A power distribution network adaptive operation and maintenance system based on 5G edge intelligence

Abstract

The application relates to the technical field of intelligent operation and maintenance, and discloses a power distribution network self-adaptive operation and maintenance system based on 5G edge intelligence. The system realizes unified expression of the operation state and the topological structure by constructing a power distribution network graph state model in the cloud through multi-source data acquisition and edge side space-time alignment processing; on the basis, state and candidate operation actions are fused and coded, a clipping distance linear attention Transformer model is introduced to predict the next state risk after the operation action is executed, a risk perception decision mechanism is constructed in combination with reinforcement learning, and an adaptive operation and maintenance control strategy is output; meanwhile, through a semi-Markov modeling mode triggered based on a dangerous event, macro-step rewards and macro-step discounts are used to realize strategy optimization across time scales. The application can effectively improve the risk prediction accuracy and decision safety, and is suitable for intelligent operation and maintenance control in a complex power distribution network environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and in particular to an adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence. Background Technology

[0002] With the expansion of distribution network scale and the access of distributed energy and various types of loads, system operation exhibits characteristics of increasingly complex topological structures and highly dynamic states. Although existing intelligent operation and maintenance technologies have incorporated deep learning and reinforcement learning methods, they still have shortcomings: At the data level, multi-source operational data lacks a unified spatiotemporal alignment and structured modeling mechanism, resulting in insufficient consistency of model inputs; at the modeling level, most methods focus on local features or single-node analysis, lacking a systematic expression of the distribution network topology and the coupling relationships between nodes, making it difficult to accurately depict the risk propagation process; at the prediction and decision-making level, existing methods are mostly based on the current state or short-term prediction results for evaluation, failing to incorporate operation and maintenance actions into state evolution modeling, and lacking integrated prediction capabilities of "state-action-risk"; at the same time, the computational complexity of attention models based on Softmax is high, making it difficult to meet the real-time requirements of the edge side, while reinforcement learning methods typically use fixed time steps and single-step reward mechanisms, making it difficult to describe the cross-time interval evolution characteristics from the triggering to the resolution of dangers; therefore, existing intelligent technologies still have room for improvement in multi-source data fusion, topology correlation modeling, action-driven risk prediction, and cross-time scale decision optimization. Summary of the Invention

[0003] This invention addresses the shortcomings of existing intelligent power distribution network operation and maintenance technologies in risk evolution prediction and cross-timescale decision modeling. It proposes a 5G edge intelligence-based adaptive operation and maintenance system for power distribution networks. By constructing an integrated "state-action-risk" modeling framework, it jointly encodes the distribution network operating state and candidate operation and maintenance actions. Furthermore, it introduces a pruned-distance linear attention Transformer model to predict the risk of the next state after the execution of an operation and maintenance action, thereby achieving a forward-looking characterization of the risk evolution process. Specifically, it replaces the traditional Softmax attention calculation with a linear proxy scoring mechanism based on pruned-distance constraints, maintaining the accuracy of the attention calculation while ensuring the accuracy of the target network. While linearizing the computation process based on the relative ranking relationship of intention and force, the model complexity is reduced and the deployment efficiency at the edge is improved. On this basis, the predicted risk prior information is fused with the action preferences output by the reinforcement learning policy network to construct a risk perception decision-making mechanism to suppress high-risk actions and strengthen the preferred actions. At the same time, a semi-Markov decision modeling method based on dangerous event triggering is introduced. The dangerous triggering and dangerous removal are used as the macrostep boundary. The long-term effect of operation and maintenance actions in the complete risk evolution process is uniformly evaluated through macrostep rewards and macrostep discounts, thereby improving the adaptability and optimization effect of the decision-making strategy to complex dynamic environments.

[0004] This invention proposes an adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence. The system includes a distribution network operation data acquisition module, an edge processing and judgment module, a graph state construction module, a physical risk prior generation module, a candidate action screening module, a reinforcement learning decision module, and an operation and maintenance execution feedback module that are connected in sequence.

[0005] The distribution network operation data acquisition module collects raw data on the operation of the distribution network.

[0006] The edge processing and judgment module performs timestamp alignment, anomaly removal, missing data completion, dimension normalization, device identity mapping, and topology consistency verification on the raw data of the distribution network operation to generate structured distribution network operation status data. Based on the structured distribution network operation status data, it calculates the load rate of each line and transformer, extracts the maximum load rate of the line cluster at the current moment as the instantaneous risk indicator, and compares the instantaneous risk indicator with the preset danger threshold to generate a danger trigger indication.

[0007] The diagram state construction module, based on the danger trigger indication, is used to map the structured distribution network operation status data into distribution network diagram state objects using a hybrid diagram modeling method of bus level and feeder level when the danger trigger conditions are met.

[0008] The physical risk prior generation module uses a graph neural network encoder to process the distribution network diagram state object, performs graph feature extraction based on edge message passing and edge-aware attention aggregation mechanisms, and jointly models and propagates information on node features and edge features to obtain a graph-level embedding representation of the current distribution network diagram state. For each executable operation and maintenance action at the current moment, the graph-level embedding representation and the corresponding action embedding vector are jointly encoded to obtain a state-action fusion feature representation. A pruning distance linear attention Transformer model is constructed, and the state-action fusion feature representation is input into the pruning distance linear attention Transformer model to predict the next state risk value after the execution of the executable operation and maintenance action. Based on the risk value of the next state, physical risk scores are generated for each executable operation and maintenance action. The construction method of the pruning distance linear attention Transformer model is as follows: based on the Transformer model, a linear proxy scoring mechanism based on pruning distance constraints is introduced into the attention calculation. A truncated linear mapping is used to replace the exponential function calculation to realize the linear expression of attention weights. The monotonicity and non-negativity of the scores are controlled by the baseline parameter and the slope parameter. While maintaining the relative ranking relationship, the attention weight aggregation is completed. The exponential attention weight calculation method in the Transformer model is linearized and optimized to obtain the pruning distance linear attention Transformer model.

