A method and system for full-cycle fault handling of distribution terminals based on NPU collaboration

Abstract

A method and system for full-cycle fault handling of distribution terminals based on NPU collaboration is disclosed. The real-time core transmits waveform data to the NPU core via an on-chip high-speed bus and shared memory. The NPU core identifies the fault; the NPU core sends a warning to the real-time core; upon receiving the warning, the real-time core pre-locates the faulty section, and the NPU core pre-calculates the power loss load level and power under different isolation boundaries in parallel; the real-time core selects the optimal isolation boundary based on minimizing the load loss area, locates the faulty section, and isolates it; the NPU core, aiming to maximize the total restored load, minimize transfer losses, and reduce the number of switching operations, solves for the optimal transfer strategy and sends a circuit breaker operation sequence containing operation order and time intervals to the real-time core; after executing one operation, the real-time core transmits waveform data to the NPU core; if the total restored load is insufficient, the NPU core determines a supplementary transfer operation, which the real-time core executes; if the target is met, the next operation is executed, achieving full-cycle fault handling of distribution terminals.

Description

Technical Field

[0001] This invention belongs to the fields of power distribution automation and artificial intelligence technology, specifically relating to a method and system for handling power distribution terminal faults throughout the entire lifecycle based on NPU (Neural Processing Unit) collaboration. Background Technology

[0002] With the expansion of distribution network scale, the high proportion of distributed power generation, and the continuous improvement of power supply reliability requirements, distribution terminals (such as data transmission units (DTUs) or feeder terminal units (FTUs)) are playing an increasingly important role in fault handling. Traditional distribution network fault handling typically involves two stages: independent protection tripping and centralized restoration at the main station. However, in practical applications, this approach has the following significant shortcomings: 1) High latency in low-level interactions hinders support for high-frequency AI early warning: Existing power distribution terminals with edge computing capabilities often use external network interfaces or file system transmission methods when interacting with the microprocessor (MCU) and AI acceleration module. This mechanism suffers from large protocol stack overhead and data copy redundancy, resulting in high transmission latency, typically in the millisecond or even second range. This cannot meet the real-time transmission requirements of 10ms-level high-frequency waveform data (such as 12.8kHz sampling rate), severely restricting the real-time inference efficiency of deep learning models on the terminal side.

[0003] 2) The early warning model is simplistic and lacks the ability to map microscopic features to macroscopic time: Traditional terminal fault early warning mainly relies on simple threshold comparisons or single artificial intelligence classification algorithms, which can only output "normal / abnormal" or simple fault type labels, and cannot quantitatively assess the urgency of the fault occurrence. In particular, for progressive fault precursors such as insulation degradation and arc discharge, the lack of a prediction mechanism that combines microscopic transient waveform characteristics with macroscopic equipment degradation evolution patterns makes it impossible to accurately estimate the "expected fault occurrence time window," making it difficult for maintenance personnel to formulate targeted troubleshooting plans.

[0004] 3) The isolation and recovery processes are disconnected, and the transfer strategy struggles to accommodate complex multidimensional constraints: When a fault occurs, traditional terminals rely solely on real-time cores for passive logical judgment and physical isolation, lacking coordination with upper-level optimization modules. This can lead to the isolation strategy sacrificing critical loads. During the power restoration (transfer) phase after a fault, existing terminals typically employ heuristic searches based on preset rules or simple topology traversal. However, actual distribution network transfer is a multi-objective mixed-integer nonlinear programming problem involving multiple priority load protection, short-term source-load fluctuations, network loss optimization, and N-1 security checks. Traditional single-layer optimization algorithms (such as pure heuristic algorithms) are prone to getting trapped in local optima when facing high-dimensional discrete decision spaces, and their computational time is long, failing to meet the requirement of generating optimal strategies within seconds.

[0005] 4) Lack of dynamic adaptive capability in the execution process: Existing transfer strategies are mostly "one-time issuance and open-loop execution", which do not take into account sudden working conditions such as actual switch failure and real-time load fluctuations. When the expected effect deviates from the actual effect, there is a lack of dynamic correction mechanism.

[0006] Therefore, there is an urgent need for a power distribution terminal fault handling technology that breaks down data barriers between heterogeneous cores and achieves deep collaboration and dynamic closed-loop control throughout the entire process of early warning, isolation, and recovery. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for full-cycle fault handling of distribution terminals based on NPU collaboration. It employs a multi-core architecture based on real-time cores and NPU cores, and through deep hardware and software collaboration between the real-time cores and NPU cores, solves the problems of high data interaction latency, insufficient early warning capabilities, weak power transfer strategy optimization capabilities, and lack of dynamic closed-loop control in existing technologies. This enables intelligent early warning, isolation, and recovery handling of distribution terminal faults throughout their entire lifecycle, and can be widely applied to intelligent operation and maintenance, rapid fault handling, and power supply reliability improvement scenarios in 10kV and below distribution networks.

