A reinforcement learning driven energy-saving load regulation method for low-voltage power distribution cabinet

Abstract

The application discloses a kind of reinforcement learning driven low-voltage distribution cabinet energy-saving load regulation and control methods, it is related to low-voltage distribution cabinet load control technical field, the present application will construction condition signal, construction load branch current operating state, electrical margin state, service arrears state and controlled switch feedback are used together for low-voltage distribution cabinet load regulation and control;After construction condition changes, dynamic control attribute is updated before opening and closing action, so that the construction load branch in hoisting, manned operation, continuous pumping or continuous drainage state is excluded from the disconnectable object;Reinforcement learning strategy model only evaluates the selected action set, so that the output action is jointly limited by branch attribute, electrical margin and switch constraint.

Description

Technical Field

[0001] This invention relates to the field of low-voltage switchgear load control technology, and in particular to a reinforcement learning-driven method for energy-saving load regulation of low-voltage switchgear. Background Technology

[0002] Temporary power supply systems at construction sites typically use low-voltage distribution cabinets, distribution boxes, and switch boxes to supply power to tower cranes, construction hoists, concrete pumps, drainage pumps, rebar processing equipment, lighting equipment, ventilation equipment, and charging equipment. These devices change their connection locations and operating status as work areas shift, construction progresses, material storage areas are adjusted, and temporary wiring changes occur. When multiple construction load branches share the same AC busbar, short-term inrush currents can easily form when motor-type equipment starts, switches operating conditions, or is restored after shutdown, causing a drop in the three-phase voltage of the AC busbar.

[0003] Existing intelligent power distribution systems at construction sites typically include data acquisition and control units in main distribution cabinets, distribution boxes, or switch boxes. These units collect data on power consumption, power status, and abnormal alarm information for each circuit and remotely control equipment to receive power, shut down during alarms, or cut off power during idle periods. This approach focuses on temporary power distribution management, power consumption statistics, and safety alarms. Branch control attributes are usually bound to equipment numbers, branch uses, or manual configuration status. When drainage pumps enter continuous drainage mode due to rising water levels, or when tower cranes or construction hoists enter the actual operation phase, the disconnection restrictions for the same branch change under different construction conditions. Fixed branch attributes may be inconsistent with the current construction status, leading to construction loads that currently require continuous power supply being included in the disconnectable categories.

[0004] Some demand response or energy-saving control methods generate load reduction, delayed access, or restored access strategies based on interruptible load capacity, electricity price, predicted load, or peak control targets, and can use reinforcement learning methods to evaluate the actions of adjustable loads. These methods are mainly aimed at relatively stable user-side load transfer or peak reduction, and the timing of actions is usually determined based on the scheduling cycle, revenue targets, or prediction results. Impact loads at construction sites are characterized by random start-up times, short durations, and significant interphase effects. If only periodic prediction or reward / penalty output branch switching actions are relied upon, unloading may only be triggered after voltage drop; after the impact ends, multiple start-up loads may cause secondary voltage drops, control jitter, and repeated switching of controlled switches due to concentrated restored access. Therefore, a reinforcement learning-driven energy-saving load control method for temporary low-voltage distribution cabinets at construction sites is urgently needed to solve these problems. Summary of the Invention

[0005] This invention provides a reinforcement learning-driven method for energy-saving load regulation of low-voltage distribution cabinets, which solves the problems of changes in the attributes of temporary power branches during construction, delayed impact unloading, and re-impact upon restoration of connection.

[0006] This invention provides a reinforcement learning-driven method for energy-saving load control of low-voltage switchgear, comprising: Collect the three-phase AC bus voltage, current of each construction load branch, status of controlled switches, thermal status of protection devices and construction condition signals of the temporary low-voltage distribution cabinet at the construction site, and determine the current connected to each phase. The dynamic control attributes are updated based on the construction condition signals and the current operating status of each construction load branch. Candidate impact segments are generated based on the sliding window correspondence between the current change sequence of the impact load branch and the three-phase voltage change sequence of the AC bus, and the impact load event is confirmed based on the switching operation matching status of the candidate impact segments and the candidate event source branch. Upon detecting the impact load event, an electrical margin state and the required unloading load are generated based on the three-phase voltage of the AC bus, the current connected to each phase, and the thermal state of the protection device. A service deficit status is generated based on the continuous disconnection data of the adjustable load branch and the power supply task status. A set of actionable actions is generated based on the dynamic control attributes, the electrical margin status, the service deficit status, and the controlled switch status; The dynamic control attributes, electrical margin status, service deficit status, controlled switch status, event status of the impact load event, and the set of actionable actions are input into the reinforcement learning strategy model, and the action priority sequence is output. The circuit breaker commands are sent branch by branch according to the action priority sequence, and the status record is updated and the set of actionable actions is regenerated after each feedback from the controlled switch is received.

[0007] In some embodiments, the AC busbar of the temporary low-voltage distribution cabinet at the construction site supplies power to multiple construction load branches via a controlled switch; Impact load branches and adjustable load branches are branch labels. The same construction load branch is allowed to have both impact load branch label and adjustable load branch label at the same time. The construction condition signals include construction equipment operation status signals, drainage demand signals, temporary power supply task signals and branch wiring change signals.

[0008] In some embodiments, the dynamic control attributes include a prohibition on disconnection attribute, a working condition protection attribute, a delayable access attribute, and a short-term disconnection attribute. When multiple dynamic control attributes are satisfied simultaneously, the action filtering attributes are determined in the following order: prohibition of disconnection attribute, operating condition protection attribute, short-term disconnection attribute, and delayed access attribute.

[0009] In some embodiments, generating the candidate impact fragment includes: The current change sequence of the impact load branch and the three-phase voltage change sequence of the AC bus are processed by sliding window to calculate the branch current slope and the bus voltage drop slope. When the branch current slope reaches the current rise threshold and the bus voltage drop slope in the current sliding window or the previous sliding window reaches the voltage drop threshold, the candidate impact window is marked. The consecutively appearing candidate impact windows are merged into candidate impact segments, and the impact load event is confirmed when the candidate matching duration reaches the preset duration and the switching operation matching status is the matched status.

[0010] In some embodiments, the electrical margin status includes the voltage margin of each phase, the current margin of each phase, the thermal margin of the protection device, and the start-up occupancy status of the branch to be connected. The voltage margin of each phase, the current margin of each phase, and the thermal margin of the protection device are converted into phase unloading current requirements, and the larger phase unloading current requirement in the same phase is taken as the required unloading load.

[0011] In some embodiments, the service deficit status includes continuous disconnection duration, remaining allowed disconnection duration, proximity of power supply task, cumulative number of disconnections, and the time of the most recent controlled switch action; A service debt score is calculated based on the service debt status, and the recovery priority is determined according to the service debt score; after the adjustable load branch is connected and reaches the minimum connection hold time, the corresponding service debt score is reduced.

[0012] In some embodiments, when generating the set of possible actions, candidate disconnection actions corresponding to prohibited disconnection attributes or operating condition protection attributes are excluded, candidate opening and closing actions that have not reached the minimum state holding time or have exceeded the upper limit of the number of actions are excluded, candidate access actions that would cause any phase current or AC bus three-phase voltage to leave the allowable operating range are excluded, and adjustable load branches whose remaining allowable disconnection time has reached the recovery boundary are added to the priority recovery set.