[0009] The candidate action filtering module is used to combine the action feasibility constraint set at the current moment, perform feasibility mask filtering on all executable operation and maintenance actions to obtain the set of actionable actions under the current dangerous state; and sort them according to the physical risk score corresponding to each action, select a preset number of low-risk preferred actions from the set of actionable actions, and construct a subset of state-related candidate actions.

[0010] The reinforcement learning decision-making module is used to input the distribution network diagram state objects into the policy value joint network, and output the policy preference value and state value estimate corresponding to each operation and maintenance action in the current state; for each candidate action in the candidate action subset, its policy preference value and corresponding physical risk score are reweighted and fused to construct the candidate action probability distribution; and the target operation and maintenance action is selected based on the candidate action probability distribution to generate the distribution network adaptive operation and maintenance control instruction.

[0011] The operation and maintenance execution feedback module sends the distribution network adaptive operation and maintenance control instructions to the corresponding execution objects via the 5G communication network; after execution, it continuously collects execution feedback data, subsequent micro-step operation data and system safety status data, and determines the danger clearance status based on the system safety status data, constructs macro-step reward information and macro-step discount information from the time of danger triggering to the time of danger clearance, and feeds it back to the reinforcement learning decision module.

[0012] Furthermore, the process of inputting the state-action fusion feature representation into the pruned distance linear attention Transformer model to predict the next state risk value after the execution of the executable operation and maintenance action specifically includes the following steps:

[0013] Step B1: Perform unified dimension mapping on the state-action fusion feature representation according to the preset feature dimensions to form a state-action joint feature sequence to characterize the coupling relationship between the current distribution network diagram state and candidate operation and maintenance actions; input the state-action joint feature sequence into the linear mapping unit to generate the corresponding query vector, key vector and value vector respectively; and divide the query vector, key vector and value vector into head sub-vector groups corresponding to multiple attention heads according to the multi-head attention structure;

[0014] Step B2: For each attention head, perform scaling dot product correlation calculation based on the query sub-vector and key sub-vector in the head sub-vector group to obtain the original attention score value corresponding to each position; and arrange and combine the original attention score values ​​according to the preset position order to construct the original attention score sequence under the attention head; and perform maximum value retrieval processing on the original attention score sequence row by row to determine the maximum score value in the corresponding row.

[0015] Step B3: For each attention head, based on the difference between the maximum score and the original attention score corresponding to each position, construct the non-negative distance value corresponding to each position; and call the head parameter set obtained by the attention head in pre-offline calibration to perform upper bound pruning processing on the non-negative distance value to obtain a pruning distance sequence limited to a preset distance range; the head parameter set includes the baseline parameter, slope parameter and maximum pruning distance parameter;

[0016] Step B4: For each attention head, based on the baseline parameters, slope parameters and pruning distance sequence corresponding to that attention head, perform proxy scoring calculation on each position according to the truncated linear mapping relationship to generate the linear proxy score sequence corresponding to that attention head;

[0017] Step B5: For each attention head, based on the linear proxy score sequence generated in Step B4, perform an accumulation operation on the linear proxy scores corresponding to each position to obtain the proxy score sum corresponding to the attention head; and construct a normalized mapping relationship based on a preset target scaling scale, perform inverse scaling on the proxy score sum to obtain the normalized scaling factor; multiply the linear proxy scores at each position with the normalized scaling factor to generate the truncated linear normalized attention weight sequence corresponding to each position under the attention head;

[0018] Step B6: Apply the truncated linearly normalized attention weight sequence corresponding to each attention head to the value sub-vector group corresponding to the head sub-vector group, and perform weighted aggregation processing to obtain the risk association representation under each attention head; then perform splicing and linear transformation on the risk association representation under each attention head to generate a multi-head fused risk representation; then input the multi-head fused risk representation into the residual connection unit, normalization unit and feedforward transformation unit in sequence to perform inter-layer progressive feature enhancement, extract the high-order coupling relationship between state features, action features and risk propagation features layer by layer, and obtain the deep risk representation vector that represents the risk evolution trend after the target operation and maintenance action is performed; input the deep risk representation vector into the risk prediction mapping unit, perform regression mapping processing, and output the next state risk value after the execution of the executable operation and maintenance action.

