A concentrator intelligent fusion terminal based on multi-source fusion data regulation

Abstract

The application discloses a concentrator intelligent fusion terminal based on multi-source fusion data regulation and control, and relates to the technical field of power distribution and utilization intelligent regulation and control. The concentrator intelligent fusion terminal comprises a twin inference module, a feedback optimization module and the like. The twin inference module is used for inputting candidate regulation action sequences into a digital twin to simulate and infer, obtaining future state trajectories and cumulative rewards corresponding to different candidate regulation action sequences, and selecting safe and reward-optimal immediate actions based on the future state trajectories and the cumulative rewards to convert into terminal control instructions and issue for execution. The feedback optimization module is used for collecting new-cycle multi-source heterogeneous real-time data after execution of the terminal control instructions, updating terminal-side dynamic attribute graphs and graph embedding vectors, generating state transition samples and action feedback samples, storing the state transition samples and the action feedback samples as interaction experience, and regularly fine-tuning policy network parameters of a deep reinforcement learning agent. The application realizes forward-looking simulation and optimization screening of regulation strategies in a digital space.

Description

Technical Field

[0001] This invention relates to the field of intelligent power distribution control technology, and in particular to an intelligent fusion terminal for concentrators based on multi-source fusion data control. Background Technology

[0002] With the accelerated development of new power infrastructure, the form and operation of power distribution networks are undergoing profound changes. The widespread integration of massive and diverse resources such as distributed photovoltaics, energy storage, and electric vehicles is driving the evolution of traditional unidirectional power flow and centralized control distribution networks towards a bidirectional, interactive, and distributed autonomous active distribution network model. Against this backdrop, concentrators, as key nodes connecting the master station with massive user-side resources, are transforming from simple data acquisition and forwarding units into "edge intelligent agents" with local data processing, state awareness, and intelligent decision-making capabilities.

[0003] However, existing technologies face core bottlenecks when dealing with high-proportion distributed energy access. First, at the perception level, current solutions largely limit the fusion of multi-source data to physical connections and shallow information aggregation. They struggle to automatically mine and characterize the complex spatiotemporal correlations and physical couplings underlying asynchronous, heterogeneous time-series data and dynamic topologies, resulting in insufficient depth of situational awareness. Second, at the decision-making level, control strategies rely heavily on pre-set rules, simple optimizations, or master station commands, lacking the ability to make forward-looking sequential decisions and autonomous collaborative optimizations in uncertain environments, making it difficult to achieve real-time, economical, and secure dynamic resource matching. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a concentrator-based intelligent fusion terminal based on multi-source fusion data regulation to address the shortcomings of existing terminals in terms of deep situational awareness and autonomous sequential decision-making capabilities.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a concentrator intelligent fusion terminal based on multi-source fusion data regulation, comprising: a data mapping module for collecting multi-source heterogeneous real-time data and constructing a dynamic attribute graph on the terminal side by combining preset transformer area topology relationships, meter box connection relationships, and link mapping relationships; a graph embedding generation module for generating graph embedding vectors based on the terminal side dynamic attribute graph through multi-round message passing, neighborhood weighted aggregation, and temporal state encoding; a reinforcement regulation module for inputting the graph embedding vectors into a deep reinforcement learning agent, outputting preliminary regulation actions, and generating a sequence of regulation actions for multiple future time steps based on a policy network; and a twin inference module. The system is used to input candidate control action sequences into a digital twin for simulation and deduction, obtain the future state trajectory and cumulative reward corresponding to different candidate control action sequences, and select the safest and most rewarding immediate action based on the future state trajectory and cumulative reward, converting it into a terminal control command for execution; the feedback optimization module is used to collect new-cycle multi-source heterogeneous real-time data after the execution of the terminal control command, update the dynamic attribute map and graph embedding vector on the terminal side, generate state transition samples and action feedback samples, and store the state transition samples and action feedback samples as interaction experience to periodically fine-tune the policy network parameters of the deep reinforcement learning agent.

[0008] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation described in this invention, the multi-source heterogeneous real-time data includes smart energy meters, terminal metering devices, communication link units, environmental monitoring units, terminal operation log units, and master station strategy interfaces.

[0009] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation according to the present invention, the specific steps for constructing the dynamic attribute map on the terminal side are as follows:

[0010] Collect multi-source heterogeneous real-time data, and clean and time-align the multi-source heterogeneous real-time data to obtain a standardized dataset;

[0011] Based on a standardized dataset, the system invokes pre-defined transformer area topology relationships, meter box connection relationships, and link mapping relationships to construct an initial topology matrix.

[0012] Spatiotemporal correlation calculations are performed using the data characteristics of each node in the initial topology matrix to generate dynamic attribute potential values ​​that characterize the fluctuation patterns of node states.

[0013] Write the dynamic attribute potential values ​​into the initial topology matrix to obtain the dynamic attribute graph on the terminal side.

[0014] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation according to the present invention, the specific steps for generating the graph embedding vector are as follows:

[0015] The dynamic attribute potential value and initial feature vector of each node are extracted from the dynamic attribute graph of the terminal side, and linear transformation projection is performed to obtain the initial feature matrix of the node.

[0016] Spatiotemporal attention weighted aggregation is performed on the initial feature matrix of the nodes, and the aggregated feature vector after multiple rounds of message passing is obtained by utilizing the dynamic attribute potential of each node.

[0017] Global pooling is performed on the aggregated feature vector, and the spatiotemporal feature information of all nodes in the dynamic attribute graph on the terminal side is fused to obtain global state features;

[0018] Based on global state features, a graph embedding vector representing the global state of the dynamic attribute graph on the terminal side is generated using the mean pooling method.

[0019] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation according to the present invention, the specific steps of the initial output regulation action are as follows:

[0020] The graph embedding vector is input into the deep reinforcement learning agent, and feature mapping is performed to obtain the hidden layer state representation;

[0021] Based on the hidden layer state representation and combined with historical action smoothing constraints, the action confidence score for each type of control action is calculated.

[0022] The action confidence score is compared with the action trigger threshold, and the action type with the highest action confidence score that is higher than the action trigger threshold is selected as the initial control action.

[0023] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation according to the present invention, the specific steps for generating the regulation action sequence for multiple future time steps are as follows:

[0024] The rolling forecast time domain is initialized based on the initial control actions, and the forecast length for multiple future time steps is set to generate an initial time series index.