[0013] In some embodiments, generating the set of possible actions further includes excluding access candidate actions that conflict with the occupancy state of the tributary to be accessed: Establish a start-up occupancy record for each adjustable load branch; when any adjustable load branch is in the start-up occupancy valid state, and the adjustable load branch to be connected has an intersection with its start-up occupancy phase, and the start-up superimposed current after connection compresses the current margin of any overlapping phase to within the current retention margin, set the corresponding access candidate action to the start-up occupancy shield state. After the start-up period ends and the valid state of the access is occupied, the corresponding access candidate action is re-entered into the access candidate action filtering process.

[0014] In some embodiments, the reinforcement learning policy model is trained using a discrete action deep Q-learning algorithm; The reinforcement learning strategy model uses dynamic control attributes, electrical margin state, service deficit state, controlled switch state, and event state of impact load event as state vectors, candidate opening and closing actions as discrete actions, and generates instant rewards based on voltage over-limit, required unloaded load, service deficit change, repeated switching, peak load, and standby energy consumption. It also uses invalid action masking to keep candidate opening and closing actions that do not belong to the set of actionable actions as unselectable. The action priority sequence is obtained by ranking the action values ​​of each candidate opening and closing action.

[0015] In some embodiments, sending opening and closing commands branch by branch includes: During the duration of the impact load event, the adjustable load branches are disconnected one by one, and the disconnection is stopped when the required unloading load is met or the three-phase voltage of the AC bus recovers to the allowable operating range; after the impact load event ends and the recovery conditions are met, the adjustable load branches are connected one by one, and the connection is suspended if the conditions for continued connection are not met. If no feedback from the controlled switch is received that matches the selected candidate opening and closing action, the new access action is stopped; if the reinforcement learning strategy model output maintains the candidate action or suspends the access action, the corresponding branch state is maintained; if the reinforcement learning strategy model output is empty, has an incorrect format, or does not belong to the set of possible actions, a deterministic action sequence is generated.

[0016] Through the above technical solution, the present invention can achieve at least the following beneficial effects: This invention utilizes construction condition signals, the current operating status of construction load branches, electrical margin status, service deficit status, and controlled switch feedback for load regulation in low-voltage distribution cabinets. When construction conditions change, dynamic control attributes are updated before switching actions, excluding construction load branches in hoisting, personnel operation, continuous pumping, or continuous drainage states from the disconnectable list. The reinforcement learning strategy model evaluates only the selected set of possible actions, subjecting output actions to the combined constraints of branch attributes, electrical margin, and switch limitations. In cases of frequent changes in the purpose of construction branches, this reduces the likelihood of incorrectly disconnecting currently necessary operating equipment and ensures that unloading actions match the actual conditions at the construction site. Simultaneously, peak load penalties and standby energy consumption penalties participate in the reinforcement learning strategy model's action evaluation, allowing adjustable load branches not yet in the work cycle to enter delayed access or short-term disconnection control under conditions that meet construction power supply constraints and recovery priority constraints, thereby reducing peak access loads and non-operational standby power consumption in temporary low-voltage distribution cabinets at the construction site.

[0017] Impact load events are confirmed by a sliding window correspondence between the current change sequence of the impact load branch and the three-phase voltage change sequence of the AC bus. The rise in branch current, the drop in the minimum phase voltage, and the switching operation matching status of the candidate event source branch all contribute to event confirmation, enabling the impact process caused by equipment startup or operating condition switching to be distinguished from ordinary load fluctuations. After the impact load event is confirmed, the required unloading load is generated based on the three-phase voltage of the AC bus, the current already connected to each phase, and the thermal state of the protection device. This ensures that the unloading object and unloading amount correspond to the current phase margin and the thermal state of the protection device, forming a priority sequence of unloading actions in the early stages of the three-phase voltage drop process of the AC bus, shortening the time from the appearance of impact characteristics to the generation of unloading actions.

[0018] Furthermore, a service deficit status is generated for adjustable load branches, and the duration of continuous disconnection, remaining allowable disconnection duration, proximity of power supply tasks, and cumulative number of disconnections are used to determine the recovery priority. After an adjustable load branch is disconnected, its recovery requirement changes with the duration of disconnection data and task status; after reconnection is completed and the minimum connection hold time is reached, the corresponding recovery priority decreases. The recovery process can take into account power supply compensation after offloading and differences in branch waiting time, reducing prolonged power outages caused by the continuous selection of the same auxiliary branch.

[0019] During the recovery phase, a start-up occupancy record is introduced to determine the start-up occupancy phase and estimated start-up current of the branch to be connected. When an adjustable load branch is already in an active start-up occupancy state, the candidate connection action that overlaps with its phase and compresses the current margin to within the reserved margin is set to a start-up occupancy shielded state. After each connection feedback is completed by the controlled switch, a new set of possible actions is generated and the electrical margin state is reassessed, thereby enabling the distributed recovery of multiple start-up loads and reducing secondary voltage drops, repeated contactor switching, and recovery process jitter caused by centralized closing. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of the present invention.

[0021] Figure 1 This is a flowchart of the reinforcement learning-driven low-voltage switchgear energy-saving load control method in the embodiments. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] All terms used in this invention (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined.

[0024] Example 1: like Figure 1 As shown, this embodiment adopts a reinforcement learning-driven low-voltage distribution cabinet energy-saving load control method. The temporary low-voltage distribution cabinet at the construction site is set on the low-voltage side of the construction temporary power system. The AC bus is a three-phase AC bus that supplies power to multiple controlled switches in the low-voltage distribution cabinet. Each construction load branch is connected to the AC bus through the corresponding controlled switch.

[0025] Energy-saving load regulation includes controlling the disconnection, maintenance, connection, or suspension of connection of construction load branches that are allowed to be delayed or temporarily disconnected, under the conditions of meeting the power supply constraints of construction equipment and the allowable operating range of AC bus.

[0026] Energy-saving load control methods for low-voltage switchgear include: Step S1: Collect the three-phase AC bus voltage of the temporary low-voltage distribution cabinet at the construction site, the current of each construction load branch, the status of the controlled switch, the thermal status of the protection device, and the construction condition signal, and determine the current connected to each phase based on the current of each construction load branch and the status of the controlled switch. In this embodiment, the current rise threshold, voltage fall threshold, preset duration, lower voltage limit, upper current limit, thermal warning interval, current reserve margin, stable holding time, and start-up transient upper limit time are determined according to the wiring phase, equipment type, protection device configuration, and historical operation records of the construction load branch. After the branch wiring change signal takes effect, the threshold, duration, and phase parameters related to the construction load branch are recalibrated according to the changed wiring phase and equipment type.