[0019] Furthermore, the continuous collection of execution feedback data, subsequent micro-step operation data, and system safety status data, and the determination of the danger clearance status based on the system safety status data, constructing macro-step reward information and macro-step discount information from the danger triggering time to the danger clearance time, and feeding this information back to the reinforcement learning decision module, specifically includes the following steps:

[0020] Step C1: After the target maintenance action is executed, the execution feedback data and subsequent micro-step operation data are continuously collected through the distribution automation terminal and field execution device; the subsequent micro-step operation data are continuously sampled according to the preset sampling period to form a micro-step operation status sequence from the time the target maintenance action is executed;

[0021] Step C2: Based on the execution feedback data, microstep operation state sequence, and system safety state data, perform real-time safety judgment on the current distribution network operation state to determine whether the danger clearance condition is met; when the danger clearance condition is met, determine the moment as the danger clearance moment; and construct the corresponding macrostep time window with the danger trigger moment as the macrostep start moment and the danger clearance moment as the macrostep end moment; based on the macrostep time window, perform interval division and state interception processing on the continuous operation process from the danger trigger moment to the danger clearance moment, and extract the operation state evolution sequence;

[0022] Step C3: Based on the running state evolution sequence, calculate the instant reward corresponding to each microstep time, and perform discount accumulation processing on the instant rewards in chronological order to construct macrostep reward information;

[0023] Step C4: Define the macrostep discount factor, and model the time span corresponding to the macrostep time window based on the macrostep discount factor to construct macrostep discount information;

[0024] Step C5: Based on the macrostep reward information and macrostep discount information mentioned above, construct semi-Markov decision feedback data and send the semi-Markov decision feedback data back to the reinforcement learning decision module to perform policy updates based on the semi-Markov decision process (SMDP).

[0025] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0026] This invention constructs an integrated "state-action-risk" modeling mechanism, realizing the joint expression and collaborative modeling of distribution network operation status and maintenance actions. This enhances the predictive ability of system risk change trends after maintenance actions are executed, solving the problem that existing intelligent methods rely solely on the current state or short-term predictions and cannot proactively assess risk evolution. By introducing a pruned distance linear attention Transformer model, traditional exponential attention calculation is transformed into linear proxy calculation, significantly reducing computational complexity while maintaining attention ranking relationships. This enhances the feasibility of deploying the model at the edge and its real-time response capability, enabling this invention to possess efficient and stable risk prediction capabilities in complex distribution network environments.

[0027] This invention introduces the predicted next-state risk as prior information of physical risk into the reinforcement learning decision-making process, achieving a deep integration of risk assessment and strategy optimization. This enhances the sensitivity and constraint capability of operation and maintenance decisions to potential risks, and solves the problems of lack of safety constraints and high-risk actions in existing methods. By reweighting and fusing action preferences and risk scores, the invention effectively enhances the selection tendency for low-risk preferred actions, thereby improving the safety, stability, and operational reliability of adaptive operation and maintenance control of the distribution network. This invention enables the output of optimized control strategies that better meet actual operational needs in complex scenarios with multiple constraints.

[0028] This invention introduces a semi-Markov decision modeling mechanism based on dangerous event triggering, achieving unified modeling of the entire process of "danger triggering - risk evolution - danger resolution," thus improving the ability to evaluate the comprehensive effects of operation and maintenance actions across time intervals. It also solves the problem that fixed time steps and single-step rewards in traditional reinforcement learning methods cannot reflect long-term temporal effects. By constructing macro-step rewards and macro-step discount information, it effectively enhances the ability of policy updates to perceive the long-term risk mitigation effect, thereby improving the convergence stability and global optimality of policy optimization. This gives the invention stronger adaptability and continuous optimization capabilities in complex and dynamic operating environments. Attached Figure Description

[0029] Figure 1 This is a graph showing the fitting distribution of the true and predicted values ​​of the clipping distance linear attention Transformer model proposed in Example 1 for the next state risk prediction task; where the horizontal axis represents the true next state risk value, the vertical axis represents the model's predicted next state risk value, the blue dashed line represents the ideal fitting line, and the orange solid line represents the linear fitting result; as can be seen from the graph, the predicted scatter points are generally closely distributed near the ideal fitting line, and the fitting curve is highly consistent with the ideal line; at the same time, the evaluation index R 2 =0.990, MAE=0.023, RMSE=0.028;

[0030] Figure 2 This is a graph showing the distribution of the true and predicted values ​​of the Transformer model proposed in Example 2 in the next-state risk prediction task. The horizontal axis represents the true next-state risk value, and the vertical axis represents the model's predicted next-state risk value. The blue dashed line represents the ideal fitting line, and the orange solid line represents the linear fitting result. As can be seen from the graph, the predicted scatter points are somewhat dispersed relative to the ideal fitting line, and the fitted curve deviates from the ideal line. Meanwhile, the evaluation index R... 2 =0.966, MAE=0.042, RMSE=0.051. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] Example 1, according to Figure 1 This invention proposes an adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence. The system is deployed in the power distribution network operation environment composed of distribution master station, edge computing node, field intelligent terminal and primary equipment of distribution network. The system includes a distribution network operation data acquisition module, an edge processing and judgment module, a graph state construction module, a physical risk prior generation module, a candidate action screening module, a reinforcement learning decision module and an operation and maintenance execution feedback module that are connected in sequence.

[0033] This embodiment uses the A power supply zone of the 10kV distribution network in City A as an example. This power supply zone consists of one 110kV substation, two 10kV busbars, four 10kV feeders, three ring main units, two tie lines, one set of SVG reactive power compensation devices, one set of station-level energy storage support devices, and multiple FTUs, DTUs, and distribution transformer monitoring terminals. The main objects involved in this operation and maintenance decision-making include: the 10kV F1 heavy-load feeder; the adjacent 10kV F2 feeder; tie switch L12; sectionalizing switches S1 and S2; the switching unit in ring main unit RMG-03; SVG reactive power compensation device Q1; energy storage converter ES-1; and the load management terminal within the load transfer section.