[0025] Based on the initial time series index, obtain the hidden state vector for each future time step;

[0026] Based on the hidden state vector, the policy network is used to calculate the control action value for each future time step with trajectory smoothing constraints;

[0027] The control action values ​​are arranged in the order of the initial time series index to generate candidate control action sequences for multiple future time steps.

[0028] As a preferred embodiment of the intelligent fusion terminal for concentrator based on multi-source fusion data regulation as described in this invention, the specific steps for inputting candidate regulation action sequences into a digital twin for simulation and deduction to obtain the future state trajectory and cumulative reward corresponding to different candidate regulation action sequences are as follows:

[0029] Based on the candidate control action sequence, initialize the digital twin environment and construct the terminal running state vector at the current moment;

[0030] Based on the terminal's operating state vector, the terminal's operating state at each future moment is recursively evolved using a neural differential equation mechanism, and combined with a causal graph structure, the future state trajectory, link load changes, and abnormal response effects are generated.

[0031] Based on future state trajectories, link load changes, and abnormal response effects, the cumulative rewards corresponding to different candidate control action sequences are obtained.

[0032] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation described in this invention, the step of selecting a safe and reward-optimal immediate action based on future state trajectories and cumulative rewards, and converting it into a terminal control command for execution, includes the following specific steps:

[0033] Based on future state trajectories and cumulative rewards, initialize the security assessment window and set dynamic security thresholds;

[0034] Calculate the safety confidence index, which characterizes the overall reliability of the action sequence, based on the safety assessment window.

[0035] Candidate action sequences that meet the dynamic safety threshold are selected based on the safety confidence index, and the first time step action corresponding to the action sequence with the highest cumulative reward is selected from the candidate action sequences as the safest and most rewarding immediate action.

[0036] The safest and most rewarding real-time actions are encoded into terminal control commands and sent to physical terminals for execution via a communication interface.

[0037] As a preferred embodiment of the intelligent fusion terminal for concentrators based on multi-source fusion data regulation according to the present invention, the specific steps for generating state transition samples and action feedback samples are as follows:

[0038] Collect new-cycle multi-source heterogeneous real-time data after the execution of terminal control commands, and perform timestamp alignment and outlier removal on the multi-source heterogeneous real-time data to obtain a standardized new-cycle data set;

[0039] The node attribute potential values ​​in the dynamic attribute graph on the terminal side are updated based on the standardized new periodic data set, and the spatiotemporal attention weighted aggregation and global pooling operations are re-executed to generate the updated graph embedding vector.

[0040] Based on the updated graph embedding vector, combined with the control actions of the previous cycle, state transition samples and action feedback samples are generated.

[0041] As a preferred embodiment of the intelligent fusion terminal based on multi-source fusion data regulation described in this invention, the specific steps for storing state transition samples and action feedback samples as interaction experience and periodically fine-tuning the policy network parameters of the deep reinforcement learning agent are as follows:

[0042] After adding timestamps and twin verification confidence labels to state transition samples and action feedback samples, they are stored in the priority experience replay buffer to form an interactive experience set.

[0043] When the fine-tuning trigger condition is met, a batch of samples are sampled from the interaction experience set, and the update vector of the policy network parameters is obtained by using the confidence label and timestamp information in the batch of samples.

[0044] Based on the update vector of the policy network parameters, the gradient ascent algorithm is used to periodically fine-tune the policy network parameters of the deep reinforcement learning agent.

[0045] The beneficial effects of this invention are as follows: Through the collaboration of the data mapping module and the graph embedding generation module, a deep structured understanding of multi-source heterogeneous data is achieved, providing a high-quality input base for advanced intelligent decision-making and achieving a fundamental improvement from information aggregation to situational understanding; By strengthening the decision-verification closed loop composed of the control module and the twin inference module, the control strategy can be prospectively simulated and optimized in the digital space, upgrading the terminal from a rule executor to an autonomous intelligent agent, and achieving closed-loop control with globally optimal safety and efficiency. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of a concentrator intelligent fusion terminal based on multi-source fusion data regulation.

[0048] Figure 2 A flowchart for generating graph embedding vectors.

[0049] Figure 3 A flowchart for selecting the safest and most rewarding immediate action.

[0050] Figure 4This is a dynamic attribute map and thermal evolution diagram of the terminal side.

[0051] Figure 5 This is a diagram illustrating the closed-loop effect of the cumulative reward-security confidence index. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-5 This is one embodiment of the present invention, which provides a concentrator intelligent fusion terminal based on multi-source fusion data regulation, including the following steps:

[0056] The data mapping module is used to collect multi-source heterogeneous real-time data and, in conjunction with preset transformer area topology relationships, meter box connection relationships, and link mapping relationships, construct a dynamic attribute map on the terminal side.

[0057] Collect multi-source heterogeneous real-time data, and clean and time-align the multi-source heterogeneous real-time data to obtain a standardized dataset.

[0058] The specific process includes multi-source heterogeneous real-time data from smart meters, terminal metering devices, communication link units, environmental monitoring units, terminal operation log units, and the master station strategy interface. Each source periodically acquires output information through its corresponding communication channel or data interface. Specifically, smart meters provide electricity consumption and voltage / current parameters; terminal metering devices provide feedback on local metering status and event records; communication link units report link quality and transmission latency indicators; environmental monitoring units collect environmental parameters such as temperature, humidity, and electromagnetic interference; terminal operation log units record internal operating status and abnormal alarm information; and the master station strategy interface receives control commands and configuration parameters from the master station. All data retains its own timestamp during acquisition and is subsequently cleaned to remove missing values, duplicates, or values ​​exceeding physically reasonable ranges. Finally, data points from different sources are aligned according to a unified time base to form a standardized data set.

[0059] Based on a standardized dataset, the system invokes pre-defined transformer area topology relationships, meter box connection relationships, and link mapping relationships to construct an initial topology matrix.

[0060] The specific process includes: based on a standardized dataset, calling pre-defined transformer area topology relationships, meter box connection relationships, and link mapping relationships; determining the hierarchy and adjacency structure between transformer areas according to the pre-defined transformer area topology relationships based on the information related to transformer area nodes, meter box units, and communication links in the standardized dataset; establishing the attribution and connection correspondence between meter boxes and energy meters or terminal metering devices based on the pre-defined meter box connection relationships; describing the communication paths between communication link units and each physical node based on the pre-defined link mapping relationships; and organizing the above structured association information into a matrix form, where rows and columns correspond to node entities in the network, and matrix elements indicate whether there is a connection between nodes and its connection type, thereby constructing an initial topology matrix.