[0027] Specifically, the three-phase AC bus voltage is obtained by a voltage sampling unit located on the bus side of the low-voltage distribution cabinet; the current of each construction load branch is obtained by the current transformer or current sampling module of the corresponding branch; the status of the controlled switch is obtained by the auxiliary contacts of the contactor, the auxiliary contacts of the circuit breaker, or the switch status acquisition terminal; the thermal status of the protection device is obtained by the output value of the thermal model of the protection device, the status quantity of the thermal relay, or the percentage of the thermal capacity of the electronic protector; and the construction condition signal is obtained by at least one of the following: the construction equipment controller, the liquid level detection device, the rainfall detection device, the temporary power supply task registration terminal, and the branch wiring confirmation terminal. Each data record includes the acquisition time, and the acquisition control unit aligns the three-phase AC bus voltage, the current of the construction load branch, the status of the controlled switch, the thermal status of the protection device, and the construction condition signal according to the acquisition time.

[0028] The sampling period for the three-phase AC bus voltage and the branch current of the construction load is determined based on the transient duration of the construction load startup, the refresh capability of the voltage sampling unit, and the controller processing cycle. In one embodiment, the sampling period for the three-phase AC bus voltage and the branch current of the construction load is 10ms to 200ms, the refresh period for the controlled switch status and construction condition signals is 50ms to 1s, and the refresh period for the thermal status of the protection device is 0.5s to 5s. The shorter electrical quantity sampling period is used to capture the startup current and the bus voltage drop process, while the longer protection device thermal status refresh period is used to match the time scale of the thermal accumulation change of the protection device. When there is a deviation in the acquisition time, the electrical quantity acquisition time is used as a reference, and the controlled switch status, protection device thermal status, and construction condition signals that are closest to this reference time and within the same control cycle are written into the same status record.

[0029] Step S2: Update the dynamic control attributes of each construction load branch based on the construction condition signal and the current operating status of each construction load branch. Step S3: Perform sliding window processing on the impact load branch current change sequence and AC bus three-phase voltage change sequence within the event identification period, calculate the branch current slope and the bus voltage drop slope determined by the minimum phase voltage in the AC bus three-phase voltage, and form a candidate impact window when the branch current slope reaches the current rise threshold and the bus voltage drop slope in the current sliding window or the previous sliding window reaches the voltage drop threshold. Step S4: Calculate the duration of candidate matching for consecutively appearing candidate impact windows, and confirm the impact load event when the duration of candidate matching reaches the preset duration, the controlled switch state of the candidate event source branch or the operation state of the construction equipment matches the start and end time of the candidate impact segment. In a preferred embodiment of Example 1, when identifying the time overlap between the current rise phase and the voltage fall phase, a sliding window processing is performed on the impact load branch current change sequence and the AC bus three-phase voltage change sequence within the event identification period.

[0030] The window length of each sliding window is set to cover the time period before the construction load starts and the current ramps up, and the window step size is set to an integer multiple of the sampling period to ensure partial overlap between adjacent windows. For the first... The impact load branch, in the first Calculate the branch current slope and bus voltage drop slope within a sliding window: , , in, Indicates the first The impact load branch is in the first Branch current slope within a sliding window Indicates the branch number of the impact load. Indicates the sliding window number, Indicates the first The impact load branch is in the first The branch current at the end of each sliding window. Indicates the first The impact load branch is in the first Branch current at the start of each sliding window Indicates the first At the end of the sliding window Indicates the length of the sliding window. Indicates the first The slope of the bus voltage drop within each sliding window Indicates separation In the AC bus voltage at the end of each sliding window Indicates separation In the AC bus voltage at the start of each sliding window Indicates that the AC busbars are separate. , , These represent the three phases of the AC bus.

[0031] When the branch current slope reaches the current rise threshold, and the bus voltage drop slope in the current or previous sliding window reaches the voltage drop threshold, the corresponding window is marked as a candidate impact window: , in, Indicates the first The impact load branch is in the first The candidate impact window markers corresponding to each sliding window Indicates the first The impact load branch is in the first The current rise indicator within the sliding window Indicates the first Bus voltage drop marker within a sliding window Indicates the first The bus voltage drop marker is within each sliding window. Matching is performed using the current sliding window and the previous sliding window to ensure the candidate impact window is in the [number]th [period]. It can be generated in real time after the sliding window ends.

[0032] The current rise marker is a binary marker, in The value is 1 when the current rise threshold is greater than or equal to the threshold value. The value is 0 when the current rise threshold is less than the current rise threshold; the bus voltage drop flag is a binary flag. The value is 1 when the voltage drop threshold is greater than or equal to the threshold value. The value is 0 when the voltage drop is less than the voltage drop threshold, and the voltage drop threshold is a positive value.

[0033] Within a sliding window where the bus voltage drop marker is set to 1, the phase with the minimum phase voltage drop slope reaching the voltage drop threshold is recorded as the voltage drop phase. When multiple phases simultaneously reach the voltage drop threshold, the phase with the larger voltage drop slope is recorded as the priority unloading phase. The voltage drop phase and the priority unloading phase are used to subsequently determine the phase identifiers for the required unloading load.

[0034] The first sliding window in the event recognition cycle does not exist. When the current rise threshold is set to 0, the voltage drop of the adjacent busbar in the missing window is marked as 0. The current rise threshold is calibrated based on the positive fluctuation range of the current slope of the corresponding impact load branch under non-starting stable conditions, and the slope margin corresponding to the current transformer sampling resolution is added. The voltage drop threshold is calibrated based on the negative fluctuation range of the three-phase voltage of the AC busbar during the period without impact load access, and the slope margin corresponding to the voltage sampling resolution is added. The lower limit of the current rise threshold is jointly limited by the current sampling resolution and the sliding window length, and the upper limit of the current rise threshold is limited by the known starting current variation range of the corresponding construction load branch. The lower limit of the voltage drop threshold is jointly limited by the voltage sampling resolution and the sliding window length, and the upper limit of the voltage drop threshold is limited by the allowable operating voltage fluctuation range of the AC busbar, so as to avoid ordinary steady-state fluctuations being misjudged as impact load events, and also to avoid typical starting impacts being filtered out.

[0035] Candidate impact windows that appear consecutively in the same impact load branch are merged into candidate impact segments, and the start window, end window, start time, and end time of the candidate impact segment are recorded.

[0036] For intervals where candidate impact windows appear consecutively, the duration of candidate matching for the corresponding candidate impact segments is calculated: , in, Indicates the first The impact load branch corresponds to the first The duration of candidate matching for each candidate impact fragment. Indicates the candidate impact fragment number, Indicates the step size of the sliding window. Indicates the first The impact load branch corresponds to the first The set of numbers of consecutive candidate impact windows within a candidate impact segment.

[0037] when When the duration is greater than or equal to the preset duration, the candidate impact segment is determined to meet the continuous confirmation condition. The preset duration is calibrated based on the start-up transient duration of the impact load branch, the sampling period, and the feedback delay of the controlled switch. The value is greater than or equal to a sliding window step size and less than or equal to the allowable identification delay of the start-up transient of this type of construction equipment.

[0038] When the preset duration is increased, the identification result is more biased towards sustained impacts; when the preset duration is decreased, the identification result is more biased towards rapid responses. When there are branch wiring change signals, sampling period adjustments, or changes in construction equipment models, the current rise threshold, voltage fall threshold, and preset duration are recalibrated. When sampling data is missing, three-phase voltage data is incomplete, or branch current data shows abnormal jumps, a single abnormal window is not used as the basis for establishing an impact load event; instead, the event state of the previous cycle is maintained, or a conservative identification state requiring confirmation from construction condition signals is entered. The aforementioned conservative identification state only applies to the impact load event identification process; the validity check of the state generated by the opening and closing execution and reinforcement learning training records is performed according to the subsequent state validity check rules.

