Machine learning-based intensive power distribution system and method for unattended power distribution room

Abstract

The invention discloses an unattended power distribution room intensive power distribution system and method based on machine learning, and the method comprises the steps: collecting the operation data of an unattended power distribution room, executing the preprocessing, and generating a standardized input data set; constructing a topology, working condition and equipment feature association relationship to form a three-dimensional coupling feature structure; inputting a hierarchical graph time sequence decision network, and fusing topology and time sequence characteristics to output candidate control actions; constructing a dynamic feasible region according to the joint features, and screening a safety action set meeting constraints; inputting into a topological reversible neural differential control module, predicting transient response and optimizing an execution action; and issuing a control instruction, collecting feedback, and triggering model increment update to realize closed-loop optimization. According to the invention, by introducing a hierarchical graph time sequence decision network and a topological reversible neural differential control mechanism, intelligent intensive scheduling and self-learning closed-loop optimization control of the unattended power distribution room are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution automation and smart power system, and particularly relates to an unattended power distribution room intensive power distribution system and method based on machine learning. BACKGROUND

[0002] With the continuous expansion of the power distribution network scale and the continuous growth of urban power load, the traditional power distribution room gradually develops towards unattended and centralized monitoring to reduce operation and maintenance costs and improve operation efficiency. The current unattended power distribution room mainly relies on fixed threshold monitoring and remote control system to collect voltage, current, temperature and other data for state monitoring and alarm response. However, such systems generally lack overall correlation analysis and intelligent coordination capabilities for multiple power distribution rooms, and cannot realize real-time optimal control under complex operating environments such as multi-station linkage, dynamic load change and equipment aging, which easily leads to local overload, uneven power distribution and increased energy loss.

[0003] The existing power distribution automation system has introduced a certain degree of algorithm optimization, such as scheduling decision based on rule engine or static model, but it lacks time sequence dependence and topology change sensitivity identification of operating data, and cannot effectively respond to abnormal disturbances and equipment health state changes. The existing system mainly focuses on single station optimization, lacks intensive scheduling mechanism across power distribution rooms, and leads to system scheduling lag, untimely response, and even transient impact problems such as voltage sag and current surge during multi-station cooperative switching, affecting power grid operation stability and safety.

[0004] Therefore, how to provide an unattended power distribution room intensive power distribution system and method based on machine learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an unattended power distribution room intensive power distribution system and method based on machine learning. The unattended power distribution room intensive power distribution method based on machine learning according to the embodiments of the present application comprises: Collecting operating data of multiple unattended power distribution rooms, pre-processing the operating data to form a standardized input data set; Based on the standardized input data set, the correlation 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, electrical impedance, thermal coupling coefficient and frequency domain correlation features are set for each connection relationship, and the correlation weights are dynamically updated according to the energy flow and operating stress calculation results between nodes to form a three-dimensional coupled feature structure; The three-dimensional coupling feature structure is input to a hierarchical graph time sequence decision network model, topological relationship features and time sequence load features are jointly learned, multi-dimensional feature information is fused through a cross attention mechanism, a joint feature vector representing the group operation state of the unattended power distribution room is generated, and a candidate control action is output; According to the joint feature vector, a dynamic feasible region is constructed, the candidate control action is subjected to dynamic feasible region projection, the candidate control action is mapped into the dynamic feasible region, and a safe action set satisfying physical and electrical safety constraints is obtained; The safe action set is input to a topological reversible neural differential control module, the execution process of the safe action is subjected to transient response prediction, current inrush and voltage sag parameters are calculated, and the execution sequence and time window are adaptively optimized, and an execution action subjected to transient optimization is generated; According to the execution action, control instructions are sent to the power distribution terminal device, power scheduling, switch switching and reactive power compensation operations are performed, execution feedback data is collected, the three-dimensional coupling feature structure and the dynamic feasible region are updated, and when there is a deviation between the prediction result and the actual feedback, the hierarchical graph time sequence decision network model is triggered for incremental update.

[0006] Optionally, the operation data includes voltage, current, active power, reactive power, temperature, humidity, device vibration signal, partial discharge signal and switch state information.

