Machine Learning-Based Integrated Power Distribution System and Method for Unmanned Substations

Abstract

This invention discloses an intensive power distribution system and method for unattended substations based on machine learning, comprising: collecting operational data from the unattended substation, performing preprocessing and generating a standardized input dataset; constructing the correlation between topology, operating conditions, and equipment features to form a three-dimensional coupled feature structure; inputting a hierarchical graph temporal decision network, fusing topology and temporal features to output candidate control actions; constructing a dynamic feasible region based on joint features, and filtering a set of safe actions that meet constraints; inputting a topologically reversible neural differential control module to predict transient responses and optimize execution actions; issuing control commands and collecting feedback to trigger incremental model updates to achieve closed-loop optimization. This invention achieves intelligent intensive scheduling and self-learning closed-loop optimization control of unattended substations by introducing a hierarchical graph temporal decision network and a topologically reversible neural differential control mechanism.

Description

Technical Field

[0001] This invention relates to the field of power distribution automation and intelligent power system technology, and in particular to an unattended power distribution substation intensive power distribution system and method based on machine learning. Background Technology

[0002] With the continuous expansion of power distribution networks and the sustained growth of urban electricity load, traditional power distribution substations are gradually evolving towards unmanned operation and centralized monitoring to reduce maintenance costs and improve operational efficiency. Currently, unmanned power distribution substations primarily rely on fixed-threshold monitoring and remote control systems, collecting data such as voltage, current, and temperature for status monitoring and alarm response. However, these systems generally lack the ability to perform overall correlation analysis and intelligent coordination among multiple power distribution substations. They cannot achieve real-time optimized control in complex operating environments such as multi-station linkage, dynamic load changes, and equipment aging, easily leading to localized overload, uneven power distribution, and increased energy loss.

[0003] While existing distribution automation systems have incorporated some algorithmic optimization, such as rule-based engines or static models for scheduling decisions, they lack sufficient ability to identify the temporal dependencies and topology changes in operational data, making them unable to effectively respond to abnormal disturbances and changes in equipment health status. Existing systems primarily focus on single-station optimization, lacking a centralized scheduling mechanism across distribution substations. This leads to system scheduling lags and untimely responses in unattended scenarios, and even transient impacts such as sags and current inrushes during multi-station coordinated switching, affecting the stability and security of the power grid.

[0004] Therefore, how to provide an intensive power distribution system and method for unattended power distribution rooms based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intensive power distribution system and method for unattended substations based on machine learning.

[0006] The machine learning-based intensive power distribution method for unattended substations according to embodiments of the present invention includes:

[0007] Collect operational data from multiple unattended power distribution rooms, perform preprocessing on the operational data, and form a standardized input dataset;

[0008] Based on a standardized input dataset, the association between three types of features—topology, operating conditions, and equipment—is constructed. For each power distribution node, a feature vector containing electrical parameters, thermal parameters, and health status is established. For each connection relationship, electrical impedance, thermal coupling coefficient, and frequency domain association features are set. Based on the calculation results of energy flow and operating stress between nodes, the association weights are dynamically updated to form a three-dimensional coupling feature structure.

[0009] The three-dimensional coupled feature structure is input into the hierarchical graph temporal decision network model, and the topological relationship features and temporal load features are jointly learned. Multi-dimensional feature information is fused through the cross attention mechanism to generate a joint feature vector representing the group operation status of unattended power distribution rooms and output candidate control actions.

[0010] Based on the joint feature vector, a dynamic feasible region is constructed. Dynamic feasible region projection is performed on the candidate control actions to map the candidate control actions into the dynamic feasible region, thereby obtaining a set of safe actions that satisfy physical and electrical safety constraints.

[0011] The set of safety actions is input into the topological reversible neural differential control module to predict the transient response of the safety action execution process, calculate the current inrush current and voltage sag parameters, and adaptively optimize the execution sequence and time window to generate the transiently optimized execution action.

[0012] Based on the execution actions, control commands are sent to the power distribution terminal equipment to perform power scheduling, switch switching and reactive power compensation operations, collect execution feedback data, update the three-dimensional coupled feature structure and dynamic feasible region, and trigger incremental updates of the hierarchical graph temporal decision network model when there is a deviation between the prediction results and the actual feedback.

[0013] Optionally, the operating data includes voltage, current, active power, reactive power, temperature, humidity, equipment vibration signal, partial discharge signal, and switch status information.

[0014] Optionally, the preprocessing of the running data includes time synchronization, outlier removal, noise filtering, and normalization to eliminate data drift and sampling errors.

[0015] Optionally, forming the three-dimensional coupled feature structure includes:

[0016] Establish data objects for power distribution nodes and connection relationships, organize the standardized input dataset in chronological order, determine the node records and connection relationship records corresponding to each time point, and assign data quality tags to each record;

[0017] Generate a node feature vector for each distribution node, which includes electrical parameters, thermal parameters, and health status;

[0018] For each connection, a connection feature vector is generated that includes electrical impedance, thermal coupling coefficient and frequency domain correlation features. The electrical impedance includes resistance and reactance, and the frequency domain correlation features form a fixed-dimensional frequency band intensity sequence based on harmonic amplitude and frequency information.

[0019] The system calculates node operating stress and energy flow. The node operating stress is characterized by a set of multiple indicators consisting of load rate, temperature rise margin, health, switching frequency, and total harmonic distortion. The energy flow is characterized by a set of multiple indicators consisting of active power flow, reactive power flow, directionality, and harmonic power flow intensity. At the same time, it generates an event impact factor to characterize the impact of recent disturbances and a potential energy level to characterize the comprehensive risk level of nodes and connection relationships.

[0020] The associated weights are dynamically updated based on the node's operating stress and energy flow:

[0021] A dual-channel mechanism is adopted, consisting of slow-channel sliding updates and fast-channel event-triggered updates;

[0022] Within the sliding window, increase the weight of low-stress, low-harmonic, and stable directional connections, and decrease the weight of high-stress, high-harmonic, directionally repetitive, or near-capacity limit connections.

