Power flow optimization system and method based on multi-level coordination and multi-layer interconnection

Abstract

The application provides a power flow optimization system and method based on multi-level cooperation and multi-layer interconnection, and relates to the technical field of power systems.The system comprises a cloud master station layer for generating planned power curves of each distribution area based on short-term prediction data;an edge-side calculation layer for generating corresponding distribution area control set values according to the planned power curves; a physical device execution layer for collecting operation state information of the corresponding distribution area; a reinforcement learning control layer for generating original control instructions based on normal mode strategies according to the operation state information and the distribution area control set values, and dynamically adjusting the original control instructions based on fault perception strategies according to fault diagnosis information when a fault occurs; and a fault inspection and protection layer for executing a protection action when a fault occurs, continuously collecting operation state information of the distribution area, and generating fault diagnosis information.The application improves the system reliability and self-healing resilience of the distribution network.

Description

Technical Field

[0001] This application relates to the field of power system technology, specifically to a power flow optimization system based on multi-level cooperation and multi-layer interconnection, and a power flow optimization method based on multi-level cooperation and multi-layer interconnection. Background Technology

[0002] With the increasing penetration of distributed energy sources such as distributed photovoltaics, energy storage systems, and electric vehicle charging stations in distribution networks, the operating characteristics of distribution networks have undergone profound changes. These distributed resources are characterized by strong randomness, intermittency, and volatility, posing serious challenges to the safe operation of distribution networks. For example: the types of faults have become more complex, with the high proportion of distributed resources leading to a shift from traditional single overcurrent faults to multiple types of faults such as overvoltage, undervoltage, and commutation timeout; the time scales of faults have diversified, with overcurrent faults requiring microsecond-level responses, overvoltage and undervoltage requiring tens of microsecond-level responses, and commutation timeout requiring hundreds of microsecond-level responses, making it difficult for traditional single-time-scale protection schemes to handle all of them; and control and protection are separated, with power flow optimization control and fault detection protection typically separated in existing systems. Optimization control does not consider fault risks, and the protection system does not participate in optimization regulation, resulting in resource waste and response delays. Summary of the Invention

[0003] The purpose of this application is to provide a power flow optimization system and a power flow optimization method based on multi-level cooperation and multi-layer interconnection to solve the above-mentioned problems.

[0004] To achieve the above objectives, the first aspect of this application provides a power flow optimization system based on multi-level cooperation and multi-layer interconnection, the system comprising: The cloud master station layer is used to generate planned power curves for each distribution station area based on short-term forecast data. The side-side calculation layer is used to generate corresponding distribution area control setpoints based on the planned power curves of each distribution area. The physical device execution layer is used to collect the operating status information of the corresponding distribution transformer area. The operating status information includes at least one of node voltage, node current, distributed resource output, energy storage charge status and local load power. The reinforcement learning control layer is used to generate the original control commands of the physical device execution layer based on the normal mode strategy, the operating status information and the control settings of the distribution radio area. The fault detection and protection layer is used to correct the original control command based on the operating status information, and control the physical device execution layer to perform corresponding operations based on the correction operation result; when a fault in the distribution transformer area is identified based on the operating status information, the physical device execution layer is controlled to perform fault protection operations; and the operating status information of the distribution transformer area is continuously acquired, and fault diagnosis information is generated based on the operating status information acquired within a preset time window. The reinforcement learning control layer is also used to dynamically adjust the original control command based on the fault diagnosis information according to the fault perception strategy and then send it to the fault inspection and protection layer.

[0005] Optionally, the short-term forecast data includes at least one of the distributed resource output, load demand, and time-of-use pricing for each distribution substation; generating planned power curves for each distribution substation based on the short-term forecast data includes: For each distribution area: Based on the output of the distributed resources and the load demand, constraints representing the power supply and demand balance of the distribution substation area are constructed. Based on the output of the distributed resources and the time-of-use pricing, a first objective function is constructed under the constraint of satisfying the power supply and demand balance of the distribution substation, with the goal of minimizing the total operating economic cost consisting of the power interaction cost between the distribution substation and the external power grid and the operating consumption cost of each distributed resource within the substation during the scheduling cycle. Solving the first objective function yields the global optimal power flow of the distribution substation, generating the planned power curve of the distribution substation. The planned power curve is used to characterize the expected power interaction between the distribution substation and the external power grid during each scheduling period within the target planning cycle.

[0006] Optionally, the first objective function includes: First component and second component; The first component is determined based on the time-of-use electricity price and the power exchanged between the distribution substation and the external power grid during the current scheduling period. The first component is used to characterize the power interaction cost generated between the distribution substation and the external power grid during the current scheduling period. The second component is determined based on the output power of each distributed resource in the distribution area during the current scheduling period and the operating cost parameters of each distributed resource. The second component is used to characterize the operating cost incurred by each distributed resource in the distribution area during the current scheduling period. A first objective function is constructed with the goal of minimizing the sum of the first component and the second component over the target planning period.

[0007] Optionally, corresponding distribution area control setpoints are generated based on the planned power curves of each distribution area, including: For each distribution area: With the objectives of minimizing the deviation of the power exchanged between the distribution substation and the external power grid relative to the planned power curve at each time step within the preset prediction time domain, and minimizing the operating costs of each distributed resource within the distribution substation at the corresponding time step, a regional model predictive control objective function is constructed. Using the switching power and output of each distributed resource within the distribution area at each time step in the prediction time domain as decision variables, the predictive control objective function of the regional model is solved by rolling optimization under the condition of satisfying the preset regional power balance and equipment operation constraints. The switching power setting value and the output setting value of each distributed resource within the current time step in the predicted time domain obtained by solving are used as the control setting value of the distribution substation.

[0008] Optionally, the original control command includes at least one of energy storage charging and discharging power, photovoltaic reactive power regulation power, and flexible interconnection device transmission power; the original control command is modified based on the operating status information, including: If the original control command includes energy storage charging and discharging power, the current physical capacity limit of the energy storage device is determined based on the current energy storage state of charge collected by the physical device execution layer. Using the current physical capacity limit as a constraint, the energy storage charging and discharging power of the original control command is limited and corrected; or If the original control command includes photovoltaic reactive power regulation, the current physical upper limit of reactive power is determined based on the current active power output and apparent power rated capacity of the photovoltaic inverter collected by the physical device execution layer. Using this current physical upper limit of reactive power as a constraint, the photovoltaic reactive power regulation of the original control command is limited and corrected; or If the original control command includes the transmission power of the flexible interconnect device, the local load power collected by the physical device execution layer is obtained. If it is determined that the difference between the local load power and the transmission power of the flexible interconnect device exceeds the power supply margin of the current distribution substation, the transmission power of the flexible interconnect device is limited and corrected based on the power supply margin; or If the current node current collected by the physical device execution layer reaches a preset current warning threshold, a limiting correction or truncation operation is performed on the power increase command in the original control command; or If the current node voltage collected by the physical device execution layer reaches the preset voltage warning threshold, the command that causes the action direction of the original control command to be superimposed with the current voltage trend is intercepted or corrected in the opposite direction, based on the action direction of the original control command and the current voltage trend.

[0009] Optionally, identifying a distribution area fault based on the operating status information includes: If the node voltage is higher than the preset overvoltage threshold for a duration of a first duration, an overvoltage fault is determined to have occurred. If the duration for which the node current is higher than the preset overcurrent threshold reaches a second duration, an overcurrent fault is determined to have occurred. If the original control command is determined to be a commutation command for the converter based on the energy storage charging / discharging power or the transmission power of the flexible interconnect device, and no zero current signal is detected within the third duration, a commutation timeout fault is determined to have occurred.

[0010] Optionally, the operating status information of the distribution transformer area is continuously acquired, and fault diagnosis information is generated based on the operating status information acquired within a preset time window, including: The polling task is executed at a preset period to continuously obtain node voltage and node current, build and continuously update the voltage history queue and current history queue, and record the commutation operation result corresponding to each commutation command to build a commutation operation result history queue. The voltage change rate and the total harmonic distortion rate of the current are calculated based on the voltage history queue and the current history queue, and the commutation failure rate is calculated based on the commutation operation result history queue. If the obtained voltage change rate is greater than the preset voltage change rate warning threshold, or if the obtained total harmonic distortion rate of the current is greater than the preset total harmonic distortion rate of the current warning threshold, or if the obtained commutation failure rate is greater than the preset commutation failure rate warning threshold, fault diagnosis information is generated.

[0011] Optionally, based on the normal mode strategy, and according to the operating status information and the control settings of the distribution radio area, the original control commands for the physical device execution layer are generated, including: When the fault diagnosis information is not detected, the original control command is output through the pre-trained normal mode strategy model, taking the operating status information and the distribution radio area control setting value as input. The normal mode strategy model is obtained by training a specified neural network based on a preset second objective function. The steps for constructing the second objective function include: The actual electricity purchase cost is determined based on the deviation between the exchange power between the distribution substation and the external power grid and the corresponding exchange power set value, and the network power loss is determined based on the exchange power. An operating cost penalty item is constructed based on the actual electricity purchase cost and the network power loss. A voltage over-limit penalty term is constructed based on the matching relationship between the current node voltage and the preset voltage safety boundary; Based on the results of the fault detection and protection layer's correction operation on the original control command, a security layer intervention penalty item is constructed; The second objective function is obtained by weighting the operating cost penalty, voltage over-limit penalty, and safety layer intervention penalty based on the preset normal mode weight parameters.