[0007] Optionally, the pre-processing of the operation data includes time synchronization, outlier elimination, noise filtering and normalization processing of the operation data, and the data drift and sampling error are eliminated.

[0008] Optionally, the three-dimensional coupling feature structure comprises: A data object of a power distribution node and a connection relationship is established, a standardized input data set is organized in time sequence, node records and connection relationship records corresponding to each time point are determined, and a data quality mark is assigned to each record; A node feature vector containing electrical parameters, thermal parameters and health degree is generated for each power distribution node; A connection relationship feature vector containing electrical impedance, thermal coupling coefficient and frequency domain correlation feature is generated for each connection relationship, the electrical impedance includes resistance and reactance, and the frequency domain correlation feature forms a fixed-dimensional frequency band intensity sequence according to harmonic amplitude-frequency information; Node operation stress and energy flow are calculated, 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, an event influence factor representing the influence of recent disturbance and a potential level representing the comprehensive risk level of the node and the connection relationship are simultaneously generated. 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; The connection relationship of low stress, low harmonic, and stable directionality is increased in weight, and the connection relationship of high stress, high harmonic, and repeated or close-to-capacity upper limit directionality is decreased in weight in the sliding window; When the event influence factor exceeds the set threshold, a quick adjustment is performed in the current time period, a freeze window and a thaw rule are set, a contraction-expansion criterion is executed according to the potential energy level, the weight fluctuation range is limited, and a three-dimensional coupling feature structure including a node feature vector, a connection relationship feature vector, and an associated weight after dynamic update is output.

[0009] Optionally, the output candidate control action comprises: A hierarchical graph time sequence decision network model is constructed, which comprises a topology-health dual-channel graph encoding module, an event-enhanced time sequence aggregation module, and a constraint-aware strategy generation module; In the topology-health dual-channel graph encoding module, the topology relationship and electrical parameter in the three-dimensional coupling feature structure are respectively taken as the topology channel input and the health channel input, the thermal parameter and the health degree and frequency domain association 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, a fixed-length sliding window is used to perform time sequence aggregation on the node embedding and the connection relationship embedding, and a short-term disturbance branch and a medium-term trend branch are set, wherein: The short-term disturbance branch uses a fixed-length short-term sliding window and event-triggered gating to process high-frequency changes, performs peak value retention, 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 uses a fixed-length medium-term sliding window and period encoding 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 timestamp, then weighted aggregated, and the causal mask and rollback placeholder are retained to obtain a joint time sequence representation; In the constraint-aware strategy generation module, constraint prior hints are generated according to device nameplate parameters, connection capacity upper limit, frequency band limit, edge importance score, and potential energy level, soft and hard masks are applied to the strategy space to impose dual-stack constraints, and the set of incombustible actions is limited, and candidate control actions including power scheduling parameters, reactive compensation parameters, and switch switching sequences are output.

[0010] Optionally, the obtaining of the safe action set satisfying the physical and electrical safety constraints comprises: establishing a basic constraint list containing voltage upper and lower limits, current upper limit, temperature rise upper limit, harmonic limit, connection capacity, switch minimum interval and sequence, and generating a constraint component library; constructing a dynamic feasible region based on the constraint component library, combining the basic boundary, the context boundary and the transient placeholder boundary by using a hierarchical deformation mechanism, introducing a boundary deformation memory and a hysteresis mechanism in the construction process, setting a recovery threshold and a frozen time window for the boundary that continuously contracts or expands in a short time, and adding a source tag and an effective period tag to each boundary; performing online self-certification and local detection on the dynamic feasible region, generating a small amplitude exploration action sequence for the node and connection relationship close to the boundary with a boundary importance score exceeding a preset threshold, using joint time sequence representation and potential energy level to quickly evaluate the exploration results, updating the active boundary index and boundary credibility level, and performing degradation or rejection on the boundary that does not meet the consistency check in the current period, and recording the source tag and causal link of the adjustment in the constraint component library; performing two-stage compliance processing on the candidate control action in the dynamic feasible region: the first stage performs continuous domain compliance on power scheduling parameters and reactive power compensation parameters according to the superposition results of the basic boundary, the context boundary and the transient placeholder boundary, forming a parameter set with clear safety margin; the second stage performs discrete sequence compliance on the switch switching sequence according to the minimum time interval, the mutual exclusion and mutual exclusion rule, the sequence, the maintenance state and the key equipment list, eliminates the actions conflicting with the transient placeholder, and fills in the necessary buffer actions, to generate a safe action set and metadata.