[0061] It should be noted that the pre-configured distribution area topology is based on the physical connection structure of the transformers, branch lines, and power supply range in the distribution network planning drawings and on-site survey records; the pre-configured meter box connection is based on the attribution and wiring relationship between the meter box and the smart meter or terminal metering equipment connected to it in the meter installation file and meter box asset ledger; and the pre-configured link mapping relationship is based on the logical communication path and communication interface correspondence between the communication link unit and each terminal node in the communication network configuration information and the equipment communication address allocation table.

[0062] By utilizing the data characteristics of each node in the initial topology matrix to perform spatiotemporal correlation calculations, a dynamic attribute potential value characterizing the fluctuation pattern of node states is generated, expressed as:

[0063] ;

[0064] in, Indicates that the node is The dynamic property potential value, This indicates the index of the node currently being computed. This represents the natural exponential function. This represents the time fluctuation weighting coefficient. Represents a node Historical data standard deviation Indicates the spatial aggregation weight coefficient. Indicates the neighbor node index. Represents a node The neighborhood set, The elements of the adjacency matrix represent nodes. and The connection strength, Representing neighboring nodes Real-time data feature values.

[0065] It should be noted that the time fluctuation weighting coefficient is a parameter used to adjust the proportion of influence of the historical data fluctuation of a node in the calculation of dynamic attribute potential value, and is configured based on the statistical analysis of historical control effects; the spatial aggregation weighting coefficient is a parameter used to adjust the proportion of influence of the connection strength and data characteristics of neighboring nodes in the calculation of dynamic attribute potential value, and is configured based on the historical evaluation results of the physical connection reliability and communication link quality of the transformer area. Representing neighboring nodes The real-time data feature value is the observation value of the smart energy meter, terminal metering device, communication link unit, environmental monitoring unit or terminal operation log unit corresponding to the neighbor node at the current moment, extracted from the standardized data set.

[0066] The specific process includes reflecting the fluctuation characteristics of each node's state over time based on the historical data change amplitude of each node in the time dimension, and reflecting the coupling characteristics of its influence by the surrounding environment by combining the correlation between the node and the current data characteristics of its neighboring nodes in the spatial dimension. The fluctuation characteristics in the time dimension and the coupling characteristics in the spatial dimension are combined to form a comprehensive index. This comprehensive index can characterize the overall fluctuation pattern of the node's state under the dual effects of time and space. Then, the comprehensive index is transformed by a nonlinear mapping function composed of natural exponential functions so that the output result falls between zero and one. This output result is the dynamic attribute potential value.

[0067] It should be noted that the dynamic attribute potential value is a normalized representation quantity that integrates the historical fluctuation characteristics of the node itself with the spatiotemporal coupling influence of the neighborhood. It is used to quantify the state activity level and anomaly sensitivity of each node in the dynamic attribute graph of the terminal side at the current moment, providing a weight basis for subsequent spatiotemporal attention aggregation and improving the perception accuracy of the grid embedding of the operating status of the grid area. The fluctuation characteristics refer to the state instability reflected by the magnitude of changes in the historical data of the node in the time dimension, reflecting the degree to which the node is affected by factors such as load fluctuations, equipment aging, or external disturbances. The coupling characteristics in the spatial dimension refer to the state mutual influence relationship between the node and the neighboring nodes through the grid area topology connection and physical interaction, which is reflected as the synergistic effect of the real-time data characteristics of the neighboring nodes on the current node state through the connection strength.

[0068] Write the dynamic attribute potential values ​​into the initial topology matrix to obtain the dynamic attribute graph on the terminal side.

[0069] The specific process includes attaching the dynamic attribute potential value corresponding to each node as a node attribute to the position of the corresponding node in the initial topology matrix, so that the initial topology matrix not only contains the connection relationship between nodes, but also contains the dynamic attribute potential value of each node representing the state fluctuation law, forming a terminal-side dynamic attribute graph that simultaneously carries structural information and dynamic attribute information.

[0070] like Figure 4 The terminal-side dynamic attribute graph heatmap shows the evolution of the dynamic attribute potential values ​​of each node over a continuous time dimension after cleaning, time alignment, and spatiotemporal correlation calculation of multi-source heterogeneous real-time data. The color changes corresponding to different nodes in the graph intuitively reflect the fluctuation patterns of node states and their coupling and propagation characteristics in the spatial neighborhood. This indicates that the present invention does not simply summarize the original data, but rather integrates the topological relationships of the transformer area, the connection relationships of the meter box, and the link mapping relationships with the dynamic attribute information of the nodes through the data mapping module, forming a terminal-side dynamic attribute graph that simultaneously carries structural information and dynamic attribute information. This provides a high-quality input basis for the subsequent graph embedding generation module to extract global state features and generate graph embedding vectors, demonstrating the improvement effect from information summarization to situational understanding.

[0071] The graph embedding generation module is used to generate graph embedding vectors based on the dynamic attribute graph on the terminal side through multi-round message passing, neighborhood weighted aggregation and temporal state encoding.

[0072] The dynamic attribute potential value and initial feature vector of each node are extracted from the dynamic attribute graph of the terminal side, and linear transformation projection is performed to obtain the initial feature matrix of the node.

[0073] The specific process includes extracting the dynamic attribute potential value and initial feature vector of each node from the dynamic attribute graph on the terminal side. The dynamic attribute potential value represents the state fluctuation pattern of the node, and the initial feature vector contains the data features corresponding to the node in the standardized dataset. The dynamic attribute potential value and the initial feature vector of each node are concatenated element by element to form an enhanced feature vector. The enhanced feature vectors of all nodes are linearly transformed and projected through the same learnable weight matrix. The projection results form a set of node representations with consistent dimensions. These node representations are arranged in a fixed order according to the nodes in the dynamic attribute graph on the terminal side to form the node initial feature matrix.

[0074] It should be noted that the enhanced feature vector refers to the fusion vector formed by splicing the dynamic attribute potential value of a node with the observation features of the new cycle, which is used to more comprehensively characterize the state characteristics of the node under the current regulation cycle.

[0075] Spatiotemporal attention weighted aggregation is performed on the initial feature matrix of the nodes, and the aggregated feature vector after multiple rounds of message passing is obtained by utilizing the dynamic attribute potential of each node.