[0012] Optionally, the matching relationship includes: If the current node voltage exceeds the voltage safety boundary, a voltage over-limit penalty value is determined based on the extent of its deviation from the lower limit or upper limit of the safety voltage boundary. If the current node voltage is within the safety voltage boundary, the voltage over-limit penalty value is determined to be a preset value. Based on the result of the fault detection and protection layer's correction operation on the original control command, a safety layer intervention penalty item is constructed, including: If it is determined, based on the result of the correction operation, that the fault detection and protection layer has corrected the original control command, the value of the safety layer intervention penalty item is set to the first value; If, based on the result of the correction operation, it is determined that the fault check and protection layer has not corrected the original control command, the value of the safety layer intervention penalty item is set to the second value.

[0013] Optionally, based on the fault perception strategy, the original control commands are dynamically adjusted according to the fault diagnosis information, including: When the fault diagnosis information is detected, the original control command is dynamically adjusted by taking the operating status information, the control setting value of the distribution radio area and the fault diagnosis information as inputs and the pre-trained fault perception strategy model as output. The fault perception strategy model is obtained by training a specified neural network based on a preset third objective function. The steps for constructing the third objective function include: An absolute penalty for safety violations is constructed based on the matching results of the current node voltage with the preset voltage safety boundary and the matching results of the current node current with the preset current threshold. A recovery time penalty term is constructed based on the fault duration, wherein the value of the recovery time penalty term is positively correlated with the fault duration; The network power loss penalty term is determined based on the exchange power between the distribution substation and the external power grid; The third objective function is obtained by weighting the absolute penalty for safety violations, the recovery time penalty, and the operational loss penalty based on the preset fault perception strategy weight parameters.

[0014] Optionally, an absolute penalty for safety violations is constructed based on the matching results of the current node voltage and the preset voltage safety boundary, and the matching results of the current node current and the preset current threshold, including: If the current node voltage exceeds the voltage safety boundary, or if the current node current is greater than the current threshold, the value of the absolute penalty for safety violation is determined to be a first value; otherwise, the value of the absolute penalty for safety violation is determined to be a second value.

[0015] Optionally, the edge computing layer is further used for: Based on the fault diagnosis information, and considering the regional topology network structure of each distribution substation corresponding to the edge computing layer and the current power flow status of each distribution substation, the minimum electrical isolation range, including the controllable switch, is determined. Send the disconnect command of the controllable switch to the physical device layer corresponding to the minimum electrical isolation range, generate a regional fault report, and send the regional fault report to the cloud master station layer; The cloud master station layer is also used for: Upon receiving the regional fault report, based on the global topology, the capacity of each line, and reactive power and voltage constraints, the power flow feasibility of each preset candidate network reconstruction path is verified, and a target path candidate set is constructed based on the verified candidate network reconstruction paths. With the goal of minimizing the impact of a fault, the optimal global recovery scheme is determined by optimizing each path in the candidate target path set based on a preset global recovery optimization objective function. This scheme includes network reconfiguration switching, backup power supply switching timing, and load transfer paths. Based on the optimal global recovery scheme, a global recovery command is generated and sent to each edge computing layer.

[0016] Optionally, the edge computing layer is further used for: Based on the global recovery instruction, control instructions are generated for each distribution radio zone corresponding to the current edge computing layer. The control instructions for each distribution radio zone are sent to the reinforcement learning control layer corresponding to each distribution radio zone, so that the reinforcement learning control layer corresponding to each distribution radio zone switches from the fault perception strategy to the normal mode strategy. Based on the normal mode strategy and the corresponding control instructions, the physical device execution layer is controlled to perform the corresponding operation.

[0017] Optionally, the fault detection and protection layer is further used for: If multiple consecutive commutation timeout faults are detected in the converter, the current pulse blocking delay is incremented according to the preset first dynamic gain ratio, and the incremented result is compared with the preset blocking delay limit value. The minimum value between the two is taken as the updated pulse blocking delay. If an overcurrent fault is detected, the switching dead time of the converter is incremented according to the preset second dynamic gain ratio, and the incremented result is compared with the preset physical limit dead time. The minimum value of the two is taken as the updated switching dead time. If multiple overvoltage faults are detected within a preset time period, the preset overvoltage threshold is lowered, or the hardware debouncing time for overvoltage detection is extended.

[0018] A second aspect of this application provides a power flow optimization method based on multi-level cooperation and multi-layer interconnection, applied to the power flow optimization system based on multi-level cooperation and multi-layer interconnection as described above, the method comprising: The cloud-based global optimization layer generates planned power curves for each distribution zone based on short-term forecast data. The side-side calculation layer generates corresponding control setpoints for each distribution area based on the planned power curve of each distribution area. The reinforcement learning control layer generates the original control commands for the physical device execution layer based on the normal mode strategy, the operating status information, and the control settings of the distribution area. The fault detection and protection layer corrects the original control command based on the operating status information, and controls the physical device execution layer to perform the corresponding operation based on the correction result; when a fault is identified in the distribution transformer area based on the operating status information, the physical device execution layer is controlled to perform a fault protection operation; and the operating status information of the distribution transformer area is continuously acquired, and fault diagnosis information is generated based on the operating status information acquired within a preset time window. In response to the received fault diagnosis information, the reinforcement learning control layer switches the normal mode strategy to a fault perception strategy. Based on the fault perception strategy, the original control command is dynamically adjusted according to the fault diagnosis information and then sent to the fault inspection and protection layer. The edge computing layer receives fault diagnosis information from each distribution station area. Based on the regional topology network structure and current power flow status of each distribution station area corresponding to the edge computing layer, it determines the minimum electrical isolation range including the controllable switch, sends the disconnection command of the controllable switch to the physical device layer corresponding to the minimum electrical isolation range, generates a regional fault report, and sends the regional fault report to the cloud master station layer. In response to the received regional fault report, the cloud master station layer determines the optimal global recovery scheme based on the global topology, the capacity of each line, and reactive power and voltage constraints. Based on the optimal global recovery scheme, it generates a global recovery command and sends the global recovery command to each side computing layer. The edge computing layer generates control instructions for each distribution radio zone corresponding to the current edge computing layer based on the global recovery instruction, and sends the control instructions for each distribution radio zone to the reinforcement learning control layer corresponding to each distribution radio zone, so that the reinforcement learning control layer corresponding to each distribution radio zone switches from the fault perception strategy to the normal mode strategy, and controls the physical device execution layer to perform corresponding operations based on the normal mode strategy and the corresponding control instructions.

[0019] This application constructs an optimized system with multi-level collaboration and multi-layer interconnection, deeply integrating control and protection. Each layer performs its own function and works collaboratively through standard interfaces. The reinforcement learning control layer generates original control commands based on state information and setpoints. The fault inspection and protection layer performs safety corrections, physical actions, and fault diagnosis feedback, and continuously polls and outputs diagnostic information to drive the reinforcement learning control layer to dynamically switch operating strategies, which greatly improves the power supply reliability and system self-healing resilience of the distribution network under complex faults.

[0020] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A schematic diagram of the system architecture of the optimized system for multi-level collaboration and multi-layer interconnection provided in the preferred embodiment of this application; Figure 2 A schematic diagram of the system workflow provided for a preferred embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0023] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0024] Currently, the main technical solutions for fault detection and protection control in power distribution networks include: (1) The traditional relay protection scheme works on the principle of hardware protection relays triggered by current and voltage thresholds, which trip and isolate faults after operation. Its disadvantages include: fixed action thresholds, which cannot adapt to changes in operation mode after the access of distributed resources; it can only achieve overcurrent protection and is insufficient for overvoltage, undervoltage, commutation timeout and other faults; manual recovery is required after operation, which affects the reliability of power supply.

[0025] (2) The fault detection scheme based on software algorithms works by polling the sampled data through software and using algorithms such as wavelet transform and neural networks to identify fault characteristics. Its disadvantages include: large software processing delay (usually >10ms), which makes it difficult to meet the microsecond-level fault response requirements; high algorithm complexity, which requires high computing power from the embedded controller.

[0026] (3) Distributed power flow optimization control scheme, the working principle of which is to realize the economic optimization operation of the distribution network through the cloud-edge-end hierarchical architecture. Its disadvantages include: the existing scheme does not consider fault detection and protection control, the optimization command may deteriorate the system operation under fault conditions; and it lacks a coordination mechanism with the underlying protection.

[0027] In summary, the existing technology has the following main technical defects: Time scale mismatch: There is a huge time gap between microsecond-level fault detection and second-level optimization control, and there is a lack of cross-scale coordination mechanism; Separation of control and protection: Power flow optimization control and fault detection protection are independent of each other, and optimization instructions may exacerbate the harm during a fault; Incomplete coverage of fault types: Traditional protection mainly targets overcurrent and is insufficient in responding to new faults after the integration of distributed power sources, such as overvoltage, undervoltage, and commutation timeout; Insufficient hardware and software coordination: Hardware detection is fast but the logic is simple, while software is intelligent but the response is slow, failing to give full play to their respective advantages; Single recovery strategy: Tripping and isolation are usually used after a fault, and there is a lack of intelligent recovery strategies based on hierarchical optimization.