[0023] When the event impact factor exceeds the set threshold, a rapid adjustment is performed in the current time period, and a freeze window and unfreeze rules are set. The contraction-expansion criterion is executed according to the potential energy level to limit the weight fluctuation range. The output is a three-dimensional coupled feature structure containing node feature vectors, connection relationship feature vectors, and dynamically updated associated weights.

[0024] Optionally, the output candidate control action includes:

[0025] A hierarchical graph temporal decision network model is constructed, which consists of a topology-health dual-channel graph encoding module, an event-enhanced temporal convergence module, and a constraint-aware strategy generation module.

[0026] In the topology-health dual-channel graph encoding module, the topological relationship and electrical parameters in the three-dimensional coupled feature structure are used as the topology channel input, and the thermal parameters, health and frequency domain correlation features are used as the health channel input. Intra-channel feature extraction and cross-channel alignment are performed. Edge selection and directionality enhancement are performed based on potential energy level and weight confidence score. The output is node embedding, connection relationship embedding and edge importance score.

[0027] In the event-enhanced time-series aggregation module, a fixed-length sliding window is used to aggregate node embeddings and connection relationship embeddings in time series, setting short-term disturbance branches and medium-term trend branches, wherein:

[0028] The short-term disturbance branch uses a fixed-length short-term sliding window and event-triggered gating to process high-frequency changes, performs peak preservation, fast reweighting of side importance and impulse response extraction, and outputs a disturbance representation including inrush intensity, slump depth and abrupt change duration;

[0029] The intermediate trend branch uses a fixed-length intermediate sliding window and periodic encoding to process smooth changes, performs denoising and smoothing, periodic pattern extraction and drift estimation, and outputs a trend representation that includes load baseline, temperature rise baseline and harmonic background.

[0030] The outputs of the two branches are aligned by timestamp and then weighted and converged, while retaining the causal mask and rollback placeholders to obtain a joint timing representation;

[0031] In the constraint-aware strategy generation module, constraint prior hints are generated based on device nameplate parameters, connection capacity limit, frequency band limit, edge importance score and potential energy level. Dual-stack constraints of soft mask and hard mask are applied to the strategy space to restrict the set of non-combinable actions. The output includes candidate control actions containing power scheduling parameters, reactive power compensation parameters and switch switching sequence.

[0032] Optionally, obtaining the set of safety actions that satisfy physical and electrical safety constraints includes:

[0033] Establish a basic constraint list that includes upper and lower voltage limits, upper current limits, upper temperature rise limits, harmonic limits, connection capacity, minimum switching intervals and order, and generate a constraint component library;

[0034] A dynamic feasible domain is constructed based on a constraint component library. A hierarchical deformation mechanism is used to combine the basic boundary, context boundary, and transient occupancy boundary. Boundary deformation memory and hysteresis mechanism are introduced during the construction process. A recovery threshold and a freeze window are set for boundaries that continuously shrink or expand within a short period of time. A source label and an effective time period label are attached to each boundary.

[0035] Online self-verification and local probing are performed on the dynamic feasible domain. For nodes and connections whose edge importance scores exceed the preset threshold and are close to the boundary, a small-amplitude probing action sequence is generated. The joint temporal representation and potential energy level are used to quickly evaluate the probing results. The active boundary index and boundary confidence level are updated. Boundaries that do not meet the consistency check are downgraded or removed in the current cycle. The source label and causal link of the adjustment are recorded in the constraint component library.

[0036] Perform two-stage compliance processing on candidate control actions within the dynamic feasible domain:

[0037] In the first stage, the power scheduling parameters and reactive power compensation parameters are made compliant in the continuous domain according to the superposition results of the basic boundary, context boundary and transient occupancy boundary, forming a parameter set with a clear safety margin.

[0038] The second stage involves making the switch switching sequence compliant by adhering to the minimum time interval, interlocking and mutual exclusion rules, sequence order, maintenance status, and list of key equipment. Actions that conflict with transient occupancy are eliminated, and necessary buffer actions are added to generate a set of safety actions and metadata.

[0039] Optionally, generating the transiently optimized execution action includes:

[0040] Receive a set of safety actions and metadata, combine the timing representation and rollback placeholder marker, and initialize the topological reversible neural differential control module, which consists of a topological reversible mapping unit, a transient coupling prediction unit, and a sequence adaptation and rollback control unit.

[0041] In the topological reversible mapping unit, based on the three-dimensional coupling feature structure and dynamic feasible region, the set of safety actions is mapped item by item to a topological state sequence that can be executed in the forward direction and can be rolled back in the reverse direction, generating a reversible mapping table containing node states, connection relationship states and energy conservation constraint labels.

[0042] In the transient coupling prediction unit, electromagnetic-thermal multi-field coupling transient response prediction is performed for each safety action in the reversible mapping table to form an action-level evaluation record for switching operation, handover and reactive power compensation, and the boundary occupancy and safety margin of each index are marked according to the active boundary index and transient occupancy margin.

[0043] In the sequence adaptation and rollback control unit, based on the action-level evaluation record, the safety action set is jointly searched by the sequential neighborhood exchange and time displacement. Compliance verification is performed according to the minimum time interval, sequence, interlock and mutual exclusion rules, maintenance status and key equipment list. Actions that conflict with transient occupancy are eliminated or replaced.

[0044] The output is the execution action after transient optimization, which includes the switch switching sequence, execution time window, power scheduling parameters and reactive power compensation parameters.

[0045] Optionally, triggering incremental updates of the hierarchical graph temporal decision network model when there is a deviation between the predicted result and the actual feedback includes:

[0046] Based on the execution actions, control commands are sent to the power distribution terminal equipment to perform power scheduling, switch switching and reactive power compensation operations, and feedback data of node voltage, current, temperature, power factor, frequency deviation and energy flow are collected in real time.

