Power distribution communication network topology optimization system based on Internet of Things

Abstract

The invention provides a power distribution communication network topology optimization system based on the Internet of Things, and relates to the technical field of power Internet of Things, which obtains and fuses topology and fault state data from a power grid SCADA system and slice performance and bearer network state data from a 5G network management system in real time. The system can construct a fusion state vector comprehensively describing the power grid-communication network joint operation situation. The vector is input into an intelligent decision-making model based on imitation-reinforcement learning training so as to generate an optimal action instruction in real time. The instructions are accurately analyzed into dynamic adjustment commands for 5G network slice resources and routing policy change instructions for key service flows. According to the concept, the problems that in an existing scheme, a communication network perceives the power grid state blindly, and slice resource configuration is rigid are solved, communication resources can be instantly and intelligently inclined and reconfigured according to the real-time and specific requirements of the power grid, and the stability and safety of power grid operation are improved.

Description

Technical Field

[0001] This application relates to the field of power Internet of Things (IoT) technology, and more specifically, to an IoT-based power distribution communication network topology optimization system. Background Technology

[0002] With the construction of new power systems, distribution networks are connecting massive amounts of distributed energy resources, charging piles, and various smart terminals, resulting in unprecedented diversity in their communication needs. For example, control services such as feeder automation require millisecond-level ultra-low latency and ultra-high reliability; electricity consumption information collection involves massive concurrent connections from a large number of terminals; and services such as video surveillance require high bandwidth. 5G network slicing technology, by virtualizing the physical network into multiple logically isolated dedicated networks, provides a feasible technical framework for meeting these differentiated quality of service requirements.

[0003] However, in the practical application of 5G slicing technology to power distribution networks, existing solutions have exposed some technical problems, especially in dealing with extreme scenarios such as sudden power grid failures. These limitations are primarily rooted in the deep disconnect between communication networks and power grid systems at the state awareness and control decision-making levels. On the one hand, network slicing resource allocation is mostly static, meaning it's pre-allocated based on long-term average service demand. This rigid strategy cannot adapt to the instantaneous dynamics of the power grid. For example, when a line fault occurs, automated terminals in the area may instantly generate high-priority signaling far exceeding normal loads, quickly exhausting pre-allocated slice resources. At this time, idle resources in other service slices cannot be urgently accessed due to strict isolation mechanisms, potentially leading to timeouts or packet loss of critical control commands, threatening power grid security. On the other hand, the 5G network management system is blind to the power grid's operational topology and the urgency of services. It can perceive communication phenomena such as sudden increases in slice traffic and latency degradation, but it cannot understand the underlying nature of the power grid events. Therefore, its response measures are often passive and suboptimal. The system lacks an intelligent decision-making core capable of deeply integrating real-time power grid status, making it unable to dynamically schedule resources between slices or intelligently reselect better logical topologies for critical business flows within a slice based on the actual needs of the power grid. Therefore, building a system that can deeply perceive the power grid status and drive joint optimization of resources and topology has become a pressing technical challenge in this field.

[0004] Therefore, an optimized power Internet of Things (IoT) resource scheduling system is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an Internet of Things-based power distribution communication network topology optimization system, comprising: The data acquisition module is used to acquire raw power grid data and raw communication data; The feature extraction and state vector construction module is used to extract features and construct state vectors from raw power grid data and raw communication data to obtain the power grid-communication network fusion state vector. The action vector acquisition module is used to input the power grid-communication network fused state vector into a pre-trained policy network to obtain action vectors; The instruction generation module is used to generate slice configuration instructions and routing policy instructions based on action vectors.

[0006] Compared with existing technologies, this application provides an IoT-based power distribution communication network topology optimization system. By acquiring and fusing topology and fault status data from the power grid SCADA system with slice performance and bearer network status data from the 5G network management system in real time, the system can construct a fused state vector comprehensively describing the joint operation of the power grid and communication network. This vector is input into an intelligent decision-making model trained using imitation-reinforcement learning to generate optimal action commands in real time. These commands are precisely parsed into dynamic adjustment commands for 5G network slice resources and routing strategy change commands for critical service flows. This concept solves the problems of blind perception of power grid status and rigid slice resource allocation in existing solutions, enabling communication resources to be instantaneously and intelligently tilted and reconfigured according to the real-time and specific needs of the power grid. This provides dynamic and reliable communication guarantees for critical power distribution services, significantly improving the stability and security of power grid operation. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a system block diagram of an IoT-based power distribution communication network topology optimization system according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of data flow in an IoT-based power distribution communication network topology optimization system according to an embodiment of this application.