[0034] The distribution network operation data acquisition module is deployed on the distribution master station side server, feeder terminal unit (FTU), substation terminal unit (DTU), switchgear control terminal, ring main unit monitoring terminal, distribution transformer monitoring terminal, intelligent switch controller, and protection and control devices. It is used to collect raw data of distribution network operation and transmit it to the distribution room edge gateway through the 5G communication network. The raw data of distribution network operation includes: feeder topology connection information, switch opening and closing status, line current, voltage, power factor, active power, reactive power, line load rate, transformer load rate, bus voltage amplitude, alarm events, protection action records, equipment interlocking status, action cooling time, communication status information, and operation and maintenance records.

[0035] The edge processing and judgment module is deployed at the edge gateway of the power distribution room. It performs timestamp alignment, anomaly removal, missing data completion, dimension normalization, device identity mapping, and topology consistency verification on the raw data of the power distribution network operation, generating structured power distribution network operation status data. Based on the structured power distribution network operation status data, it calculates the load rate of each line and transformer, extracts the maximum load rate of the line cluster at the current moment as the instantaneous risk indicator, compares the instantaneous risk indicator with the preset danger threshold, and generates a danger trigger indicator. When the danger trigger indicator indicates that the current power distribution network operation status has reached the danger trigger condition, the adaptive operation and maintenance control process is initiated; when the danger trigger indicator indicates that the current power distribution network operation status has not reached the danger trigger condition, the default operation strategy is maintained.

[0036] The graph state construction module is deployed on the main station server. Based on the danger trigger indication, it is used to map the structured distribution network operation status data into distribution network graph state objects using a hybrid graph modeling method of bus level and feeder level when the danger trigger conditions are met. The distribution network graph state objects include node sets, edge sets, node features, and edge features. Among them, nodes are used to represent substation bus nodes, feeder nodes, ring network nodes, tie nodes, transformer area nodes, and distribution transformer nodes. Edges are used to represent feeder segment connection relationships, switch connection relationships, tie line connection relationships, and transformer coupling relationships. Node features include node active power injection, node reactive power injection, voltage estimate, load level, equipment cooling time, and node operation constraint features. Edge features include line load rate, directed power flow information, edge voltage information, edge state variables, switch state variables, and edge cooling constraint features, thereby constructing a dynamic graph state representation that reflects the current distribution network topology, electrical operation status, and operational constraints.

[0037] The hybrid modeling method of busbar and feeder level includes: establishing busbar node representations for the double busbar structure, segmented busbar structure, and busbar structure formed by electrical equivalence within the station; establishing line edge representations for feeder trunk line segments, branch line segments extended from the trunk, and tie line segments used for interconnection between feeders; updating the edge connection relationship in real time based on the switch opening and closing status; and performing dynamic removal processing on the corresponding edge relationships for lines in the open state, tie lines not in operation, and equipment in the locked state, in order to characterize the dynamic changes in node connection relationships and edge state characteristics during the distribution network topology reconfiguration process.

[0038] The physical risk prior generation module is deployed on the main server. It uses a graph neural network encoder to process the distribution network diagram state objects, performing graph feature extraction based on edge message passing and edge-aware attention aggregation mechanisms. It jointly models and propagates information from node and edge features to obtain a graph-level embedding representation of the current distribution network state. For each executable maintenance action at the current moment, the graph-level embedding representation and the corresponding action embedding vector are jointly encoded to obtain a state-action fusion feature representation. A pruned distance linear attention Transformer model is constructed, and the state-action fusion feature representation is input into the pruned distance linear attention Transformer model to predict the next state risk value after the execution of the executable maintenance action. Based on the next state risk value, each executable maintenance action is generated. The physical risk score corresponds to the action; the physical risk score is used to characterize the ability of each executable operation and maintenance action to mitigate the current overload risk, limit violation risk, or security risk; the construction method of the pruning distance linear attention Transformer model is as follows: based on the Transformer model, a linear proxy scoring mechanism based on pruning distance constraints is introduced into the attention calculation, and a truncated linear mapping is used to replace the exponential function calculation to realize the linear expression of attention weights; the monotonicity and non-negativity of the score are controlled by the baseline parameter and slope parameter, and the attention weight aggregation is completed while maintaining the relative ranking relationship. The exponential attention weight calculation method in the Transformer model is linearized and optimized, thereby constructing the pruning distance linear attention Transformer model;

[0039] In this embodiment, the physical risk prior generation module uses a graph neural network encoder to process the graph state object; the graph neural network encoder adopts a 3-layer edge message passing structure, with hidden dimensions of 64, 64 and 128 for each layer; through edge-aware attention aggregation, node features and edge features are jointly modeled to obtain a graph-level embedding representation with a dimension of 256.

[0040] At this moment, eight operational actions can be identified: A1: Close the tie switch L12 to transfer the load from F1 to F2; A2: Open the sectionalizing switch S2 to disconnect the general load section at the end of F1; A3: Activate the SVG reactive power compensation device Q1; A4: Start the energy storage ES-1 discharge support; A5: Close L12 and simultaneously activate Q1; A6: Close L12 and start ES-1; A7: Adjust the power distribution of the flexible interconnection device; A8: Issue peak shaving instructions only to the load management terminal.

[0041] In this embodiment, the next state risk value and corresponding physical risk score output by the model are described as follows:

[0042] ;

[0043] Among them, the higher the physical risk score, the stronger the ability to mitigate the current overload risk, over-limit risk and safety risk; as can be seen from the above results, A5, A6 and A1 have better risk mitigation effects.