[0011] Optionally, the generating of the execution action after transient optimization comprises: receiving the safe action set, the metadata, the joint time sequence representation and the rollback placeholder mark, initializing a topologically reversible neural differential control module, the topologically reversible neural differential control module being 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, the safe action set is mapped into a topological state sequence corresponding to the consistent forward execution and reverse rollback according to the three-dimensional coupling feature structure and the dynamic feasible region, and a reversible mapping table containing node state, connection relationship state and energy conservation constraint tag is generated; In the transient coupling prediction unit, the transient response prediction of electromagnetic-thermal multi-field coupling is performed on each item of the reversible mapping table to form an action level evaluation record for switch operation, tie switching and reactive compensation, and the boundary occupation and safety margin of each index are marked according to the active boundary index and the transient occupation margin; In the sequence adaptation and rollback control unit, based on the action level evaluation record, the joint search of sequence neighborhood exchange and time displacement is performed on the safety action set, the compliance check is performed according to the minimum time interval, the order, the interlocking and mutual exclusion rules, the maintenance state and the key equipment list, and the actions in conflict with the transient occupation are eliminated or replaced; The execution action after transient optimization is output, and the execution action includes switch switching sequence, execution time window, power scheduling parameter and reactive compensation parameter.

[0012] Optionally, the triggering of the hierarchical graph timing decision network model incremental update when the prediction result deviates from the actual feedback comprises: Control instructions are sent to the power distribution terminal device according to the execution action, power scheduling, switch switching and reactive 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 and the prediction result before execution are compared, the node level and the connection level are calculated, the three-dimensional coupling feature structure and the dynamic feasible region are updated according to the deviation result, 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 timing decision network model is triggered, the local parameter adjustment of the topology-health dual-channel graph encoding module, the event-enhanced timing aggregation module and the constraint-aware strategy generation module is performed, and the version and confidence level of the hierarchical graph timing decision network model are updated.

[0013] The unmanned power distribution room intensive power distribution system based on machine learning according to the embodiment of the application comprises the following modules: The data acquisition and preprocessing module is used for collecting operation data and performing preprocessing to generate a standardized input data set; The feature construction and correlation analysis module is used for constructing the correlation relationship of topology, working condition and equipment features based on the standardized input data set to generate a three-dimensional coupling feature structure; The hierarchical graph timing decision network module is used for receiving the three-dimensional coupling feature structure, fusing topology and timing load features, and outputting candidate control actions of power scheduling, reactive compensation and switch switching; The dynamic feasible region construction module is used for constructing a dynamic feasible region, projecting the candidate control actions, and screening and outputting a safety action set meeting physical and electrical constraints; Reversible neural differential control module for transient response prediction and optimization of safety actions, calculating surge, voltage sag and energy disturbance parameters, generating optimized execution actions; Execution and self-learning module for sending control instructions to power distribution terminals according to execution actions, collecting feedback data and updating three-dimensional coupled feature structure and dynamic feasible region.

[0014] The beneficial effects of the present application are: The present application realizes adaptive scheduling and fine control of unattended power distribution rooms under complex power operating conditions by introducing hierarchical graph time sequence decision network and topological reversible neural differential control mechanism. The system can form a unified data correlation model in multiple power distribution rooms, automatically identify topology change and load migration characteristics, dynamically adjust power distribution strategy, and realize intensive operation and management of cross-station collaboration. Compared with traditional control methods based on fixed rules or static thresholds, the present application can maintain the stability and safety of power grid operation in dynamic scenes such as load fluctuation, temperature rise change and harmonic interference, and improve the real-time response ability and operation efficiency of the power distribution system.