[0010] Figure 3 This is a block diagram of the feature extraction and state vector construction module in the Internet of Things-based power distribution communication network topology optimization system according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of a feature fusion unit in an IoT-based power distribution communication network topology optimization system according to an embodiment of this application. Detailed Implementation

[0012] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0013] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0014] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0015] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0016] Currently, when applying 5G network slicing technology to distribution communication networks, there are common problems such as insufficient real-time state awareness of the power grid and rigid allocation of slice resources. This results in the communication network being unable to adaptively adjust in emergency situations such as sudden power grid failures, potentially jeopardizing the reliable transmission of critical services and the safety and stability of the power grid. Therefore, this application proposes an IoT-based distribution communication network topology optimization system. The system first uses a data acquisition module to collect raw power grid data from the power grid SCADA system and raw communication data from the 5G network management system in real time. Subsequently, a feature extraction and state vector construction module performs deep processing on these two types of heterogeneous data: it uses graph neural networks and other technologies to extract power grid topology features from the raw power grid data, forming a power grid embedding vector; simultaneously, it performs performance feature engineering on the raw communication data to obtain a communication feature vector. Next, through a dependency-based gating fusion mechanism, these two vectors are combined into a fused state vector that comprehensively represents the joint operation status of the power grid and communication network. This fused state vector is input into a policy network pre-trained with expert data to generate an optimal action vector containing resource adjustment and routing information in real time. Finally, the instruction generation module precisely parses this action vector into specific slice configuration instructions and routing policy instructions. The former is used to dynamically adjust the resource allocation of each network slice, while the latter is used to optimize the transmission paths of critical service flows. Through this series of steps, this solution enables communication resources to be intelligently tilted and dynamically reassigned according to the actual, instantaneous needs of the power grid, thereby solving the bottlenecks of existing technologies.

[0017] Figure 1 This is a system block diagram of an IoT-based power distribution communication network topology optimization system according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in an IoT-based power distribution communication network topology optimization system according to an embodiment of this application. Figure 1 and Figure 2 As shown, the IoT-based power distribution communication network topology optimization system 100 according to an embodiment of this application includes: a data acquisition module 110 for acquiring raw power grid data and raw communication data; a feature extraction and state vector construction module 120 for performing feature extraction and state vector construction on the raw power grid data and raw communication data to obtain a power grid-communication network fused state vector; an action vector acquisition module 130 for inputting the power grid-communication network fused state vector into a pre-trained policy network to obtain action vectors; and an instruction generation module 140 for generating slice configuration instructions and routing policy instructions based on the action vectors.

[0018] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the data acquisition module 110 is used to acquire raw power grid data and raw communication data. It should be understood that due to the physical and logical separation between the communication network and the power grid system, the management and scheduling decisions of the communication network lack direct perception of the real-time operating status of the power grid, making it unable to make optimal or even correct resource allocations in response to dynamic changes in the power grid, especially sudden faults. Therefore, in the technical solution of this application, raw power grid data and raw communication data are acquired to provide comprehensive and real-time cross-domain status input for subsequent intelligent decision-making. This fundamentally solves the information barrier between the two domains, enabling the topology optimization and resource scheduling of the communication network to closely align with the actual business needs and urgency of the power grid, laying the foundation for truly achieving grid-driven business operations.

[0019] More specifically, in a specific example of this application, the data acquisition module performs the data acquisition process including the following steps. First, the data acquisition module accesses the data bus of the distribution automation master station through an industry-standard interface and initiates a data subscription request to acquire raw power grid data reflecting the power grid topology and operating status in real time. For example, when switch S12 on feeder F05 trips due to overload, the SCADA system immediately generates and publishes a remote signaling change message containing the device ID, the switch remote signaling status changing from closed to open, and the associated fault alarm identifier. This message is captured by the data acquisition module in real time. At the same time, the data acquisition module, as a network management client, initiates periodic data polling to the 5G network management and orchestration system and the SDN controller of the bearer network through a standardized northbound interface. This polling request specifies the key performance indicators to be monitored, thereby acquiring raw communication data. For example, it acquires performance indicator data such as the physical resource block occupancy rate of the uRLLC network slice carrying distribution automation services reaching 90% and the queue latency being 12 milliseconds under base station Cell_A covering the fault area.

[0020] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the feature extraction and state vector construction module 120 is used to extract features and construct state vectors from the original power grid data and original communication data to obtain a power grid-communication network fusion state vector. It should be understood that, due to fundamental differences in data structure, format, and physical meaning between the original power grid data and communication data, they are heterogeneous and high-dimensional, and cannot be directly understood and processed by the decision model. This raw data format cannot quantitatively reveal the intrinsic correlation between the operating state of the power grid and the performance state of the communication network. Therefore, in the technical solution of this application, feature extraction and state vector construction are further performed on the original power grid data and original communication data to obtain a power grid-communication network fusion state vector. This transforms the discrete power grid topology alarm information and multi-dimensional communication network performance indicators into a standardized, fixed-dimensional mathematical vector, which can quantitatively characterize the joint operating status of the entire system. This provides a precise, information-rich, and machine-readable decision-making basis for subsequent policy networks, enabling the system not only to perceive isolated events but also to understand the causal relationships between events. For example, it can directly reflect from the vector that the fault of feeder F05 is the root cause of the congestion in the uRLLC slice performance of base station Cell_A, providing the necessary prerequisite for generating optimal control commands with power grid situation awareness.

