs on the edge server, constructs and dynamically updates the virtual space based on the uploaded measurement data, and realizes a two-dimensional characterization of the physical state: first, it realizes a high-precision real-time mirror of the current physical state through digital twins; second, it realizes multi-step forward prediction of future states without real-time observation through world models.

[0030] The control layer deploys a deep reinforcement learning agent, which performs joint optimization decision-making based on the enhanced state estimation provided by the virtual space, integrating synesthesia and computation.

[0031] The cloud-based business layer enhances the observability of the power grid based on power business data, thereby supporting the reliable operation of core businesses such as distributed photovoltaic dispatch, load forecasting, and energy storage charging and discharging dispatch. Among these, the improvement of power grid observability has strict requirements for data timeliness.

[0032] Specifically, each edge server interacts with IoT devices within its coverage area, using the precise characterization and forward-looking prediction of physical states in virtual space to dynamically adjust perception triggering strategies, communication resource allocation, and computing resource scheduling. This ensures the timeliness of business operations while enhancing the cognitive confidence in virtual space, achieving global optimization for long-term performance, and improving the observability of the power grid.

[0033] This invention proposes a virtual space-driven integrated optimization method for synesthesia and computation, the specific process of which is as follows: Figure 2 As shown.

[0034] S1 constructs a virtual space-driven integrated power grid observation enhancement system model, including a virtual space measurement data perception model, a power business data transmission model, an edge computing model, an edge-side virtual space model based on measurement data, and a power business data timeliness loss model.

[0035] S1.1 Virtual Space Measurement Data Perception Model Using a discrete time-slot model, the total time is divided into... There are 1 time slot, and the time slot set is 1 .consider There are 1 edge servers, and their set is 1 Edge server Within the scope of service The set of terminals is .

[0036] In the time slot ,terminal Virtual space measurement data needs to be uploaded to update the digital twin and world model. Define the terminal. In the time slot The virtual space measurement perception variable is ,when At that time, the terminal The system senses and uploads virtual space measurement data. At this point, there is a communication resource contention between the uploading of power service data and the uploading of virtual space measurement data. Therefore, a Time Division Multiple Access (TDMA) transmission method is adopted, and the terminal... The transmission of virtual space measurement data will be completed first, and the remaining time slots will be used to transmit power business data.

[0037] S1.2 Power Business Data Transmission Model Because the volume of power business data is much larger than that of virtual space measurement data, it is prone to transmission congestion and increased latency under the condition of limited wireless channel resources. Therefore, a transmission data compression mechanism is introduced to reduce transmission load and improve channel utilization. Based on the transmission data compression ratio, a transmission model is constructed to achieve synergistic optimization of compression efficiency and successful decoding.

[0038] terminal The data compression ratio is defined as This represents the ratio of the compressed power service data volume to the uncompressed data volume. The upper and lower bounds of the data compression ratio are defined as follows: and Discretize the range of values ​​for the transmitted data compression ratio into... Each level is represented as (1) In the formula, Index for data compression ratio levels.

[0039] The data compression rate is expressed as (2) In the formula, For the terminal CPU computing resources For the terminal Use data compression ratio The number of CPU cycles required to compress a unit of data per second.

[0040] According to Shannon's theorem, the terminal In the time slot The theoretical transmission rate of business data is expressed as (3) In the formula, , , and Terminals With edge servers Transmission bandwidth, transmit power, channel gain, and electromagnetic interference power between them. This represents the power of Gaussian white noise.

[0041] The actual transmission rate of power business data is limited by the theoretical transmission rate and compression rate, expressed as: (4) In the time slot ,terminal The amount of raw power service data successfully transmitted was (5) In the formula, This represents the time slot length. and Representing terminals respectively The amount of virtual space measurement data transmitted and the amount of power business data collected. This represents the amount of power service data retransmitted due to decoding failure. Queue backlog is represented as (6) S1.3 Edge Computing Model Construct mapping function , indicating edge server In the time slot Processing terminal No. Each power service data packet is in a time slot Sent. Assume the data size of the packet before compression is... The decoding delay of this data packet is expressed as: (7) In the formula, For edge servers The number of CPU cycles required to decode a unit of data. For edge servers In the time slot Allocation for decoding terminals Computational resources for data.