[0044] The candidate action filtering module is deployed on the main station server. It is used to combine the action feasibility constraint set at the current moment to perform feasibility mask filtering on all executable operation and maintenance actions to obtain the set of actionable actions under the current dangerous state. It sorts the actionable actions according to the physical risk score corresponding to each action, selects a preset number of low-risk preferred actions from the set of actionable actions, and constructs a subset of state-related candidate actions to compress the original action space into a candidate control space oriented towards the current dangerous state.

[0045] The reinforcement learning decision-making module is deployed on the main station server. It is used to input the distribution network diagram state objects into the policy value joint network and output the policy preference value and state value estimate corresponding to each maintenance action in the current state. For each candidate action in the candidate action subset, its policy preference value and corresponding physical risk score are reweighted and fused to construct the candidate action probability distribution. Based on the candidate action probability distribution, the target maintenance action is selected and the distribution network adaptive maintenance control command is generated. At the same time, the reinforcement learning decision-making module updates the network parameters of the policy value joint network based on the semi-Markov decision feedback data returned by the maintenance execution feedback module to achieve adaptive policy optimization for the risk evolution process. The target maintenance actions include network reconfiguration actions, tie switch switching actions, sectional switch switching actions, load transfer actions, flexible regulation equipment control actions, reactive power compensation device switching actions, energy storage support actions, and fault isolation auxiliary actions.

[0046] In this embodiment, the calculated probability distribution of candidate actions is as follows: A5: 0.336; A6: 0.291; A1: 0.245; A4: 0.128; the system selects A5 as the target maintenance action.

[0047] The operation and maintenance execution feedback module is deployed in the main station control system, distribution automation control terminal, and field execution devices. It sends distribution network adaptive operation and maintenance control commands to the corresponding execution objects via the 5G communication network, controlling the corresponding execution objects to perform operation and maintenance actions. The execution objects include feeder switches, tie switches, sectionalizing switches, ring main unit switches, reactive power compensation devices, flexible interconnection devices, energy storage converters, load management terminals, and fault isolation devices. After execution, it continuously collects execution feedback data, subsequent micro-step operation data, and system safety status data. Based on the system safety status data, it determines the danger clearance status and constructs macro-step reward information and macro-step discount information from the danger triggering time to the danger clearance time. It feeds this information back to the reinforcement learning decision module to realize a semi-Markov adaptive operation and maintenance control closed loop based on the danger triggering time. This part constructs a semi-Markov reinforcement learning feedback mechanism based on dangerous event triggering. It models the risk evolution process across time intervals through macro-step rewards and macro-step discounts, realizing a unified evaluation of the long-term effects of operation and maintenance actions, thus distinguishing it from traditional reinforcement learning methods based on fixed time steps and single-step rewards.

[0048] The target maintenance action A5 is: close the tie switch L12 to implement load transfer from F1 to F2, and simultaneously activate the SVG reactive power compensation device Q1 to support the voltage of the transferred section; based on this target maintenance action, the reinforcement learning decision module generates the following distribution network adaptive maintenance control instructions:

[0049] 1. Communication switch control command:

[0050] Instruction number: CMD-20260718-183512-01;

[0051] Execution object: Interchange switch L12; Control content: Switching from open to closed;

[0052] Pre-operation verification requirements: voltage difference between the two busbars not greater than 0.35 kV; phase angle difference not greater than 8°; frequency difference not greater than 0.15 Hz; switch mechanism lockout status = 0; operation cooling time = 0 s; operation execution time limit: completed within 2 s;

[0053] Action result feedback requirements: return of open / closed position status, execution success flag, action delay, and summary of mechanism current curve.

[0054] 2. Load transfer control command:

[0055] Instruction number: CMD-20260718-183512-02;

[0056] Target of execution: RMG-03 ring main unit switchgear and related feeder end switches;

[0057] Control measures: Implement interconnection and transfer of a total of 1.62 MW of load from areas A and B at the end of F1 to F2;

[0058] Distribution areas: Area T-12: 0.43 MW; Area T-13: 0.51 MW; Area T-16: 0.36 MW; Commercial load group CL-08: 0.32 MW;

[0059] Control constraints: The predicted load rate after F2 power transfer must not exceed 98%; critical power supply loads must not lose power; the power transfer path must not pass through interlocked equipment.

[0060] Execution order: First verify the contact path, then perform ring network reconstruction, and finally confirm power flow redistribution.

[0061] 3. SVG commissioning control instructions:

[0062] Instruction number: CMD-20260718-183512-03;

[0063] Execution target: SVG reactive power compensation device Q1;

[0064] Control settings: Put into operation, target reactive power support capacity set at +1.8 Mvar;

[0065] Voltage regulation target: Increase the minimum voltage at the end of F1 from 9.21 kV to no less than 9.45 kV;

[0066] Regulation mode: Voltage closed loop + reactive power fast tracking mode;

[0067] Response time requirement: Enter the commissioning state within 300 ms.

[0068] 4. Monitor and transmit control commands:

[0069] Instruction number: CMD-20260718-183512-04;

[0070] Targets: FTUs of feeders F1 and F2, RMG-03 terminals, and related distribution transformer monitoring terminals;

[0071] Control content: Starting from the automatic execution time, the subsequent micro-step operation data will be continuously transmitted back at a sampling period of 2 seconds for a duration of 120 seconds;

[0072] Data items to be returned: line current, voltage, power flow direction, load rate, switch status, alarm status, protection action status, and cool-down time update value.