[0028] To address the aforementioned problems, the first aspect of this application provides a power flow optimization system based on multi-level cooperation and multi-layer interconnection, the system comprising: The cloud master station layer is used to generate planned power curves for each distribution substation based on short-term forecast data; the edge computing layer is used to generate corresponding control setpoints for each distribution substation based on its planned power curves; the physical device execution layer is used to collect operating status information of the corresponding distribution substation, including at least one of node voltage, node current, distributed resource output, energy storage state of charge, and local load power; the reinforcement learning control layer is used to generate original control commands for the physical device execution layer based on the normal mode strategy, operating status information, and distribution substation control setpoints; the fault detection and protection layer is used to correct the original control commands based on the operating status information and control the physical device execution layer to perform corresponding operations based on the correction results; when a distribution substation fault is identified based on the operating status information, the physical device execution layer is controlled to perform fault protection operations; and the layer continuously acquires operating status information of the distribution substation and generates fault diagnosis information based on the operating status information acquired within a preset time window; the reinforcement learning control layer is also used to dynamically adjust the original control commands based on the fault diagnosis information according to the fault perception strategy and then send them to the fault detection and protection layer.

[0029] Thus, as Figure 1 As shown, this application constructs an optimized system with multi-level collaboration and multi-layer interconnection, deeply integrating control and protection. Each layer performs its own function and works collaboratively through standard interfaces. The reinforcement learning control layer generates original control commands based on state information and set values. The fault inspection and protection layer performs safety corrections, physical actions, and fault diagnosis feedback, and continuously polls and outputs diagnostic information to drive the reinforcement learning control layer to dynamically switch operating strategies, which greatly improves the power supply reliability and system self-healing resilience of the distribution network under complex faults.

[0030] Specifically, the five-layer collaborative control architecture of this application includes an L1 cloud master station layer, an L2 edge computing layer, an L3 reinforcement learning control layer, an L4 fault detection and protection layer, and an L5 physical device execution layer. In this application, the cloud master station layer can be composed of a cloud server cluster, with a timescale of hours. It is used for global optimal power flow calculation, day-ahead planning, network-wide fault diagnosis and global recovery strategy optimization, as well as receiving regional fault reports, network-wide fault location, global network reconstruction, and issuing recovery commands. There are multiple edge computing layers, each corresponding to multiple distribution substations. Each edge computing layer is composed of edge computing units, with a timescale of minutes. It is used for regional autonomy, tracking plans, multi-substation power mutual assistance, regional fault diagnosis and location, regional isolation strategies, regional recovery decisions, and reporting fault information. Each distribution substation deploys a reinforcement learning control layer, which is composed of local controllers with a timescale of seconds. It is used for economic optimization, security layer correction, receiving fault status, switching control modes, outputting fault perception actions, and adjusting optimization targets. Each distribution transformer substation deploys a fault detection and protection layer, composed of FPGA+ARM, with a timescale of microseconds / milliseconds. This layer is used for hardware interrupt detection (microsecond level): overvoltage / undervoltage / overcurrent / commutation timeout, priority coding, fault snapshot, software polling detection (millisecond level), trend analysis (dV / dt / THD), fault prediction, log recording, and protection execution: pulse blocking / crowbar / circuit breaker, rapid detection of four types of faults, emergency protection actions, fault snapshot recording, and early warning information generation. Each substation also deploys a physical device execution layer, composed of field devices such as photovoltaic inverters, energy storage PCS, SNOP / UPF, and charging piles, with a timescale of milliseconds. This layer is used for executing control commands, status feedback, sensor data acquisition, commutation execution, and status feedback; it serves as both the protected object and the source of the fault. This application's system organically integrates microsecond-level hardware interrupt detection, millisecond-level software polling, and second-level optimized control, forming a complete fault response chain.

[0031] like Figure 2 As shown, in stage 0, i.e., the system is operating normally, the cloud master station layer acquires ultra-short-term forecast data of the entire distribution network on an hourly basis, including the photovoltaic output forecast, load demand forecast, and external grid time-of-use tariff for each distribution substation. Then, for each distribution substation, a planned power curve is generated based on the short-term forecast data, specifically including: Step 1: Construct constraints representing the power supply and demand balance of a distribution substation based on distributed resource output and load demand. For example, the constraints for the power supply and demand balance of a distribution substation can be expressed as: ,in, To optimize the decision variables, let represent the main grid exchange power, that is, the power absorbed by the distribution substation from the upper-level main grid at time t. Positive values ​​represent purchasing electricity from the main grid, and negative values ​​represent selling electricity to the main grid. The range is from negative values ​​to transformer capacity. This represents the output of the i-th distributed resource, such as the active power generated by photovoltaics, energy storage, or micro-gas turbines. Positive values ​​represent power generation, while negative values ​​represent load (such as energy storage charging). M is the total number of resources. Total load demand represents the total demand of all electrical equipment (critical loads, non-critical loads, etc.) within the transformer area at time t; This refers to network line loss, which is the power lost as heat is generated when power flows through lines and transformers.

[0032] Step 2: Based on the output of distributed resources and time-of-use pricing, construct a first objective function that minimizes the total economic cost of operation, consisting of the power interaction cost between the distribution substation and the external power grid and the operating costs of each distributed resource within the substation during the dispatching cycle, under the constraint of satisfying the power supply and demand balance of the distribution substation.

[0033] The first objective function includes a first component and a second component. The first component is determined based on time-of-use pricing and the power exchanged between the distribution transformer area and the external power grid during the current scheduling period. The first component characterizes the power interaction cost between the distribution transformer area and the external power grid during the current scheduling period, and can be expressed as follows: ,in, The first component represents the electricity purchase price at time t, the price at which electricity is purchased from the upstream main grid, typically a time-of-use price, with prices varying at different times. The second component is determined based on the output power of each distributed resource within the distribution transformer area during the current dispatch period and the operating cost parameters of each distributed resource. The second component characterizes the operating costs incurred by each distributed resource within the distribution transformer area during the current dispatch period, and can be expressed as: Where M represents the total number of distributed resources, P i (t) represents the output of the i-th resource at time t, active power; positive indicates output, negative indicates absorption (e.g., energy storage charging). i ( Let represent the cost function of the i-th resource; the first objective function is constructed with the objective of minimizing the sum of the first and second components over the target planning period. Specifically, the first objective function can be expressed as: Where t is the time index, representing the t-th time period in the optimization time domain, and T is the total number of time periods, the total number of time periods included in the planning cycle. For example, the daily plan T=24 (hours), and the intraday plan T=96 (15 minutes).

[0034] Step 3: Solve the first objective function to obtain the globally optimal power flow of the distribution substation, generating the planned power curve for the distribution substation. The planned power curve represents the expected power exchange between the distribution substation and the external power grid during each scheduling period within the target planning cycle. By solving the objective function to obtain the optimal power flow, the planned exchange power curve P for each distribution substation for the next 24 hours is generated. grid (t) distributes the planned curve and pre-trained reinforcement learning model parameters to each side fusion terminal.

[0035] After receiving the planned power curve, the side-side calculation layer solves the regional model predictive control problem at a period of 5-15 minutes. Based on the planned power curve of each distribution area, it generates the corresponding control setpoints for each distribution area. For each distribution area, the specific steps include: Step 1: With the objectives of minimizing the deviation of the power exchanged between the distribution substation and the external power grid relative to the planned power curve at each time step within the preset prediction time domain, and minimizing the operating costs of each distributed resource within the distribution substation at the corresponding time step, a regional model predictive control objective function is constructed. The objective function can be expressed as: Where H represents the prediction time domain length, such as the next 10 control cycles or the next 5-15 minutes; k represents the step size, i.e., the sampling period, for example, k=0 represents the current time. Indicates the next 5-15 minutes. This indicates the next 5-15 minutes, and so on. P represents the expected planned switching power issued by the cloud. grid (t+k) represents the actual exchange power at this step size; λ is the weighting coefficient.

[0036] Step 2: Using the switching power and output of each distributed resource within the distribution area at each time step in the prediction time domain as decision variables, the predictive control objective function of the regional model is solved by rolling optimization under the conditions of satisfying the preset regional power balance and equipment operation constraints.

[0037] For example, constraints may include: The output power of each distributed power source within the distribution area must be limited to between its preset minimum physical output and maximum physical output, i.e. ; Within the distribution area, the state of charge of energy storage devices at each step in the predicted time domain must be maintained between the preset safe charging upper and lower limits to avoid overcharging or over-discharging. ); The power flowing through each feeder within the distribution transformer area must be less than the thermal stability limit (current carrying capacity limit) of that line, i.e. ; The power exchange between this distribution substation and the external main grid must be strictly limited within the boundaries of the transformer's rated capacity and reverse power transmission capability. The operating voltage of each node in this distribution substation area must be maintained within a preset safe operating voltage range, such as 0.95~1.05 pu.