[0042] S1.4 Edge-side Virtual Space Model Based on Measurement Data Edge server Based on static data, virtual space measurement data, and learnable parameters, a digital twin model is constructed to characterize the current operational status of the network and its environmental impact. (In time slots) Edge server The digital twin mapping state is represented as (8) In the formula, For digital twins, it is a nonlinear mapping function. It is a static data set for the digital twin model, including time-invariant data such as terminal types, hardware parameters, and communication interface specifications within the management scope. It is configured only during the initialization phase and does not require real-time updates thereafter. It is the latest set of measurement data received by the digital twin model, including time-varying data such as channel state information, interference power, noise power, terminal location, and terminal queue backlog, reflecting the real-time status of the network. This is the set of learnable parameters for the mapping function.

[0043] definition For edge servers Time slots calculated based on virtual space measurement data Real physical state. Define the Euclidean distance between the real physical state and the digital twin mapping state. To minimize the digital twin error, the target parameters of the digital twin are iteratively updated based on gradients along the direction of minimizing the digital twin error. , represented as (9) In the formula, The learning rate is used to adaptively adjust the update step size.

[0044] To avoid oscillations or abrupt changes during parameter updates, a soft update strategy is adopted for time slots. The original parameter set Next slot parameters updated based on digital twin error Perform a weighted fusion update, represented as (10) In the formula, This is the smoothing coefficient.

[0045] The core function of a world model is to perform forward-looking projections of future states based on historical operational data and system dynamics, even in scenarios lacking real-time measurement data. It involves multi-step projections using state transition functions. The future state is derived recursively and iteratively. The world model applies to time slots. Predicted state Represented as (11) In the formula, For when Time-world model for time slots The predicted state. Forward steps, Dynamically adjustable maximum number of forward steps. State transition function The set of learnable parameters.

[0046] Define the Euclidean distance between the physical reality and the world model's predicted state. For the world model error, update the objective parameters of the world model iteratively based on gradients along the direction of minimizing the world model error. Real-time updates can easily lead to overfitting and increase computational costs. Therefore, the world model employs a low-frequency fine-tuning mechanism, updating every... The world model is updated in each time slot. The next objective parameter update is represented as: (12) In the formula, The learning rate is used to ensure stable convergence during the update process.

[0047] The world model parameter update process is represented as follows: (13) In the formula, This is the smoothing coefficient.

[0048] Cognitive inconsistency This is used to quantify the degree of discrepancy between digital twins and world models in their descriptions of the physical system's state, serving as a trigger for measurement data perception. When When the digital twin is small, it aligns with the world model's cognition, reducing the need for uploading measurement data to free up communication resources; when... When the magnitude is large, measurement data should be collected and uploaded in a timely manner to eliminate cognitive discrepancies and ensure the accuracy of the physical state description. Represented as (14) Cognitive confidence varies with cognitive inconsistency The increase and decrease are expressed as: (15) In the formula, This is the adjustment coefficient.

[0049] S1.5 Timeliness Loss Model of Power Business Data Based on an updated virtual space model, synergistic computing decisions are guided, thereby reducing the timeliness loss of power business data and improving grid observability. Peak information age is used to quantify the timeliness loss of power business data. PAoI can accurately characterize the maximum information aging degree of business data throughout the entire process from compression, queuing, transmission, retransmission to successful decoding.

[0050] To quantify the impact of decoding failures caused by compression on the timeliness of power business data, a decoding state variable is defined. When the terminal In the time slot The When a power business data packet is successfully decoded ;on the contrary .