[0015] The present application builds a dynamic feasible region and a safety constraint double-layer mechanism, so that the control instructions are verified in real time before being issued, avoiding invalid scheduling, out-of-limit operation or transient impact problems that may occur in traditional systems. The topological reversible neural differential control module can perform transient prediction and optimization of voltage sag, current surge and energy disturbance in the control execution phase, ensuring energy conservation and safe operation of equipment during scheduling, not only improving the anti-disturbance ability of the system in complex switching scenarios, but also enabling unattended power distribution rooms to complete adaptive regulation and risk prevention without relying on manual intervention.

[0016] The present application introduces a feedback-driven incremental self-learning mechanism, which can automatically trigger hierarchical graph time sequence decision network model updating when there is a deviation between the prediction result and the actual feedback, continuously optimizing system parameters and control strategies, effectively overcoming the limitations of traditional algorithm models being static and not having continuous evolution ability, so that the system can maintain high precision and high reliability for a long time. By combining the closed-loop system of data self-sensing, model self-evolution and control self-correction, the present application realizes intelligent, dynamic and adaptive management of unattended power distribution rooms, significantly improving the safety, flexibility and overall operation intelligence level of the power distribution system. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1The flow chart of the unattended power distribution room intensive power distribution method based on machine learning proposed in the present application; Figure 2 The structural schematic diagram of the unattended power distribution room intensive power distribution system based on machine learning proposed in the present application. DETAILED DESCRIPTION

[0018] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only schematically illustrate the basic structure of the present application, and thus only show the components related to the present application.

[0019] REFERENCE Figure 1 The unattended power distribution room intensive power distribution method based on machine learning, comprising: 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, electrical impedance, thermal coupling coefficient and frequency domain association features are set for each connection relationship, and 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; The three-dimensional coupled feature structure is input into a hierarchical graph time sequence decision network model, the topology relationship features and time sequence load features are jointly learned, the multi-dimensional feature information is fused through a cross-attention mechanism, a joint feature vector representing the group operation state of the unattended power distribution room is generated, and a candidate control action is output; 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 satisfies the physical and electrical safety constraints; The safe action set is input into a topology reversible neural differential control module, the execution process of the safe action is transient response predicted, the current inrush and voltage sag parameters are calculated, and the execution sequence and time window are adaptively optimized 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, execution feedback data is collected, the three-dimensional coupled feature structure and the dynamic feasible region are updated, and the hierarchical graph time sequence decision network model is triggered for incremental update when there is a deviation between the prediction result and the actual feedback.

[0020] In this embodiment, the operation data includes voltage, current, active power, reactive power, temperature, humidity, equipment vibration signal, partial discharge signal and switch state information.

[0021] In the embodiment, the preprocessing of the operation data includes time synchronization, outlier elimination, noise filtering and normalization processing, and eliminates data drift and sampling error.

[0022] In the embodiment, the forming of the three-dimensional coupling feature structure includes: A data object of the power distribution node and the connection relationship is established, the standardized input data set is organized in time sequence, the node record and the connection relationship record corresponding to each time point are determined, and a data quality mark is given to each record; A node feature vector including electrical parameters, thermal parameters and health degree is generated for each power distribution node, the electrical parameters include voltage, current, active power and reactive power, the thermal parameters include temperature and temperature rise margin, and the health degree includes equipment degradation degree and maintenance state identification; A connection relationship feature vector including electrical impedance, thermal coupling coefficient and frequency domain correlation feature is generated for each connection relationship, the electrical impedance includes resistance and reactance, and the frequency domain correlation feature forms a fixed-dimension frequency band intensity sequence according to harmonic amplitude-frequency information; Node operation stress and energy flow are calculated, the node operation stress is represented by a multi-index set composed of load rate, temperature rise margin, health degree, switching 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, an event influence factor for representing the influence of recent disturbance and a potential level for representing the comprehensive risk level of the node and the connection relationship are simultaneously generated; The associated weight is dynamically updated according to the node operation stress and the energy flow: A double-channel mechanism of slow channel sliding update and fast channel event triggered update is adopted; The weight of the connection relationship with low stress, low harmonic and stable directionality is increased in the sliding window, and the weight of the connection relationship with high stress, high harmonic, repeatedly changing direction or close to the capacity upper limit is decreased; When the event influence factor exceeds the set threshold, a quick adjustment is performed in the current time period, a freeze window and a thawing rule are set, a contraction-expansion criterion is executed according to the potential level, the weight fluctuation range is limited, a three-dimensional coupling feature structure including the node feature vector, the connection relationship feature vector and the dynamically updated associated weight is output, and the thawing rule is that when the event influence factor is lower than the lower limit of the threshold and remains stable in the continuous monitoring period, the corresponding freeze window constraint is released, the dynamic adjustment ability of the node and the connection relationship is restored by time decay weight; wherein the recovery rate is adaptively set according to the node potential level and the associated edge importance, the high potential node is preferentially thawed, and the low potential node is delayed thawed, so as to avoid frequent fluctuations; when the global potential gradient tends to be stable and the associated weight change rate is lower than the preset convergence threshold, the frozen state is terminated and the regular dynamic update is restored.