[0008] The present invention adopts the following technical solution.

[0009] This invention proposes a method for full-cycle fault handling of distribution terminals based on NPU collaboration. The distribution terminal adopts a heterogeneous multi-core architecture, which includes a real-time core and an NPU core. The method includes: The real-time core collects electrical quantity data from the distribution network and records waveforms. The real-time core transmits the waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. The NPU core determines whether a fault has occurred based on the waveform data. The waveform recording of the real-time core and the fault identification of the NPU core are performed simultaneously. When a fault is detected, the NPU core sends a fault warning to the real-time core. When the real-time core receives a fault warning, it pre-locates the fault section. The NPU core pre-calculates the level and power of the power loss load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power loss area as the optimization objective, and finally locates the fault section and performs physical isolation based on the optimal isolation boundary. The NPU core aims to maximize the total restored load, minimize distribution network transfer losses, and minimize the number of switching operations as fault recovery objectives. Under distribution network constraints, it solves for these objectives to obtain the optimal transfer strategy. This optimal strategy is then distributed to the real-time core in the form of a circuit breaker operation sequence, which includes the operation order and the time interval between adjacent operations. After executing one operation, the real-time core transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it uses real-time global topology and power flow data of the distribution network to solve for the fault recovery objectives under distribution network constraints to obtain a supplementary transfer operation, which the real-time core executes. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

[0010] After the optimal power transfer strategy is executed, the master station receives the local power transfer results uploaded by each power distribution terminal; The main station performs global multi-terminal coordination optimization based on the global network topology, power margin, and all received local power transfer results in order to detect and eliminate conflicts between different terminal power transfer schemes in terms of contact switch status and backup capacity allocation. The master station will send the globally coordinated instructions generated after coordination and optimization to the corresponding power distribution terminals to correct or confirm the local power transfer strategies of each power distribution terminal.

[0011] Preferably, the steps of real-time acquisition of electrical quantity data of the distribution network and recording waveforms include: The real-time kernel has multiple built-in abnormal condition triggering mechanisms, which will trigger waveform recording when any of the following conditions are met: 1) The rate of change of current is greater than the set threshold for the rate of change of current; 2) The voltage level after the drop is less than the set voltage threshold; 3) The harmonic content is greater than the set content threshold; 4) The zero-sequence current is greater than the set zero-sequence current threshold; Among them, various thresholds are adaptively adjusted based on historical statistical data, and the adjustment method for each threshold is the same, as shown in the following formula:

[0012] In the formula, For the updated threshold, The threshold before the update. This is a static threshold. This is the smoothing coefficient.

[0013] Preferably, the step of the real-time core transmitting waveform data to the NPU core via an on-chip high-speed bus combined with a shared memory mechanism includes: A circular buffer is divided in the shared memory. The real-time core writes the waveform data into the buffer through direct memory access, and the NPU core reads the data by maintaining read and write pointers through atomic operations. The real-time core configures the NPU core's inference priority register to high-throughput mode and starts the NPU core's parallel inference task through the cooperative control instruction set; Based on this configuration, the NPU core executes multiple device status assessment models in parallel and processes the waveform data of multiple power distribution terminals concurrently.

[0014] Preferably, the step of the NPU core determining whether a fault has occurred based on the recorded waveform data includes: The NPU core uses a pre-trained first neural network model to extract features and recognize patterns from the waveform data, and outputs a probability distribution of fault types; the first neural network model is a fusion of a one-dimensional convolutional neural network and a long short-term memory network. When the probability distribution of the fault type meets the preset warning conditions, a fault warning message is generated and fed back to the real-time kernel; the fault warning message includes the fault type, probability value and expected occurrence time window.

[0015] Preferably, the step of pre-locating the faulty section when a fault warning is received in real time includes: When a fault is confirmed, the fault characteristic information of the distribution terminal is extracted in real time, and the status information of the adjacent distribution terminals is obtained through horizontal peer-to-peer communication. The fault characteristic information includes fault current, current direction, voltage amplitude and other characteristic information, and the status information includes line switch status, load distribution and network topology. Real-time verification locates the fault section based on the fault characteristics of the current power distribution terminal and the status information of adjacent power distribution terminals.

[0016] Preferably, when the real-time core detects a fault characteristic and determines that it has entered the isolation phase, the real-time core immediately updates the NPU core's inference priority register to the lowest latency mode. After completing the register update, the real-time core sends a high-priority preemptive interrupt to the NPU core through a bidirectional deterministic interrupt mechanism. After receiving an interrupt, the NPU core saves the complete context of the currently executing non-urgent task, loads the emergency inference model related to the current fault location, and begins to execute the single-path fault location inference task.