[0039] After a candidate impact segment is established, the corresponding branch is designated as a candidate event source branch, and the controlled switch status and construction equipment operation status of that branch are read.

[0040] The switch operation matching status is used to represent the matching result between the controlled switch state change time, the construction equipment operation state change time, and the start and end times of the candidate impact segment in the candidate event source branch. The switch operation matching status includes a matched state and a matched state. A matched state is considered successful when the controlled switch state change time falls within the start and end time range of the candidate impact segment, or when the construction equipment operation state change time falls within the start and end time range of the candidate impact segment, or when the deviation between the aforementioned state change time and the start and end time range is within the controlled switch feedback delay range. A matched state is considered unsuccessful when neither of the two types of state change times satisfies the above time relationship.

[0041] When the switching operation matching status of a candidate event source branch is in the matched state, and the candidate matching duration of the corresponding candidate impact segment reaches the preset duration, the candidate event source branch is confirmed as an event source branch, and the impact load event is confirmed. The confirmation result of the impact load event is jointly determined by the candidate impact segment, the effective overlap length, and the switching operation matching status.

[0042] When historical feedback records exist, the feedback delay range of the controlled switch is calibrated based on the upper limit of the historical delay between the time the controlled switch opens and closes the command and the time when the feedback state stabilizes. When historical feedback records are insufficient, the range is calibrated based on the communication timeout configured by the controller. When feedback jitter exists, the earliest time when the same feedback state is obtained consecutively is taken as the time when the feedback state stabilizes.

[0043] The bus voltage drop window time deviation is the absolute value of the difference between the end time of the sliding window where the current rise marker of the impact load branch is located and the end time of the sliding window where the bus voltage drop marker is located. If multiple branches simultaneously meet the above conditions, the branch with the largest current rise slope is selected as the event source branch; when the current rise slopes of multiple branches are the same, the branch with the smallest time deviation from the bus voltage drop window is selected as the event source branch; the remaining branches that meet the conditions are retained as associated impact branches for subsequent electrical margin status updates.

[0044] Step S5: When an impact load event is detected, an electrical margin state and the required unloading load are generated based on the three-phase voltage of the AC bus, the current connected to each phase, and the thermal state of the protection device. Step S6: Generate a service deficit status based on the continuous disconnection data of the adjustable load branch and the power supply task status; Step S7: Based on dynamic control attributes, electrical margin status, service deficit status and controlled switch status, candidate opening and closing actions are filtered to generate a set of actionable actions; Step S8: Input the dynamic control attributes, electrical margin status, service deficit status, event status of impact load events and action set into the reinforcement learning strategy model, and output the action priority sequence. Step S9: Send opening and closing commands to each branch according to the action priority sequence, and update the status record and regenerate the set of actionable actions after receiving feedback from the controlled switch each time.

[0045] Both disconnection feedback and connection feedback belong to controlled switch feedback. Controlled switch feedback includes feedback time, feedback branch number, and feedback on / off status; when the feedback on / off status corresponds to the sent opening and closing command, a consistent status feedback is formed.

[0046] Furthermore, the AC busbar of the temporary low-voltage distribution cabinet at the construction site supplies power to multiple construction load branches via controlled switches. Impulse load branches and adjustable load branches are tagged with branch labels; the same construction load branch is allowed to have both impulse load branch labels and adjustable load branch labels simultaneously. The impulse load branch label is determined by the equipment type corresponding to the construction load branch, historical startup peak value, historical operating condition switching current peak value, and branch wiring change signal; the adjustable load branch label is determined by the allowable delay time of the task corresponding to the construction load branch, the allowable disconnection duration, and the current construction condition. Construction condition signals include construction equipment operating status signals, drainage demand signals, temporary power supply task signals, and branch wiring change signals; the current operating status includes branch on / off status, branch current status, and equipment operating stage.

[0047] Specifically, the construction equipment operation status signal indicates whether the construction equipment is in standby, start-up, running, continuous operation, stopped, or shut down due to a fault; the drainage demand signal indicates the liquid level status, rainfall status, and continuous drainage conditions; the temporary power supply task signal indicates the start time, expected end time, and urgency of the charging, ventilation, lighting, processing, or drainage task; and the branch wiring change signal indicates a change in the wiring phase, connected equipment type, or controlled switch correspondence of the construction load branch. After the branch wiring change signal takes effect, the dynamic control attributes, start-up phase, estimated start-up current, and action count of the corresponding construction load branch will be based on the changed branch information.

[0048] The current operating status is jointly determined by the branch on / off status, branch current status, and equipment operation stage. The branch on / off status is determined by the controlled switch status; the branch current status is determined by the current range, including zero current range, rising range, stable range, and abnormal jump range; the equipment operation stage is determined by the operating status signal of the construction equipment. When the branch on / off status and branch current status are inconsistent, the current operating status is confirmed by the controlled switch feedback and continuously sampled branch current; when controlled switch feedback is missing, the current operating status is generated by the most recent valid controlled switch status and continuously sampled branch current, and the reason for the status is written in the status record.

[0049] Furthermore, the dynamic control attributes include prohibition of disconnection, operating condition protection, deferred access, and short-term disconnection. When multiple dynamic control attributes are satisfied simultaneously, the action filtering attributes are determined in the order of prohibition of disconnection, operating condition protection, short-term disconnection, and deferred access. When a construction load branch is set to the prohibition of disconnection attribute, its disconnection candidate actions are excluded from the set of possible actions. When a construction load branch is set to the operating condition protection attribute, it does not participate in the unloading action filtering until the corresponding operating condition protection condition is released. When a construction load branch is set to the short-term disconnection or deferred access attribute, it enters the action filtering process for adjustable load branches.

[0050] When updating dynamic control attributes, construction load branches in the hoisting, manned operation, or continuous pumping phases are set to the "disconnect prohibited" attribute; construction load branches corresponding to drainage equipment are set to the "operational condition protection" attribute when the water level or rainfall conditions meet the continuous drainage conditions, and are set to the "short-term disconnectable" attribute after the continuous drainage conditions are lifted; branches corresponding to charging equipment not in the work cycle, non-continuous ventilation equipment, and auxiliary electrical equipment in non-operation areas are set to the "delayed access" or "short-term disconnectable" attribute; after the construction condition signal changes, a set of actionable actions is generated based on the changed dynamic control attributes.

[0051] Furthermore, the electrical margin status includes the voltage margin of each phase, the current margin of each phase, the thermal margin of the protection device, and the start-up occupancy status of the branch to be connected. The voltage margin of each phase is calculated from the difference between the AC bus voltage and the lower voltage limit of the corresponding phase. The current margin of each phase is calculated from the difference between the upper current limit of the corresponding phase and the current already connected to each phase. The thermal margin of the protection device is determined from the difference between the thermal warning value of the protection device and the thermal state of the protection device. The start-up occupancy status of the branch to be connected is determined by the start-up occupancy record and is used to indicate whether the adjustable load branch to be connected forms a phase and start-up current superposition relationship with the adjustable load branch that is already in the start-up process.