[0023] In this embodiment, the output candidate control action includes: The hierarchical graph time series decision network model is constructed, and the hierarchical graph time series decision network model is composed of a topology-health dual-channel graph encoding module, an event-enhanced time series aggregation module, and a constraint-aware strategy generation module; In the topology-health dual-channel graph encoding module, the topological relationship and electrical parameters in the three-dimensional coupled feature structure are respectively taken as the topology channel input, the thermal parameters and the health degree and the frequency domain correlation features 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 level and weight credibility score, and the node embedding, connection relationship embedding and edge importance score are outputted; In the event-enhanced time series aggregation module, a fixed-length sliding window is used to perform time series aggregation on the node embedding and the connection relationship embedding, and a short-term disturbance branch and a medium-term trend branch are set, wherein: The short-term disturbance branch uses a fixed-length short-term sliding window and an event trigger gate to process high-frequency changes, performs peak value preservation, edge importance rapid reweighting, and impulse response extraction, and outputs a disturbance representation containing surge intensity, sag depth, and mutation duration, wherein the short-term sliding window is set to 5 consecutive sampling periods, corresponding to a time span of 10 seconds; The medium-term trend branch uses a fixed-length medium-term sliding window and a period encoding to process smooth changes, performs denoising smoothing, cycle pattern extraction, and drift estimation, and outputs a trend representation containing a load baseline, a temperature rise baseline, and a harmonic background, wherein the medium-term sliding window is set to 60 consecutive sampling periods, corresponding to a time span of 120 seconds, and the denoising smoothing, cycle pattern extraction, and drift estimation are performed as follows: The medium-term trend data is denoised and smoothed using a moving average and an adaptive weighted filtering method to suppress random disturbances and short-term fluctuations while retaining the main trend; The cycle pattern features of the load, voltage, and temperature are extracted based on a cycle decomposition and a spectrum clustering method to identify the main cycle components and phase relationships, forming a cycle pattern vector; The long-term drift amount is calculated using a time series residual regression and a shift detection mechanism to evaluate the trend shift direction and amplitude, and dynamically correct the baseline reference to generate stable load baseline, temperature rise baseline, and harmonic background trend representation; The two branches are aligned by timestamp, weighted and aggregated, and the causal mask and rollback placeholder are retained to obtain a joint time series representation; In the constraint-aware strategy generation module, constraint priori hints are generated according to device nameplate parameters, connection capacity upper limit, frequency band limit, edge importance score and potential level, soft mask and hard mask double-stack constraints are applied to the strategy space to limit the set of incombustible actions, and candidate control actions containing power scheduling parameters, reactive compensation parameters and switch switching sequences are output. The soft mask dynamically probabilistically constrains the actionable actions in the strategy space, and the sampling priority of different actions is adjusted by introducing a continuous weight coefficient, so that the actions maintain exploration flexibility under the premise of meeting safety and economic requirements. The hard mask is used to strictly shield action combinations that do not meet physical constraints or safety boundaries, and directly sets actions that exceed device rated capacity, violate electrical interlocking rules or break frequency band limits to be unselectable. The execution order of the double-stack constraint is to first apply the hard mask to exclude non-compliant actions, and then apply the soft mask to weight and select the remaining actions, thereby achieving hierarchical pruning and safety guidance of the strategy space.