[0017] Preferably, the NPU core pre-calculates the level and power of the power outage load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network; the real-time core selects the optimal isolation boundary with the minimum load power outage area as the optimization objective, and the steps of finally locating the fault section and performing physical isolation based on the optimal isolation boundary include: In the lowest latency mode, the NPU core uses known global topology and real-time power flow data to pre-calculate the total power outage load and the impact assessment of important loads under different potential isolation boundaries in parallel; the pre-calculation results are quickly sent to the real-time core in the form of "suggested isolation cut-off points"; After obtaining the "suggested isolation cut-off point" in real time, the optimal isolation boundary is selected with the minimum load power loss area as the optimization objective, and the fault section isolation is performed. The real-time core synchronously transmits the status data after fault isolation to the NPU core. The status data includes the fault segment number, the isolated network topology matrix, the circuit breaker status, and the load distribution data. The status data is transmitted in a structured data serialization format.

[0018] Preferably, the NPU core takes maximizing the total restored load, minimizing the distribution network transfer loss, and minimizing the number of switching operations as the fault recovery objectives. The steps for solving the fault recovery objectives to obtain the optimal transfer strategy under distribution network constraints include: The NPU core obtains status data, load priority, and short-term source-load forecast data from the real-time core and local cache. Under distribution network constraints, a multi-objective power transfer optimization model is established based on the fault recovery objective. The objective function of the model includes: 1) Maximize the total weighted recovery load, where the weighting coefficient is set according to the load priority; 2) Minimize power transfer losses in the distribution network; 3) Minimize the number of switching actions; The NPU core uses a first optimization algorithm to perform a preliminary solution to the multi-objective power transfer optimization model and generate a set of candidate power transfer schemes; the first optimization algorithm includes a heuristic swarm optimization algorithm. The NPU core uses a pre-trained second neural network model to finely optimize the set of candidate transfer schemes and generate the optimal transfer strategy; the second neural network model includes a deep reinforcement learning model.

[0019] Preferably, the objective function that maximizes the total weighted recovery load is... As shown in the following formula:

[0020] In the formula, For load nodes The recovery power; For load nodes Priority weights; For load nodes The decision variable has a value of 1 indicating that power has been restored and a value of 0 indicating that power has not been restored. This represents the number of load nodes.

[0021] Preferably, the objective function for minimizing the distribution network transfer losses is... As shown in the following formula:

[0022] In the formula, , Branch roads The current and resistance, E represents the duration of the transfer, and E represents the set of branches.

[0023] Preferably, the objective function is to minimize the number of switching actions. As shown in the following formula:

[0024] In the formula, For switch The state variable is denoted by 0, which indicates that the switch is open, and 1, which indicates that the switch is closed; S is the set of switches.

[0025] Preferably, the optimal transfer strategy is issued to the real-time core in the form of a circuit breaker operation sequence. The circuit breaker operation sequence includes the steps of operation order and time interval between two adjacent operations: The NPU core performs a security check on the optimal transfer strategy, and after the check passes, it sends the strategy to the real-time core for execution. Security checks include: 1) Calculate the three-phase / single-phase short-circuit current of each node after the power transfer, and ensure that it does not exceed the switching capacity of the switch; 2) Verify the coordination of protection settings between upper and lower levels to ensure that the time difference is not less than the set value; The time interval between two consecutive operations shall not be less than 2 seconds.

[0026] In another aspect, this invention proposes a full-cycle fault handling system for power distribution terminals based on NPU collaboration. The power distribution terminal adopts a heterogeneous multi-core architecture, which includes a real-time core and an NPU core. The real-time core collects electrical quantity data from the distribution network and records waveforms. The real-time core transmits the waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. The NPU core determines whether a fault has occurred based on the waveform data. The waveform recording of the real-time core and the fault identification of the NPU core are performed simultaneously. When a fault is detected, the NPU core sends a fault warning to the real-time core. When the real-time core receives a fault warning, it pre-locates the fault section. The NPU core pre-calculates the level and power of the power loss load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power loss area as the optimization objective, and finally locates the fault section and performs physical isolation based on the optimal isolation boundary. The NPU core aims to maximize the total restored load, minimize distribution network transfer losses, and minimize the number of switching operations as fault recovery objectives. Under distribution network constraints, it solves for these objectives to obtain the optimal transfer strategy. This optimal strategy is then distributed to the real-time core in the form of a circuit breaker operation sequence, which includes the operation order and the time interval between adjacent operations. After executing one operation, the real-time core transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it uses real-time global topology and power flow data of the distribution network to solve for the fault recovery objectives under distribution network constraints to obtain a supplementary transfer operation, which the real-time core executes. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