[0038] Step 3: Use the obtained switching power setpoint and output setpoint of each distributed resource at the current time step within the prediction time domain as the control setpoint for the distribution substation. It is understandable that in the mathematical optimization model of Model Predictive Control (MPC), the decision variables (control variables) are the sequence of control quantities at each step within the entire prediction time domain H. In this application, the decision variables of MPC include the actual main grid switching power. The sequence of each step in the prediction time domain, and the output of each internal distributed resource. The sequence at each step within the prediction time domain. The L2 layer MPC algorithm, after solving this sequence, only takes the current time step (…). The power and voltage setpoints are sent to the L3 layer, and the calculation is repeated in the next cycle. The voltage setpoint is the optimal voltage target calculated by the L2 solver when calculating the power setpoint, after simultaneously satisfying the power flow equations and voltage safety boundaries.

[0039] In this application, the reinforcement learning control layer reads the current state s collected by the physical device execution layer at a 1-second interval. t (Node voltage, photovoltaic output, load power, energy storage SOC, time characteristics, etc.) In normal mode, with power setpoint and voltage setpoint as inputs and economic efficiency as the goal, the original action is output through forward inference of a pre-trained neural network. , where π normal (s t ) represents a pre-trained neural network, i.e., a normal pattern policy model, a raw This refers to the initial action, i.e., the initial control command, including the energy storage charging and discharging power P. ess Photovoltaic reactive power Q pv Flexible interconnect device transmission power P snop The generated raw actions are then sent to the L4 fault detection and protection layer.

[0040] The fault detection and protection layer performs hardware sampling at the microsecond / millisecond level and software polling trend analysis at 1ms. The software part runs a polling task at a 1ms cycle to perform trend analysis, including but not limited to: updating the voltage history queue (data from the most recent 100ms), updating the current history queue (data from the most recent 100ms), calculating the voltage change rate dV / dt, calculating the current harmonic distortion rate THD, and calculating the commutation failure rate (for the converter).

[0041] The fault detection and protection layer corrects the received original control commands based on the collected operating status information, including: If the original control command includes energy storage charging and discharging power, the current physical capacity limit of the energy storage device is determined based on the current state of charge of the energy storage collected by the physical device execution layer. Using this current physical capacity limit as a constraint, the energy storage charging and discharging power of the original control command is limited and corrected. For example, the original control command is corrected to the current physical capacity limit of the energy storage device; or... If the original control command includes photovoltaic reactive power regulation, the current physical upper limit of reactive power is determined based on the current active power output and apparent power rated capacity of the photovoltaic inverter collected by the physical device execution layer. Using this physical upper limit as a constraint, the photovoltaic reactive power regulation in the original control command is limited and corrected. For example, the original control command is corrected to the current physical upper limit of reactive power of the photovoltaic inverter; or... If the original control command includes the transmission power of the flexible interconnect device, the local load power collected by the physical device execution layer is obtained. If it is determined that the difference between the local load power and the transmission power of the flexible interconnect device exceeds the power supply margin of the current distribution substation, the transmission power of the flexible interconnect device is limited and corrected according to the power supply margin. For example, if the original control command is for the flexible interconnect device to transmit 300kW of power from substation A to substation B, after receiving the original control command, L4 directly reads the preset rated power limit from the hardware configuration file of the device. For example, if the rated capacity of the device is 200kW, then L4 forcibly intercepts the command and corrects it to 200kW; or If the current node current collected by the physical device execution layer reaches the preset current warning threshold, the power increase command in the original control command will be limited or truncated. For example, if the current node current exceeds the preset current warning threshold, and the original control command contains control commands such as increasing the cross-regional transmission power of flexible interconnection devices, increasing the discharge power of energy storage, or significantly increasing the reactive power of photovoltaic inverters to support voltage, these power increase commands will be intercepted or the power increase value will be reduced; or If the current node voltage collected by the physical device execution layer reaches the preset voltage warning threshold, the command that causes the action direction of the original control command and the current voltage trend to be superimposed in the same direction will be intercepted or corrected in the opposite direction, based on the action direction of the original control command and the current voltage trend. For example, if the original control command includes the inverter outputting capacitive reactive power, and the current node voltage has reached the voltage warning threshold such as 1.04 pu, approaching the 1.05 overvoltage hardware interrupt threshold, and the voltage is still rising, then the original control command and the current voltage trend are superimposed in the same direction, which will trigger an overvoltage hardware interrupt. At this time, the control command for the inverter to output capacitive reactive power will be intercepted or corrected in the opposite direction, for example, by correcting it to the inverter outputting inductive reactive power to pull the node voltage back to the safe zone.

[0042] The L4 fault detection and protection layer sends verified safety action commands to the L5 physical device execution layer. L5 executes control through PWM regulation, switching actions, etc., and samples physical quantities such as voltage, current, and temperature in real time using sensors, feeding the measured data back to the L3 and L4 layers. Simultaneously, L4 records the correction or interception status as a safety layer intervention state. (0 if not intercepted, 1 if intercepted), to be fed back to L3 for reinforcement learning online / offline penalty updates.

[0043] In Phase 1, the fault occurrence phase (t=0μs), electrical quantities change abruptly due to distributed resource fluctuations, grid disturbances, or equipment anomalies. For example, a sudden drop in photovoltaic power can cause a voltage slump, a line overload can cause a current rise, or an abnormal commutation process can occur. The fault electrical quantities are transmitted to the L4 layer in real time via sensors.

[0044] In Phase 2, the hardware interrupt detection phase (t=0-50μs), L4 identifies faults by comparing node voltage, current, and other data collected in real time by L5. Specifically, the fault detection and protection layer identifies distribution area faults based on operational status information, including: Overvoltage and undervoltage detection: If the node voltage is higher than the preset overvoltage threshold for a duration of a first duration, an overvoltage fault is determined. For example, L4 inputs the collected voltage sample value into the hardware comparator and compares it with the preset threshold (1.05 times the rated voltage). If the comparator trigger time is <1μs, a 20μs hardware debouncing counter is started to prevent false triggering. If overvoltage is still detected after 20μs, the overvoltage flag is set. As another example, when performing undervoltage detection, the voltage sample value is input into the hardware comparator and compared with the preset threshold (0.95 times the rated voltage). If the comparator trigger time is <1μs, a 20μs hardware debouncing counter is started. If undervoltage is still detected after 20μs, the undervoltage flag is set.

[0045] Overcurrent detection: If the node current exceeds the preset overcurrent threshold for a duration of a second duration, an overcurrent fault is determined to have occurred. For example, the current sample value is input into a hardware comparator and compared with a preset threshold (such as 1.2 times the rated current). If the comparator trigger time is <1μs, a 1μs hardware debouncing counter is started (overcurrent is more dangerous and the response is faster). If overcurrent is still detected after 1μs, the overcurrent flag is set.

[0046] Commutation timeout detection: If the original control command is determined to be a commutation command for the converter based on the energy storage charging / discharging power or the transmission power of the flexible interconnect device, and no zero current signal is detected within the third duration, a commutation timeout fault is determined to have occurred. For example, the commutation start signal triggers the hardware state machine to enter the COMMUTATING state from the IDLE state, and starts a 100μs hardware timer to start counting. If a zero current signal is detected within 100μs, the state machine returns to IDLE, and the commutation is normal; if no zero current is detected within 100μs, the state machine enters the TIMEOUT state, the commutation timeout flag is set, and a timeout interrupt signal is generated.

[0047] In this application, interrupt priorities can be predetermined. For example, the interrupt priority encoding can be set to: overcurrent > overvoltage > undervoltage > commutation timeout. All fault flags are input to the interrupt priority encoder, with the priority order (from high to low) as: overcurrent > overvoltage > undervoltage > commutation timeout. The encoder outputs the highest priority interrupt request and sends a hardware interrupt signal IRQ_Fault to the ARM processor. The interrupt response time is <10μs. For example, if the hardware interrupt priority encoder recognizes the overcurrent flag as the highest priority, it immediately outputs the highest priority interrupt request IRQ signal to the L4 ARM processor. Simultaneously, a fault snapshot is saved according to timestamp + voltage + current + fault type. At the moment a fault occurs, the hardware circuit automatically stores the 64-bit timestamp, 12-bit current sample value, 12-bit voltage sample value, 4-bit fault type code, and 8-bit device status lock at that instant into a FIFO register for software retrieval.

[0048] In Phase 3, the emergency protection execution phase (t=50-100μs), L4 executes corresponding emergency protection actions according to the fault type. Specifically, for overcurrent faults, pulse blocking protection is executed (response time <1μs), immediately blocking all IGBT trigger pulses and disconnecting the main circuit; for overvoltage faults, Crowbar protection is executed (response time <10μs), activating the Crowbar thyristor protection circuit to consume overvoltage energy; for undervoltage faults, power reduction is executed (response time <50μs), reducing the power setpoint to 50% to reduce the impact on the power grid; for commutation timeout faults, forced commutation is executed (response time <100μs), forcibly completing the commutation process and recording the timeout event.