[0021] Figure 3 This is a block diagram of the feature extraction and state vector construction module in an IoT-based power distribution communication network topology optimization system according to an embodiment of this application. Figure 3 As shown, the feature extraction and state vector construction module 120 includes: a topology feature extraction unit 121, used to extract grid topology features from the original grid data to obtain a grid embedding vector; a communication performance feature engineering unit 122, used to perform communication performance feature engineering on the original communication data to obtain a communication feature vector; and a feature fusion unit 123, used to perform feature fusion on the grid embedding vector and the communication feature vector to obtain a grid-communication network fused state vector.

[0022] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the topology feature extraction unit 121 is used to extract power grid topology features from the original power grid data to obtain a power grid embedding vector. It should be understood that, since the original power grid data is essentially graph-structured data describing the physical connections and electrical relationships between devices, its discrete remote signaling states cannot directly and completely reflect the global operating status of the power grid, especially the complex impact of changes in the state of a single device on its neighboring devices and even the topological connectivity of the entire region. Therefore, in the technical solution of this application, a power grid topology graph object is further extracted from the original power grid data, and the power grid topology graph object is input into a graph neural network to obtain a final set of node embedding vectors. Graph pooling is then performed on the final set of node embedding vectors to obtain the power grid embedding vector. This maps and compresses the high-dimensional, discrete, and non-Euclidean power grid topology information into a low-dimensional, dense real-number vector that can globally represent the network structure and state. In this way, the system can not only learn about the isolated event of the state change of switch S12, but also quantitatively understand the structural changes caused by this change to feeder F05 and even the entire regional power supply path through the final generated power grid embedding vector. This provides key quantitative basis for subsequent decision-making models to accurately assess the severity and scope of the power grid event.

[0023] More specifically, in a specific example of this application, the topology feature extraction unit is configured to: extract a power grid topology graph object from the original power grid data; input the power grid topology graph object into a graph neural network to obtain a final set of node embedding vectors; and perform graph pooling processing on the final set of node embedding vectors to obtain the power grid embedding vector.

[0024] Accordingly, a power grid topology map object is extracted from the original power grid data, and this object is input into a graph neural network to obtain the final set of node embedding vectors. It should be understood that the discrete device state list presented by the original power grid data cannot explicitly express the complex topological dependencies between devices, nor can it reveal the chain reaction of a single device state change on the local or even global operation of the power grid. Therefore, in the technical solution of this application, a power grid topology map object is further extracted from the original power grid data, and this object is input into a graph neural network to obtain the final set of node embedding vectors. This transforms the physical structure and operating state of the power grid into a feature space rich in high-dimensional semantic information, where the feature representation of each device contains information about its neighboring topological environment. This allows the system's understanding of the power grid state to deepen from a single, superficial event perception to a profound insight into the structured impact behind the event. For example, the system not only knows that switch S12 is open, but also understands the changes in network connectivity caused by its disconnection as a key node on feeder F05 through its embedding vector.

[0025] More specifically, in a concrete example of this application, the feature extraction and state vector construction module first constructs a power grid topology graph object from the raw power grid data. This module parses the received remote signaling messages and pre-set static connection relationship data of the power grid, creating a corresponding node in the graph data structure for each power device, such as switches S12, S13, and S14. Simultaneously, it generates an initial feature vector for each node, which numerically encodes the static attributes and dynamic states of the device. Subsequently, based on the line connection relationships, the module establishes edges between the nodes representing switch S12 and its adjacent devices, thereby completely constructing a graph object that accurately reflects the current electrical topology and operating state of feeder F05 and its surrounding area. Next, this graph object is fed as input into a pre-trained graph neural network model. This model iteratively updates the feature vector of each node by performing multi-layer neighborhood information aggregation operations. In each layer, the vector of a node is calculated by aggregating the vectors of all its direct neighboring nodes and combining them with the vectors of its previous layer. After all levels of computation, the model output nodes are finally embedded in a set of vectors, where each vector is a highly condensed and abstract representation of the topology and state of the corresponding device and its multi-hop neighborhood.

[0026] Accordingly, the final set of node embedding vectors is processed by graph pooling to obtain the power grid embedding vector. It should be understood that since the final set of node embedding vectors output by the graph neural network is a discrete feature representation for each device node in the power grid, its quantity and content vary with the scale and state of the analyzed power grid topology. This variable-size, distributed feature set cannot be directly used as a stable input for subsequent decision-making models. Therefore, in the technical solution of this application, the final set of node embedding vectors is further processed by graph pooling to obtain the power grid embedding vector. This aggregates and refines multiple vector information representing the states of various local devices into a unique, fixed-length vector that can represent the global state of the entire power grid graph. In this way, the complex and dynamically changing power grid topology state can be compressed into a standardized, global feature fingerprint, providing the system with a macroscopic, quantitative description of the overall operating status of the power grid, thereby ensuring the stability and effectiveness of the input for subsequent fusion and decision-making steps.