[0047] The feedback data is compared with the prediction results before execution. The node level and the connection level are calculated. The three-dimensional coupling feature structure and dynamic feasible region are updated according to the deviation results. The node features, connection relationships and association weights are adjusted. The basic boundary, context boundary and transient occupancy boundary are reconstructed. The active constraint set, constraint residual table and safety confidence score are updated.

[0048] When the deviation exceeds the preset threshold, the incremental update of the hierarchical graph temporal decision network model is triggered. Local parameter adjustments are performed on the topology-health dual-channel graph encoding module, the event-enhanced temporal convergence module, and the constraint-aware strategy generation module to update the version and confidence level of the hierarchical graph temporal decision network model.

[0049] According to an embodiment of the present invention, a machine learning-based unattended substation intensive power distribution system includes the following modules:

[0050] The data acquisition and preprocessing module is used to collect runtime data and perform preprocessing to generate standardized input datasets;

[0051] The feature construction and correlation analysis module is used to construct the correlation between topology, operating conditions and equipment features based on a standardized input dataset, and generate a three-dimensional coupled feature structure;

[0052] The hierarchical graph timing decision network module is used to receive the three-dimensional coupled feature structure, fuse the topology and timing load features, and output candidate control actions for power scheduling, reactive power compensation and switching.

[0053] The dynamic feasible region construction module is used to construct the dynamic feasible region, perform projection on candidate control actions, filter and output a set of safe actions that meet physical and electrical constraints;

[0054] The reversible neural differential control module is used to predict and optimize the transient response of safety actions, calculate inrush current, voltage sag and energy disturbance parameters, and generate optimized execution actions.

[0055] The execution and self-learning module is used to send control commands to the power distribution terminal based on the execution actions, collect feedback data, and update the three-dimensional coupled feature structure and dynamic feasible domain.

[0056] The beneficial effects of this invention are:

[0057] This invention introduces a hierarchical graph-based temporal decision network and a topologically reversible neural differential control mechanism to achieve adaptive scheduling and refined control of unattended substations under complex power conditions. The system can form a unified data association model across multiple substations, automatically identify topology changes and load migration characteristics, dynamically adjust power allocation strategies, and achieve intensive operation management through cross-site collaboration. Compared with traditional control methods based on fixed rules or static thresholds, this invention can maintain the stability and safety of the power grid operation in dynamic scenarios such as load fluctuations, temperature rise changes, and harmonic interference, improving the real-time response capability and operating efficiency of the power distribution system.

[0058] This invention constructs a two-layer mechanism of dynamic feasible domain and safety constraints, enabling real-time safety verification of control commands before issuance. This avoids ineffective scheduling, over-limit operations, or transient impacts that may occur in traditional systems. The topologically reversible neural differential control module can transiently predict and pre-optimize voltage sags, current inrushes, and energy disturbances during the control execution phase, ensuring energy conservation and safe equipment operation during scheduling. This not only improves the system's anti-disturbance capability in complex switching scenarios but also enables unattended power distribution rooms to achieve adaptive control and risk prevention without human intervention.

[0059] This invention introduces a feedback-driven incremental self-learning mechanism. When there is a deviation between the predicted result and the actual feedback, it automatically triggers the hierarchical graph time-series decision network model update, continuously optimizing system parameters and control strategies. This effectively overcomes the limitations of traditional algorithm models being static and lacking continuous evolution capabilities, enabling the system to maintain high accuracy and reliability over a long period. By combining a closed-loop system of data self-sensing, model self-evolution, and control self-correction, this invention achieves intelligent, dynamic, and adaptive management of unattended power distribution rooms, significantly improving the safety, flexibility, and overall operational intelligence level of the power distribution system. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of the machine learning-based intensive power distribution method for unattended substations proposed in this invention.

[0062] Figure 2 This is a schematic diagram of the structure of the machine learning-based unattended power distribution room integrated power distribution system proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1 A machine learning-based method for intensive power distribution in unattended substations includes:

[0065] Collect operational data from multiple unattended power distribution rooms, perform preprocessing on the operational data, and form a standardized input dataset;

[0066] Based on a standardized input dataset, the association between three types of features—topology, operating conditions, and equipment—is constructed. For each power distribution node, a feature vector containing electrical parameters, thermal parameters, and health status is established. For each connection relationship, electrical impedance, thermal coupling coefficient, and frequency domain association features are set. Based on the calculation results of energy flow and operating stress between nodes, the association weights are dynamically updated to form a three-dimensional coupling feature structure.

[0067] The three-dimensional coupled feature structure is input into the hierarchical graph temporal decision network model, and the topological relationship features and temporal load features are jointly learned. Multi-dimensional feature information is fused through the cross attention mechanism to generate a joint feature vector representing the group operation status of unattended power distribution rooms and output candidate control actions.

[0068] Based on the joint feature vector, a dynamic feasible region is constructed. Dynamic feasible region projection is performed on the candidate control actions to map the candidate control actions into the dynamic feasible region, thereby obtaining a set of safe actions that satisfy physical and electrical safety constraints.

[0069] The set of safety actions is input into the topological reversible neural differential control module to predict the transient response of the safety action execution process, calculate the current inrush current and voltage sag parameters, and adaptively optimize the execution sequence and time window to generate the transiently optimized execution action.

[0070] Based on the execution actions, control commands are sent to the power distribution terminal equipment to perform power scheduling, switch switching and reactive power compensation operations, collect execution feedback data, update the three-dimensional coupled feature structure and dynamic feasible region, and trigger incremental updates of the hierarchical graph temporal decision network model when there is a deviation between the prediction results and the actual feedback.

[0071] In this embodiment, the operating data includes voltage, current, active power, reactive power, temperature, humidity, equipment vibration signal, partial discharge signal, and switch status information.