[0073] Example 2, according to Figure 2The difference between this embodiment and Embodiment 1 is that the physical risk prior generation module uses a pruned distance linear attention Transformer model to model the state-action fusion feature representation in order to predict the next state risk value after each executable operation and maintenance action is executed. However, in this embodiment, the pruned distance constraint and linear attention mechanism are no longer introduced. Specifically, in this embodiment, the physical risk prior generation module is constructed as a Transformer model structure, which takes the state-action fusion feature representation as input, models the relationship between state and action through a multi-head attention mechanism and a feedforward network, and outputs the next state risk value after each executable operation and maintenance action is executed.

[0074] Example 3, based on Example 1, involves inputting the state-action fusion feature representation into a pruned distance linear attention Transformer model to predict the next state risk value after the execution of an executable operation and maintenance action. The specific steps include:

[0075] Step B1: Perform unified dimension mapping on the state-action fusion feature representation according to the preset feature dimensions to form a state-action joint feature sequence to represent the coupling relationship between the current distribution network diagram state and candidate operation and maintenance actions; input the state-action joint feature sequence into the linear mapping unit to generate corresponding query vectors, key vectors and value vectors respectively; and according to the multi-head attention structure, divide the query vector, key vector and value vector into head sub-vector groups corresponding to multiple attention heads, so that different attention heads perform parallel association modeling for state topology information, action semantic information, operation constraint information and risk propagation information respectively;

[0076] Step B2: For each attention head, perform scaling dot product correlation calculation based on the query sub-vector and key sub-vector in the head sub-vector group to obtain the original attention score value corresponding to each position; and arrange and combine the original attention score values ​​according to the preset position order to construct the original attention score sequence under the attention head; and perform maximum value retrieval processing on the original attention score sequence row by row to determine the maximum score value in the corresponding row.

[0077] Step B3: For each attention head, based on the difference between the maximum score and the original attention score corresponding to each position, construct the non-negative distance value corresponding to each position; and call the head parameter set obtained by the attention head in pre-offline calibration to perform upper bound pruning on the non-negative distance value to obtain a pruned distance sequence limited to a preset distance range; the head parameter set includes the baseline parameter, slope parameter and maximum pruning distance parameter, which are used to adapt each attention head to its corresponding score distribution characteristics and ensure that the subsequent linear proxy scoring results satisfy the monotonicity, boundedness and non-negativity constraints;

[0078] The non-negative distance value is calculated using the following clipping function:

[0079] ;

[0080] in, Indicates the index of the attention head. Indicates the position index in the attention sequence; Indicates the first The first thing to pay attention to. The clipping distance value corresponding to each position is used to characterize the distance metric of that position relative to the position with the maximum response; Indicates the first The maximum value among all original attention scores at all locations under each attention head is used as the response intensity at the reference location. Indicates the first The first thing to pay attention to. The original attention score corresponding to each position is calculated by the scaled dot product between the query subvector and the key vector; Indicates the first The maximum clipping distance parameter corresponding to each attention head is used to limit the upper bound of the distance value;

[0081] The corresponding set of header parameters is denoted as:

[0082] ;

[0083] in, Indicates the first The head parameter set corresponding to each attention head is used to characterize the parameter configuration of the attention head in the process of truncating linear proxy mapping; Indicates the first The baseline parameters corresponding to each attention head are used to determine the initial baseline value of the attention head in the linear surrogate scoring function. The baseline parameters are used to control the upper bound of the surrogate score when the clipping distance value is zero. Indicates the first The slope parameter corresponding to each attention head is used to characterize the influence of the change in the clipping distance value on the linear surrogate score. The slope parameter is used to control the magnitude by which the surrogate score decreases as the distance increases.

[0084] Step B4: For each attention head, based on the baseline parameters, slope parameters, and pruning distance sequence corresponding to that attention head, perform proxy scoring calculations on each position according to the truncated linear mapping relationship to generate a linear proxy score sequence corresponding to that attention head; wherein, the smaller the pruning distance, the larger the linear proxy score, and the larger the pruning distance, the smaller the linear proxy score, thereby maintaining the relative ranking relationship of the original attention scores without performing exponential operations; and through preset head parameter constraints, the linear proxy scores corresponding to each position are naturally kept non-negative, avoiding the introduction of additional position-by-position zero-value rectification operations;

[0085] The linear proxy scores satisfy the following calculation relationship:

[0086] ;

[0087] in, Indicates the first A person's attention is focused on the position. Linear agent score at the location;

[0088] Step B5: For each attention head, based on the linear proxy score sequence generated in Step B4, perform an accumulation operation on the linear proxy scores corresponding to each position to obtain the proxy score sum corresponding to the attention head; and construct a normalized mapping relationship based on a preset target scaling scale, perform inverse scaling on the proxy score sum to obtain a normalized scaling factor; perform a multiplication mapping between the linear proxy scores at each position and the normalized scaling factor to generate the truncated linear normalized attention weight sequence corresponding to each position under the attention head; the truncated linear normalized attention weight sequence satisfies the normalization constraint and maintains a fixed-point or integer form, thus replacing the exponential mapping and floating-point normalization process in traditional Softmax with addition, subtraction, pruning, multiplication-addition and inverse scaling methods;

[0089] Step B6: Apply the truncated linearly normalized attention weight sequence corresponding to each attention head to the value sub-vector group corresponding to the head sub-vector group, and perform weighted aggregation processing to obtain the risk association representation under each attention head; then perform splicing and linear transformation on the risk association representation under each attention head to generate a multi-head fused risk representation; then input the multi-head fused risk representation into the residual connection unit, normalization unit and feedforward transformation unit in sequence to perform inter-layer progressive feature enhancement, extract the high-order coupling relationship between state features, action features and risk propagation features layer by layer, and obtain the deep risk representation vector that represents the risk evolution trend after the target operation and maintenance action is performed; input the deep risk representation vector into the risk prediction mapping unit, perform regression mapping processing, and output the next state risk value after the execution of the executable operation and maintenance action.