[0076] The specific process includes obtaining the attention weights between each node and its neighboring nodes in each round of message passing based on the node connection relationships defined in the dynamic attribute graph on the terminal side. The attention weights simultaneously consider the state evolution information in the time dimension and the adjacency dependency relationship in the spatial dimension. The dynamic attribute potential of each node is used to adjust the attention weights so that nodes with more significant state fluctuations receive more attention during the aggregation process. The features of the neighboring nodes are weighted and summed according to the adjusted attention weights to update the representation of the current node. After multiple rounds of message passing, an aggregated feature vector of each node that integrates spatiotemporal context information is obtained.

[0077] It should be noted that state evolution information refers to the trajectory of the terminal's operating state vector changing over time in the continuous time domain, reflecting the dynamic behavior of each device and unit in the distribution area under physical causal logic constraints; adjacency dependency refers to the state interdependence between a node and its directly connected neighbor nodes in the dynamic attribute graph of the terminal side due to power supply connection, communication interaction or physical influence.

[0078] Global pooling is performed on the aggregated feature vector, and the spatiotemporal feature information of all nodes in the dynamic attribute graph on the terminal side is fused to obtain the global state feature.

[0079] The specific process includes aligning the aggregated feature vectors of each node in the dynamic attribute graph on the terminal side with the same feature dimension, and then obtaining the average value of each feature dimension along the node index direction. The average value of each dimension can reflect the overall spatiotemporal characteristics of the entire dynamic attribute graph on the terminal side in that dimension. Through this process, the spatiotemporal feature information of all nodes is fused into a single vector. The dimension of this single vector is the same as that of a single aggregated feature vector, but the result no longer depends on the number of nodes. The final vector obtained is the global state feature.

[0080] It should be noted that the global state feature is a vector obtained by summarizing the spatiotemporal aggregated features of all nodes in the dynamic attribute graph on the terminal side through global pooling operations, which represents the overall operating status of the entire substation at the current moment.

[0081] Based on global state features, a graph embedding vector representing the global state of the dynamic attribute graph on the terminal side is generated using the mean pooling method.

[0082] The specific process includes: based on global state features, averaging the features of each node using the mean pooling method to generate a graph embedding vector representing the global state of the terminal-side dynamic attribute graph. In this process, the features of each node are considered equally important. The features of all nodes are aggregated and normalized according to the total number of nodes to form a single vector that reflects the global state features of the entire terminal-side dynamic attribute graph. Finally, the features of all nodes are fused into a graph embedding vector that integrates overall network information, accurately depicting the global characteristics of the terminal-side dynamic attribute graph.

[0083] The enhanced regulation module is used to input graph embedding vectors into the deep reinforcement learning agent, output preliminary regulation actions, and generate a sequence of regulation actions for multiple future time steps based on the policy network. The regulation actions include sampling frequency adjustment actions, buffer window adjustment actions, data reporting granularity adjustment actions, communication link switching actions, and local resampling trigger actions.

[0084] The graph embedding vector is input into the deep reinforcement learning agent, and feature mapping is performed to obtain the hidden layer state representation.

[0085] The specific process includes inputting graph embedding vectors into a deep reinforcement learning agent. The policy network inside the deep reinforcement learning agent performs a linear transformation on the graph embedding vectors. The linear transformation is jointly completed by a learnable weight matrix and a learnable bias vector. The number of rows in the weight matrix corresponds to the dimension of the hidden layer, and the number of columns is the same as the dimension of the graph embedding vector. The dimension of the bias vector is the same as the dimension of the hidden layer. The result of the linear transformation is fed into the bias vector element by element and then fed into the ReLU activation function. The ReLU activation function performs a non-negation operation on each element, that is, sets all negative values ​​to zero and keeps positive values ​​unchanged. The output vector is the hidden layer state representation.

[0086] It should be noted that the hidden layer state representation is a compressed encoding of the current environmental observation information by the deep reinforcement learning agent during the decision-making process. It integrates the key features of historical regulation experience and the current terminal operating state, and is used to support the policy network in reasoning and generating future actions, thereby improving the coherence and adaptability of regulation decisions.

[0087] Furthermore, a deep reinforcement learning agent refers to an intelligent decision-making entity used to generate regulatory actions based on environmental states. It is based on a policy network, which consists of a multi-layer fully connected neural network, including an input layer, one or more hidden layers, and an output layer. The input layer receives graph embedding vectors, the hidden layers transform the input features layer by layer using weight matrices, bias vectors, and nonlinear activation functions, and the output layer generates confidence scores or action values ​​for each regulatory action type. The deep reinforcement learning agent accumulates state transition samples and action feedback samples through interaction with the environment, and uses these interaction experiences to fine-tune the policy network parameters, thereby continuously improving the performance of regulatory decisions.

[0088] Based on the hidden layer state representation and combined with historical action smoothing constraints, the action confidence score for each type of controlled action is calculated, expressed as follows:

[0089] ;

[0090] in, Indicates the first Confidence score of regulatory actions, This represents the Sigmoid function. An index representing the type of control action. Indicates the first Weight matrix of regulatory actions, This represents the hidden layer state representation. Indicates the first The bias vector of the control action. This represents the smoothing penalty coefficient. Represents the hyperbolic tangent function. Indicates the first Predicted values ​​of regulatory actions, This indicates the actual value of the control measures implemented in the previous cycle.

[0091] It should be noted that the smoothing penalty coefficient is a parameter used to control the degree of influence of the change in the control action on the action confidence score. It is configured based on the stability requirements of the transformer area and the historical control action jump tolerance analysis. Indicates the first The predicted value of the class-controlled action is the original action output obtained by the deep reinforcement learning agent under the current hidden layer state representation, which is obtained by the weight matrix and bias vector corresponding to the output layer of the policy network acting on the hidden layer state representation and then passing it through a nonlinear activation function. This indicates that the actual execution value of the control action in the previous cycle is the numerical representation of the immediate action that was selected and converted into a terminal control command in the previous control cycle and then sent to the physical terminal for execution.