[0024] In this embodiment, the safe action set satisfying the physical and electrical safety constraints includes: A basic constraint list containing 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; A dynamic feasible region is constructed based on the constraint component library, a hierarchical deformation mechanism is used to combine the basic boundary, the context boundary and the transient placeholder boundary, boundary deformation memory and hysteresis mechanism are introduced during construction, a recovery threshold and a frozen time window are set for boundaries that continuously contract or expand within a short time, and a source tag and an effective period tag are attached to each boundary, wherein the basic boundary is determined by device nameplate parameters and operating procedures, the context boundary is contracted or expanded according to potential level, health degree and edge importance score, and the transient placeholder boundary reserves an execution window and a capacity margin according to rollback placeholder markers and recent event intensity; On-line self-certification and local detection are performed on the dynamic feasible region, small-amplitude exploratory action sequences are generated for nodes and connection relationships with edge importance scores exceeding a preset threshold and approaching the boundary, joint time sequence representation and potential level are used to quickly evaluate the exploratory results, active boundary index and boundary credibility level are updated, boundaries that do not meet consistency checking are degraded or excluded in the current period, and the source tag and causal link of the adjustment are recorded in the constraint component library, wherein the preset threshold is set to 0.7 and the edge importance score takes a value in the range of 0 to 1; Two-stage compliance processing is performed on the candidate control actions in the dynamic feasible region: The first stage performs continuous domain compliance on the power scheduling parameters and the reactive compensation parameters according to the superposition results of the basic boundary, the context boundary and the transient occupation boundary, to form a parameter set with a clear safety margin; The second stage performs discrete sequence compliance on the switch switching sequence according to the minimum time interval, the interlocking and mutual exclusion rules, the sequence order, the maintenance state and the key equipment list, eliminates the actions conflicting with the transient occupation, and supplements the necessary buffer actions, to generate a safe action set and metadata.

[0025] In the embodiment, the generated execution action after transient optimization comprises: Receiving the safe action set and metadata, jointly representing the time sequence and the rollback occupation mark, initializing the topologically reversible neural differential control module, which 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 item by item into a topological state sequence corresponding to the forward execution and the reverse rollback, to generate a reversible mapping table containing node state, connection relationship state and energy conservation constraint label; In the transient coupling prediction unit, the transient response prediction of electromagnetic-thermal multi-field coupling is performed on each item of the safe action in the reversible mapping table, to form an action level evaluation record for switch operation, tie switching and reactive compensation, and according to the active boundary index and the transient occupation margin, the boundary occupation and safety margin of each index are marked, and the transient response prediction of electromagnetic-thermal multi-field coupling is performed on each item of the safe action in the reversible mapping table, specifically: Before execution, the voltage, current and temperature change are calculated synchronously according to the node electrical parameters, device thermal parameters and connection relationship involved in the safe action, and the electromagnetic disturbance and thermal response process at the action triggering moment are analyzed; For different types of actions such as switch operation, tie switching and reactive compensation, the current rise rate, voltage drop amplitude, energy release intensity and short-time temperature rise value are calculated, and the high-risk period and high-stress part are identified; The electromagnetic response and thermal change results are integrated in time sequence, the key indicators are extracted and compared with the active boundary index and the transient occupation margin, the boundary occupation and safety margin of each index are marked, and the action level transient evaluation results for scheduling optimization are formed; In the sequence adaptive and rollback control unit, based on the action level evaluation record, the sequence neighborhood exchange and time displacement joint search are performed on the safe action set, and the compliance check is performed according to the minimum time interval, the sequence order, the interlocking and mutual exclusion rules, the maintenance state and the key equipment list, and the actions conflicting with the transient occupation are eliminated or replaced; The outputted execution action after transient optimization includes switch switching sequence, execution time window, power scheduling parameter and reactive compensation parameter.

[0026] In the embodiment, the triggering of the hierarchical graph timing decision network model incremental update when the prediction result deviates from the actual feedback comprises: According to the execution action, control instructions are sent to the power distribution terminal device to perform power scheduling, switch switching and reactive compensation operations, and feedback data of node voltage, current, temperature, power factor, frequency deviation and energy flow are collected in real time; The feedback data is compared with the prediction result before execution, and 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 result, 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, and 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 timing decision network model is triggered, the local parameter adjustment of the topology-health dual-channel graph encoding module, the event-enhanced timing aggregation module and the constraint-aware strategy generation module is performed, and the version and confidence level of the hierarchical graph timing decision network model are updated.