[0090] Example 4, based on Example 3, continuously collects execution feedback data, subsequent microstep operation data, and system safety status data. Based on the system safety status data, it determines the danger clearance status and constructs macrostep reward information and macrostep discount information from the danger triggering time to the danger clearance time. This process is then fed back to the reinforcement learning decision module, specifically including the following steps:

[0091] Step C1: Feedback Acquisition and Microstep Modeling: After the target maintenance action is executed, execution feedback data and subsequent microstep operation data are continuously collected through the distribution automation terminal and field execution device. The execution feedback data includes: switch switching results, execution success flags, and execution delay, used to characterize the execution status of the target maintenance action. Subsequent microstep operation data includes: continuous operation status data and equipment status data. Continuous operation status data includes line current, voltage, load rate, and power flow distribution; equipment status data includes switch status changes, equipment interlocking status, and cooling time updates. Subsequent microstep operation data is continuously sampled according to a preset sampling period to form a microstep operation status sequence from the moment the target maintenance action is executed.

[0092] Step C2: Hazard Determination and Macrostep Construction: Based on execution feedback data, microstep operation state sequence, and system safety status data, perform real-time safety determination on the current distribution network operation status to determine whether the hazard clearance conditions are met. When the hazard clearance conditions are met, the moment is designated as the hazard clearance moment. A corresponding macrostep time window is constructed, using the hazard trigger moment as the macrostep start moment and the hazard clearance moment as the macrostep end moment. Based on the macrostep time window, interval division and state truncation processing are performed on the continuous operation process from the hazard trigger moment to the hazard clearance moment to extract the operation state evolution sequence. The microstep operation state sequence is used to characterize the changes in current, voltage, and load rate of lines and transformers, while the system safety status data is used to characterize over-limit states, alarm states, and protection action states. The hazard clearance conditions include: all line and transformer load rates are below the preset safety threshold, there are no voltage or equipment over-limit states, and the overall system operation status has returned to the preset safe operation range.

[0093] Step C3: Macrostep reward information construction: Based on the operational state evolution sequence, calculate the instant reward corresponding to each microstep moment, and perform discount accumulation processing on the instant rewards in chronological order to construct macrostep reward information, which is used to characterize the comprehensive effect of the target operation and maintenance action in the entire risk evolution process;

[0094] Macro step reward information is obtained by accumulating micro step rewards over time, and its expression is:

[0095] ;

[0096] in, This indicates the start time of the macrostep, i.e., the moment the danger is triggered. This represents the cumulative reward value of macrosteps, used to characterize the overall impact of the target maintenance actions on the system's operating status from the moment the danger is triggered to the moment the danger is resolved. This represents the time step offset relative to the start of the macrostep. This indicates the length of the macrostep time window, which is the number of microsteps taken from the moment the danger is triggered to the moment the danger is cleared. Indicates the first The time discount weight corresponding to each microstep moment is used to weight the instantaneous reward at different time positions within the macrostep time window; Indicates the first The instant reward for each step is used to characterize the quality of the system's operating status at that moment. The instant reward for each step can be constructed based on the following factors: changes in line and transformer load rates, voltage or equipment over-limit situations, alarm status and protection action status, as well as the stability of system operation and the degree of risk mitigation.

[0097] Step C4: Macrostep Discount Information Construction: Define the macrostep discount factor and model the time span corresponding to the macrostep time window based on the macrostep discount factor to construct macrostep discount information to characterize the degree of decision impact across time intervals; wherein, the macrostep discount factor is used to characterize the cumulative discount effect of the time step from the moment the danger is triggered to the moment the danger is cleared.

[0098] Define the macrostep discount factor as:

[0099] ;

[0100] in, This represents the macrostep discount factor, used to characterize the overall time discount result for the entire macrostep process from the moment the danger is triggered to the moment the danger is cleared; This represents the overall discount factor corresponding to the entire macrostep time window, used to characterize the impact of the time span of the macrostep process on value propagation;

[0101] Step C5: Semi-Markov Feedback Construction and Backhaul: Based on the macrostep reward information and macrostep discount information mentioned above, construct semi-Markov decision feedback data and backhaul the semi-Markov decision feedback data to the reinforcement learning decision module for executing policy updates based on the semi-Markov decision process (SMDP). Specifically, the hazard trigger event serves as the starting point for decision updates, and the hazard resolution event serves as the ending point. Instead of a fixed-time-step, incremental update method, a one-time value backhaul and policy optimization is performed based on the complete macrostep process, thereby achieving unified evaluation and optimization updates of the long-term effects of operational actions across time intervals.