[0027] The present invention is also a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0028] The present invention is also a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0029] The beneficial effects of this invention are that, compared with the prior art, the method proposed in this invention achieves the following technical effects through the deep integration of the real-time core and the NPU core at the hardware bus layer, algorithm model layer, and decision control layer: 1) High-precision intelligent early warning: Utilizes NPU to run deep learning models to intelligently analyze waveform data, identify complex fault precursor characteristics, achieve accurate prediction of fault types, and significantly reduce false alarm rate and missed alarm rate; 2) Multi-constraint optimal power transfer: Taking into account multi-dimensional constraints such as load priority, source load fluctuation, network loss optimization, and N-1 security, a hybrid intelligent optimization algorithm is adopted to formulate the optimal power transfer strategy, thereby improving the recovery speed of important loads and the economic efficiency of power grid operation; 3) Rapid response processing: Fully leverage the hardware acceleration capabilities of the NPU to achieve second-level response for fault prediction and strategy optimization, meeting the rapid processing needs of the distribution network; 4) Adaptive optimization capability: Through online learning and experience accumulation, continuously optimize early warning models and decision-making strategies to adapt to dynamic changes in the power grid operating environment; 5) Full-process collaborative closed loop: Connect the information flow and decision-making flow in the three stages of pre-fault, during-fault, and post-fault to achieve effective collaboration between the real-time core, NPU core, and master station. Attached Figure Description

[0030] Figure 1 This is a flowchart of a full-cycle fault handling method for power distribution terminals based on NPU collaboration proposed in this invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0032] This invention proposes a full-cycle fault handling method for power distribution terminals based on NPU collaboration. The power distribution terminal adopts a heterogeneous multi-core architecture, which includes a real-time core and an NPU core. The real-time core is used for high-real-time data acquisition, threshold monitoring, and protection tripping logic; the NPU core is used for computationally intensive deep learning inference and multi-objective optimization. The real-time core and the NPU core interact with each other through an on-chip high-speed bus and a shared memory mechanism to achieve deep collaboration.

[0033] Traditional power distribution terminals are mostly single-core or homogeneous multi-core, employing a centralized model of data acquisition → uploading → master station decision-making → waiting for instructions; or although there are edge computing units at the terminal, logic control and AI computing need to share computing resources, leading to mutual contention and interference. This invention systematically maps a heterogeneous multi-core architecture of real-time core + NPU core to the entire lifecycle of power distribution terminal fault handling: early warning → isolation → recovery. The NPU core's instruction set includes a collaborative control instruction group, enabling the NPU core to directly call collaborative control instructions after completing deep learning inference, generating control operation codes for the power distribution terminal fault early warning, isolation, or recovery stages. The NPU core front end is also equipped with an inference priority register, which is dynamically configured by the real-time core based on the stage of fault handling: in the fault early warning stage, the real-time core configures the register to high-throughput mode, allowing the NPU core to process multiple device status assessment models in parallel; when the real-time core detects fault characteristics and determines that the isolation stage has begun, the real-time core immediately updates the register to the lowest latency mode, allowing the NPU core to focus all acceleration resources on a single-path fault location inference model. After the real-time core completes the register update, it sends a high-priority preemptive interrupt to the NPU core through a bidirectional deterministic interrupt mechanism. This forces the NPU core to save the context of the current non-urgent task and immediately load and run the emergency inference model related to the current fault. This ensures that the end-to-end delay from fault perception to isolation decision meets the millisecond-level deterministic requirement, achieving deep integration and collaboration between deterministic control and complex intelligent computing.

[0034] At the power distribution terminal level, a dedicated heterogeneous architecture was proposed and implemented, consisting of a high-reliability real-time control unit based on a real-time core and a high-performance AI computing unit based on an NPU core. The two are physically isolated but interconnected at high speed. By assigning high real-time, deterministic, and safety-critical tasks to the real-time core and computationally intensive, data-driven, and non-deterministic tasks to the NPU core, the existing technology model of everything being calculated by the master station has been changed. A closed-loop collaborative relationship of "perception-decision-execution-feedback" has been formed between the real-time core and the NPU core.

[0035] The full-cycle handling method proposed in this invention includes three stages: pre-fault early warning, fault isolation, and post-fault recovery; such as Figure 1 As shown, the method includes the following steps: Step S1, pre-fault warning stage.

[0036] The real-time core collects electrical quantity data from the distribution network and records waveforms. The real-time core transmits the waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. The NPU core determines whether a fault has occurred based on the waveform data. The waveform recording of the real-time core and the fault identification of the NPU core are performed simultaneously. When a fault is determined to have occurred, the NPU core sends a fault warning back to the real-time core.

[0037] The core task of the pre-fault warning stage is to identify early warning signs through intelligent analysis. Specifically, step S1 includes: Step S1.1, Real-time monitoring and anomaly triggering.

[0038] Specifically, the system collects electrical quantity data of the power distribution network in real time and triggers waveform recording when abnormal triggering conditions are detected.