[0092] It should be noted that the historical action smoothing constraint refers to the regularization term formed by adding the square of the difference between the current action and the action at the previous moment to the loss function when the strategy network outputs control actions. This constraint term is obtained by weighting the trajectory smoothing coefficient and is used to suppress drastic jumps in the control action sequence. The distribution area operation stability requirement refers to the operational safety boundary set for indicators such as voltage fluctuation, load balancing, communication reliability, and equipment failure rate. It is determined by operation and maintenance experience based on the distribution network operation procedures, equipment technical specifications, and historical fault statistical analysis. The historical control action jump tolerance refers to the maximum limit of the allowable change in the amplitude between adjacent control actions. It is set based on the mechanical response capability of the distribution area equipment, the control command execution delay, and the operational stability requirements, combined with the statistical distribution analysis of the action differences in the historical control records.

[0093] The specific process includes: transforming the hidden layer state representations linearly through the weight matrix and bias vector of the corresponding control action type, and then inputting them into the Sigmoid function to obtain the basic confidence score; calculating the difference between the predicted value of the control action type and the actual execution value of the control action in the previous cycle; feeding the difference into the hyperbolic tangent function; and then interacting the result with the smoothing penalty coefficient and adjusting it to form a smoothing adjustment factor. Finally, interacting the basic confidence score with the smoothing adjustment factor, the action confidence score of the control action type is obtained.

[0094] The action confidence score is compared with the action trigger threshold, and the action type with the highest action confidence score that is higher than the action trigger threshold is selected as the initial control action.

[0095] The specific process includes comparing the action confidence score corresponding to each type of control action with a preset action trigger threshold, retaining all control action types with action confidence scores greater than the action trigger threshold, selecting the control action type with the highest action confidence score as the initial control action if multiple control action types meet the conditions, and not generating an initial control action if no control action type has an action confidence score higher than the action trigger threshold.

[0096] It should be noted that the action trigger threshold is preset based on the power distribution terminal's control sensitivity requirements and malfunction tolerance level, combined with the statistical analysis of historical control effects. An exemplary value range is between 0.3 and 0.7.

[0097] The rolling forecast time domain is initialized based on the initial control actions, and the forecast lengths for multiple future time steps are set to generate an initial time series index.

[0098] The specific process includes initializing the rolling prediction time domain based on the initial control action, taking the current control decision time as the time zero point, determining the number of future time steps to be predicted based on the prediction length of multiple future time steps pre-configured, each time step corresponding to a fixed time interval, and generating consecutive time points starting from the current time in sequence according to the time interval, and arranging these time points in order to form the initial time series index.

[0099] It should be noted that the predicted lengths of the pre-configured future time steps are set based on the control response cycle of the power distribution terminal, communication delay characteristics, and the effective time window of the control action, combined with the analysis of the load fluctuation frequency and control stability requirements of the distribution area.

[0100] Based on the initial time series index, obtain the hidden state vector for each future time step.

[0101] The specific process includes generating the hidden state vector corresponding to each future time step by recursively through the policy network of the deep reinforcement learning agent, starting from the current hidden state representation based on the initial time series index. The hidden state vector of each time step is obtained by transforming the hidden state vector of the previous time step through the fully connected layer and activation function in the policy network. All generated hidden state vectors are arranged in the order of the initial time series index.

[0102] Based on the hidden state vector, the policy network is used to calculate the control action value for each future time step with trajectory smoothing constraints. The expression is as follows:

[0103] ;

[0104] in, Indicates the future number The adjustment action value of the time step, Index representing a future time step, Represents the Gaussian error linear unit function. Indicates the future number The policy weight matrix of the time step, Indicates the future number The hidden state vector of the time step. Indicates the trajectory smoothing coefficient. This represents the hidden state vector of the previous time step.

[0105] It should be noted that the trajectory smoothing coefficient is a preset parameter used to adjust the degree of drastic change in the timing of the control action sequence. It is configured after analyzing the response characteristics of the transformer area equipment and the requirements for the smoothness of the control action execution.

[0106] The specific process includes: based on the hidden state vector, using the policy network to perform linear transformation and Gaussian error linear unit function activation on the hidden state vector of each time step to obtain the basic control action value; then calculating the squared Euclidean distance between the hidden state vector of the current time step and the hidden state vector of the previous time step; inputting the squared Euclidean distance into an exponential function and scaling it with a negative trajectory smoothing coefficient to form a trajectory smoothing factor; and then combining the basic control action value with the trajectory smoothing factor to obtain the control action value for each future time step with trajectory smoothing constraints.

[0107] Furthermore, the training process of the policy network is as follows: Based on the set of interactive experiences, when the fine-tuning trigger condition is met, a batch of samples are sampled from the priority experience replay buffer. Each sample contains state transition samples, action feedback samples, timestamps, and Siamese verification confidence labels. The loss function of each sample is weighted using the Siamese verification confidence labels, and the temporal order reflected by the timestamps is preserved. The update vector of the policy network parameters is obtained by calculating the policy gradient with historical action smoothing constraints. The gradient ascent algorithm is used to fine-tune the policy network parameters periodically, so that the policy network continuously optimizes the cumulative reward and safety confidence index of the control actions under the guidance of high-confidence experience, thereby improving the decision-making stability and adaptability of the deep reinforcement learning agent in the dynamic environment of the transformer area.

[0108] It should be noted that the fine-tuning trigger condition is a threshold set based on the cumulative number of control cycles in the distribution area or the degree of decline in strategy performance indicators. It is determined by combining the stability analysis of historical control effects and the trend of changes in action confidence with operation and maintenance experience.

[0109] The control action values ​​are arranged in the order of the initial time series index to generate candidate control action sequences for multiple future time steps.

[0110] The specific process involves organizing the control action values ​​for each future time step with trajectory smoothing constraints according to the chronological order of the time points in the initial time series index, so that the control action value corresponding to the first time point is located at the beginning of the sequence, and the control action values ​​corresponding to each subsequent time point are arranged in sequence following the control action value of the previous time point, forming an ordered action list that is strictly aligned with the initial time series index and unfolded in the order of time steps. This ordered action list is the candidate control action sequence for multiple future time steps.

[0111] The twin simulation module is used to input candidate control action sequences into the digital twin for simulation and simulation, obtain the future state trajectory and cumulative reward corresponding to different candidate control action sequences, and select the safest and most rewarding immediate action based on the future state trajectory and cumulative reward, convert it into a terminal control command for execution.

[0112] Based on the candidate control action sequence, the digital twin environment is initialized and the terminal running state vector at the current moment is constructed.