[0102] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A power distribution network adaptive operation and maintenance system based on 5G edge intelligence, characterized in that, The system includes: The edge processing and judgment module acquires structured distribution network operation status data, calculates the load rate of each line and transformer, and generates danger trigger indication quantities. The graph state construction module maps structured distribution network operation status data into distribution network graph state objects based on the danger trigger indication quantity and using graph modeling. The physical risk prior generation module uses a graph neural network encoder to extract the graph-level embedding representation of the distribution network diagram state objects and co-encodes it with the corresponding action embedding vector to obtain the state-action fusion feature representation; it constructs a pruning distance linear attention Transformer model, inputs the state-action fusion feature representation into the pruning distance linear attention Transformer model, predicts the next state risk value, and generates a physical risk score. The candidate action filtering module constructs a subset of candidate actions. The reinforcement learning decision-making module processes the distribution network diagram state objects through the policy value joint network and outputs policy preference values. For each candidate action in the candidate action subset, its policy preference value and physical risk score are reweighted and fused to select the target operation and maintenance action and generate distribution network adaptive operation and maintenance control instructions. The operation and maintenance execution feedback module sends the distribution network adaptive operation and maintenance control instructions to the execution objects; it constructs macro-step reward information and macro-step discount information from the time of danger triggering to the time of danger clearance, and feeds them back to the reinforcement learning decision module.

2. The adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence according to claim 1, characterized in that: The construction method of the clipping distance linear attention Transformer model is as follows: Based on the Transformer model, a linear proxy scoring mechanism based on clipping distance constraints is introduced to linearize and optimize the exponential attention weight calculation method in the Transformer model, thereby constructing the clipping distance linear attention Transformer model.

3. The adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence according to claim 2, characterized in that: The process of inputting the state-action fusion feature representation into a clipped distance linear attention Transformer model to predict the risk value of the next state includes the following steps: Step B1: Perform a unified dimension mapping on the state-action fusion feature representation to form a state-action joint feature sequence; perform a linear mapping on the state-action joint feature sequence to generate query vector, key vector and value vector respectively; and divide the query vector, key vector and value vector into head sub-vector groups corresponding to the attention heads according to the multi-head attention structure. Step B2: For each attention head, perform scaling dot product correlation calculation based on the query sub-vector and key sub-vector in the head sub-vector group to obtain the original attention score and construct the original attention score sequence; and perform maximum value retrieval processing on the original attention score sequence row by row to determine the maximum score. Step B3: Based on the difference between the maximum score and the original attention score, construct non-negative distance values; and call the head parameter set to perform upper bound pruning on the non-negative distance values ​​to obtain the pruned distance sequence. Step B4: Based on the baseline parameters, slope parameters, and clipping distance sequence corresponding to the attention head, perform proxy scoring calculation to generate a linear proxy score sequence corresponding to the attention head; Step B5: Perform an accumulation operation on the linear proxy scores corresponding to each position in the linear proxy score sequence to obtain the proxy score sum; and construct a normalized mapping relationship based on a preset target scaling scale, perform reciprocal scaling on the proxy score sum to obtain the normalized scaling factor; multiply the linear proxy scores at each position with the normalized scaling factor to generate a truncated linear normalized attention weight sequence. Step B6: Apply the truncated linearly normalized attention weight sequence corresponding to each attention head to the value sub-vector group corresponding to the head sub-vector group, perform weighted aggregation and fusion to generate a multi-head risk representation, extract high-order features through residual connection and feedforward transformation, and output the next state risk value corresponding to the operation and maintenance action through mapping.

4. The adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence according to claim 3, characterized in that: The head parameter set includes baseline parameters, slope parameters, and maximum clipping distance parameters.

5. The adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence according to claim 1, characterized in that: The operation and maintenance execution feedback module specifically works as follows: it sends the adaptive operation and maintenance control instructions of the distribution network to the corresponding execution objects via the 5G communication network; it continuously collects execution feedback data, subsequent micro-step operation data and system safety status data, and determines the danger clearance status based on the system safety status data, constructs macro-step reward information and macro-step discount information from the time of danger triggering to the time of danger clearance, and feeds it back to the reinforcement learning decision module.

6. The adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence according to claim 5, characterized in that: The process of constructing macrostep reward and macrostep discount information from the moment the danger is triggered to the moment the danger is resolved, and feeding this information back to the reinforcement learning decision module, includes the following steps: Step C1: Continuously sample subsequent micro-step operation data according to the preset sampling period to form a micro-step operation state sequence; Step C2: Based on the execution feedback data, microstep operation state sequence, and system safety state data, perform real-time safety judgment to determine whether the danger clearance condition is met; when the danger clearance condition is met, determine that moment as the danger clearance moment; construct a macrostep time window with the danger trigger moment as the macrostep start moment and the danger clearance moment as the macrostep end moment; based on the macrostep time window, perform interval division and state interception processing on the continuous operation process from the danger trigger moment to the danger clearance moment, and extract the operation state evolution sequence; Step C3: Based on the running state evolution sequence, calculate the instant reward corresponding to each microstep time, and perform discount accumulation processing on the instant rewards in chronological order to construct macrostep reward information; Step C4: Model the time span corresponding to the macro step time window and construct macro step discount information; Step C5: Based on the macrostep reward information and macrostep discount information, construct semi-Markov decision feedback data and send the semi-Markov decision feedback data back to the reinforcement learning decision module.

7. The adaptive operation and maintenance system for power distribution networks based on 5G edge intelligence according to claim 1, characterized in that: The objects to be implemented include feeder switches, tie switches, sectionalizing switches, ring main unit switches, reactive power compensation devices, flexible interconnection devices, energy storage converters, load management terminals, and fault isolation devices.

Images