[0039] In this embodiment, the real-time kernel continuously collects electrical quantity data of the distribution network at a high-frequency sampling period of 10ms, including three-phase voltage, zero-sequence voltage, three-phase current, zero-sequence current, frequency, active power, reactive power, power factor and harmonic content (3rd, 5th and 7th). After the collected electrical quantity data is processed by low-pass filtering and amplitude / phase calibration, the real-time kernel performs online anomaly monitoring.

[0040] The real-time kernel has multiple built-in abnormal condition triggering mechanisms, which will trigger waveform recording when any of the following conditions are met: 1) Current change rate is greater than the set current change rate threshold: Calculate the current change rate of adjacent sampling points, and trigger waveform recording when the current change rate is greater than 30% / s of the rated current; 2) The voltage drop level is less than the set voltage threshold: When the voltage of any phase drops to below 85% of the rated voltage and the duration exceeds 2 cycles (40ms), waveform recording is triggered; 3) Harmonic content is greater than the set content threshold: The 3rd, 5th and 7th harmonic components are extracted by Fast Fourier Transform (FFT). When the content of any specified harmonic exceeds 5% of the fundamental frequency, the recording is triggered. 4) Zero-sequence current is greater than the set zero-sequence current threshold: Zero-sequence current is the core characteristic quantity of ground fault. When the zero-sequence current is greater than 0.2 times the secondary rated current, waveform recording is triggered.

[0041] The aforementioned thresholds for triggering various anomalies are not fixed but are adaptively adjusted based on historical statistical data. A sliding window is maintained in real time, and the thresholds are updated using an exponentially weighted moving average method. The adjustment method for all thresholds is the same, as shown in the following formula:

[0042] In the formula, For the updated threshold, The threshold before the update. This is a static threshold. The value is 0.9, which is used as a smoothing coefficient in this example to ensure the stability of threshold changes.

[0043] Step S1.2: Acquisition of waveform data and high-speed transmission within the chip.

[0044] Specifically, the real-time core generates waveform data; in this embodiment, when any abnormal condition is triggered, the real-time core immediately starts waveform recording. The waveform recording window is set to 500ms, including 400ms before the trigger and 100ms after the trigger; in order to capture high-frequency transient components, the sampling frequency is set to 12.8kHz, that is, 256 sampling points per power frequency cycle, including 8 channels, to collect the time-series data of three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current respectively.

[0045] The real-time core transmits waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. A circular buffer is partitioned within the shared memory. The real-time core, acting as a producer, writes waveform data to the buffer via DMA (Direct Memory Access), while the NPU core, acting as a consumer, reads data through atomic operations maintaining read / write pointers. During this stage, the real-time core configures the NPU core's inference priority register to high-throughput mode and initiates parallel inference tasks for the NPU core via a cooperative control instruction set. Based on this configuration, the NPU core executes multiple device status assessment models in parallel, performing concurrent analysis of waveform data from multiple power distribution terminals to identify faults, significantly improving processing efficiency during the early warning phase. The data transmission mode between the real-time core and the NPU core proposed in this invention avoids the protocol stack overhead and copy latency associated with traditional file or network transmissions. Transmission bandwidth reaches hundreds of MB / s, latency is stabilized at the microsecond level, and waveform data is simultaneously shared between the real-time core and the NPU core, ensuring data real-time performance. By dividing a circular buffer in shared memory and combining it with atomic operations, the lock-free data structure from the field of high-performance computing is introduced into the inter-core communication of the power distribution terminal, realizing lock-free synchronization, avoiding producer-consumer blocking, and ensuring data integrity. Moreover, it breaks the traditional edge computing mode of NPU cores passively waiting for data, enabling NPU cores to read data immediately when the recorded waveform data is ready without polling, thereby significantly reducing power consumption and improving response speed.

[0046] Moreover, based on the above mechanism, after the real-time core triggers the NPU core to perform fault type inference, it can continue to collect electrical quantity data of the distribution network without waiting for the inference result of the NPU core. At this time, the NPU model inference and the continuous collection of the real-time core are carried out in parallel, without the risk of data blocking.

[0047] Step S1.3, NPU core deep learning model inference.

[0048] Specifically, the NPU core uses a pre-trained first neural network model to extract features and recognize patterns from the recorded waveform data, and outputs a probability distribution of fault types.

[0049] In this embodiment, after receiving the waveform data, the NPU core first performs normalization and wavelet transform noise reduction preprocessing. Then, the preprocessed data is input into a pre-trained first neural network model. The first neural network model is a fusion of a one-dimensional convolutional neural network and a long short-term memory network; the one-dimensional convolutional neural network contains multiple convolutional layers to extract local spatial features of the waveform data; the long short-term memory network is used to capture the temporal dependencies of the convolutionally processed data.