[0113] The specific process includes: based on the candidate control action sequence, the first time step control action value in the candidate control action sequence is used as input to the digital twin environment. The digital twin environment generates a multi-dimensional vector reflecting the current operating status of the power distribution terminal based on the control action value and the real-time observation data of the smart energy meter, terminal metering device, communication link unit, environmental monitoring unit and terminal operation log unit obtained from the standardized data set. The multi-dimensional vector includes voltage and current parameters, metering status, link quality, environmental conditions and operation log characteristics, which constitute the terminal operating status vector at the current moment.

[0114] Based on the terminal's operating state vector, the terminal's operating state at each future moment is recursively evolved using a neural differential equation mechanism. Combined with a causal graph structure, this generates future state trajectories, link load changes, and abnormal response effects.

[0115] The specific process includes: based on the terminal operating state vector, using a neural differential equation mechanism to recursively evolve the terminal operating state vector in the continuous time domain; describing the change of state over time through ordinary differential equations parameterized by a neural network; taking the terminal operating state vector at the current moment as the initial condition; combining the variable dependency directions formed by the transformer area topology relationship, meter box connection relationship and link mapping relationship defined in the causal graph structure; constraining the mutual influence path of each dimension of features during the state evolution process; generating future state trajectories that conform to physical causal logic; and simultaneously deriving the load change trend of the communication link unit at each future moment and the response effect of the terminal metering device or smart energy meter under abnormal disturbances.

[0116] It should be noted that the neural differential equation mechanism is a method that combines neural networks with ordinary differential equations. By parameterizing the derivative function of the state that changes continuously with time through neural networks, it realizes continuous modeling and evolution of the dynamic process at any point in time. The causal graph structure is a directed graph built based on the physical connection and communication relationship of the transformer area. Nodes represent various terminal units, and edges represent the causal dependency direction between them, which is used to constrain the variable interaction logic in the state evolution. Physical causal logic refers to the causal dependency relationship between various devices and units in the transformer area, which is determined by the power supply topology, communication links and control commands, and conforms to the power operation mechanism and information flow direction.

[0117] Based on future state trajectories, link load changes, and abnormal response effects, the cumulative rewards corresponding to different candidate control action sequences are obtained.

[0118] The specific process includes evaluating the terminal operating status of each candidate control action sequence at each time step in the rolling prediction time domain based on future state trajectories, link load changes, and abnormal response effects. The instant reward for each time step is obtained based on indicators such as voltage stability, communication link availability, equipment anomaly rate, and control target achievement. The instant rewards of all time steps are accumulated to obtain the cumulative reward corresponding to the candidate control action sequence.

[0119] Based on future state trajectories and cumulative rewards, initialize the security assessment window and set dynamic security thresholds.

[0120] The specific process includes selecting several consecutive time points from the current moment in the future state trajectory based on the future state trajectory and the cumulative reward to form a safety assessment window. The safety assessment window covers the key time period used for safety verification in the rolling prediction time domain. The dynamic safety threshold is determined according to the numerical level of the cumulative reward and the safety specifications for the operation of the transformer area. The dynamic safety threshold is adaptively adjusted according to the level of the cumulative reward and is used to screen candidate control action sequences that meet the safety requirements in the subsequent process.

[0121] It should be noted that the exemplary range of the dynamic safety threshold is between 0.4 and 0.85; the critical period refers to the time interval in the rolling prediction time domain that is most sensitive to the safe operation of the terminal. It is determined based on the physical characteristics of the equipment response delay, fault propagation delay and control action effective cycle, combined with the state evolution law analysis when historical abnormal events occur.

[0122] Based on the safety assessment window, the safety confidence index, which characterizes the overall reliability of the action sequence, is calculated. The expression is as follows:

[0123] ;

[0124] in, This represents the security confidence index. This indicates the total number of sampling points within the safety assessment window. This indicates the index of the sampling point within the security assessment window. Indicates the first The state vector of each sampling point Represents the desired safe state vector. This represents the safety tolerance parameter.

[0125] It should be noted that the safety tolerance parameter is a constant used to control the sensitivity of the safety confidence index to state deviation, reflecting the acceptable range of terminal state deviation from the expected safe state during the operation of the distribution area.

[0126] The specific process includes: extracting the terminal running state vector corresponding to each time point within the safety assessment window from the future state trajectory according to the safety assessment window; measuring the difference between each terminal running state vector and the expected safety state vector; using a Gaussian kernel function to perform nonlinear mapping on the difference results at each time point to obtain the safety similarity at each time point; and calculating the arithmetic mean of the safety similarity at all time points within the safety assessment window. The arithmetic mean is the safety confidence index that characterizes the overall reliability of the action sequence.

[0127] Candidate action sequences that meet the dynamic safety threshold are selected based on the safety confidence index, and the first time step action corresponding to the action sequence with the highest cumulative reward is selected from the candidate action sequences as the safest and most rewarding immediate action.

[0128] The specific process includes: screening candidate action sequences that meet the dynamic safety threshold based on the safety confidence index; comparing the safety confidence index corresponding to all candidate action sequences with the dynamic safety threshold; retaining candidate action sequences whose safety confidence index is not lower than the dynamic safety threshold; selecting the candidate action sequence with the highest cumulative reward from these retained candidate action sequences; and extracting the adjustment action value of the first time step in the candidate action sequence as the safest and most rewarding immediate action.

[0129] The safest and most rewarding real-time actions are encoded into terminal control commands and sent to physical terminals for execution via a communication interface.

[0130] The specific process includes encoding the safest and most rewarding immediate actions according to the format specified in the power distribution terminal communication protocol, generating terminal control instructions that conform to the protocol specifications. The terminal control instructions include the target device identifier, action type, action parameters, and verification information. The terminal control instructions are sent to the corresponding physical terminal through a standardized communication interface, and the physical terminal parses and executes the operation specified by the terminal control instructions.

[0131] like Figure 5 The cumulative reward-security confidence index closed-loop effect diagram illustrates the coupling relationship between the cumulative reward, security confidence index, and dynamic security threshold of candidate control action sequences after simulation and deduction in a digital twin. The diagram clearly shows that this invention does not simply use the maximum cumulative reward as the sole decision-making criterion. Instead, it first constructs a security assessment window based on the future state trajectory, further calculates the security confidence index representing the comprehensive reliability of the action sequence, and then compares it with the dynamic security threshold. From the candidate action sequences that meet the security constraints, the first time-step action corresponding to the highest cumulative reward is selected as the immediate action. This diagram highlights the technical effect of this invention in achieving synergistic optimization between security and efficiency by strengthening the decision-making-verification closed loop formed by the control module and the twin deduction module.