[0027] More specifically, in a concrete example of this application, after receiving the final set of node embedding vectors related to feeder F05 from the graph neural network output, the feature extraction and state vector construction module immediately performs graph pooling processing. This set contains multiple high-dimensional feature vectors that correspond to switches S12, S13, S14, and other related power equipment, each already containing rich neighborhood information. The module applies mean pooling to this vector set, that is, it calculates the arithmetic mean for each dimension of all vectors in the set. This process aggregates the feature information of all nodes without discrimination, ultimately generating a single vector with the same dimension as the embedding vector of a single node, but whose value in each dimension represents the average feature of the entire power grid graph in that dimension. This single vector is the power grid embedding vector, which no longer describes the specific state of any single device, but rather quantifies the global characteristics of the power grid topology and operating state of the entire feeder F05 region caused by the switch S12 disconnection event.

[0028] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the communication performance feature engineering unit 122 is used to perform communication performance feature engineering on the raw communication data to obtain a communication feature vector. It should be understood that since the raw communication data consists of multi-dimensional performance indicators with different dimensions and scales, such as resource utilization, latency, and packet loss rate, these discrete raw values ​​cannot directly and comprehensively quantify the network's service quality and congestion status, making them difficult for decision-making models to directly use as a basis for judgment. Therefore, in the technical solution of this application, the raw communication data is further subjected to communication performance feature engineering to obtain a communication feature vector. This refines and transforms the complex and heterogeneous performance indicator data into a standardized set of quantitative features that comprehensively reflects the operational health of each network slice and each network node. This allows the system's evaluation of the communication network status to shift from reading scattered performance parameters to directly determining a congestion index or service level index with clear physical meaning, thereby providing accurate and easily understood communication-side status input for subsequent decision-making models.

[0029] More specifically, in a concrete example of this application, after receiving the raw communication data, the feature extraction and state vector construction module first filters and parses out the performance indicators most relevant to the current scenario. For example, it extracts specific values ​​from the data stream, such as the physical resource block occupancy rate of the uRLLC slice under base station Cell_A being 90%, the queue latency being 12 milliseconds, and the resource occupancy rates of the mMTC and eMBB slices under the same cell being at a relatively low level. Next, the module calculates a normalized congestion index for each slice, which is a weighted value that integrates resource, latency, and reliability indicators. It normalizes the 90% PRB occupancy rate and the 12-millisecond latency value with preset total resources and a 15-millisecond latency alarm threshold, respectively, and then performs a weighted sum according to preset weights, thereby calculating a high congestion index close to saturation for the uRLLC slice. Finally, the module concatenates the calculated congestion indices of all slices, including uRLLC, mMTC, and eMBB, in Cell_A and other related cells, as well as the bandwidth utilization of critical links in the bearer network, in a predetermined order to form a structured, fixed-length communication feature vector for subsequent fusion processing.

[0030] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the feature fusion unit 123 is used to fuse the power grid embedding vector and the communication feature vector to obtain a power grid-communication network fused state vector. It should be understood that the original state fusion mechanism, which directly concatenates the feature vectors of the power grid and the communication network, is essentially an information-preserving but relation-neutral operation. The technical problem with this method is that it treats the power grid feature vector and the communication feature vector as two independent entities, placing them side-by-side in a higher-dimensional vector space, thus completely ignoring the profound and nonlinear special relationship between these two physically tightly coupled systems. Specifically, this special relationship manifests as the statistical correlation and causal dependence between their states. In this coupled power grid-communication network system, a specific power grid state, such as a feeder failure, and a specific communication state, such as a surge in uRLLC slice traffic in the cell to which the feeder's terminal belongs, inevitably have a very strong conditional probabilistic correlation. Simple vector concatenation cannot explicitly measure or encode the inherent characteristics of this joint probability distribution; it shifts the entire burden of learning this complex correlation to the downstream deep reinforcement learning model. This approach inevitably leads to low learning efficiency, requiring the model to consume more training iterations to implicitly uncover cross-domain dependencies from raw, unrefined features. More seriously, in rare but crucial fault scenarios, the key signals describing these dependencies may be relatively weak and easily drowned out by high-variance feature noise within the two vectors, preventing the model from effectively capturing decisive cross-domain associations and thus posing a decision-making risk. Therefore, in this application's technical solution, feature fusion is further performed on the power grid embedding vector and communication feature vector to obtain a power grid-communication network fused state vector. This explicitly models and quantifies the dependencies between these two domains at the feature level, generating a unified state representation that inherently contains cross-domain causal association information. This significantly reduces the learning difficulty and decision-making risk of the subsequent policy network, ensuring that the decisive cross-domain association signal of communication congestion caused by feeder faults is not drowned out, thereby enabling the system to generate more accurate and efficient optimization instructions based on a deep understanding of the nature of the event.