[0050] In a preferred embodiment of the present invention, the model adopts an architecture that fuses a one-dimensional convolutional neural network (1D-CNN) with a bidirectional long short-term memory network, as detailed below: Feature extraction layer (1D-CNN): Contains 4 convolutional layers with kernel sizes of 32, 16, 8, and 4, and kernel numbers of 16, 32, 64, and 128 respectively. The activation function is ReLU, followed by max pooling. To mitigate gradient vanishing, each convolutional block uses a residual connection structure. After global average pooling, the 5000-point input sequence is compressed into a 128-dimensional feature vector, which encodes the local morphological features of the waveform. Timing analysis layer: A two-layer Bi-LSTM (128 hidden units per layer) is used to capture the forward and backward timing dependencies of the waveform and output 128-dimensional timing features. Classification Output Layer: After passing through a fully connected layer and a Softmax activation function, the probability distribution of eight fault and state types (single-phase grounding, two-phase short circuit, three-phase short circuit, open circuit, overload, harmonic pollution, voltage sag, and normal state) is output, as shown in the following formula:

[0051] In the formula, For the first The probability of each fault and state type For the first The output (logits) of the fully connected layer under various fault and state types.

[0052] Step S1.4, fault warning and time window estimation.

[0053] Specifically, when the probability distribution of fault types meets the preset early warning conditions, fault early warning information is generated and fed back to the real-time kernel.

[0054] The probability value of any fault type exceeds the preset probability threshold; the fault warning information includes the fault type, probability value, and expected occurrence time window; the expected occurrence time window is estimated by the NPU core after matching the extracted waveform features with the local historical degradation feature database.

[0055] In the embodiment, when the probability of a certain type of fault exceeds a preset probability threshold (preferably 0.75), it is determined to be a potential fault.

[0056] For predicting the expected occurrence time window (5-30 minutes in the future), this invention employs a feature mapping matching method: the NPU core performs a similarity search between the currently extracted 128-dimensional high-dimensional waveform feature vector and a locally stored historical degradation evolution database. This database records the time span from the appearance of similar microscopic precursor features to the eventual macroscopic failure of past devices. By calculating cosine similarity, the K most similar historical samples are found, and their time spans are weighted and averaged to scientifically estimate the expected occurrence time window of the current failure. After the warning information is generated, it is fed back to the real-time core (increasing the sampling frequency to 5ms) and uploaded to the main station for a network-wide risk assessment.

[0057] In addition, the NPU core adopts an online incremental learning approach. Every 100 new samples, it uses mini-batch gradient descent (batch size=32) and cosine annealing learning rate strategies to fine-tune the parameters and achieve model adaptation.

[0058] Step S2, fault isolation phase.

[0059] When the real-time core receives a fault warning, it pre-locates the fault section. The NPU core pre-calculates the level and power of the power loss load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power loss area as the optimization objective, and finally locates the fault section and performs physical isolation based on the optimal isolation boundary.

[0060] Among them, the optimization objective is to minimize the load loss area, which satisfies the following relationship:

[0061] In the formula, To optimize the objective, For power failure load The power; For power failure load The priority weights corresponding to the levels are as follows: the priority weight of level 1 load is 3.0, the priority weight of level 2 load is 2.0, and the priority weight of level 3 load is 1.0. This represents the number of power-loss loads.

[0062] Specifically, step S2 includes: Step S2.1, lateral peer-to-peer communication and global situational awareness.

[0063] Specifically, when a fault is confirmed, the fault characteristic information of the distribution terminal is extracted in real time, and the status information of the adjacent distribution terminals is obtained through horizontal peer-to-peer communication. The fault characteristic information includes fault current, current direction, voltage amplitude and other characteristic information, and the status information includes line switch status, load distribution and network topology.

[0064] The real-time core locates the fault section based on the fault characteristics of the power distribution terminal and the status information of adjacent power distribution terminals, so as to perform physical isolation actions. At the same time, it also provides data support for the NPU to build a global operational status.

[0065] Step S2.2: Dynamically configure the NPU core to the lowest latency mode in real time.

[0066] When the real-time core detects a fault and determines that it needs to enter the isolation phase, it immediately updates the NPU core's inference priority register to the lowest latency mode. This register configuration instruction is written via the on-chip high-speed bus in a single-cycle instruction manner, ensuring the determinism of the configuration update (latency <1μs).

[0067] Step S2.3: Bidirectional deterministic interrupt and NPU core task preemption.

[0068] After completing the register update, the real-time core sends a high-priority preemptive interrupt to the NPU core via a bidirectional deterministic interrupt mechanism. This interrupt path is directly connected in hardware, with an interrupt response time of less than 100μs. Upon receiving the interrupt, the NPU core immediately: 1) Save the complete context of currently executing non-urgent tasks (such as parallel evaluation models in the early warning phase); 2) Load the emergency reasoning model related to the current fault location; 3) Begin executing the single-path fault location reasoning task.

[0069] The total time for context saving and model loading is controlled within 500μs, ensuring that the end-to-end latency from fault detection to the start of critical inference by the NPU core does not exceed 1ms.