[0072] In this embodiment, the preprocessing of the running data includes time synchronization, outlier removal, noise filtering, and normalization of the running data to eliminate data drift and sampling errors.

[0073] In this embodiment, forming the three-dimensional coupling feature structure includes:

[0074] Establish data objects for power distribution nodes and connection relationships, organize the standardized input dataset in chronological order, determine the node records and connection relationship records corresponding to each time point, and assign data quality tags to each record;

[0075] For each power distribution node, a node feature vector is generated that includes electrical parameters, thermal parameters, and health status. The electrical parameters include voltage, current, active power, and reactive power. The thermal parameters include temperature and temperature rise margin. The health status includes equipment degradation degree and maintenance status indicator.

[0076] For each connection, a connection feature vector is generated that includes electrical impedance, thermal coupling coefficient and frequency domain correlation features. The electrical impedance includes resistance and reactance, and the frequency domain correlation features form a fixed-dimensional frequency band intensity sequence based on harmonic amplitude and frequency information.

[0077] The system calculates node operating stress and energy flow. The node operating stress is characterized by a set of multiple indicators consisting of load rate, temperature rise margin, health, switching frequency, and total harmonic distortion. The energy flow is characterized by a set of multiple indicators consisting of active power flow, reactive power flow, directionality, and harmonic power flow intensity. At the same time, it generates an event impact factor to characterize the impact of recent disturbances and a potential energy level to characterize the comprehensive risk level of nodes and connection relationships.

[0078] The associated weights are dynamically updated based on the node's operating stress and energy flow:

[0079] A dual-channel mechanism is adopted, consisting of slow-channel sliding updates and fast-channel event-triggered updates;

[0080] Within the sliding window, increase the weight of low-stress, low-harmonic, and stable directional connections, and decrease the weight of high-stress, high-harmonic, directionally repetitive, or near-capacity limit connections.

[0081] When the event impact factor exceeds a set threshold, a rapid adjustment is performed within the current time period, and a freeze window and unfreeze rules are set. A contraction-expansion criterion is applied based on the potential energy level to limit weight fluctuations. The output is a three-dimensional coupled feature structure containing node feature vectors, connection relationship feature vectors, and dynamically updated associated weights. The unfreeze rule is as follows: when the event impact factor remains below the lower threshold and stable within a continuous monitoring period, the corresponding freeze window constraint is released, and the dynamic adjustment capability of nodes and connections is restored according to time-decaying weights. The restoration rate is adaptively set based on the node potential energy level and the importance of associated edges; high-potential-energy nodes are unfrozen first, while low-potential-energy nodes are unfrozen later to avoid frequent fluctuations. When the overall potential energy gradient tends to stabilize and the rate of change of associated weights is lower than a preset convergence threshold, the freeze state is terminated, and regular dynamic updates are resumed.

[0082] In this embodiment, the output candidate control action includes:

[0083] A hierarchical graph temporal decision network model is constructed, which consists of a topology-health dual-channel graph encoding module, an event-enhanced temporal convergence module, and a constraint-aware strategy generation module.

[0084] In the topology-health dual-channel graph encoding module, the topological relationship and electrical parameters in the three-dimensional coupled feature structure are used as the topology channel input, and the thermal parameters, health and frequency domain correlation features are used as the health channel input. Intra-channel feature extraction and cross-channel alignment are performed. Edge selection and directionality enhancement are performed based on potential energy level and weight confidence score. The output is node embedding, connection relationship embedding and edge importance score.

[0085] In the event-enhanced time-series aggregation module, a fixed-length sliding window is used to aggregate node embeddings and connection relationship embeddings in time series, setting short-term disturbance branches and medium-term trend branches, wherein:

[0086] The short-term disturbance branch uses a fixed-length short-term sliding window and event-triggered gating to process high-frequency changes, performs peak preservation, fast reweighting of side importance and impulse response extraction, and outputs a disturbance representation including inrush intensity, slump depth and abrupt change duration. The short-term sliding window is set to 5 consecutive sampling periods, corresponding to a time span of 10 seconds.

[0087] The intermediate trend branch employs a fixed-length intermediate sliding window and periodic encoding to process stable changes, performing denoising smoothing, periodic pattern extraction, and drift estimation. The output includes a trend representation encompassing the load baseline, temperature rise baseline, and harmonic background. The intermediate sliding window is set to 60 consecutive sampling periods, corresponding to a time span of 120 seconds. The denoising smoothing, periodic pattern extraction, and drift estimation are specifically performed as follows:

[0088] Moving average and adaptive weighted filtering methods are used to denoise and smooth the medium-term trend data, suppress random disturbances and short-term fluctuations, and retain the main trend.

[0089] Based on the periodic decomposition and spectral clustering method, the periodic pattern features of load, voltage and temperature are extracted, the main periodic components and phase relationships are identified, and periodic pattern vectors are formed.

[0090] The long-term drift is calculated using time series residual regression and drift detection mechanisms, the direction and magnitude of trend drift are evaluated, and the baseline reference is dynamically corrected to generate stable load baseline, temperature rise baseline and harmonic background trend representation.

[0091] The outputs of the two branches are aligned by timestamp and then weighted and converged, while retaining the causal mask and rollback placeholders to obtain a joint timing representation;

[0092] In the constraint-aware strategy generation module, constraint priors are generated based on device nameplate parameters, connection capacity limits, frequency band restrictions, edge importance scores, and potential energy levels. Dual-stack constraints of soft and hard masks are applied to the strategy space to restrict the set of non-combinable actions. The module outputs candidate control actions containing power scheduling parameters, reactive power compensation parameters, and switching sequences. The dual-stack constraints of soft and hard masks applied to the strategy space refer to:

[0093] Soft masking applies dynamic probabilistic constraints to possible actions in the policy space. By introducing continuous weight coefficients to adjust the sampling priority of different actions, it enables actions to maintain exploration flexibility while meeting safety and economic requirements.