[0031] Figure 4 This is a block diagram of a feature fusion unit in an IoT-based power distribution communication network topology optimization system according to an embodiment of this application. Figure 4As shown, the feature fusion unit 123 includes: a probability integral transformation unit 1231, used to perform a probability integral transformation based on edge distribution on the power grid embedding vector and the communication feature vector to obtain a uniformly distributed power grid vector and a uniformly distributed communication vector; a joint dependency density estimation unit 1232, used to perform joint dependency density estimation on the uniformly distributed power grid vector and the uniformly distributed communication vector to obtain a dependency score vector; and a gated fusion unit 1233, used to perform gated fusion based on dependency score on the power grid embedding vector and the communication feature vector to obtain a power grid-communication network fusion state vector.

[0032] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the probability integral transformation unit 1231 is used to perform a probability integral transformation based on edge distribution on the power grid embedding vector and the communication feature vector to obtain a uniformly distributed power grid vector and a uniformly distributed communication vector. It should be understood that since the components of the power grid embedding vector and the communication feature vector have drastically different physical meanings, dimensions, and statistical distributions—for example, some features may exhibit skewed distributions with significant differences in their value ranges and scales—direct correlation analysis cannot accurately isolate and measure the pure dependency structure between different feature variables, as the analysis results will be severely interfered with by the edge distribution patterns of each variable itself. Therefore, in the technical solution of this application, a probability integral transformation based on edge distribution is further performed on the power grid embedding vector and the communication feature vector to obtain a uniformly distributed power grid vector and a uniformly distributed communication vector. This allows all original feature values ​​to be uniformly mapped to a standard probability space by applying a probability integral transformation to each feature component of the input power grid embedding vector and communication feature vector. In this way, all original feature values ​​with different physical meanings and dimensions can be transformed into a unified probability space. The normalized probability values ​​within the interval generate two new vectors, each of which has a component that follows a standard uniform distribution. This creates an unbiased and standardized data foundation for the next step of purely dependent structure modeling.

[0033] More specifically, in a concrete example of this application, after obtaining the power grid embedding vector representing the fault state of feeder F05 and the communication feature vector containing the high congestion index of the uRLLC slice under base station Cell_A, the feature extraction and state vector construction module independently performs a probability integral transformation based on the empirical cumulative distribution function for each component of these two vectors. First, for a specific component in the power grid embedding vector, the module retrieves historical sample values ​​of that component from the historical database for the past N observation periods. Then, it calculates the number of samples in these N historical sample values ​​that are less than or equal to the current observation value, and divides that number by the total number of samples N, thereby obtaining the empirical cumulative probability of the current observation value in the historical distribution of that feature. Similarly, for a key component in the power grid embedding vector representing the network topology anomaly, its current value is 0.85, and 995 of the N=1000 historical samples have values ​​less than or equal to 0.85, then the output value of this component after transformation is 0.995. For example, if the current congestion index of the uRLLC slice in the communication feature vector is 0.92, and in the historical N=1000 observation samples, there are 980 values ​​less than or equal to 0.92, then the output value of this feature component after transformation is 0.98. This can be expressed by the formula: Here, It is the transformed output value, i.e., the components of the uniformly distributed power grid embedding vector / evenly distributed communication feature vector; It is a feature The empirical cumulative distribution function; is the original feature value of the input, i.e., the component of the power grid embedding vector / communication feature vector; N is the total number of historical observation samples; It is an indicator function that takes the value 1 when the condition is true and 0 otherwise. This is the k-th observation of this feature in the historical dataset. This module repeats this process for all components of the two input vectors, ultimately outputting a uniformly distributed power grid vector and a uniformly distributed communication vector, where all components of these two new vectors are... The probability values ​​within the interval objectively reflect the extreme degree of each feature in its respective historical data. This standardization process is crucial, as it ensures that subsequent joint dependency density estimation can be performed in a fair and unbiased probability space, thereby accurately revealing the true correlation between features from different domains, rather than being misled by their respective original numerical scales.

[0034] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the joint dependency density estimation unit 1232 is used to estimate the joint dependency density of uniformly distributed power grid vectors and uniformly distributed communication vectors to obtain a dependency fraction vector. It should be understood that the two uniformly distributed vectors after probability integral transformation only eliminate the interference of the marginal distribution patterns of each feature variable, but do not reveal or quantify the inherent, complex joint dependency structure between the two domains of power grid state and communication state. The system still cannot determine whether the currently observed cross-domain state combination conforms to the normal pattern of historical data, that is, it cannot determine whether the simultaneous occurrence of a specific power grid event and a specific communication phenomenon is common or abnormal. Therefore, in the technical solution of this application, the joint dependency density of the uniformly distributed power grid vectors and uniformly distributed communication vectors is further estimated to obtain a dependency fraction vector, thereby using the Gaussian Copula function to estimate the joint occurrence probability density between the current power grid state and communication state, that is, to quantify the degree of dependency between them. In this way, a highly condensed intelligent index can be generated, whose score directly reflects the normality or abnormality of the currently observed cross-domain state combination. An extremely low score can become a strong anomalous signal, indicating that the system may be in a rare and dangerous state, thus providing a high-order statistical assessment of whether the current system state is normal or not for subsequent decision-making.