[0049] The pulse blocking circuit operates as follows: the hardware-level pulse blocking signal is directly connected to the IGBT drive circuit, unaffected by software interference, ensuring absolute reliability. The blocking signal remains until the software confirms safety and resets. The crowbar protection circuit operates as follows: the thyristor quickly turns on, diverting overvoltage energy to the energy-dissipating resistor, protecting power electronic devices from overvoltage breakdown, and works in conjunction with the pulse blocking circuit.

[0050] Phase 4, the software processing phase (t=1-5ms), includes steps such as responding to the IRQ, reading the fault snapshot register, fault classification and recording, sending fault status notification to L3, reporting fault information to L2, and polling the task to continue running for 1ms.

[0051] The interrupt service routine responds to the IRQ as follows: The ARM processor responds to the hardware interrupt IRQ_Fault, reads the fault type register (address 0x00) and the fault snapshot register (address 0x04) through the AXI-Lite bus, obtains detailed fault information, performs emergency protection confirmation (if it has already been performed by the hardware, then confirmation is sufficient), wakes up the fault handling task (through the task notification mechanism), clears the interrupt flag, and exits the interrupt service routine.

[0052] Fault classification and logging include obtaining detailed fault information after being woken up by an interrupt.

[0053] Fault classification and identification include: single-phase fault / three-phase fault, transient fault / permanent fault, fault severity assessment, etc.

[0054] Fault recording and log storage includes: writing fault snapshots to Flash storage, recording the fault occurrence time, type, electrical quantity, and updating fault statistics counts.

[0055] Among them, the 1ms polling task continues to run without interruption. It continues to perform trend analysis, updates the historical queue based on the data after the fault, and continuously performs fault prediction and early warning judgment.

[0056] In this application, the fault detection and protection layer continuously acquires the operating status information of the distribution transformer area, and generates fault diagnosis information based on the operating status information acquired within a preset time window. This includes: executing a polling task at a preset cycle to continuously acquire node voltage and node current, constructing and continuously updating the voltage history queue and current history queue, and recording the commutation operation result corresponding to each commutation command, and constructing a commutation operation result history queue; calculating the voltage change rate and current total harmonic distortion rate based on the voltage history queue and current history queue, and calculating the commutation failure rate based on the commutation operation result history queue; generating fault diagnosis information if the obtained voltage change rate is greater than a preset voltage change rate warning threshold, or if the obtained current total harmonic distortion rate is greater than a preset current total harmonic distortion rate warning threshold, or if the obtained commutation failure rate is greater than a preset commutation failure rate warning threshold.

[0057] Specifically, an independent 1ms-period polling task in L4 continuously runs, constantly reading real-time sampled values, maintaining a voltage / current history queue for the most recent 100ms, and calculating the voltage change rate. The total harmonic distortion (THD) and commutation failure rate of the current were analyzed and determined to determine whether the fault was transient or permanent. Among them, the voltage change rate is calculated from the latest sampled value and the sampled value of the previous millisecond in the voltage history queue, and is used to reflect the severity of the voltage drop; the total harmonic distortion rate of the current is calculated by extracting the most recent 20 sampling points from the current history queue (based on the 50Hz power grid frequency, the 1ms sampling rate just covers a complete power frequency cycle of 20ms), performing spectrum analysis on the extracted 20 points based on Fast Fourier Transform (FFT), extracting the amplitude of the fundamental component and the amplitude of each harmonic component, and calculating it using the existing THD calculation formula, which is used to reflect the purity of the current waveform and predict whether there is abnormal nonlinear load, resonance or insulation degradation of the equipment; the converter commutation failure rate is calculated by counting the number of commutation timeouts reported by the converter hardware state machine in the most recent 100 consecutive commutation operations. For example, if in the most recent 100 IGBT current commutation operations, 6 operations triggered hardware timeout records because zero current was not detected within 100μs, then the commutation failure rate = 6 / 100 = 6%.

[0058] L4 aggregates hardware snapshots and software diagnostic results, generates fault diagnosis information (including fault type, severity level, and electrical quantity snapshots), and sends it to the L3 reinforcement learning optimization control layer.

[0059] In Phase 5, the fault perception and control phase (t=5-10ms), L3 receives the fault status, receives the fault status notification from L4, obtains information such as fault type, severity, and fault snapshot, and switches from normal mode to fault perception mode.

[0060] In normal mode, the reinforcement learning control layer generates the original control commands for the physical device execution layer based on the normal mode strategy, according to the operating status information and the control setpoint of the distribution station area. This includes: when no fault diagnosis information is detected, the original control commands are output through the pre-trained normal mode strategy model, with the operating status information and the control setpoint of the distribution station area as input. The normal mode strategy model is obtained by training a specified neural network based on a preset second objective function. The steps for constructing the second objective function include: Step 1: Determine the actual electricity purchase cost based on the deviation between the exchange power between the distribution substation and the external power grid and the corresponding setpoint for the exchange power. Also determine the network power loss based on the exchange power. Construct an operating cost penalty term based on the actual electricity purchase cost and the network power loss. The operating cost penalty term can be expressed as: [α Loss t +β C grid,t ], α is the network loss weighting coefficient, representing the penalty intensity per unit of network loss, with a value ranging from 0.1 to 0.5; Loss t Let be the network line loss at time t, representing the power loss caused by current flowing through the line; β is the electricity purchase cost weighting coefficient, ranging from 0.1 to 0.5, representing the penalty intensity per unit of electricity purchase cost; C grid,t C represents the cost of purchasing electricity at time t. grid,t =c grid (t) P grid (t), c grid (t) represents the time-of-use electricity price at time t (yuan / kWh), P grid (t) represents the power (kW) purchased from the main grid at time t. For areas with severe grid losses, α>β, so priority is given to reducing grid losses; if the electricity price fluctuates greatly, then α<β, so priority is given to peak-valley arbitrage. For balanced optimization, α≈β, taking both into account. Typical values: α=0.3, β=0.3.

[0061] Step 2: Construct a voltage over-limit penalty term based on the matching relationship between the current node voltage and the preset voltage safety boundary. The matching relationship includes: if the current node voltage exceeds the voltage safety boundary, determine the voltage over-limit penalty value based on the deviation of its deviation from the lower or upper limit of the safety voltage boundary; if the current node voltage is within the safety voltage boundary, determine the voltage over-limit penalty term as the preset value.

[0062] Specifically, the voltage over-limit penalty term can be expressed as: Where γ is the voltage safety weighting coefficient, typically 5-20, representing the overall strength of the voltage over-limit penalty. It increases when the safety requirements are high, and the weighting relationship is γ>α+β to ensure that safety takes precedence over economy; N is the total number of system nodes, i.e., the total number of distribution substations; P(Vi,t) is the voltage over-limit penalty function for node i, which is a piecewise function, with the penalty increasing as the boundary moves further away.

[0063] Where V represents the current node voltage, Vmin = 0.95pu (per unit), Vmax = 1.05pu.

[0064] When the voltage is within the safe range, the penalty is 0. When the voltage exceeds the limit, the penalty is proportional to the square of the over-limit magnitude. The further away from the boundary, the greater the penalty, thus forming a safe potential field.

[0065] Step 3: Construct a safety layer intervention penalty item based on the result of the fault detection and protection layer's correction operation of the original control command, including: if it is determined from the correction operation result that the fault detection and protection layer has corrected the original control command, set the value of the safety layer intervention penalty item to the first value; if it is determined from the correction operation result that the fault detection and protection layer has not corrected the original control command, set the value of the safety layer intervention penalty item to the second value.

[0066] Specifically, the security layer intervention penalty term can be expressed as η. Here, η is the intervention penalty weight coefficient, typically ranging from 0.1 to 1.0, representing the intensity of the penalty for intervention at the safety layer. It can be dynamically adjusted during the training phase. For example, in the early stages of training, η is relatively small, ranging from 0.1 to 0.3, to allow exploration; in the middle stages of training, η is relatively medium, ranging from 0.3 to 0.6, to guide the learning of safety boundaries; and in the later stages of training, η is relatively large, ranging from 0.6 to 1.0, to achieve refined control and reduce intervention.

[0067] II. `intervention` is the intervention indication function, which takes the value 0 or 1. 1 indicates intervention by the safety layer. For example, if the fault detection and protection layer modifies the original control command, it is set to 1. The intervention indication function is expressed as:

[0068] Whenever the security layer intervenes to correct an action, it imposes additional penalties, thereby encouraging the agent to autonomously learn the security boundary and reduce its dependence on the security layer. This is the core mechanism for achieving security awareness training.

[0069] Step 4: Based on preset normal mode weight parameters, the operating cost penalty, voltage over-limit penalty, and safety layer intervention penalty are weighted and combined to obtain the second objective function, which can be expressed as follows: , where r t This represents the total penalty value of the normal mode policy model.

[0070] During the training of the normal mode strategy model, the second objective function is used as the optimization objective, the collected operating status information and distribution station control settings are used as the input to the model, and the original control commands are used as the output to train neural networks such as convolutional neural networks and temporal neural networks. The model parameters are adjusted according to the value of the second objective function. After the algorithm converges, the normal mode strategy model is obtained. The training samples can be constructed from the actual historical data collected or from the relevant data generated by simulation. There are no restrictions on this here.