[0051] terminal In the time slot The The PAoI of a power service data packet is represented as follows: (16) In the formula, the first term Indicates the first The latency from packet generation to decoding includes compression latency, end-side queuing latency, and edge-end transmission latency. Definition For the first The second term of the formula represents the number of decoding attempts required to successfully decode a data packet. The total delay of the decoding. The third term in the formula is the total delay of the decoding. Decoding delay.

[0052] terminal In the time slot PAoI is represented as (17) S2 Synesthesia-Computation Integration Optimization Problem Modeling This invention aims to address the key scientific problems of resource allocation decision mismatch, significant timeliness losses in power services, and low system observability caused by inaccurate virtual space characterization of physical states. This is achieved through joint optimization of virtual space measurement and sensing variables. Data compression ratio With edge computing resource allocation This aims to minimize the weighted sum of timeliness loss in power business data and cognitive confidence in the virtual space model, thereby improving system observability. The optimization problem is summarized as follows: (18) In the formula, and These are the weighting coefficients for the timeliness loss of power business data and cognitive confidence, respectively. Represents edge server The allocated computing resources shall not exceed its own computing resource limit. . This indicates the long-term PAoI constraint of improved grid observability on terminal service data. This is the maximum PAoI threshold for terminal service data. This represents a long-term cognitive confidence constraint. The minimum cognitive confidence threshold. This represents a long-term time-averaged bounded constraint on the backlog of terminal queues, preventing data from accumulating indefinitely and causing system instability.

[0053] question The core difficulty stems from the inherent contradiction between long-run time averaging constraints and short-run optimization. - want , and While the optimization problem satisfies thresholds and stability criteria over long timescales, it performs resource optimization by maximizing the objective function within each time slot. This inconsistency causes short-term optimality to constrain future states through inter-slot coupling and cumulative effects, potentially leading to violations of long-term constraints. Therefore, the optimization problem... Transform into

[0054] (19) In the formula, , These are the PAoI virtual queue deficit and the cognitive confidence virtual queue deficit, respectively. The weighting coefficients are used to adjust the priority of the penalty terms in the original optimization objective. By introducing the PAoI virtual queue deficit and the cognitive confidence virtual queue deficit, the long-term average constraint is transformed into an instantaneous optimization penalty term, which is minimized in each time slot. The goal is to control the accumulation of constraint deviations on a time-slot basis, so that long-term constraints are naturally satisfied; at the same time, the constraints are fully preserved. The optimization orientation of timeliness and cognitive confidence can achieve instantaneous optimization while stably achieving the optimization goal of the original problem, effectively resolving the inherent contradiction between short-term optimization and long-term constraints.

[0055] Will The problem is broken down into two sub-problems: SP1, "Joint Optimization of Virtual Space Measurement Sensing Variables and Transmission Data Compression Ratio," and SP2, "Optimization of Computational Resource Allocation." For these sub-problems, this invention proposes a virtual space-driven integrated sensing-computing optimization algorithm. First, a co-optimization method of perception-communication driven by cognition and guided by world model trajectory is proposed. This method enhances the state perception capability of the DRL by dynamically adjusting the look-ahead inference step size of the world model and fusing the cognitive trajectory sequence and cognitive confidence in the virtual space. Furthermore, multi-step sequence sampling and cumulative rewards based on trajectory similarity are used to jointly optimize the virtual space measurement sensing variables and transmission data compression ratio. Second, an edge computing resource allocation method for minimizing the timeliness loss of power business is proposed. The function smoothly and continuously approximates the PAoI evolution, adaptively allocating edge computing resources to accelerate data decoding and reduce information aging, thereby effectively reducing the timeliness loss of power services and improving the observability of the system. The specific implementation will be detailed in step S3, including two parts: "a perception-communication optimization algorithm driven by cognition and guided by world model trajectory" and "an edge computing resource allocation method for minimizing timeliness loss in power services." The specific implementation steps are as follows.