[0052] The required unloading load is expressed as phase-specific unloading current demand or phase-specific unloading power demand. When unloading is triggered by the current margin of each phase, the phase-specific unloading current demand is calculated from the current already connected, the upper limit of the current, and the current reserve margin of the corresponding phase. When unloading is triggered by the voltage margin of each phase, the phase-specific unloading current demand is first calculated from the voltage deficit and voltage-load sensitivity mapping table of the corresponding phase; the voltage-load sensitivity mapping table is established from the field commissioning records or historical control records, recording the correspondence between the phase-specific voltage change and the phase-specific current change before and after disconnecting the known branch load. When unloading is triggered by the thermal margin of the protection device, the unloading current demand of the corresponding downstream branch of the protection device is calculated from the thermal state of the protection device, the thermal warning interval, and the thermal model of the protection device. When there are multiple triggering sources for the same phase, the larger value among the converted phase-specific unloading current demands is taken as the required unloading load for that phase.

[0053] Furthermore, the service deficit status includes continuous disconnection duration, remaining allowable disconnection duration, proximity of power supply tasks, cumulative number of disconnections within a specified time range, and the time of the most recent controlled switch action. When the continuous disconnection duration increases, the remaining allowable disconnection duration decreases, the proximity of power supply tasks increases, or the cumulative number of disconnections increases, the recovery priority of the corresponding adjustable load branch is increased. After the adjustable load branch is connected and reaches the minimum connection holding time, the recovery priority of the corresponding adjustable load branch is decreased.

[0054] Service deficit status is recorded for each adjustable load branch. When the corresponding controlled switch reports an open state, the continuous disconnection duration is updated, and the remaining allowable disconnection duration is obtained by subtracting from the allowable disconnection duration of that adjustable load branch. The proximity of the power supply task is calculated based on the time difference between the current control time and the start or end time of the power supply task. The cumulative number of disconnections within a specified time range is determined by the number of disconnection feedbacks for the corresponding adjustable load branch within that time range. The continuous disconnection duration, remaining allowable disconnection duration, proximity of the power supply task, and cumulative number of disconnections are normalized and then weighted and summed to obtain the service deficit score. The time of the most recent controlled switch action is used for sequential resolution when service deficit scores are the same. In the deterministic action sequence, the recovery order is generated from high to low service deficit scores; when service deficit scores are the same, the recovery order is generated from farthest to nearest time based on the most recent controlled switch action time.

[0055] Furthermore, the reinforcement learning strategy model includes a shared state encoding network, an unloading action evaluation network, and a recovery action evaluation network. The shared state encoding network encodes dynamic control attributes, electrical margin status, service deficit status, event status of the impact load event, controlled switch status, and the set of possible actions. During the duration of the impact load event, the unloading action evaluation network ranks the disconnection candidate actions and the hold candidate actions in the set of possible actions. After the impact load event ends, the recovery action evaluation network ranks the access candidate actions and the pause access action in the set of possible actions. Candidate opening and closing actions that do not belong to the set of possible actions are set to an unselectable state.

[0056] Furthermore, the reinforcement learning strategy model is trained using a discrete action deep Q-learning algorithm. The state vector includes bus-level and branch-level state fields. Bus-level state fields include phase voltage margin, phase current margin, protection device thermal margin, and event status of impact load events. Branch-level state fields include dynamic control attributes, controlled switch status, normalized branch current value, service arrears score, minimum state hold remaining time, and the start-up occupancy status of the branch to be connected. Each continuous state field is normalized according to the corresponding branch's rated current, lower voltage limit, upper current limit, allowable disconnection duration, or protection device thermal warning value. Discrete state fields are numbered and coded.

[0057] Discrete actions are represented by action numbers, which include disconnecting a specified adjustable load branch, connecting to a specified adjustable load branch, maintaining the current branch status, and suspending connection. Candidate opening and closing actions in the set of actionable actions have corresponding action numbers; candidate opening and closing actions that do not belong to the set of actionable actions are set to an unselectable state through an invalid action mask.

[0058] During training, the state before the action is executed, the selected candidate opening and closing actions, the immediate reward, the state after the action is executed, and the set of possible actions after the action is executed are written into the reinforcement learning training record. The immediate reward is obtained by weighting the voltage over-limit penalty, the required unloaded load satisfaction, the service deficit change, the controlled switch repeated switching penalty, the peak load penalty, and the standby energy consumption penalty. The peak load penalty is obtained from the current connected to each phase or the total power of the AC bus in the current control cycle; the standby energy consumption penalty is obtained from the power of the adjustable load branch that is not in the working cycle but is still connected.

[0059] The reinforcement learning strategy model consists of an online network and a target network. The online network outputs the action value of each candidate opening / closing action based on the state before the action is executed; the target network calculates the temporal difference objective based on the state after the action is executed and the set of possible actions after the action is executed. The training loss is obtained by squared the difference between the action value of the selected candidate opening / closing action and the temporal difference objective. The model parameters are updated using the training record set, and the validation record set is used to determine the discount factor, learning rate, reward weight, and training stopping time. During inference, the online network outputs the action value of candidate opening / closing actions within the set of possible actions and forms an action priority sequence according to the action value from high to low; the current branch state and pause access are maintained as stop actions in the action priority sequence.

[0060] The unloadable load corresponding to the disconnected adjustable load branch is determined by the branch current, connection phase, and three-phase AC bus voltage before the adjustable load branch is disconnected, indicating the phase current or phase power that can be released after the branch is disconnected. Furthermore, when sending opening and closing commands branch by branch, the adjustable load branches are disconnected one by one according to the action priority sequence during the duration of the impact load event. After receiving disconnection feedback from the corresponding controlled switch, the three-phase AC bus voltage and branch current are re-acquired, the status record is updated, and a new set of actionable actions is generated. Disconnection stops when the unloadable load corresponding to the disconnected adjustable load branch meets the required unloading load, or when the three-phase AC bus voltage returns to the allowable operating range. After the impact load event ends and the three-phase voltage of the AC bus meets the recovery conditions, the adjustable load branches are connected one by one according to the action priority sequence. After receiving the connection feedback of the corresponding controlled switch, the set of actionable actions is regenerated, and the connection is suspended when the updated electrical margin state does not meet the conditions for continued connection.

[0061] Recovery conditions include the three-phase voltage of the AC bus continuously being within the allowable operating range, the event status of the impact load event being in the terminated state, and the thermal status of the protection device being below the thermal warning range. Continued connection conditions include the current margin of each phase of the adjustable load branch to be connected meeting the estimated starting current requirements, the voltage margin of each phase meeting the voltage retention requirements, the thermal margin of the protection device meeting the thermal capacity requirements after connection, and the corresponding connection candidate action not being in a startup occupancy shielded state.