[0132] The feedback optimization module is used to collect new-cycle multi-source heterogeneous real-time data after the execution of terminal control commands, update the dynamic attribute map and graph embedding vector on the terminal side, generate state transition samples and action feedback samples, and store the state transition samples and action feedback samples as interaction experience to periodically fine-tune the policy network parameters of the deep reinforcement learning agent.

[0133] After the execution of terminal control commands, new-cycle multi-source heterogeneous real-time data is collected, and timestamp alignment and outlier removal are performed on the multi-source heterogeneous real-time data to obtain a standardized new-cycle data set.

[0134] The specific process includes collecting multi-source heterogeneous real-time data for the new cycle after the execution of terminal control commands. The multi-source heterogeneous real-time data for the new cycle includes observation data generated by smart energy meters, terminal metering devices, communication link units, environmental monitoring units, and terminal operation log units within the new control cycle. The multi-source heterogeneous real-time data for the new cycle is timestamped according to a unified time base, and data from different sources are mapped to the same time grid. Then, outliers are identified and removed according to a preset physical reasonable range and statistical outlier criteria, and finally a standardized new cycle data set with consistent structure, synchronized time, and no abnormal observations is formed.

[0135] It should be noted that the preset physical reasonable range is determined based on the engineering specifications and physical limits of the operating parameters of the power distribution equipment; the statistical outlier criterion is set based on the distribution characteristics of historical operating data using the interquartile range method or the three-standard deviation method.

[0136] The node attribute potential values ​​of the old period in the dynamic attribute graph on the terminal side are updated based on the standardized new period data set, and the spatiotemporal attention weighted aggregation and global pooling operations are re-executed to generate the updated graph embedding vector.

[0137] The specific process includes: based on the standardized new cycle data set, extracting the new cycle observation features corresponding to each node in the standardized new cycle data set; combining the topology of the dynamic attribute graph on the terminal side, recalculating the dynamic attribute potential value of each node according to the dynamic attribute potential value calculation expression to update the node attribute potential value of the old cycle in the dynamic attribute graph on the terminal side; and based on the enhanced feature vector formed by concatenating the updated node attribute potential value with the new cycle observation features, performing spatiotemporal attention weighted aggregation on all nodes, and then performing a global pooling operation on the aggregation result to generate the updated graph embedding vector.

[0138] It should be noted that the new cycle observation characteristics are extracted from the standardized new cycle data set and are the real-time operation indicators generated by smart energy meters, terminal metering devices, communication link units, environmental monitoring units and terminal operation log units for each node in the current control cycle. After timestamp alignment and outlier removal, they are obtained directly.

[0139] Based on the updated graph embedding vector, combined with the control actions of the previous cycle, state transition samples and action feedback samples are generated.

[0140] The specific process includes using the updated graph embedding vector as the state representation of the current cycle, using the control action of the previous cycle as the action taken in the previous state, pairing the updated graph embedding vector with the control action of the previous cycle to form a state transition sample, and obtaining the corresponding action feedback sample based on the running effect reflected in the standardized new cycle data set after the execution of the terminal control command. The action feedback sample includes the cumulative reward and the safety confidence index, thus constituting an experience replay buffer unit for updating the parameters of the deep reinforcement learning agent.

[0141] After adding timestamps and twin verification confidence labels to state transition samples and action feedback samples, they are stored in the priority experience replay buffer to form an interactive experience set.

[0142] The specific process includes attaching a timestamp of the current control cycle to the state transition samples and action feedback samples, and using the security confidence index generated during the digital twin simulation as a twin verification confidence label. The state transition samples and action feedback samples with timestamps and twin verification confidence labels are stored in a priority experience replay buffer. The priority experience replay buffer assigns different sampling priorities to the samples according to the level of twin verification confidence labels, forming an interactive experience set for training deep reinforcement learning agents.

[0143] When the fine-tuning trigger condition is met, a batch of samples is sampled from the interaction experience set, and the update vector of the policy network parameters is obtained by using the confidence labels and timestamp information in the batch of samples.

[0144] The specific process includes: when the fine-tuning trigger condition is met, a batch of samples is extracted from the interaction experience set according to the sampling priority of the priority experience replay buffer. The batch of samples includes state transition samples, action feedback samples, timestamps and Siamese verification confidence labels. The Siamese verification confidence labels in the batch of samples are used as weights to weight the loss function of each sample. At the same time, the samples are sorted in chronological order in combination with the timestamp information to preserve temporal dependencies. The gradient of the weighted loss function is calculated with respect to the policy network parameters. The resulting gradient is processed by the optimizer to form the update vector of the policy network parameters.

[0145] It should be noted that the fine-tuning trigger conditions are set based on the analysis of the stability of the control effect accumulated by the deep reinforcement learning agent in historical interactions and the trend of changes in action confidence.

[0146] Based on the update vector of the policy network parameters, the gradient ascent algorithm is used to periodically fine-tune the policy network parameters of the deep reinforcement learning agent.

[0147] The specific process includes adding the updated vector of the policy network parameters to the current policy network parameters to obtain the updated policy network parameters, which are then executed when the fine-tuning trigger condition is met. By increasing the expected cumulative reward of the policy network on high-confidence interaction experience, the regulatory decision-making ability of the deep reinforcement learning agent can continuously adapt to the evolution law of the dynamic attribute graph on the terminal side.

[0148] It should be noted that the gradient ascent algorithm is a parameter optimization method. It obtains the gradient of the objective function with respect to the parameters and adjusts the parameter values ​​along the gradient direction so that the objective function gradually increases in each iteration and eventually converges to a local or global maximum. It is often used to maximize the expected cumulative reward of the policy network in reinforcement learning.

[0149] In summary, this invention achieves deep structured cognition of multi-source heterogeneous data through the synergy of the data mapping module and the graph embedding generation module, providing a high-quality input foundation for advanced intelligent decision-making and achieving a fundamental improvement from information aggregation to situational understanding. By strengthening the decision-verification closed loop composed of the control module and the twin inference module, it realizes the forward-looking simulation and optimization screening of control strategies in the digital space, upgrading the terminal from a rule executor to an autonomous intelligent agent, and achieving closed-loop control with globally optimal safety and efficiency.