[0056] S3 virtual space-driven integrated synesthetic computing optimization algorithm S3.1 Cognitive-Driven Perception-Communication Optimization Algorithm Guided by World Model Trajectory (1) Reinforced state space construction driven by cognitive confidence 1) State space: Defined as real-time mapping state of digital twin and The cognitive trajectory sequence pieced together from the predicted states of the step-world model, i.e. (20) The virtual space significantly enhances the state awareness capabilities of DRL. Specifically, the enhanced DRL state space integrates edge-side queue backlog, PAoI virtual queue deficit, cognitive confidence virtual queue deficit, virtual space cognitive confidence, and virtual space cognitive trajectory sequence. Edge server In the time slot The perceived state space is constructed as follows: (twenty one) This invention employs a cognitive confidence-driven dynamic look-ahead step size adjustment mechanism to address the inefficiency of fixed-step-size prediction in traditional world models. This mechanism adaptively adjusts the look-ahead step size based on cognitive confidence. , represented as (twenty two) in and For cognitive confidence threshold, and These represent the minimum and maximum lookahead step sizes, respectively. This mechanism uses a small step size when cognitive confidence is high (model self-consistency). To ensure precise and stable convergence; when cognitive confidence is low (significant model bias), a large step size is used. Rapid corrections are used to avoid persistent deviations; when cognitive confidence is at a moderate level, a linearly increasing step size is adopted to achieve a dynamic balance between "convergence speed" and "computational overhead".

[0057] 2) Action Space: Defined as the combination of virtual space measurement data perception variables and business data transmission compression ratio, expressed as... (twenty three) 3) Reward Function: Price the reward function as the optimization objective. The negative value is represented as (twenty four) 4) DAC Architecture: This invention employs the Actor-Critic framework to achieve collaborative optimization of policy generation and value evaluation. The Actor network enhances the state... As input, the mean of the perceived variables of the virtual space measurement data is output. and variance and the average compression ratio of business data transmission. and variance .in, This represents the parameters of the Actor network.

[0058] Critic network parameters are used to estimate state values. Evaluate the long-term benefits of the action. The evaluation function of the Critic network is represented. This represents the parameters of the Critic network.

[0059] 5) Sensor-based motion generation mechanism: based on enhanced state With Actor network parameters Construct a mean of variance is The data follows a normal distribution. Subsequently, by sampling from this normal distribution, a perception decision corresponding to the measurement data is generated. The initial action is as follows: (25) Similarly, the initial actions corresponding to the business data transmission compression ratio are as follows: (26) Furthermore, regarding the initial action and Cutting process (27) (28) In the formula, For the floor function, This is the clipping function.

[0060] By introducing stochastic exploration into a deterministic policy, this mechanism enhances the efficient exploration of the measurement-aware action space, thereby improving optimization robustness and global search capability. The generated action is then executed. , so that the state changes from Transfer to Based on observed rewards , construct samples And store it in the experience replay buffer. .

[0061] (2) Multi-step sequence sampling guided by the world model prospective trajectory This invention utilizes a cognitive trajectory sequence constructed based on a world model to achieve multi-step sequence sampling based on trajectory similarity.

[0062] In each training round, the current time slot will be... Cognitive trajectory sequence In the playback pool The cosine similarity of the historical sample trajectory sequences generated in the time slots is calculated and represented as follows: (29) This invention determines the sampling probability by measuring the relative fit of each historical sample with other samples. Definition The sampling probability of samples generated in the time slot is (30) In the formula, For sampling weights, This represents the size of the experience base. When cognitive confidence is low, it tends to extract historical samples that are highly similar to the current trajectory to enhance the learning ability in high-uncertainty scenarios and quickly correct model biases. When cognitive confidence is high, it performs uniform sampling to enhance sample diversity and improve the robustness of the strategy by encouraging exploration.