[0062] Furthermore, it also includes generating reinforcement learning training records based on controlled switch feedback: After receiving feedback that the controlled switch's opening and closing action is consistent with the selected candidate's action, the state before the action is executed, the selected candidate's opening and closing action, the state after the action is executed, and the change in the service deficit state are written into the reinforcement learning training record. Feedback tags are generated based on whether the three-phase voltage of the AC bus after the action is executed is within the allowable operating range, whether the required unloaded load has been met, whether the service deficit state continues to increase, and whether the controlled switch has undergone repeated switching. Feedback tags are only used to generate the corresponding reward items in the immediate reward and are not directly used as supervised learning labels.

[0063] If no consistent status feedback is received, new access actions are stopped, and the branch on / off status is determined based on the feedback from the controlled switch, the current of the construction load branch, and the most recent valid status record. If the branch on / off status can be confirmed by continuously sampled branch current, and the three-phase voltage of the AC bus or the current connected to each phase leaves the allowable operating range, a conservative disconnect action is permitted. If the branch on / off status cannot be confirmed, the current power supply status of the branch corresponding to the prohibited disconnect attribute and operating condition protection attribute is maintained, and the system waits for feedback from the controlled switch to restore operation. When the construction operating condition signal is missing, the construction load branch involving hoisting, personnel operation, continuous pumping, continuous drainage, or temporary power supply tasks is placed under the dynamic control attribute after protection, and a deterministic action sequence is generated.

[0064] Specifically, before generating a deterministic action sequence, a state validity check is performed on the current state record. The state validity check includes field integrity checks, data acquisition continuity checks, electrical quantity validity checks, and switch feedback consistency checks. Field integrity checks confirm that the three-phase AC bus voltage, construction load branch current, controlled switch status, protection device thermal status, and construction condition signals all have valid values; dynamic control attributes, electrical margin status, and service deficit status are generated after the above fields pass the checks; data acquisition continuity checks confirm that electrical quantities, switch status, and construction condition signals within the same control cycle are aligned according to the data acquisition time; electrical quantity validity checks confirm that the three-phase AC bus voltage and construction load branch current are within the range of the sampling device; and switch feedback consistency checks confirm that the controlled switch feedback status is consistent with the sent opening and closing commands.

[0065] The status validity check result is written to the status field in the status record. The status reason is used to indicate field missing, discontinuous timing, electrical quantity exceeding limits, inconsistent switch feedback, or missing construction condition signal. When different status reasons occur but the processing result is the same, they are merged into the same status field. For status records with missing fields or discontinuous acquisition timing, the writing of that status record into the reinforcement learning training record is stopped; for status records with electrical quantity exceeding limits or inconsistent switch feedback, the subsequent actions determined by the reinforcement learning strategy model are stopped, and a deterministic action sequence is generated based on the dynamic control attribute, electrical margin status, and service deficit status of the most recently passed status validity check; for status records with missing construction condition signals, the construction load branches involving continuous operation, manned operation, continuous drainage, or temporary power supply tasks are kept under the dynamic control attribute after protection, and a deterministic action sequence is generated.

[0066] When the reinforcement learning policy model outputs a candidate action or pauses an access action, the corresponding branch state is maintained, and the output is written as a valid action into the state record. When the reinforcement learning policy model outputs an empty value, an incorrect output format, or an output action that does not belong to the set of actionable actions, the model output is recorded as abnormal, and a deterministic action sequence is generated based on the current dynamic control attributes, electrical margin status, and service deficit status.

[0067] The deterministic action sequence includes a deterministic action sequence for the unloading phase and a deterministic action sequence for the recovery phase. During the duration of the impact load event, a deterministic action sequence for the unloading phase is generated. This sequence first excludes candidate disconnection actions corresponding to the prohibited disconnection attribute and the operating condition protection attribute. Then, from the construction load branches corresponding to the short-time disconnection attribute, disconnection actions are generated in the following order: remaining allowable disconnection duration from longest to shortest, unloadable load of the same phase from largest to smallest, and the time of the most recent controlled switch action from oldest to most recent. After each disconnection action is executed, the three-phase voltage of the AC bus and the current of the construction load branch are re-acquired, and the electrical margin status is regenerated. The deterministic action sequence for the unloading phase ends when the required unloading load is met or the three-phase voltage of the AC bus recovers to the allowable operating range.

[0068] After the impact load event ends, a deterministic action sequence for the recovery phase is generated. This sequence first processes adjustable load branches in the priority recovery set, and then processes the remaining adjustable load branches with high service deficit status. Adjustable load branches in the priority recovery set generate access actions in the following order: remaining allowable disconnection time from shortest to longest, proximity of power supply tasks from highest to lowest, and cumulative disconnection counts within a specified time range from most to least. The remaining adjustable load branches generate access actions in the following order: service deficit status from highest to lowest. After each access action is executed, a new set of actionable actions is generated. If the updated electrical margin status does not meet the conditions for continued access, or if the corresponding access candidate action is in a start-up, occupied, or shielded state, the access action is suspended, and the corresponding service deficit status is retained for reordering in the next control cycle.

[0069] The control period for the status validity check is determined based on the controlled switch feedback period, the electrical quantity sampling period, and the allowable response time for impact load events. In one implementation, the control period for the status validity check is 0.5s to 5s, the controlled switch feedback timeout is 1s to 10s, and the threshold for the number of consecutively missing electrical quantity samples is 2 to 10. If the control period is too short, temporary sampling jitter may frequently trigger deterministic action sequences; if the control period is too long, the unloading response during the duration of the impact load event may be delayed. If the controlled switch feedback timeout is too short, normal communication delays may be judged as inconsistent switch feedback; if the controlled switch feedback timeout is too long, the waiting time under abnormal conditions will be prolonged.

[0070] Example 2: Based on Embodiment 1, this embodiment provides a specific method for excluding access candidate actions that conflict with the start-up occupancy state of the branch to be accessed in the energy-saving load control method of low-voltage distribution cabinet; The minimum state holding time includes the minimum on-time holding time and the minimum off-time holding time; the minimum on-time holding time is used to limit the construction load branch from being disconnected again after being connected, and the minimum off-time holding time is used to limit the construction load branch from being connected again after being disconnected.

[0071] When generating the set of possible actions, for disconnection candidate actions, disconnection candidate actions corresponding to prohibited disconnection attributes or operating condition protection attributes are excluded, disconnection candidate actions for construction load branches that have not reached the minimum connection holding time are excluded, and disconnection candidate actions that would cause the corresponding controlled switch to exceed the upper limit of the number of actions within the specified time range are excluded. For connection candidate actions, connection candidate actions for construction load branches that have not reached the minimum disconnection holding time are excluded, and connection candidate actions that, according to the electrical margin status, would cause any phase current or AC bus three-phase voltage to leave the allowable operating range are excluded. Adjustable load branches whose remaining allowable disconnection time has reached the recovery boundary are added to the priority recovery set. The priority recovery set is used to give priority to adjustable load branches whose remaining allowable disconnection time has reached the recovery boundary when entering the connection candidate action screening. Adjustable load branches in the priority recovery set enter the connection candidate action screening when the electrical margin status requirements are met. Adjustable load branches entering the connection candidate action screening are still screened according to the start-up occupancy shield status, minimum disconnection holding time, and upper limit of the number of actions of the controlled switch.