[0070] Step S2.4, NPU collaborative pre-computation and rapid isolation.

[0071] Specifically, while the real-time core locates the faulty section, the NPU core plays a collaborative role. In the lowest latency mode, it uses known global topology and real-time power flow data to pre-calculate the total amount of power loss load and the impact assessment of important loads under different potential isolation boundaries in parallel. The pre-calculation results are quickly sent to the real-time core in the form of "suggested isolation cut-off points" to assist the real-time core in selecting the optimal isolation boundary so that the power loss area of ​​important loads is minimized. The real-time core combines its own judgment and NPU suggestions to accurately locate and execute the faulty section isolation. The entire action time is strictly controlled within 100ms.

[0072] Step S2.5: Synchronize the status data after isolation.

[0073] Specifically, the real-time core synchronously transmits the status data after fault isolation to the NPU core.

[0074] During the fault isolation phase, the status data includes the fault section number, the isolated network topology matrix, the circuit breaker status, and the load distribution data; the status data is transmitted in a structured data serialization format.

[0075] In this embodiment, after the isolation action is completed, the real-time core synchronously transmits the fault segment number, the isolated network topology, the circuit breaker status, and the load distribution data to the NPU core. To adapt to heterogeneous core interaction, the data is encoded using the Protocol Buffers (Protobuf) structured data serialization format and transmitted through the internal high-speed interface with a delay of no more than 50ms.

[0076] Step S3, post-fault recovery phase.

[0077] The NPU core aims to maximize the total restored load, minimize distribution network transfer losses, and minimize the number of switching operations as fault recovery objectives. Under distribution network constraints, it solves for these objectives to obtain the optimal transfer strategy. This optimal strategy is then distributed to the real-time core in the form of a circuit breaker operation sequence, which includes the operation order and the time interval between adjacent operations. After executing one operation, the real-time core transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it uses real-time global topology and power flow data of the distribution network to solve for the fault recovery objectives under distribution network constraints to obtain a supplementary transfer operation, which the real-time core executes. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

[0078] After fault isolation, the power-loss area needs to be restored through power transfer. Power transfer decision-making is a complex optimization problem with multiple constraints and objectives.

[0079] Specifically, step S3 includes: Step S3.1, Input Information Acquisition.

[0080] Specifically, the NPU core obtains status data, load priority, short-term source load forecast data, and other information from the real-time core and local cache, as follows: 1) Topological matrix T (n×n adjacency matrix); 2) Load power vector (active power P and reactive power Q at each node); 3) Load priority vector L ( These correspond to uninterrupted Level 1 loads, interruptible Level 2 loads for 5 minutes, and interruptible Level 3 loads for 30 minutes, respectively. 4) Power node capacity vector C: Capacity limitations of transformers and feeders. Indicates power node Maximum power supply capacity (unit: kVA); 5) Short-term load forecast: A lightweight LSTM time series model runs within the NPU core, combining historical data from the same period over the past 15 days with meteorological data to predict load and distributed power output fluctuations for the next 15 minutes. 6) Circuit breaker operable set S: All circuit breaker numbers available for transfer operation, excluding circuit breakers under maintenance or faulty.

[0081] In step S3.2, the NPU core constructs a multi-objective power transfer optimization model based on status data, load priority, short-term source load forecast data, and distribution network constraints.

[0082] The objective functions of the multi-objective power transfer optimization model include: maximizing the total weighted recovery load, where the weighting coefficients are set according to load priority; minimizing distribution network power transfer losses; and minimizing the number of switching operations. The NPU core establishes a multi-objective power transfer optimization model based on fault recovery objectives, and the model's objective functions include: 1) First objective function : Objective function to maximize the total weighted recovery load As shown in the following formula:

[0083] In the formula, For load nodes The recovery power; For load nodes The priority weights are as follows: the priority weight for level 1 loads is 3.0, the priority weight for level 2 loads is 2.0, and the priority weight for level 3 loads is 1.0. For load nodes The decision variable has a value of 1 indicating that power has been restored and a value of 0 indicating that power has not been restored. This represents the number of load nodes.

[0084] 2) Second objective function : Objective function to minimize distribution network transfer losses As shown in the following formula:

[0085] In the formula, , Branch roads The current and resistance, E represents the duration of the transfer, and E represents the set of branches.

[0086] 3) Third objective function : Objective function: Minimize the number of switching actions As shown in the following formula:

[0087] In the formula, For switch The state variable is denoted by 0, which indicates that the switch is open, and 1, which indicates that the switch is closed; S is the set of switches.

[0088] Distribution network constraints include radial network constraints, power supply capacity constraints, node voltage constraints, and N-1 security constraints; distribution network constraints include: 1) Radial Network Constraint: The network after power transfer maintains a tree structure, and the number of connected components equals the number of power source nodes. Mathematically, this is expressed as: Let the network after power transfer be a graph... ,in, This is the set of nodes after the transfer. Given the set of branches after the transfer, then:

[0089] In the formula, This refers to the number of nodes in the node set after the transfer. The number of branches in the branch set after the transfer is denoted as m, where m is the number of power supply nodes and the number of connected components.