[0094] Hard masks are used to strictly shield combinations of actions that do not meet physical constraints or safety boundaries, and directly set actions that exceed the rated capacity of the equipment, violate electrical interlock rules, or break frequency band limits to an unselectable state.

[0095] The execution order of dual-stack constraints is as follows: first, a hard mask is applied to remove non-compliant actions, and then a soft mask is used to perform weighted filtering of the remaining actions, thereby achieving layered pruning and security guidance of the policy space.

[0096] In this embodiment, obtaining the set of safety actions that satisfy physical and electrical safety constraints includes:

[0097] Establish a basic constraint list that includes upper and lower voltage limits, upper current limits, upper temperature rise limits, harmonic limits, connection capacity, minimum switching intervals and order, and generate a constraint component library;

[0098] A dynamic feasible domain is constructed based on a constraint component library. A hierarchical deformation mechanism is adopted to combine the basic boundary, context boundary, and transient occupancy boundary. Boundary deformation memory and hysteresis mechanism are introduced during the construction process. Recovery thresholds and freeze windows are set for boundaries that continuously shrink or expand within a short period of time. Each boundary is also labeled with a source label and an effective period label. The basic boundary is determined by the equipment nameplate parameters and operating procedures. The context boundary shrinks or expands based on the potential energy level, health degree, and edge importance score. The transient occupancy boundary reserves an execution window and capacity margin based on the rollback occupancy mark and the intensity of recent events.

[0099] Online self-verification and local probing are performed on the dynamic feasible domain. For nodes and connections whose edge importance scores exceed the preset threshold and are close to the boundary, a small-amplitude probing action sequence is generated. The probing results are quickly evaluated using joint temporal representation and potential energy level. The active boundary index and boundary confidence level are updated. Boundaries that do not meet the consistency check are downgraded or removed in the current cycle. The source label and causal link of the adjustment are recorded in the constraint component library. The preset threshold is set to 0.7, and the edge importance score ranges from 0 to 1.

[0100] Perform two-stage compliance processing on candidate control actions within the dynamic feasible domain:

[0101] In the first stage, the power scheduling parameters and reactive power compensation parameters are made compliant in the continuous domain according to the superposition results of the basic boundary, context boundary and transient occupancy boundary, forming a parameter set with a clear safety margin.

[0102] The second stage involves making the switch switching sequence compliant by adhering to the minimum time interval, interlocking and mutual exclusion rules, sequence order, maintenance status, and list of key equipment. Actions that conflict with transient occupancy are eliminated, and necessary buffer actions are added to generate a set of safety actions and metadata.

[0103] In this embodiment, generating the transiently optimized execution action includes:

[0104] Receive a set of safety actions and metadata, combine the timing representation and rollback placeholder marker, and initialize the topological reversible neural differential control module, which consists of a topological reversible mapping unit, a transient coupling prediction unit, and a sequence adaptation and rollback control unit.

[0105] In the topological reversible mapping unit, based on the three-dimensional coupling feature structure and dynamic feasible region, the set of safety actions is mapped item by item to a topological state sequence that can be executed in the forward direction and can be rolled back in the reverse direction, generating a reversible mapping table containing node states, connection relationship states and energy conservation constraint labels.

[0106] In the transient coupling prediction unit, electromagnetic-thermal multi-field coupling transient response prediction is performed on each safety action in the reversible mapping table to form an action-level evaluation record for switching operations, handover, and reactive power compensation. The boundary occupancy and safety margin of each indicator are marked according to the active boundary index and transient occupancy margin. Specifically, the electromagnetic-thermal multi-field coupling transient response prediction for each safety action in the reversible mapping table is performed as follows:

[0107] Before execution, based on the electrical parameters of the nodes involved in the safety action, the thermal parameters of the equipment, and the connection relationships, the changes in voltage, current, and temperature are calculated simultaneously, and the electromagnetic disturbance and thermal response process at the moment the action is triggered is analyzed.

[0108] For different types of actions such as switching operations, handover switching, and reactive power compensation, calculate the current rise rate, voltage sag, energy release intensity, and short-time temperature rise value to identify high-risk periods and high-stress areas.

[0109] The electromagnetic response and thermal change results are integrated in time series, key indicators are extracted and compared with active boundary index and transient occupancy margin, the boundary occupancy and safety margin of each indicator are marked, and action-level transient evaluation results for scheduling optimization are formed.

[0110] In the sequence adaptation and rollback control unit, based on the action-level evaluation record, the safety action set is jointly searched by the sequential neighborhood exchange and time displacement. Compliance verification is performed according to the minimum time interval, sequence, interlock and mutual exclusion rules, maintenance status and key equipment list. Actions that conflict with transient occupancy are eliminated or replaced.

[0111] The output is the execution action after transient optimization, which includes the switch switching sequence, execution time window, power scheduling parameters and reactive power compensation parameters.

[0112] In this embodiment, triggering the incremental update of the hierarchical graph temporal decision network model when there is a deviation between the prediction result and the actual feedback includes:

[0113] Based on the execution actions, control commands are sent to the power distribution terminal equipment to perform power scheduling, switch switching and reactive power compensation operations, and feedback data of node voltage, current, temperature, power factor, frequency deviation and energy flow are collected in real time.

[0114] The feedback data is compared with the prediction results before execution. The node level and the connection level are calculated. The three-dimensional coupling feature structure and dynamic feasible region are updated according to the deviation results. The node features, connection relationships and association weights are adjusted. The basic boundary, context boundary and transient occupancy boundary are reconstructed. The active constraint set, constraint residual table and safety confidence score are updated.

[0115] When the deviation exceeds the preset threshold, the incremental update of the hierarchical graph temporal decision network model is triggered. Local parameter adjustments are performed on the topology-health dual-channel graph encoding module, the event-enhanced temporal convergence module, and the constraint-aware strategy generation module to update the version and confidence level of the hierarchical graph temporal decision network model.