[0072] Furthermore, when excluding candidate access actions that conflict with the start-up occupancy state of the branch to be accessed, a start-up occupancy record is established for each adjustable load branch. The start-up occupancy record includes the start-up occupancy start time, start-up occupancy end time, start-up occupancy phase, and estimated start-up current value. The start-up occupancy start time is determined by the earliest of the following: the controlled switch access feedback time, the construction equipment start command time, and the time when the branch current enters the rising range from the zero current range. The start-up occupancy end time is determined by the time after the corresponding branch current enters the stable range and remains stable for a certain period of time. When any adjustable load branch is in an effective start-up occupancy state, and the branch to be accessed... If the start-occupied phase of the adjustable load branch overlaps with the start-occupied phase of the adjustable load branch, and the superimposed start-up current after connection compresses the current margin of any overlapping phase to within the current retention margin, the access candidate action corresponding to the adjustable load branch to be connected is set to start-occupied shielded state and excluded from the set of available actions; after the start-occupied effective state ends, the access candidate action that was previously in the start-occupied shielded state is re-entered into the access candidate action screening; when the regenerated set of available actions meets the electrical margin state requirements, minimum disconnection holding time and the upper limit of the number of controlled switch actions, the access candidate action is restored to the selectable state.

[0073] In a preferred embodiment of Example 2, the start-up occupancy record includes the start-up occupancy start time, the start-up occupancy end time, the start-up occupancy phase, and the estimated start-up current. The start-up occupancy start time is determined by the earliest of the following: the time when the controlled switch is connected to feedback, the time when the construction equipment starts, and the time when the branch current enters the rising range from the zero current range. The start-up occupancy end time is determined by the moment when the corresponding branch current enters the stable range and remains there continuously for the stable holding time. The stable range is calibrated based on the average current and steady-state fluctuation band of the corresponding branch after start-up. The stable holding time is calibrated based on the sampling period and the upper limit of the transient start-up time of the construction equipment, and is taken as the time length that can cover multiple consecutive sampling points. When the corresponding branch current remains in the stable range for the stable holding time, the end time of the stable holding time is taken as the start-up occupancy end time.

[0074] The start-up transient upper limit time is determined based on the equipment nameplate start-up characteristics, the preset high percentile value of the historical start-up duration of similar equipment, or on-site commissioning records. After the adjustable load branch is connected, the start-up occupancy start time plus the start-up transient upper limit time is initially used as the provisional start-up occupancy end time. When the corresponding branch current enters the stable range and remains there continuously for the stable holding time, the start-up occupancy end time is corrected to the end time of the stable holding time. For connected branches that have not yet formed a stable current, the provisional start-up occupancy end time continues to be used.

[0075] For the The effective state of startup occupancy of an adjustable load branch that has been connected and is in the startup process is determined by the following formula at the current control moment: , in, Indicates the first The effective state of startup and occupancy of each adjustable load branch at the current control moment. This indicates the number of an adjustable load branch that is already connected and may be in the startup process. Indicates the first The start-up time of an adjustable load branch Indicates the current moment of regulation. Indicates the first The start-up and end times of an adjustable load branch.

[0076] when When taking 1, the first... The starting phase of each adjustable load branch is considered as the occupied phase, and its estimated starting current is included in the phase occupancy.

[0077] For the For each adjustable load branch to be connected, calculate the superimposed starting current in each phase after its connection: , in, Indicates the first After each adjustable load branch is connected, in the phase... The starting superposition current formed above, Indicates the branch number of the adjustable load to be connected. Following the aforementioned definition, Indicates separation The current base current is the current of the construction load branch that is currently connected and not in an active occupancy state. The combined current on, This represents the set of adjustable load branches that are currently in use and have not yet been released. Indicates the first Does each adjustable load branch occupy a phase? The phase occupancy markers, Indicates the first Estimated starting current for each adjustable load branch. Indicates the first Does each adjustable load branch waiting to be connected occupy a phase? The phase occupancy markers, Indicates the first Estimated starting current of each adjustable load branch to be connected.

[0078] The phase occupancy marker is a binary marker, corresponding to the branch and phase. When an electrical connection exists, take 1, corresponding to the branch and phase. Take 0 when there is no electrical connection. The set of start-up occupied phases for each adjustable load branch is denoted as . , No. The set of start-up occupancy phases for each adjustable load branch to be connected is denoted as . The starting phases of a single-phase load are its own connection phases; the starting phases of a three-phase load are the three phases of the AC bus; and the starting phases of a cross-phase load are the two phases actually connected to it.

[0079] The estimated starting current is determined by combining the peak current of the branch in the most recent connection processes, the rated current of the branch, and the starting multiple of similar construction equipment. When historical connection data is sufficient, the conservative quantile of the historical starting peak value is used. When historical connection data is missing, the product of the equipment rated current and the preset starting multiple is used.

[0080] The lower limit of the estimated starting current is the rated current of the corresponding equipment, and the upper limit is the smaller of the allowable instantaneous current of the protection device or the design upper limit of the branch starting current. When a branch wiring change signal appears, the equipment model is changed, or the current transformer range is adjusted, the starting occupied phase and the estimated starting current are recalibrated. When sampling is missing or the controlled switch feedback is abnormal, the upper limit of the estimated starting current is used, and the time when the starting occupation ends is postponed to the time after the start time of the starting occupation and the upper limit of the starting transient time of this type of construction equipment.

[0081] For the The access candidate action corresponding to the adjustable load branch to be connected is the shield boundary, which is the current margin determination boundary for triggering the start-up of the shield state. When there is a first... The adjustable load branch in the active state of the start-up occupancy is the first one. The set of start-up occupied phases of each adjustable load branch and the first If the sets of start-up occupied phases of the adjustable load branches to be connected intersect, and the start-up superimposed current compression current margin of any overlapping phase after connection reaches the shielding boundary, then the candidate connection action is set to the start-up occupied shielding state: If it exists make and ,and ,but ; If the above conditions are not met simultaneously, then ; in, Indicates the first The activation, occupancy, and masking status of the corresponding access candidate actions for each adjustable load branch to be connected. A value of 1 indicates that the access candidate action is in the start-up, occupancy, and masking state. A value of 0 indicates that the access candidate action is not in the start-up, occupancy, or masking state. Indicates separation The upper limit of current, Indicates separation The current is given a margin, and other parameters follow the definitions above.

[0082] when When the value is 1, it indicates that the candidate access action is in the start-up occupancy masking state. The candidate access action will overlap with the ongoing start-up occupancy process and compress the current margin of at least one related phase to within the reserve margin. Therefore, the candidate access action is excluded from the set of available actions.

[0083] The current reserve margin is calibrated based on the operating characteristics of the protection device, the AC bus voltage fluctuation range, and the sampling error range. Increasing the current reserve margin expands the shielding range, making the connection process more biased towards distributed startup; decreasing the current reserve margin shrinks the shielding range, making the connection process more biased towards power restoration. The lower limit of the current reserve margin is jointly limited by the sampling error range and the thermal state calculation error of the protection device, while the upper limit is limited by the available difference between the upper limit of the corresponding phase current and the current base current.