[0035] More specifically, in a concrete example of this application, after obtaining the uniformly distributed power grid vector and communication vector representing the F05 fault and Cell_A congestion state, the feature extraction and state vector construction module first maps each component of these two vectors through the inverse cumulative distribution function of the standard normal distribution to obtain a standard normal space vector, i.e., the intermediate vector. Subsequently, this module uses an empirical correlation matrix, calculated based on historical data and describing the linear correlation between all features, to calculate the probability density function value of the Gaussian Copula. The empirical correlation matrix is ​​pre-calculated based on large-scale historical normal operation data; it solidifies the interdependence and linear correlation patterns between all feature variables of the power grid and communication network under normal conditions in matrix form. This calculation process takes the intermediate vector and the empirical correlation matrix as input and substitutes them into the Gaussian Copula probability density function formula for calculation. The calculation formula is as follows: in, It is the calculated dependency score; Represented by the correlation matrix The Gaussian Copula density function with parameters in The value of the point; It is the determinant of the correlation matrix; Through The resulting standard normal space vector, i.e., the intermediate vector, It is the standard normal quantile function; It is the inverse of the correlation matrix; It is the identity matrix. Due to a severe fault in feeder F05, which directly caused a surge in traffic to the uRLLC slice to its terminal to near saturation, this strong causal relationship is likely a rare event compared to the normal patterns captured in historical data. Therefore, the calculated probability density function value, i.e., the final dependency score, will be an extremely low value. Mathematically, this is because there is a significant deviation between the current state (represented by the intermediate vector) and the historical norm (represented by the empirical correlation matrix), resulting in a significant deviation in the exponential term of the Gaussian Copula density function. The value of becomes extremely large, causing the final exponential function value (i.e., the dependency score) to approach zero. This physically captures precisely the low-probability, high-risk system state resulting from the simultaneous occurrence of two historically rare, extreme events: feeder failure and uRLLC slice congestion. This module uses this dependency score as a key component of the dependency score vector to explicitly encode the high anomaly of the current system state.

[0036] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the gating fusion unit 1233 is used to perform dependency-score-based gating fusion of the power grid embedding vector and the communication feature vector to obtain a power grid-communication network fusion state vector. It should be understood that since the obtained dependency score vector is an independent intelligent indicator, simply concatenating it with the power grid and communication feature vectors would lose the dynamic guidance information provided by the score, making it impossible to adjust the importance of different features in the final state representation. This approach ignores the need to give higher attention to certain specific features (such as faulty line status and congestion slice delay) when the system state is abnormal. Therefore, in the technical solution of this application, the power grid embedding vector and the communication feature vector are further gating fusion based on the dependency score vector to obtain a power grid-communication network fusion state vector. This utilizes the dependency score as a dynamic control signal, employing a gating mechanism to calculate the weights of power grid and communication features in the final state representation. In this way, a fusion state vector with extremely high information density and rich cross-domain dependency context can be generated. This vector not only contains the independent states of each subsystem, but also embeds dynamic and quantitative evaluations of the relationships between them. When the system state is abnormal, this mechanism can play the role of an attention regulator, amplifying or weakening the influence of a certain domain feature, thereby guiding the downstream model to focus on the most critical information at present and providing high-quality input for subsequent intelligent decision-making.

[0037] More specifically, in a particular example of this application, the feature extraction and state vector construction module performs fusion through a dependency score-based gating mechanism. The calculation process of this mechanism can be represented by the following formula: In the above formula, It is the gating signal vector; It is the Sigmoid activation function; and These are learnable weights and bias parameters; It is the converged state vector of the power grid and communication network; Represents element-wise product; This represents a vector concatenation operation; and These are the input power grid and communication feature vectors, respectively. The weights W and bias b, as learnable parameters, are automatically optimized and learned through backpropagation based on the reward signal of the final task during the end-to-end training of the entire downstream deep reinforcement learning model. This enables the gating signal generation process to adaptively learn how to adjust according to different anomaly levels. To appropriately focus on or ignore specific input features. During execution, the dependency score calculated due to a fault in feeder F05... Extremely low, indicating the system is in a highly anomalous state, therefore the gated signal vector generated by the first formula... The values ​​of each component will correspondingly approach 0 or 1, forming a weighted controller with strong filtering capabilities. Subsequently, the module utilizes this gating signal vector... For the original power grid embedding vector and communication feature vectors Element-level weighting is performed to dynamically amplify grid topology features closely related to faults while suppressing non-critical communication features. For example, if the dimension corresponding to the severity of line faults in the gating signal vector g is 0.98, the original value of this grid feature will be almost completely preserved in the final fused state vector; while the gating signal corresponding to the communication feature dimension of the non-critical area meter reading service slice bandwidth will be reduced. The value might be as low as 0.05, thus significantly weakening the influence of this feature. Finally, the module combines the weighted power grid vector, the weighted communication vector, and the dependency score vector as the anomalous signal itself. The three are concatenated to form the final power grid-communication network integrated state vector. .