[0027] Reference Figure 2 , the unmanned power distribution room intensive power distribution system based on machine learning includes the following modules: The data acquisition and preprocessing module is used for collecting operation data and performing preprocessing to generate a standardized input data set; The feature construction and correlation analysis module is used for constructing the correlation between topology, working condition and device features based on the standardized input data set, and generating a three-dimensional coupled feature structure; The hierarchical graph timing decision network module is used for receiving the three-dimensional coupled feature structure, fusing topology and timing load features, and outputting candidate control actions of power scheduling, reactive compensation and switch switching; The dynamic feasible region construction module is used for constructing a dynamic feasible region, projecting the candidate control actions, and selecting and outputting a safe action set that meets the physical and electrical constraints; The reversible neural differential control module is used for transient response prediction and optimization of the safe action, calculation of inrush current, voltage sag and energy disturbance parameters, and generation of optimized execution action; The execution and self-learning module is used for sending control instructions to the power distribution terminal according to the execution action, collecting feedback data and updating the three-dimensional coupled feature structure and the dynamic feasible region. Embodiment 1:

[0028] To verify the feasibility of the application in practice, the application is applied to three unattended distribution rooms under a certain energy dispatching center, each of which is composed of a 35kV main transformer, a 10kV bus and multiple outgoing circuit loops, and serves residential areas, industrial parks and commercial areas respectively. Due to the significant difference in power consumption characteristics, the traditional system has obvious problems during the summer high-temperature load period: A station often has line current over-limit, B station is in low-load operation state for a long time, and C station is prone to voltage sag during switching operation. The dispatching relies on fixed threshold alarms and manual decision-making, lacks multi-station collaborative optimization mechanism, resulting in local overload, low energy utilization and high operation risk.

[0029] In this scenario, the machine learning-based unattended distribution room intensive power distribution system proposed by the application is deployed. The data acquisition and preprocessing module collects voltage, current, active power, temperature and equipment vibration signals of a total of 102 monitoring points in A, B and C stations in real time, synchronized once every 5 seconds. After time alignment, abnormality elimination and normalization, the system generates a standardized input data set. Then, the feature construction and correlation analysis module extracts topology, working condition and equipment health features based on the data set to form a three-dimensional coupled feature structure. In actual operation, the system dynamically identifies the high-temperature node of the main transformer in A station (operating temperature 81.6°C) and the redundant low-load node in B station (load rate 42.3%), automatically adjusts the correlation weight, and realizes real-time modeling of energy flow.

[0030] The hierarchical graph time series decision network module is trained based on 30 days of historical load data, uses topology-health dual-channel graph encoding, event-enhanced time series aggregation and constraint-aware strategy generation mechanism to generate joint feature vectors, and automatically outputs candidate control actions. Taking August 15, 2025, 14:30 as an example, when the system monitors that the load of A station reaches 94% of the rated capacity, the model predicts that the power transfer path is "A station to B station-C station", and outputs the dispatching instruction to reduce the load of A station to 79%, while making B station rise from 47% to 65%, and the voltage fluctuation remains within ±1.8%.

[0031] The dynamic feasible region construction module monitors the constraint parameters such as current, voltage and temperature rise in real time. When the temperature of the reactive power compensation device in C station reaches 80°C, the system automatically shrinks its dynamic feasible region and adjusts the safety action set, so that the power switching process remains within the safety threshold. The topology reversible neural differential control module further predicts and optimizes the transient response of the safety action execution, predicting the current inrush current to be 210A before the dispatching action is executed, and the actual execution is monitored to be 203A, with a deviation of less than 3.5%.

[0032] The self-learning module automatically issues operation commands to each terminal switch device according to 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 incremental updating of the model. For example, on August 20, 2025, a temperature rise prediction deviation of 4.2% was detected at B station, and the system automatically adjusted the health degree channel weight, reducing the deviation to 1.6% after updating, and the model confidence increased from 0.88 to 0.94.