[0071] When L3 receives a fault status notification from L4, it obtains information such as fault type, severity, and fault snapshot through fault diagnosis information, and switches from normal mode to fault awareness mode, adjusting the optimization objective to prioritize safety. Specifically, based on the fault awareness strategy, the original control commands are dynamically adjusted according to the fault diagnosis information. This includes: when fault diagnosis information is detected, taking operating status information, distribution area control setpoints, and fault diagnosis information as input, and outputting dynamically adjusted original control commands through a pre-trained fault awareness strategy model; the fault awareness strategy model is obtained by training a specified neural network based on a preset third objective function. The steps for constructing the third objective function include: Step 1: Construct an absolute penalty for safety violations based on the matching results of the current node voltage and the preset voltage safety boundary and the matching results of the current node current and the preset current threshold. This includes: if the current node voltage exceeds the voltage safety boundary, or if the current node current is greater than the current threshold, determine the value of the absolute penalty for safety violations as the first value; otherwise, determine the value of the absolute penalty for safety violations as the second value.

[0072] Specifically, the absolute penalty for safety violations can be expressed as: λ safe 1 unsafe , where λ safe This is the safety weighting coefficient, representing the intensity of punishment for unsafe behavior, with a value ranging from 100 to 1000; 1 unsafe This is an unsafe state indicator function, taking the value 0 or 1, where 1 indicates unsafety and 0 indicates safety. The function can be represented as: , where V i V represents the current node voltage of the i-th node. min V represents the lower limit of the voltage safety boundary. max I represents the upper limit of the voltage safety boundary. i I represents the current node current of the i-th node. max This indicates the current threshold.

[0073] In normal mode, safety penalties allow for minor overstepping of limits, but impose continuous penalties to guide the agent away from the boundary; while in fault mode, no unsafe state is allowed, and severe penalties are imposed once it occurs. The safety penalty design of this application ensures that the agent prioritizes safety over economy during fault periods.

[0074] Step 2: Construct a recovery time penalty term based on the fault duration, where the value of the recovery time penalty term is positively correlated with the fault duration. Specifically, the recovery time penalty term can be expressed as... λ recovery t recovery , where λ recovery The recovery time weighting coefficient represents the penalty for each second of extended recovery time, with a value ranging from 1 to 10; t recovery The fault duration represents the time from the occurrence of a fault to the restoration of normal operation. It starts from the time the fault flag is set until all electrical quantities return to normal and the fault flag is cleared. It includes: protection action time + fault isolation time + recovery control time. It encourages intelligent agents and control systems to eliminate faults and restore normal operation as soon as possible. The longer the fault duration, the greater the impact on power supply to users and the heavier the penalty. It is directly related to power system reliability indicators (such as SAIDI). Typical values ​​for different fault levels are shown in Table 1.

[0075] Table 1

[0076] Step 3: Determine the network power loss penalty term based on the exchange power between the distribution substation and the external power grid. Specifically, the network power loss penalty term can be expressed as... λloss Loss t Where λloss is the loss weighting coefficient, ranging from 0.1 to 0.5, representing the penalty intensity per unit loss; Loss t The operating losses at time t include network losses and equipment losses. Compared to the normal mode, the failure mode still considers economic efficiency, but with a reduced weight. Under the premise of ensuring safety and rapid recovery, losses are minimized as much as possible, thereby avoiding unreasonable control (such as over-adjustment) caused by completely ignoring economic efficiency.

[0077] Step 4: Based on the preset fault perception strategy weight parameters, the absolute penalty for safety violations, the penalty for recovery time, and the penalty for operational losses are weighted and combined to obtain the third objective function, which can be specifically expressed as follows: ,in, This represents the total penalty value of the fault perception strategy model.

[0078] Similarly, during the training of the fault perception strategy model, the third objective function is used as the optimization objective. The collected operating status information, distribution station control setpoints, and fault diagnosis information are used as the model inputs, and fault perception control commands are used as the model outputs. Neural networks such as convolutional neural networks and temporal neural networks are trained. The model parameters are adjusted according to the value of the third objective function. After the algorithm converges, the fault perception strategy model is obtained. Training samples can be constructed from actual historical data or relevant data generated through simulation; this is not limited here. Understandably, in the offline training phase, the optimization objective is to minimize cost. That is, when updating gradients, the policy network will move towards actions with larger values ​​(i.e., smaller penalty costs), thereby achieving, in engineering terms, minimizing the system's operating cost and physical loss while satisfying safety boundaries. Understandably, model training can be performed in L1, with L1 distributing the pre-trained reinforcement learning model parameters to L3.

[0079] After invoking the fault perception strategy model, L3 reads the current state s. t (Including fault indicators), using the fault-aware policy network π fault Output control actions tend to reduce power and increase reactive power support to avoid aggravating the fault. The fault detection action is sent to the L4 layer, which performs instruction verification and correction operations again. The L4 then sends the corrected safety physical action instructions to the L5 layer to execute physical protection actions (PWM modulation, energy storage network support, etc.). At the same time, the L3 layer stores the fault experience in the experience pool and updates the policy network in the background to learn the optimal response strategy under fault conditions.

[0080] In phase 6, L2 receives fault diagnosis information from one or more L4 layers and performs regional fault diagnosis (t=10-100ms). The edge computing layer is also used to: determine the minimum electrical isolation range, including controllable switches, based on fault diagnosis information, the regional topology network structure of each distribution substation corresponding to the edge computing layer, and the current power flow status of each distribution substation. Specifically, it performs comprehensive analysis based on multiple fault reports, uses regional power grid topology information to determine the location of the fault, and distinguishes the fault source (equipment fault / line fault / external fault); and issues a disconnect command for the controllable switch to the physical equipment layer corresponding to the minimum electrical isolation range. Specifically, it determines whether electrical isolation is required based on the fault location and severity. If isolation is required, it determines the isolation range (minimum isolation area), generates an isolation command (such as disconnecting the corresponding switch), and issues the isolation command to the affected L4 layer. For example, if the fault is only in a local branch, L2 issues a command to disconnect only the circuit breaker of that branch, generates a regional fault report, and sends the regional fault report to the cloud master station layer. For example, the regional fault report includes assessing the impact of the fault on regional operation, formulating temporary operation strategies (such as switching power supply paths, activating backup power), and considering recovery time and economic trade-offs.

[0081] In phase 7, the cloud master station layer L1 receives regional fault reports from L2 and performs network-wide fault diagnosis and global recovery strategy optimization. Specifically, in this application, the cloud master station layer is also used for: Upon receiving regional fault reports, based on the global topology, line capacity, and reactive power and voltage constraints, the system performs power flow feasibility verification on each of the pre-defined candidate network reconfiguration paths. Based on the verified candidate paths, a target path candidate set is constructed. For example, multiple recovery paths are extracted from the pre-defined candidate network reconfiguration path library, such as path A (cross-regional mutual assistance recovery) and path B (high-voltage backup grid transfer recovery). Power flow simulation calculations are performed on each path to verify whether the current flowing through the line is less than the thermal stability limit and whether the voltage of each node is within the 0.95~1.05 pu safety range, eliminating paths that do not meet the constraints. Furthermore, the system comprehensively analyzes faults in each region, identifies cross-regional fault modes, determines the impact of faults on the safe and stable operation of the entire network, and assesses the fault severity level, such as Level I / II / III / IV.

[0082] After obtaining the target path candidate set, L1 further optimizes each path in the candidate set based on a preset global recovery optimization objective function, aiming to minimize the impact of the fault. This determines the optimal global recovery scheme, including network reconfiguration switching, backup power supply switching timing, and load transfer paths. A global recovery command is then generated based on the optimal global recovery scheme and sent to each edge computing layer. For feasible candidate paths that pass verification, their estimated recovery time, load loss, and other fault impact indicators are calculated. The optimal global recovery scheme is selected according to the principle of minimizing the overall fault impact. Specifically, in this application, the global recovery optimization objective function can be expressed as: The optimization variables are network reconfiguration scheme, backup power supply switching, and load transfer path; the constraints are power flow constraints, equipment capacity constraints, and voltage constraints. After generating a global recovery command, L1 sends a global recovery command to the relevant L2, including the network reconfiguration scheme, backup power supply switching sequence, load recovery priority, etc., and updates the planned curves for subsequent time periods.

[0083] In Phase 8, the edge computing layer receives the global recovery command issued by L1, decomposes it into a region-level recovery action sequence, generates control commands for each distribution station corresponding to the current edge computing layer, and sends the control commands for each distribution station to the reinforcement learning control layer corresponding to each distribution station. Upon receiving the control commands from L2, the reinforcement learning control layer corresponding to each distribution station switches from a fault-aware strategy to a normal mode strategy, calls the normal mode strategy model to output control commands to control the physical device execution layer to perform the corresponding operations. Upon receiving the recovery control command, the corresponding physical device execution layer gradually restores to the normal power level and feeds back the status data during the recovery process.