[0116] refer to Figure 2 The machine learning-based unmanned substation integrated power distribution system includes the following modules:

[0117] The data acquisition and preprocessing module is used to collect runtime data and perform preprocessing to generate standardized input datasets;

[0118] The feature construction and correlation analysis module is used to construct the correlation between topology, operating conditions and equipment features based on a standardized input dataset, and generate a three-dimensional coupled feature structure;

[0119] The hierarchical graph timing decision network module is used to receive the three-dimensional coupled feature structure, fuse the topology and timing load features, and output candidate control actions for power scheduling, reactive power compensation and switching.

[0120] The dynamic feasible region construction module is used to construct the dynamic feasible region, perform projection on candidate control actions, filter and output a set of safe actions that meet physical and electrical constraints;

[0121] The reversible neural differential control module is used to predict and optimize the transient response of safety actions, calculate inrush current, voltage sag and energy disturbance parameters, and generate optimized execution actions.

[0122] The execution and self-learning module is used to send control commands to the power distribution terminal based on the execution actions, collect feedback data, and update the three-dimensional coupled feature structure and dynamic feasible domain. Example 1:

[0123] To verify the feasibility of this invention in practice, it was applied to three unmanned substations under an energy dispatch center. Each substation consists of a 35kV main transformer, a 10kV busbar, and multiple outgoing circuits, serving residential areas, industrial parks, and commercial areas respectively. Due to significant differences in power consumption characteristics, the traditional system exhibits obvious problems during high-temperature load periods in summer: Substation A frequently experiences line current exceeding limits, Substation B operates under low load conditions for extended periods, and Substation C is prone to voltage dips during switching operations. Dispatch relies on fixed threshold alarms and manual decision-making, lacking a multi-station collaborative optimization mechanism, leading to localized overloads, low energy utilization, and high operational risks.

[0124] In this scenario, the machine learning-based unmanned substation integrated power distribution system proposed in this invention is deployed. The data acquisition and preprocessing module collects voltage, current, active power, temperature, and equipment vibration signals from 102 monitoring points across three stations (A, B, and C) in real time, synchronizing every 5 seconds. After time alignment, anomaly removal, and normalization, the system generates a standardized input dataset. Subsequently, the feature construction and correlation analysis module extracts topology, operating conditions, and equipment health features based on this dataset, forming a three-dimensional coupled feature structure. In actual operation, the system dynamically identifies the high-temperature node (operating temperature 81.6℃) of the main transformer at station A and the redundant low-load node (load rate 42.3%) at station B, automatically adjusting the correlation weights to achieve real-time modeling of energy flow.

[0125] The hierarchical graph temporal decision network module is trained based on 30 days of historical load data. It employs topology-health dual-channel graph encoding, event-enhanced temporal convergence, and constraint-aware strategy generation mechanisms to generate joint feature vectors and automatically output candidate control actions. Taking 14:30 on August 15, 2025 as an example, when the system detects that the load of station A has reached 94% of its rated capacity, the model predicts that the power transfer path is "station A to station B to station C". It outputs a scheduling command to reduce the load of station A to 79%, while increasing the load of station B from 47% to 65%, keeping the voltage fluctuation within ±1.8%.

[0126] The dynamic feasible region construction module monitors constraint parameters such as current, voltage, and temperature rise in real time. When the temperature of the reactive power compensation device at station C reaches 80℃, the system automatically shrinks its dynamic feasible region and adjusts the set of safe actions to keep the power switching process within the safe threshold. The topology reversible neural differential control module further predicts and optimizes the transient response of the safe action execution. Before the scheduling action is executed, the predicted inrush current is 210A, and the actual monitored current during execution is 203A, with a deviation of less than 3.5%.

[0127] The execution and self-learning module automatically issues operation commands to each terminal switching device based on the optimized control instructions. After execution, feedback data is collected and compared with the model prediction results. When the deviation exceeds the threshold, the system triggers an incremental model update. For example, on August 20, 2025, a 4.2% deviation in Bilibili's temperature rise prediction was detected. The system automatically adjusted the weight of the health channel, and after the update, the deviation decreased to 1.6%, and the model confidence increased from 0.88 to 0.94.

[0128] After 60 days of continuous operation and testing, the system achieved real-time power balance and risk prevention through multi-station collaboration, significantly shortened dispatch response time, reduced equipment operating temperature, and significantly improved energy utilization.

[0129] Table 1. Comparison of the effects of centralized dispatching and operation of unattended power distribution rooms As can be seen from the data in Table 1, the machine learning-based unattended power distribution method for centralized power distribution in the present invention has achieved significant results in actual operation. Regarding load allocation, the average power allocation error decreased from ±4.9% before optimization to ±1.7%, a reduction of 65.3%. This indicates that the system's power scheduling in multiple power distribution rooms is more accurate, enabling dynamic balancing and real-time optimization, thus improving the load balance of the power distribution network. The scheduling response time was shortened from 5.1 seconds to 3.2 seconds, an improvement of 37.3%. This demonstrates that the system achieves rapid command generation and issuance through a hierarchical graph-based time-series decision network, significantly reducing manual intervention and communication delays, and significantly improving scheduling efficiency.

[0130] Regarding power quality and equipment safety, voltage sag decreased from 6.5% to 3.4%, and peak inrush current decreased from 315A to 203A, representing improvements of 47.7% and 35.6%, respectively. This indicates that the topological reversible neural differential control module played a crucial role in switching transient prediction and execution optimization, successfully suppressing inrush current and transient fluctuations in power distribution operations and effectively reducing mechanical and thermal stress on equipment. The average equipment temperature rise decreased from 79.4℃ to 74.1℃, and the heat load decreased by 5.3℃, further validating the system's thermal safety advantages under multi-point parallel operation and ensuring long-term stable operation of equipment in unattended conditions.