[0033] After 60 days of continuous operation test, the system realizes real-time power balance and risk prevention of multi-station cooperation, significantly shortens the scheduling response time, reduces the device operating temperature, and significantly improves the energy utilization rate.

[0034] Table 1 Comparison of intensive scheduling operation effects of unattended power distribution room As can be seen from the data in Table 1, the unattended power distribution room intensive power distribution method based on machine learning of the present application has achieved remarkable results in actual operation. In terms of load distribution, the average power distribution error decreased from ±4.9% before optimization to ±1.7%, with an error reduction rate of 65.3%, indicating that the system's power scheduling in multiple power distribution rooms is more accurate, enabling dynamic balance and real-time optimization, and improving the load balancing of the power distribution network. The scheduling response time was shortened from 5.1 seconds to 3.2 seconds, with a response speed improvement of 37.3%, indicating that the system achieves fast instruction generation and issuance through the hierarchical graph time sequence decision network, significantly reducing manual intervention and communication delay, and significantly improving scheduling efficiency.

[0035] In terms of power quality and device safety, the voltage sag amplitude decreased from 6.5% to 3.4%, and the current surge peak decreased from 315A to 203A, with improvements of 47.7% and 35.6%, respectively. This indicates that the topology reversible neural differential control module plays a key role in switching transient prediction and execution optimization, successfully suppressing the impact current and transient fluctuations in power distribution operations, effectively reducing the mechanical and thermal stress of devices. The average device temperature rise decreased from 79.4℃ to 74.1℃, with a thermal load reduction of 5.3℃, further verifying the thermal safety advantage of the system under multi-point parallel operation, ensuring the long-term stable operation of devices in unattended state.

[0036] The intelligent learning and self-optimization capability of the system is also verified. The integrated coefficient of performance (Pout / Pin) is improved from 0.86 to 0.93, the energy utilization rate is increased by 8.1%, the number of abnormal alarms is reduced from 122 times per month to 71 times, which is reduced by 41.8%, which shows that the system is more stable and reliable in operation scheduling and safety control. The mean of model prediction deviation is reduced from 3.8% to 1.4%, and the self-learning model confidence is improved from 0.88 to 0.94, which shows that the combination of hierarchical graph time series decision network and feedback self-learning mechanism enables the system to continuously optimize the prediction accuracy and control strategy, and realizes self-correction and long-term stable evolution. The method has significant intelligentization, stability and energy saving effect in the unattended power distribution scene, and provides a feasible engineering path for the intelligentization upgrade of future urban power distribution system.

[0037] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A machine learning-based method for intensive power distribution in unattended substations, characterized in that: include: Collect operational data from multiple unattended power distribution rooms, perform preprocessing on the operational data, and form a standardized input dataset; 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. 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. 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. 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. 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.

2. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The operating data includes voltage, current, active power, reactive power, temperature, humidity, equipment vibration signals, partial discharge signals, and switch status information.

3. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The preprocessing of the running data includes time synchronization, outlier removal, noise filtering, and normalization to eliminate data drift and sampling errors.

4. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The formation of the three-dimensional coupled feature structure includes: 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; Generate a node feature vector for each distribution node, which includes electrical parameters, thermal parameters, and health status; 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. 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. The associated weights are dynamically updated based on the node's operating stress and energy flow: A dual-channel mechanism is adopted, consisting of slow-channel sliding updates and fast-channel event-triggered updates; 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. 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.

5. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The output candidate control action includes: 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. 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. 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: 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 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. 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; 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.

6. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The set of safety actions that satisfy physical and electrical safety constraints includes: 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; 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. 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. Perform two-stage compliance processing on candidate control actions within the dynamic feasible domain: 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. 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.

7. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The generation of the transiently optimized execution action includes: 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. 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. 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. 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. 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.

8. The machine learning-based intensive power distribution method for unattended substations according to claim 1, characterized in that, The step of triggering incremental updates to the hierarchical graph temporal decision network model when there is a deviation between the predicted result and the actual feedback includes: 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. 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. 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.

9. A machine learning-based unattended substation centralized power distribution system, executing the machine learning-based unattended substation centralized power distribution method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect runtime data and perform preprocessing to generate standardized input datasets; 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; 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. 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; 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. 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.

Images