[0063] Define edge server In the time slot From the experience replay pool Medium sampling yields a small batch of sample sets. ,in, Indicates the first One sample, The total number of samples collected. The start time slot of the trajectory stored in each sample is defined as To overcome the short-sighted decision-making problem caused by the traditional single-step reward mechanism of DRL, the first step is defined as... The cumulative reward for the trajectory corresponding to each sample is: , indicating from time slot Forward The cumulative reward for each step is expressed as (31) In the formula, This is the discount factor.

[0064] Define sample The timing difference error is (32) in, This is a discount factor used to balance the importance of current and future returns.

[0065] The loss function is defined as (33) Edge server Update based on gradient descent method and

[0066] (34) (35) in and These are the learning step sizes for the Actor network and the Critic network, respectively.

[0067] S3.2 Edge computing resource allocation algorithm for minimizing the timeliness loss of power business information Edge servers allocate computing resources to the business data uploaded by decoding terminals, ensuring the timeliness of power business information. Based on the sensor-based decisions, the optimization problem is restructured into an edge computing resource allocation optimization problem, which is summarized as follows: (36) definition For time slots The set of sequence numbers of successfully processed data packets. Then PAoI is updated to... (37) The last item represents the decoding success indicator function, which can be rewritten as follows: (38) Furthermore, by introducing the hyperbolic tangent function to approximate equation (37) as a continuous function, it can be expressed as follows: (39) because The function is smooth and monotonic, and equation (39) is related to the decision variables. Continuously differentiable. Combined with linear constraints. ,question It is transformed into a convex optimization problem, which can be solved using the KKT method.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A virtual space-driven integrated sensor-computer interface method for improving the observability of a power grid, characterized in that, include: S1. Construct a virtual space-driven integrated power grid observation enhancement system model, which includes a virtual space measurement data perception model, a power business data transmission model, an edge computing model, an edge-side virtual space model, and a power business data timeliness loss model. S2. Based on the system model, establish an integrated optimization problem of sensing and computing, with the weighted sum of minimizing the timeliness loss of power business data and the cognitive confidence of the virtual space model as the optimization objective. The long-term average constraint is transformed into an instantaneous optimization penalty term, and decomposed into a joint sensing-communication optimization sub-problem and a computing resource allocation optimization sub-problem. S3. Execute the virtual space-driven integrated optimization algorithm for sensing and computing, and use the perception-communication optimization algorithm driven by cognition and guided by world model trajectory to solve the joint optimization sub-problem of perception and communication. Use the edge computing resource allocation algorithm oriented towards minimizing the timeliness loss of power business to solve the optimization sub-problem of computing resource allocation, so as to realize the collaborative scheduling of sensing and computing resources and improve the observability of the power grid.

2. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 1, characterized in that, The construction of the edge-side virtual space model includes: Based on static data, virtual space measurement data, and learnable parameters, a digital twin model is constructed to achieve a high-precision real-time mirror of the physical power grid state. The digital twin error is minimized through gradient iteration and soft update strategies. A world model is established based on historical operational data and system dynamics. Multi-step forward-looking state simulations are performed under conditions without real-time observation. The parameters of the world model are updated using a low-frequency fine-tuning mechanism. The cognitive inconsistency between the digital twin and the world model is calculated, and a cognitive confidence level is defined as an adaptive trigger for measuring data perception.

3. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 2, characterized in that, The cognitive inconsistency refers to the degree of discrepancy between the digital twin and the world model in their descriptions of the physical system's state. It serves as the trigger for sensing measurement data. The cognitive confidence decreases as cognitive inconsistency increases, and is used to dynamically adjust the frequency of measurement data uploads and the allocation of communication resources.

4. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 1, characterized in that, The power business data timeliness loss model uses peak information age (PAoI) quantification, where PAoI is used to characterize the maximum information aging degree of business data throughout the entire process from compression, queuing, transmission, retransmission to decoding.

5. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 1, characterized in that, The cognitive-driven and world model trajectory-guided perception-communication optimization algorithm includes: The world model's forward inference step size is dynamically adjusted based on cognitive confidence, and an enhanced state space is constructed by integrating cognitive trajectory sequences and cognitive confidence. Using the virtual space measurement and sensing variables and the compression ratio of business data transmission as the action space, an Actor-Critic architecture is adopted to generate sensing decisions. Multi-step sequence sampling is performed based on trajectory similarity, and cumulative rewards are used instead of single-step rewards to overcome the short-sightedness problem in decision-making in deep reinforcement learning, thereby achieving joint optimization of perception strategy and compression ratio.

6. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 5, characterized in that, The dynamic look-ahead step size adjustment mechanism is as follows: When the cognitive confidence level is higher than the high threshold, the minimum look-ahead step size is used. When the cognitive confidence level is below a low threshold, the maximum look-forward step size is used. When cognitive confidence is in the middle range, use a linear growth step size.

7. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 5, characterized in that, The multi-step sequence sampling determines the sampling probability by calculating the cosine similarity between the current cognitive trajectory and the historical sample trajectory. When the cognitive confidence is low, similar trajectories are sampled first, and when the cognitive confidence is high, uniform sampling is performed.

8. The virtual space-driven integrated sensing and computing power grid observation enhancement method according to claim 1, characterized in that, The edge computing resource allocation algorithm includes: The evolution of PAoI is smoothly and continuously approximated by the hyperbolic tangent function, transforming the non-convex optimization problem into a convex optimization problem; Edge computing resources are allocated adaptively based on KKT conditions, prioritizing the data decoding and processing of latency-sensitive power services, and reducing the risk of information aging and transmission congestion.

9. A virtual space-driven integrated sensing and computing power grid observation enhancement system, used to execute the virtual space-driven integrated sensing and computing power grid observation enhancement method according to any one of claims 1-8, characterized in that, include: The equipment layer is used to collect power business data and virtual space measurement data; The network layer, consisting of 6G base stations and edge servers, is used to provide communication and computing services; The virtual space layer, deployed on edge servers, is used to build a self-consistent virtual space that coordinates digital twins and world models, enabling real-time mirroring of physical states and multi-step forward-looking simulations. The control layer deploys deep reinforcement learning agents to perform integrated synesthesia-computation joint optimization decisions. The cloud-based business layer is used to support distributed photovoltaic dispatching, load forecasting, and energy storage charging and discharging dispatching services based on optimized grid observation capabilities.

10. A virtual space-driven integrated sensing and computing power grid observation enhancement device, used to implement the virtual space-driven integrated sensing and computing power grid observation enhancement method according to any one of claims 1-8, characterized in that, include: The model building unit is used to build a virtual space-driven integrated power grid observation enhancement system model, which includes a virtual space measurement data perception model, a power business data transmission model, an edge computing model, an edge-side virtual space model, and a power business data timeliness loss model. The optimization modeling unit is used to establish an integrated optimization problem of sensing and computing based on the system model. The optimization objective is to minimize the weighted sum of the timeliness loss of power business data and the cognitive confidence of the virtual space model. The long-term average constraint is transformed into an instantaneous optimization penalty term and decomposed into a joint optimization sub-problem of sensing and communication and a sub-problem of computing resource allocation. The algorithm execution unit is used to execute the virtual space-driven integrated optimization algorithm for sensing and computing. It uses a perception and communication optimization algorithm driven by cognition and guided by world model trajectory to solve the joint optimization sub-problem of perception and communication, and uses an edge computing resource allocation algorithm oriented towards minimizing the timeliness loss of power business to solve the optimization sub-problem of computing resource allocation, so as to realize the collaborative scheduling of sensing and computing resources and improve the observability of the power grid.

Images