[0084] The fault detection and protection layer dynamically updates the protection parameters based on the received fault diagnosis information. Specifically, the fault detection and protection layer is also used to: if multiple consecutive commutation timeout faults are detected in the converter, increment the current pulse blocking delay according to a preset first dynamic gain ratio, compare the increment calculation result with a preset blocking delay limit value, and take the minimum value of the two as the updated pulse blocking delay. For example, if there are 3 consecutive commutation timeouts, increase the blocking delay T. block =min(T block ×1.2, 50μs); If an overcurrent fault is detected, the switching dead time of the converter is incremented according to the preset second dynamic gain ratio, and the incremented result is compared with the preset physical limit dead time. The minimum value between the two is taken as the updated switching dead time. For example, if circulating current is detected, the switching dead time T is increased. dead =min(T dead×1.2, 100μs); If multiple overvoltage faults are detected within the preset time period, reduce the preset overvoltage threshold or extend the hardware debouncing time for overvoltage detection. For example, if overvoltage is frequent, appropriately reduce the voltage threshold or increase the debouncing time. After updating the protection parameters, clear the fault flag and prepare for the next protection.

[0085] In Phase 9, all levels return to normal operation, continuing to optimize the control loop layer by layer, and the protection parameters are optimized based on fault experience. After fault recovery, the cloud master station layer updates the prediction model based on fault data, the edge computing layer updates the regional fault diagnosis rules, the reinforcement learning control layer uses fault experience for strategy optimization, the fault inspection and protection layer executes the protection strategy using the updated protection parameters, and the physical device execution layer continues to provide status feedback.

[0086] Thus, the system of this application forms a complete time chain by progressively advancing the time scale, from microsecond-level hardware response to hour-level global optimization; each layer collaborates in a hierarchical manner, with each layer performing its specific function and working together through standard interfaces; control and protection are integrated, with optimized control and fault protection deeply integrated to avoid separation; intelligent self-adaptation is achieved by dynamically adjusting parameters based on fault history and continuously optimizing; and a hierarchical recovery strategy can minimize the impact of faults. The system of this application can achieve optimal economic operation while ensuring absolute safety, and possesses self-learning and self-adaptive capabilities.

[0087] A second aspect of this application provides a power flow optimization method based on multi-level cooperation and multi-layer interconnection, applied to the power flow optimization system based on multi-level cooperation and multi-layer interconnection as described above. The method includes: The cloud-based global optimization layer generates planned power curves for each distribution zone based on short-term forecast data. The side-side calculation layer generates corresponding control setpoints for each distribution area based on the planned power curve of each distribution area. The reinforcement learning control layer generates the original control commands for the physical device execution layer based on the normal mode strategy, the operating status information and the control settings of the distribution area. The fault detection and protection layer corrects the original control commands based on the operating status information, and controls the physical equipment execution layer to perform corresponding operations based on the correction results; when a fault is identified in the distribution transformer area based on the operating status information, the physical equipment execution layer is controlled to perform fault protection operations; and the operating status information of the distribution transformer area is continuously acquired, and fault diagnosis information is generated based on the operating status information acquired within a preset time window. The reinforcement learning control layer responds to the received fault diagnosis information by switching the normal mode strategy to the fault perception strategy. Based on the fault perception strategy, the original control commands are dynamically adjusted according to the fault diagnosis information and then sent to the fault inspection and protection layer. The edge computing layer receives fault diagnosis information from each distribution transformer area. Based on the regional topology network structure and current power flow status of each distribution transformer area corresponding to the edge computing layer, it determines the minimum electrical isolation range including the controllable switch, sends the controllable switch disconnection command to the physical device layer corresponding to the minimum electrical isolation range, generates a regional fault report, and sends the regional fault report to the cloud master station layer. In response to the received regional fault report, the cloud master station layer determines the optimal global recovery scheme based on the global topology, the capacity of each line, and reactive power and voltage constraints. Based on the optimal global recovery scheme, it generates a global recovery command and sends the global recovery command to each side computing layer. The edge computing layer generates control commands for each distribution radio zone corresponding to the current edge computing layer based on the global recovery command. It then sends the control commands for each distribution radio zone to the reinforcement learning control layer corresponding to each distribution radio zone. This enables the reinforcement learning control layer corresponding to each distribution radio zone to switch from the fault perception strategy to the normal mode strategy. Based on the normal mode strategy and the corresponding control commands, it controls the physical device execution layer to perform the corresponding operations.

[0088] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0090] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A power flow optimization system based on multi-level cooperation and multi-layer interconnection, characterized in that, The system includes: The cloud master station layer is used to generate planned power curves for each distribution station area based on short-term forecast data. The side-side calculation layer is used to generate corresponding distribution area control setpoints based on the planned power curves of each distribution area. The physical device execution layer is used to collect the operating status information of the corresponding distribution transformer area. The operating status information includes at least one of node voltage, node current, distributed resource output, energy storage charge status and local load power. The reinforcement learning control layer is used to generate the original control commands of the physical device execution layer based on the normal mode strategy, the operating status information and the control settings of the distribution radio area. The fault detection and protection layer is used to correct the original control command based on the operating status information, and control the physical device execution layer to perform corresponding operations based on the correction operation result; when a fault in the distribution transformer area is identified based on the operating status information, the physical device execution layer is controlled to perform fault protection operations; and the operating status information of the distribution transformer area is continuously acquired, and fault diagnosis information is generated based on the operating status information acquired within a preset time window. The reinforcement learning control layer is also used to dynamically adjust the original control command based on the fault diagnosis information according to the fault perception strategy and then send it to the fault inspection and protection layer.

2. The power flow optimization system based on multi-level collaboration and multi-layer interconnection according to claim 1, characterized in that, The short-term forecast data includes at least one of the distributed resource output, load demand, and time-of-use pricing for each distribution substation; based on the short-term forecast data, a planned power curve for each distribution substation is generated, including: For each distribution area: Based on the output of the distributed resources and the load demand, constraints representing the power supply and demand balance of the distribution substation area are constructed. Based on the output of the distributed resources and the time-of-use pricing, a first objective function is constructed under the constraint of satisfying the power supply and demand balance of the distribution substation, with the goal of minimizing the total operating economic cost consisting of the power interaction cost between the distribution substation and the external power grid and the operating consumption cost of each distributed resource within the substation during the scheduling cycle. Solving the first objective function yields the global optimal power flow of the distribution substation, generating the planned power curve of the distribution substation. The planned power curve is used to characterize the expected power interaction between the distribution substation and the external power grid during each scheduling period within the target planning cycle.

3. The power flow optimization system based on multi-level collaboration and multi-layer interconnection according to claim 2, characterized in that, The first objective function includes: First component and second component; The first component is determined based on the time-of-use electricity price and the power exchanged between the distribution substation and the external power grid during the current scheduling period. The first component is used to characterize the power interaction cost generated between the distribution substation and the external power grid during the current scheduling period. The second component is determined based on the output power of each distributed resource in the distribution area during the current scheduling period and the operating cost parameters of each distributed resource. The second component is used to characterize the operating cost incurred by each distributed resource in the distribution area during the current scheduling period. A first objective function is constructed with the goal of minimizing the sum of the first component and the second component over the target planning period.

4. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 1, characterized in that, Based on the planned power curve of each distribution area, corresponding control setpoints for each distribution area are generated, including: For each distribution area: With the objectives of minimizing the deviation of the power exchanged between the distribution substation and the external power grid relative to the planned power curve at each time step within the preset prediction time domain, and minimizing the operating costs of each distributed resource within the distribution substation at the corresponding time step, a regional model predictive control objective function is constructed. Using the switching power and output of each distributed resource within the distribution area at each time step in the prediction time domain as decision variables, the predictive control objective function of the regional model is solved by rolling optimization under the condition of satisfying the preset regional power balance and equipment operation constraints. The switching power setting value and the output setting value of each distributed resource within the current time step in the predicted time domain obtained by solving are used as the control setting value of the distribution substation.

5. The power flow optimization system based on multi-level collaboration and multi-layer interconnection according to claim 1, characterized in that, The original control commands include at least one of the following: energy storage charging and discharging power, photovoltaic reactive power regulation power, and flexible interconnection device transmission power; the original control commands are modified based on the operating status information, including: If the original control command includes energy storage charging and discharging power, the current physical capacity limit of the energy storage device is determined based on the current energy storage state of charge collected by the physical device execution layer. Using the current physical capacity limit as a constraint, the energy storage charging and discharging power of the original control command is limited and corrected; or If the original control command includes photovoltaic reactive power regulation, the current physical upper limit of reactive power is determined based on the current active power output and apparent power rated capacity of the photovoltaic inverter collected by the physical device execution layer. Using this current physical upper limit of reactive power as a constraint, the photovoltaic reactive power regulation of the original control command is limited and corrected; or If the original control command includes the transmission power of the flexible interconnect device, the local load power collected by the physical device execution layer is obtained. If it is determined that the difference between the local load power and the transmission power of the flexible interconnect device exceeds the power supply margin of the current distribution substation, the transmission power of the flexible interconnect device is limited and corrected based on the power supply margin; or If the current node current collected by the physical device execution layer reaches a preset current warning threshold, a limiting correction or truncation operation is performed on the power increase command in the original control command; or If the current node voltage collected by the physical device execution layer reaches the preset voltage warning threshold, the command that causes the action direction of the original control command to be superimposed with the current voltage trend is intercepted or corrected in the opposite direction, based on the action direction of the original control command and the current voltage trend.

6. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 5, characterized in that, Based on the aforementioned operating status information, a fault in the distribution radio area was identified, including: If the node voltage is higher than the preset overvoltage threshold for a duration of a first duration, an overvoltage fault is determined to have occurred. If the duration for which the node current is higher than the preset overcurrent threshold reaches a second duration, an overcurrent fault is determined to have occurred. If the original control command is determined to be a commutation command for the converter based on the energy storage charging / discharging power or the transmission power of the flexible interconnect device, and no zero current signal is detected within the third duration, a commutation timeout fault is determined to have occurred.

7. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 6, characterized in that, Continuously acquire the operating status information of the distribution radio area, and generate fault diagnosis information based on the operating status information acquired within a preset time window, including: The polling task is executed at a preset period to continuously obtain node voltage and node current, build and continuously update the voltage history queue and current history queue, and record the commutation operation result corresponding to each commutation command to build a commutation operation result history queue. The voltage change rate and the total harmonic distortion rate of the current are calculated based on the voltage history queue and the current history queue, and the commutation failure rate is calculated based on the commutation operation result history queue. If the obtained voltage change rate is greater than the preset voltage change rate warning threshold, or if the obtained total harmonic distortion rate of the current is greater than the preset total harmonic distortion rate of the current warning threshold, or if the obtained commutation failure rate is greater than the preset commutation failure rate warning threshold, fault diagnosis information is generated.

8. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 4, characterized in that, Based on the normal mode strategy, and according to the operating status information and the control settings of the distribution substation, the original control commands for the physical device execution layer are generated, including: When the fault diagnosis information is not detected, the original control command is output through the pre-trained normal mode strategy model, taking the operating status information and the distribution radio area control setting value as input. The normal mode strategy model is obtained by training a specified neural network based on a preset second objective function. The steps for constructing the second objective function include: The actual electricity purchase cost is determined based on the deviation between the exchange power between the distribution substation and the external power grid and the corresponding exchange power set value, and the network power loss is determined based on the exchange power. An operating cost penalty item is constructed based on the actual electricity purchase cost and the network power loss. A voltage over-limit penalty term is constructed based on the matching relationship between the current node voltage and the preset voltage safety boundary; Based on the results of the fault detection and protection layer's correction operation on the original control command, a security layer intervention penalty item is constructed; The second objective function is obtained by weighting the operating cost penalty, voltage over-limit penalty, and safety layer intervention penalty based on the preset normal mode weight parameters.

9. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 8, characterized in that, The matching relationships include: If the current node voltage exceeds the voltage safety boundary, a voltage over-limit penalty value is determined based on the extent of its deviation from the lower limit or upper limit of the safety voltage boundary. If the current node voltage is within the safety voltage boundary, the voltage over-limit penalty value is determined to be a preset value. Based on the result of the fault detection and protection layer's correction operation on the original control command, a safety layer intervention penalty item is constructed, including: If it is determined, based on the result of the correction operation, that the fault detection and protection layer has corrected the original control command, the value of the safety layer intervention penalty item is set to the first value; If, based on the result of the correction operation, it is determined that the fault check and protection layer has not corrected the original control command, the value of the safety layer intervention penalty item is set to the second value.

10. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 4, characterized in that, Based on the fault perception strategy, the original control commands are dynamically adjusted according to the fault diagnosis information, including: When the fault diagnosis information is detected, the original control command is dynamically adjusted by taking the operating status information, the control setting value of the distribution radio area and the fault diagnosis information as inputs and the pre-trained fault perception strategy model as output. The fault perception strategy model is obtained by training a specified neural network based on a preset third objective function. The steps for constructing the third objective function include: An absolute penalty for safety violations is constructed based on the matching results of the current node voltage with the preset voltage safety boundary and the matching results of the current node current with the preset current threshold. A recovery time penalty term is constructed based on the fault duration, wherein the value of the recovery time penalty term is positively correlated with the fault duration; The network power loss penalty term is determined based on the exchange power between the distribution substation and the external power grid; The third objective function is obtained by weighting the absolute penalty for safety violations, the recovery time penalty, and the operational loss penalty based on the preset fault perception strategy weight parameters.

11. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 10, characterized in that, Based on the matching results of the current node voltage and the preset voltage safety boundary, and the matching results of the current node current and the preset current threshold, an absolute penalty for safety violations is constructed, including: If the current node voltage exceeds the voltage safety boundary, or if the current node current is greater than the current threshold, the value of the absolute penalty for safety violation is determined to be a first value; otherwise, the value of the absolute penalty for safety violation is determined to be a second value.

12. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 1, characterized in that, The side-side computing layer is also used for: Based on the fault diagnosis information, and considering the regional topology network structure of each distribution substation corresponding to the edge computing layer and the current power flow status of each distribution substation, the minimum electrical isolation range, including the controllable switch, is determined. Send the disconnect command of the controllable switch to the physical device layer corresponding to the minimum electrical isolation range, generate a regional fault report, and send the regional fault report to the cloud master station layer; The cloud master station layer is also used for: Upon receiving the regional fault report, based on the global topology, the capacity of each line, and reactive power and voltage constraints, the power flow feasibility of each preset candidate network reconstruction path is verified, and a target path candidate set is constructed based on the verified candidate network reconstruction paths. With the goal of minimizing the impact of a fault, the optimal global recovery scheme is determined by optimizing each path in the candidate target path set based on a preset global recovery optimization objective function. This scheme includes network reconfiguration switching, backup power supply switching timing, and load transfer paths. Based on the optimal global recovery scheme, a global recovery command is generated and sent to each edge computing layer.

13. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 12, characterized in that, The side-side computing layer is also used for: Based on the global recovery instruction, control instructions are generated for each distribution radio zone corresponding to the current edge computing layer. The control instructions for each distribution radio zone are sent to the reinforcement learning control layer corresponding to each distribution radio zone, so that the reinforcement learning control layer corresponding to each distribution radio zone switches from the fault perception strategy to the normal mode strategy. Based on the normal mode strategy and the corresponding control instructions, the physical device execution layer is controlled to perform the corresponding operation.

14. The power flow optimization system based on multi-level cooperation and multi-layer interconnection according to claim 6, characterized in that, The fault detection and protection layer is also used for: If multiple consecutive commutation timeout faults are detected in the converter, the current pulse blocking delay is incremented according to the preset first dynamic gain ratio, and the incremented result is compared with the preset blocking delay limit value. The minimum value between the two is taken as the updated pulse blocking delay. If an overcurrent fault is detected, the switching dead time of the converter is incremented according to the preset second dynamic gain ratio, and the incremented result is compared with the preset physical limit dead time. The minimum value of the two is taken as the updated switching dead time. If multiple overvoltage faults are detected within a preset time period, the preset overvoltage threshold is lowered, or the hardware debouncing time for overvoltage detection is extended.

15. A power flow optimization method based on multi-level cooperation and multi-layer interconnection, applied to the power flow optimization system based on multi-level cooperation and multi-layer interconnection as described in any one of claims 1-14, characterized in that, The method includes: The cloud-based global optimization layer generates planned power curves for each distribution zone based on short-term forecast data. The side-side calculation layer generates corresponding control setpoints for each distribution area based on the planned power curve of each distribution area. The reinforcement learning control layer generates the original control commands for the physical device execution layer based on the normal mode strategy, the operating status information, and the control settings of the distribution radio area. The fault detection and protection layer corrects the original control command based on the operating status information, and controls the physical device execution layer to perform the corresponding operation based on the correction result; when a fault is identified in the distribution transformer area based on the operating status information, the physical device execution layer is controlled to perform a fault protection operation; and the operating status information of the distribution transformer area is continuously acquired, and fault diagnosis information is generated based on the operating status information acquired within a preset time window. In response to the received fault diagnosis information, the reinforcement learning control layer switches the normal mode strategy to a fault perception strategy. Based on the fault perception strategy, the original control command is dynamically adjusted according to the fault diagnosis information and then sent to the fault inspection and protection layer. The edge computing layer receives fault diagnosis information from each distribution station area. Based on the regional topology network structure and current power flow status of each distribution station area corresponding to the edge computing layer, it determines the minimum electrical isolation range including the controllable switch, sends the disconnection command of the controllable switch to the physical device layer corresponding to the minimum electrical isolation range, generates a regional fault report, and sends the regional fault report to the cloud master station layer. In response to the received regional fault report, the cloud master station layer determines the optimal global recovery scheme based on the global topology, the capacity of each line, and reactive power and voltage constraints. Based on the optimal global recovery scheme, it generates a global recovery command and sends the global recovery command to each side computing layer. The edge computing layer generates control instructions for each distribution radio zone corresponding to the current edge computing layer based on the global recovery instruction, and sends the control instructions for each distribution radio zone to the reinforcement learning control layer corresponding to each distribution radio zone, so that the reinforcement learning control layer corresponding to each distribution radio zone switches from the fault perception strategy to the normal mode strategy, and controls the physical device execution layer to perform corresponding operations based on the normal mode strategy and the corresponding control instructions.

Images