[0038] Through the aforementioned technical means, this optimized feature fusion mechanism achieves a fundamental transformation from simple feature juxtaposition to complex probability-dependent fusion. The resulting fused state vector is no longer a simple list of original information, but a deeply processed and refined intelligent representation containing high-order statistical correlation information. This technical effect directly serves its core technical objective: significantly improving the learning efficiency and decision-making performance of downstream deep reinforcement learning models. Based on this more informative state vector, the model can identify key dynamic correlations between the power grid and communication networks more quickly and accurately, especially in scenarios with low probability but significant impact from power grid faults, enabling more timely and precise resource scheduling and topology optimization decisions. Ultimately, this mechanism aims to construct an intelligent agent with a deeper understanding of the power grid-communication network coupled system situation, thereby achieving more robust and responsive adaptive network management. Its fundamental purpose is to ensure the safe and stable operation of the power system and improve the utilization efficiency of communication resources.

[0039] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the action vector acquisition module 130 is used to input the power grid-communication network fused state vector into a pre-trained policy network to obtain action vectors. It should be understood that although the power grid-communication network fused state vector contains rich system situational information, it is not itself a decision command. The system needs an intelligent decision-making core capable of rapidly and accurately transforming this high-dimensional, complex state representation into specific, executable network control actions. Furthermore, allowing an untrained decision model to learn from scratch in a real or simulated environment is not only inefficient but may also generate numerous erroneous or even dangerous exploratory actions in the early stages, posing a threat to power grid security. Therefore, in the technical solution of this application, the power grid-communication network fused state vector is further input into a pre-trained policy network to obtain action vectors. This utilizes a deep neural network model that has absorbed expert knowledge and experience through imitation learning to quickly reason about the current system state and directly map it to a composite action that includes optimal resource scheduling and routing selection. This ensures that when faced with various complex states, including feeder F05 failure, the system can generate a reasonable control strategy that conforms to expert logic and aims to maximize system safety in real time, significantly improving the real-time performance, accuracy, and safety of decision-making.

[0040] More specifically, in a concrete example of this application, before performing this step, the action vector acquisition module's embedded policy network has undergone an offline pre-training phase. The training process of the pre-trained policy network includes: acquiring an expert demonstration dataset; using the expert demonstration dataset as a training set, training the policy network in a supervised learning manner to obtain the pre-trained policy network, with the training objective being to minimize the difference between the actions predicted by the policy network and the actual actions of the experts.

[0041] More specifically, in this stage, firstly, various typical scenarios, including feeder faults and load fluctuations, are simulated on a highly realistic power grid-communication network co-simulation platform. Power grid dispatch experts or validated optimization algorithms then handle these scenarios, resulting in a large number of state-action pairs and forming an expert demonstration dataset. Each record in this dataset contains a fused power grid-communication network state vector for a specific scenario and the optimal action vector taken by the expert for that state. Subsequently, using this expert demonstration dataset as a training set, a deep neural network structure is trained using supervised learning. The training objective is to minimize the difference between the action predicted by the policy network based on the input state vector and the actual action taken by the experts in the dataset. Once the training converges, the pre-trained policy network is obtained. During actual system operation, when the module receives the fused state vector representing feeder F05 fault and base station Cell_A communication congestion, it uses this vector as input to perform a forward propagation calculation in the pre-trained policy network. Based on the weight parameters learned from expert experience, the policy network can quickly infer that the optimal response in this specific state is to prioritize critical services. Therefore, the network ultimately outputs a structured action vector that explicitly contains targeted control instructions. For example, its resource adjustment section specifies that the uRLLC slice resources of Cell_A cell will be increased by 50% while the mMTC slice resources will be decreased by 30%, and its routing section specifies that the remote control command stream sent to switch S13 will be switched to a better backup bearer path.

[0042] In the aforementioned IoT-based power distribution communication network topology optimization system 100, the instruction generation module 140 is used to generate slice configuration instructions and routing strategy instructions based on action vectors. It should be understood that since the action vector output by the policy network is a high-level abstract, mathematical decision representation, it is not a standard instruction format that network devices or management systems can directly parse and execute. Therefore, there is a semantic gap between intelligent decision-making and physical network execution. Thus, in the technical solution of this application, slice configuration instructions and routing strategy instructions are further generated based on action vectors to accurately translate abstract control intentions into specific, executable configuration commands that conform to the interface specifications of 5G network management and orchestration systems and software-defined network controllers. This effectively bridges the last mile from intelligent decision-making to network execution, truly implementing the optimization strategies generated by the upper-level model onto the communication infrastructure, thereby completing the closed loop of the entire adaptive optimization control and achieving precise communication assurance for power grid services.