[0131] The system's intelligent learning and self-optimization capabilities were also verified. The overall energy efficiency ratio (Pout / Pin) improved from 0.86 to 0.93, energy utilization increased by 8.1%, and the number of abnormal alarms decreased from 122 per month to 71, a reduction of 41.8%, indicating that the system is more robust and reliable in terms of operation scheduling and safety control. The mean model prediction bias decreased from 3.8% to 1.4%, and the confidence level of the self-learning model increased from 0.88 to 0.94, demonstrating that the combination of hierarchical graph temporal decision network and feedback self-learning mechanism enables the system to continuously optimize prediction accuracy and control strategy, achieving self-correction and long-term stable evolution. The method of this invention has significant intelligent, stable, and energy-saving effects in unattended power distribution scenarios, providing a feasible engineering path for the intelligent upgrading of future urban power distribution systems.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An unattended power distribution house intensification power distribution method based on machine learning, characterized in that, The method comprises the following steps: Collecting operation data of a plurality of unattended power distribution rooms, pre-processing the operation data to form a standardized input data set; Based on the standardized input data set, the association between the three types of features of topology, working condition and equipment is constructed, a feature vector containing electrical parameters, thermal parameters and health degree is established for each power distribution node, and electrical impedance, thermal coupling coefficient and frequency domain association features are set for each connection relationship. The association weight is dynamically updated according to the energy flow and operation stress calculation results between nodes to form a three-dimensional coupled feature structure; Input the three-dimensional coupled feature structure into the hierarchical graph time sequence decision network model, jointly learn the topology relationship features and time sequence load features, fuse multi-dimensional feature information through cross attention mechanism, generate a joint feature vector representing the group operation state of the unattended power distribution room, and output a candidate control action; According to the joint feature vector, a dynamic feasible region is constructed, and the candidate control action is projected into the dynamic feasible region to obtain a safe action set that meets the physical and electrical safety constraints; The safe action set is input into the topology reversible neural differential control module to predict the transient response of the execution process of the safe action, calculate the current inrush and voltage sag parameters, and adaptively optimize the execution sequence and time window to generate an execution action after transient optimization; According to the execution action, control instructions are sent to the power distribution terminal equipment to perform power scheduling, switch switching and reactive power compensation operations, collect execution feedback data, update the three-dimensional coupled feature structure and the dynamic feasible region, and trigger incremental update of the hierarchical graph time sequence decision network model when the predicted result deviates from the actual feedback. 2.The method of claim 1, wherein, The operation data includes voltage, current, active power, reactive power, temperature, humidity, device vibration signal, partial discharge signal and switch state information. 3.The method of claim 1, wherein, The pre-processing of the operation data includes time synchronization, outlier removal, noise filtering and normalization of the operation data to eliminate data drift and sampling error. 4.The method of claim 1, wherein, The three-dimensional coupled feature structure is formed by: Establishing data objects of power distribution nodes and connection relationships, organizing the standardized input data set in time sequence, determining the node record and connection relationship record corresponding to each time point, and assigning data quality marks to each record; Generating a node feature vector containing electrical parameters, thermal parameters and health degree for each power distribution node; Generating a connection relationship feature vector containing electrical impedance, thermal coupling coefficient and frequency domain association features for each connection relationship, the electrical impedance includes resistance and reactance, and the frequency domain association features form a fixed-dimensional frequency band intensity sequence according to harmonic amplitude-frequency information; Calculate the node operation stress and energy flow, the node operation stress is represented by a multi-index set composed of load rate, temperature rise margin, health degree, switch action frequency and total harmonic distortion, the energy flow is represented by a multi-index set composed of active power flow, reactive power flow, directionality and harmonic power flow intensity, and an event influence factor representing the influence of recent disturbance and a potential level representing the comprehensive risk level of nodes and connection relationships are generated simultaneously; The associated weight is dynamically updated according to the node running stress and energy flow: A dual-channel mechanism of slow-channel sliding update and fast-channel event-triggered update is adopted; In the sliding window, the weight of the connection relationship with low stress, low harmonic and stable directionality is increased, and the weight of the connection relationship with high stress, high harmonic, repeated or close to the capacity upper limit of directionality is reduced; When the event influence factor exceeds the set threshold, the fast adjustment is performed in the current time period, the freeze window and the unfreezing rule are set, the contraction-expansion criterion is executed according to the potential energy level, the weight fluctuation range is limited, and the three-dimensional coupling feature structure including the node feature vector, the connection relationship feature vector and the dynamically updated associated weight is output. 5.The method of claim 1, wherein, The output candidate control action includes: A hierarchical graph time sequence decision network model is constructed, and the hierarchical graph time sequence decision network model is composed of a topology-health dual-channel graph coding module, an event enhanced time sequence aggregation module and a constraint perception strategy generation module; In the topology-health dual-channel graph coding module, the topology relationship and the electrical parameter in the three-dimensional coupling feature structure are respectively taken as the topology channel input and the health channel input, the heat parameter, the health degree and the frequency domain associated feature are taken as the health degree channel input, the channel feature extraction and cross-channel alignment are performed, the edge selection and directionality enhancement are performed based on the potential energy level and the weight credibility score, and the node embedding, the connection relationship embedding and the edge importance score are output; In the event enhanced time sequence aggregation module, the node embedding and the connection relationship embedding are aggregated in time sequence by using a fixed length sliding window, a short-term disturbance branch and a medium-term trend branch are set, and the short-term disturbance branch and the medium-term trend branch are set. The short-term disturbance branch adopts a fixed length short-term sliding window and an event triggered gate to process high frequency changes, performs peak value preservation, edge importance rapid reweighting and impulse response extraction, and outputs a disturbance representation including surge intensity, temporary drop depth and mutation duration; The medium-term trend branch adopts a fixed length medium-term sliding window and a period coding to process smooth changes, performs denoising smoothing, period pattern extraction and drift estimation, and outputs a trend representation including load baseline, temperature rise baseline and harmonic background; The outputs of the two branches are aligned by time stamp, weighted aggregated, and the causal mask and rollback placeholder are retained to obtain a joint time sequence representation; In the constraint perception strategy generation module, constraint prior hints are generated according to the device nameplate parameters, the connection capacity upper limit, the frequency band limit, the edge importance score and the potential energy level, the soft mask and the hard mask are applied to the strategy space to impose a double-stack constraint, the incombustible action set is limited, and the candidate control action including the power scheduling parameter, the reactive compensation parameter and the switch switching sequence is output. 