[0150] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A concentrator-based intelligent fusion terminal for multi-source fusion data regulation, characterized in that, include: The data mapping module is used to collect multi-source heterogeneous real-time data and, in combination with preset transformer area topology relationships, meter box connection relationships and link mapping relationships, construct a dynamic attribute map on the terminal side. The graph embedding generation module is used to generate graph embedding vectors based on the dynamic attribute graph on the terminal side through multi-round message passing, neighborhood weighted aggregation and temporal state encoding. The enhanced regulation module is used to input graph-embedded vectors into the deep reinforcement learning agent, output initial regulation actions, and generate a sequence of regulation actions for multiple future time steps based on the policy network. The twin simulation module is used to input candidate control action sequences into the digital twin for simulation and simulation, obtain the future state trajectory and cumulative reward corresponding to different candidate control action sequences, and select the safest and most rewarding immediate action based on the future state trajectory and cumulative reward, convert it into a terminal control command for execution. The feedback optimization module is used to collect new-cycle multi-source heterogeneous real-time data after the execution of terminal control commands, update the dynamic attribute map and graph embedding vector on the terminal side, generate state transition samples and action feedback samples, and store the state transition samples and action feedback samples as interaction experience to periodically fine-tune the policy network parameters of the deep reinforcement learning agent.

2. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 1, characterized in that, The multi-source heterogeneous real-time data includes smart energy meters, terminal metering devices, communication link units, environmental monitoring units, terminal operation log units, and master station policy interfaces.

3. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 1, characterized in that, The specific steps for constructing the dynamic attribute graph on the terminal side are as follows: Collect multi-source heterogeneous real-time data, and clean and time-align the multi-source heterogeneous real-time data to obtain a standardized dataset; Based on a standardized dataset, the system invokes pre-defined transformer area topology relationships, meter box connection relationships, and link mapping relationships to construct an initial topology matrix. Spatiotemporal correlation calculations are performed using the data characteristics of each node in the initial topology matrix to generate dynamic attribute potential values ​​that characterize the fluctuation patterns of node states. Write the dynamic attribute potential values ​​into the initial topology matrix to obtain the dynamic attribute graph on the terminal side.

4. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 1 or 3, characterized in that, The specific steps for generating the graph embedding vector are as follows: The dynamic attribute potential value and initial feature vector of each node are extracted from the dynamic attribute graph of the terminal side, and linear transformation projection is performed to obtain the initial feature matrix of the node. Spatiotemporal attention weighted aggregation is performed on the initial feature matrix of the nodes, and the aggregated feature vector after multiple rounds of message passing is obtained by utilizing the dynamic attribute potential of each node. Global pooling is performed on the aggregated feature vector, and the spatiotemporal feature information of all nodes in the dynamic attribute graph on the terminal side is fused to obtain global state features; Based on global state features, a graph embedding vector representing the global state of the dynamic attribute graph on the terminal side is generated using the mean pooling method.

5. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 1, characterized in that, The initial output control action involves the following steps: The graph embedding vector is input into the deep reinforcement learning agent, and feature mapping is performed to obtain the hidden layer state representation; Based on the hidden layer state representation and combined with historical action smoothing constraints, the action confidence score for each type of control action is calculated. The action confidence score is compared with the action trigger threshold, and the action type with the highest action confidence score that is higher than the action trigger threshold is selected as the initial control action.

6. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 1 or 5, characterized in that, The specific steps for generating the sequence of regulatory actions for multiple future time steps are as follows: The rolling forecast time domain is initialized based on the initial control actions, and the forecast length for multiple future time steps is set to generate an initial time series index. Based on the initial time series index, obtain the hidden state vector for each future time step; Based on the hidden state vector, the policy network is used to calculate the control action value for each future time step with trajectory smoothing constraints; The control action values ​​are arranged in the order of the initial time series index to generate candidate control action sequences for multiple future time steps.

7. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 6, characterized in that, The specific steps for inputting candidate control action sequences into a digital twin for simulation and deduction to obtain the future state trajectory and cumulative reward corresponding to different candidate control action sequences are as follows: Based on the candidate control action sequence, initialize the digital twin environment and construct the terminal running state vector at the current moment; Based on the terminal's operating state vector, the terminal's operating state at each future moment is recursively evolved using a neural differential equation mechanism, and combined with a causal graph structure, the future state trajectory, link load changes, and abnormal response effects are generated. Based on future state trajectories, link load changes, and abnormal response effects, the cumulative rewards corresponding to different candidate control action sequences are obtained.

8. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 7, characterized in that, The process of selecting a safe and reward-optimal immediate action based on future state trajectories and cumulative rewards, and then converting it into a terminal control command for execution, involves the following steps: Based on future state trajectories and cumulative rewards, initialize the security assessment window and set dynamic security thresholds; Calculate the safety confidence index, which characterizes the overall reliability of the action sequence, based on the safety assessment window. Candidate action sequences that meet the dynamic safety threshold are selected based on the safety confidence index, and the first time step action corresponding to the action sequence with the highest cumulative reward is selected from the candidate action sequences as the safest and most rewarding immediate action. The safest and most rewarding real-time actions are encoded into terminal control commands via protocol and then sent to physical terminals for execution through the communication interface.

9. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 1 or 8, characterized in that, The specific steps for generating state transition samples and action feedback samples are as follows: Collect new-cycle multi-source heterogeneous real-time data after the execution of terminal control commands, and perform timestamp alignment and outlier removal on the multi-source heterogeneous real-time data to obtain a standardized new-cycle data set; The node attribute potential values ​​in the dynamic attribute graph on the terminal side are updated based on the standardized new periodic data set, and the spatiotemporal attention weighted aggregation and global pooling operations are re-executed to generate the updated graph embedding vector. Based on the updated graph embedding vector, combined with the control actions of the previous cycle, state transition samples and action feedback samples are generated.

10. The intelligent fusion terminal for concentrators based on multi-source fusion data regulation as described in claim 9, characterized in that, The specific steps for storing state transition samples and action feedback samples as interaction experience and periodically fine-tuning the policy network parameters of the deep reinforcement learning agent are as follows: After adding timestamps and twin verification confidence labels to state transition samples and action feedback samples, they are stored in the priority experience replay buffer to form an interactive experience set. When the fine-tuning trigger condition is met, a batch of samples are sampled from the interaction experience set, and the update vector of the policy network parameters is obtained by using the confidence label and timestamp information in the batch of samples. Based on the update vector of the policy network parameters, the gradient ascent algorithm is used to periodically fine-tune the policy network parameters of the deep reinforcement learning agent.

Images