[0084] After the adjustable load branch that the controlled switch has fed back the connection action has been connected, the adjustable load branch is recorded in the start-up occupancy record and continues to maintain its start-up occupancy valid state until the corresponding branch current enters the stable range, or reaches the start-up occupancy end time extrapolated from the upper limit time of the equipment start-up transient.

[0085] After the start-up occupancy period ends, the start-up occupancy phase corresponding to the branch is released, the estimated start-up current value is stopped from being included in the current phase occupancy, and a new set of actionable actions is generated. Previously, the access candidate actions that were in the start-up occupancy shielded state due to phase overlap and start-up current superposition will re-enter the access candidate action screening after the start-up occupancy period ends. When the updated electrical margin state meets the conditions for continued access, the minimum disconnection holding time has been reached, and the number of controlled switch actions has not exceeded the upper limit of the number of actions, the access candidate action is restored to the selectable state.

[0086] If both a branch in the priority recovery set and a branch that has been activated and blocked exist simultaneously, the priority recovery set only increases the recovery priority of the corresponding branch, without bypassing the electrical margin status and activation blocking criteria. Under the condition that the electrical margin status requirements are met and activation blocking is not triggered, branches whose remaining allowable disconnection time has reached the recovery boundary are restored first. If the adjustable load branch in the priority recovery set still conflicts with the activation blocking status, its access action is suspended and the service deficit status continues to accumulate for reordering in the next control cycle.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

[0088] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of the invention and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in the Background section is intended only to enhance the understanding of the general background of the invention and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A reinforcement learning-driven method for energy-saving load control of low-voltage switchgear, characterized in that, include: Collect the three-phase AC bus voltage, current of each construction load branch, status of controlled switches, thermal status of protection devices and construction condition signals of the temporary low-voltage distribution cabinet at the construction site, and determine the current connected to each phase. The dynamic control attributes are updated based on the construction condition signals and the current operating status of each construction load branch. Candidate impact segments are generated based on the sliding window correspondence between the current change sequence of the impact load branch and the three-phase voltage change sequence of the AC bus, and the impact load event is confirmed based on the switching operation matching status of the candidate impact segments and the candidate event source branch. Upon detecting the impact load event, an electrical margin state and the required unloading load are generated based on the three-phase voltage of the AC bus, the current connected to each phase, and the thermal state of the protection device. A service deficit status is generated based on the continuous disconnection data of the adjustable load branch and the power supply task status. A set of actionable actions is generated based on the dynamic control attributes, the electrical margin status, the service deficit status, and the controlled switch status; The dynamic control attributes, electrical margin status, service deficit status, controlled switch status, event status of the impact load event, and the set of actionable actions are input into the reinforcement learning strategy model, and the action priority sequence is output. The circuit breaker commands are sent branch by branch according to the action priority sequence, and the status record is updated and the set of actionable actions is regenerated after each feedback from the controlled switch is received.

2. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, The AC busbar of the temporary low-voltage distribution cabinet at the construction site supplies power to multiple construction load branches via a controlled switch. Impact load branches and adjustable load branches are branch labels. The same construction load branch is allowed to have both impact load branch label and adjustable load branch label at the same time. The construction condition signals include construction equipment operation status signals, drainage demand signals, temporary power supply task signals and branch wiring change signals.

3. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, The dynamic control attributes include a prohibition on disconnection attribute, a working condition protection attribute, a delayed access attribute, and a short-term disconnection attribute. When multiple dynamic control attributes are satisfied simultaneously, the action filtering attributes are determined in the following order: prohibition of disconnection attribute, operating condition protection attribute, short-term disconnection attribute, and delayed access attribute.

4. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, Generating the candidate impact fragments includes: The current change sequence of the impact load branch and the three-phase voltage change sequence of the AC bus are processed by sliding window to calculate the branch current slope and the bus voltage drop slope. When the branch current slope reaches the current rise threshold and the bus voltage drop slope in the current sliding window or the previous sliding window reaches the voltage drop threshold, the candidate impact window is marked. The consecutively appearing candidate impact windows are merged into candidate impact segments, and the impact load event is confirmed when the candidate matching duration reaches the preset duration and the switching operation matching status is the matched status.

5. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, The electrical margin status includes the voltage margin of each phase, the current margin of each phase, the thermal margin of the protection device, and the start-up and occupancy status of the branch to be connected. The voltage margin of each phase, the current margin of each phase, and the thermal margin of the protection device are converted into phase unloading current requirements, and the larger phase unloading current requirement in the same phase is taken as the required unloading load.

6. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, The service deficit status includes continuous disconnection duration, remaining allowed disconnection duration, proximity of power supply task, cumulative number of disconnections, and the time of the most recent controlled switch action. A service debt score is calculated based on the service debt status, and the recovery priority is determined according to the service debt score; after the adjustable load branch is connected and reaches the minimum connection hold time, the corresponding service debt score is reduced.

7. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, When generating the set of possible actions, candidate disconnection actions corresponding to prohibited disconnection attributes or operating condition protection attributes are excluded, candidate opening and closing actions that have not reached the minimum state holding time or have exceeded the upper limit of the number of actions are excluded, candidate access actions that would cause any phase current or AC bus three-phase voltage to leave the allowable operating range are excluded, and adjustable load branches whose remaining allowable disconnection time has reached the recovery boundary are added to the priority recovery set.

8. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 7, characterized in that, Generating the set of possible actions also includes excluding access candidate actions that conflict with the occupancy state of the tributary to be accessed: Establish a start-up occupancy record for each adjustable load branch; when any adjustable load branch is in the start-up occupancy valid state, and the adjustable load branch to be connected has an intersection with its start-up occupancy phase, and the start-up superimposed current after connection compresses the current margin of any overlapping phase to within the current retention margin, set the corresponding access candidate action to the start-up occupancy shield state. After the start-up period ends and the valid state of the access is occupied, the corresponding access candidate action is re-entered into the access candidate action filtering process.

9. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, The reinforcement learning strategy model is trained using a discrete action deep Q-learning algorithm. The reinforcement learning strategy model uses dynamic control attributes, electrical margin state, service deficit state, controlled switch state, and event state of impact load event as state vectors, candidate opening and closing actions as discrete actions, and generates instant rewards based on voltage over-limit, required unloaded load, service deficit change, repeated switching, peak load, and standby energy consumption. It also uses invalid action masking to keep candidate opening and closing actions that do not belong to the set of actionable actions as unselectable. The action priority sequence is obtained by ranking the action values ​​of each candidate opening and closing action.

10. The reinforcement learning-driven low-voltage switchgear energy-saving load control method according to claim 1, characterized in that, Sending opening and closing commands to each branch includes: During the duration of the impact load event, the adjustable load branches are disconnected one by one, and the disconnection is stopped when the required unloading load is met or the three-phase voltage of the AC bus recovers to the allowable operating range; after the impact load event ends and the recovery conditions are met, the adjustable load branches are connected one by one, and the connection is suspended if the conditions for continued connection are not met. If no feedback from the controlled switch is received that matches the selected candidate opening and closing action, the new access action is stopped; if the reinforcement learning strategy model output maintains the candidate action or suspends the access action, the corresponding branch state is maintained; if the reinforcement learning strategy model output is empty, has an incorrect format, or does not belong to the set of possible actions, a deterministic action sequence is generated.

Images