6.The method of claim 1, wherein, The safe action set satisfying the physical and electrical safety constraints is obtained, including: A basic constraint list including voltage upper and lower limits, current upper limit, temperature rise upper limit, harmonic limit, connection capacity, switch minimum interval and sequence is established, and a constraint component library is generated. The dynamic feasible region is constructed based on a constraint component library, a layered deformation mechanism is used to combine the basic boundary, the context boundary and the transient placeholder boundary, a boundary deformation memory and a hysteresis mechanism are introduced in the construction process, a recovery threshold and a frozen time window are set for the boundary that continuously shrinks or expands in a short time, and a source label and an effective period label are attached to each boundary; The dynamic feasible region is executed online self-proving and local detection, the node and the connection relationship close to the boundary are generated for the edge importance score exceeding the preset threshold, a small amplitude exploration action sequence is generated, a joint time sequence representation and a potential energy level are used to quickly evaluate the exploration results, the active boundary index and the boundary credibility level are updated, the boundary that does not meet the consistency check is executed in the current period, and the source label and the causal link of the adjustment are recorded in the constraint component library; The candidate control action is executed in the dynamic feasible region Two-stage compliance processing: The first stage is to continuously domain-compliant the power scheduling parameters and the reactive power compensation parameters according to the superposition results of the basic boundary, the context boundary and the transient placeholder boundary, and to form a parameter set with clear safety margin; The second stage is to discrete sequence-compliant the switch switching sequence according to the minimum time interval, the interlock and exclusion rule, the order, the maintenance state and the key equipment list, to eliminate the actions conflicting with the transient placeholder and to fill in the necessary buffer actions, to generate a safe action set and metadata. 7.The method of claim 1, wherein, The generation of the execution action after transient optimization includes: Receiving the safe action set, the metadata, the joint time sequence representation and the rollback placeholder label, initializing the topologically reversible neural differential control module, the topologically reversible neural differential control module is composed of a topologically reversible mapping unit, a transient coupling prediction unit and a sequence adaptive and rollback control unit; In the topologically reversible mapping unit, according to the three-dimensional coupling feature structure and the dynamic feasible region, the safe action set is mapped into a topological state sequence corresponding to the consistent forward execution and reverse rollback one by one, and a reversible mapping table containing node state, connection relationship state and energy conservation constraint label is generated; In the transient coupling prediction unit, the transient response prediction of electromagnetic-thermal multi-field coupling is performed on each safe action in the reversible mapping table, forming an action level evaluation record for switch operation, tie switching and reactive power compensation, and according to the active boundary index and the transient placeholder margin, the boundary occupation and safety margin of each index are marked; In the sequence adaptive and rollback control unit, based on the action level evaluation record, the safe action set is executed joint search of order neighborhood exchange and time displacement, and the compliance check is performed according to the minimum time interval, the order, the interlock and exclusion rule, the maintenance state and the key equipment list, and the actions conflicting with the transient placeholder are eliminated or replaced; The output of the execution action after transient optimization includes the switch switching sequence, the execution time window, the power scheduling parameter and the reactive power compensation parameter. 8.The method of claim 1, wherein, The hierarchical graph time sequence decision network model is updated incrementally when the prediction result deviates from the actual feedback, including: The control instructions are sent to the power distribution terminal according to the execution actions, power scheduling, switch switching and reactive power compensation operations are performed, and feedback data of node voltage, current, temperature, power factor, frequency deviation and energy flow are collected in real time; The feedback data are compared with the prediction results before execution, the node level and the connection level are calculated, the three-dimensional coupled feature structure and the dynamic feasible region are updated according to the deviation results, the node features, the connection relationship and the associated weight are adjusted, the basic boundary, the context boundary and the transient occupation boundary are reconstructed, the active constraint set, the constraint margin table and the safety confidence score are updated; When the deviation exceeds the preset threshold, the incremental update of the hierarchical graph time sequence decision network model is triggered, the local parameter adjustment of the topology-health dual-channel graph encoding module, the event-enhanced time sequence aggregation module and the constraint-aware strategy generation module is performed, and the version and confidence level of the hierarchical graph time sequence decision network model are updated.

9. The unattended power distribution house intensification power distribution system based on machine learning according to any one of claims 1 to 8, wherein the unattended power distribution house intensification power distribution method based on machine learning according to any one of claims 1 to 8 is executed. The method comprises the following modules: A data acquisition and preprocessing module is configured to collect operation data and perform preprocessing to generate a standardized input data set; A feature construction and correlation analysis module is configured to construct the correlation between topology, working condition and equipment features based on the standardized input data set, and generate a three-dimensional coupled feature structure; A hierarchical graph time sequence decision network module is configured to receive the three-dimensional coupled feature structure, fuse the topology and time sequence load features, and output candidate control actions of power scheduling, reactive power compensation and switch switching; A dynamic feasible region construction module is configured to construct a dynamic feasible region, project the candidate control actions, and output a safe action set meeting the physical and electrical constraints; A reversible neural differential control module is configured to perform transient response prediction and optimization on the safe actions, calculate surge, voltage sag and energy disturbance parameters, and generate optimized execution actions; An execution and self-learning module is configured to send control instructions to the power distribution terminal according to the execution actions, collect feedback data and update the three-dimensional coupled feature structure and the dynamic feasible region.

Images