[0043] More specifically, in a concrete example of this application, after receiving the action vector output by the policy network, which is designed to address the feeder F05 fault, the instruction generation module first parses it and distributes it to different processing units. Firstly, the instruction generation module converts the resource adjustment portion of the action vector into slice configuration instructions. It extracts the resource adjustment parameters for base station Cell_A, namely, increasing uRLLC slice resources by 50% and decreasing mMTC slice resources by 30%, and converts these percentage changes into specific modification requests for network slice subnet instance resources. This module, following the ETSIMANO standard interface specification, encapsulates an API call containing parameters such as the target cell ID, target slice ID, and updated resource quotas (e.g., guaranteed bit rate, scheduling priority weight), such as a RESTful API request, and sends it to the slice management function entity of the 5G core network to trigger real-time updates to radio resource scheduling. Secondly, the instruction generation module converts the routing portion of the action vector into routing policy instructions. It extracts the decision that the remote control service flow destined for switch S13 needs to be switched to an alternative path and converts it into instructions recognizable by the SDN controller. This module generates a new segment routing policy that defines a series of network nodes (segment list) along the alternative path. Then, it distributes the policy to the SDN controller of the bearer network via the PCEP protocol or BGP-LS. The controller is responsible for updating the forwarding table entries on the relevant network devices, thereby accurately guiding the critical service flow to the newly planned, higher-performance transmission path.

[0044] In summary, the IoT-based power distribution communication network topology optimization system according to the embodiments of this application is explained. By acquiring and fusing topology and fault status data from the power grid SCADA system with slice performance and bearer network status data from the 5G network management system in real time, the system can construct a fused state vector that comprehensively describes the joint operation status of the power grid and communication network. This vector is input into an intelligent decision-making model based on imitation-reinforcement learning training to generate optimal action commands in real time. These commands are precisely parsed into dynamic adjustment commands for 5G network slice resources and routing strategy change commands for critical service flows. This concept solves the problems of blind perception of power grid status and rigid slice resource allocation in existing solutions, enabling communication resources to be instantaneously and intelligently tilted and reconfigured according to the real-time and specific needs of the power grid, thereby providing dynamic and reliable communication guarantees for critical power distribution services and significantly improving the stability and security of power grid operation.

[0045] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A power distribution communication network topology optimization system based on the Internet of Things, characterized in that, include: The data acquisition module is used to acquire raw power grid data and raw communication data; The feature extraction and state vector construction module is used to extract features and construct state vectors from raw power grid data and raw communication data to obtain the power grid-communication network fusion state vector. The action vector acquisition module is used to input the power grid-communication network fused state vector into a pre-trained policy network to obtain action vectors; The instruction generation module is used to generate slice configuration instructions and routing policy instructions based on action vectors.

2. The power distribution communication network topology optimization system based on the Internet of Things according to claim 1, characterized in that, The training process of the pre-trained policy network includes: Obtain the expert demonstration dataset; Using an expert demonstration dataset as the training set, a pre-trained policy network is trained using supervised learning. The training objective is to minimize the difference between the actions predicted by the policy network and the actual actions of the experts.

3. The power distribution communication network topology optimization system based on the Internet of Things according to claim 1, characterized in that, The feature extraction and state vector construction module includes: The topology feature extraction unit is used to extract grid topology features from the raw grid data to obtain the grid embedding vector; The communication performance feature engineering unit is used to perform communication performance feature engineering on the raw communication data to obtain a communication feature vector. The feature fusion unit is used to fuse the power grid embedding vector and the communication feature vector to obtain the power grid-communication network fused state vector.

4. The power distribution communication network topology optimization system based on the Internet of Things according to claim 3, characterized in that, The topological feature extraction unit is used for: Extract power grid topology objects from raw power grid data; Input the power grid topology graph object into a graph neural network to obtain the final set of node embedding vectors; The final set of node embedding vectors is processed by graph pooling to obtain the power grid embedding vector.

5. The power distribution communication network topology optimization system based on the Internet of Things according to claim 3, characterized in that, The feature fusion unit includes: The probability integral transformation unit is used to perform a probability integral transformation based on edge distribution on the power grid embedding vector and the communication feature vector to obtain a uniformly distributed power grid vector and a uniformly distributed communication vector. The joint dependency density estimation unit is used to perform joint dependency density estimation on uniformly distributed power grid vectors and uniformly distributed communication vectors to obtain dependency score vectors. The gated fusion unit is used to perform dependency-based gating fusion of the power grid embedding vector and the communication feature vector to obtain the power grid-communication network fusion state vector.

6. The power distribution communication network topology optimization system based on the Internet of Things according to claim 5, characterized in that, The gated fusion unit is used for: Generate gated fusion vectors based on dependency score vectors; Based on the gated fusion vector, the power grid embedding vector and the communication feature vector are gatedly fused using the following formula to obtain the power grid-communication network fusion state vector: in, and These are the power grid embedding vector and the communication feature vector. For gated fusion vectors, For dependent score vectors, Represents element-wise product. This represents a vector concatenation operation. This is the state vector for the convergence of the power grid and communication network.

7. The power distribution communication network topology optimization system based on the Internet of Things according to claim 1, characterized in that, The instruction generation module is used for: Convert the resource adjustment portion of the action vector into slice configuration instructions; The route selection part in the action vector is converted into a route policy instruction.

Images