[0090] 2) Capacity constraint: The total load supplied by each power node must not exceed its capacity limit.

[0091]

[0092] In the formula, For power nodes The set of load nodes supplying power. For load nodes power, For load nodes The standard deviation of the prediction For power nodes The capacity is 0.9, which is the safety margin factor.

[0093] 3) Voltage constraints: Voltage at all nodes It should be within the allowable range, 0.93pu≤ ≤1.07pu.

[0094] 4) Switch capacity constraints: Some switches (such as load switches) have rated current limits, and the current passing through the switch must not exceed its rated value.

[0095]

[0096] In the formula, , For switch The current and rated current.

[0097] 5) N-1 safety constraint: If any line or transformer in the network is disconnected after the transfer of power, the system still meets the constraint.

[0098] Step S3.3, solve using a hybrid optimization strategy.

[0099] To solve the above nonconvex mixed-integer nonlinear programming problem, the NPU core adopts a hybrid intelligent strategy of using a first optimization algorithm and a second neural network model. The specific solution process is as follows: 1) The NPU core uses the first optimization algorithm to perform a preliminary solution to the multi-objective power transfer optimization model and generate a set of candidate power transfer schemes.

[0100] The first optimization algorithm is a heuristic swarm optimization algorithm; during the initial solution process, the NPU core adopts a linear decreasing inertial weight strategy and a tournament selection method to retain the Pareto optimal solution set in order to generate a candidate transfer scheme set.

[0101] Phase 1 (Initial Solution): An improved Particle Swarm Optimization (PSO) algorithm is employed. In this preferred embodiment, the population size is set to 50, the maximum number of iterations is 100, the inertia weight decreases linearly from 0.9 to 0.4, and the acceleration factors c1=c2=2.0. A tournament selection method is used to retain the Pareto optimal solution set, and the top 10 candidate transfer schemes that satisfy the constraints are output.

[0102] 2) The NPU core uses a pre-trained second neural network model to finely optimize the candidate transfer scheme set and generate the optimal transfer strategy.

[0103] The second stage (refined optimization): The candidate schemes obtained from PSO are used as the initial state and input into a second neural network model based on deep reinforcement learning (DRL) for refined optimization. The second neural network model is a deep reinforcement learning model; the deep reinforcement learning model uses the candidate power transfer schemes as the initial state, uses topological features, load distribution, voltage level and network loss value to form the state space, uses the circuit breaker operation sequence to form the action space, and uses the weighted sum of load restoration amount, network loss value and number of switch operations to form the reward function. After a preset number of iterations, the optimal power transfer strategy is output.

[0104] This preferred embodiment employs the dual-delay deep deterministic strategy gradient algorithm (TD3). Its state space includes topological features, load distribution, voltage level, and network loss value; the action space is defined as the circuit breaker operation sequence; and the reward function comprehensively considers the weighted sum of load restoration amount, network loss, and the number of switching operations. After 50 iterations, the globally optimal power transfer strategy is output.

[0105] Step S3.4, security verification and dynamic policy distribution.

[0106] Specifically, the NPU core performs a security check on the optimal transfer strategy, and once the check passes, it sends the strategy to the real-time core for execution.

[0107] In this embodiment, the NPU core performs strict safety checks before outputting the strategy, including: calculating the three-phase / single-phase short-circuit current of each node after the power transfer to ensure that it does not exceed the switching capacity, and checking the coordination of the upper and lower level protection settings to ensure that the time difference is not less than the set value.

[0108] After successful verification, the NPU core will send the supply strategy to the real-time core in the form of a circuit breaker operation sequence, including the operation order and a time interval of no less than 2 seconds.

[0109] Step S3.5: After the real-time core completes one operation, it transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it solves the fault recovery target under the constraints of the distribution network based on the real-time global topology and power flow data of the distribution network to obtain the power transfer supplement operation, and the real-time core executes the power transfer supplement operation. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

[0110] After the real-time core executes the circuit breaker operations in sequence, it feeds back the actual execution results to the NPU core. The NPU core compares the actual execution results with the expected results. When the deviation exceeds the preset threshold, it re-triggers step S3, the post-fault recovery stage, based on the current actual state of the power grid, and dynamically adjusts the strategy.

[0111] The real-time core executes operations sequentially, collecting actual grid data and feeding it back to the NPU core after each operation. The NPU core compares the actual effect with the expected effect, such as whether the actual load restoration meets the standard. If the deviation exceeds the preset threshold, the NPU core, based on the current actual topology and load status, re-triggers the optimization calculations from S3.2 to S3.3 and dynamically issues supplementary operation instructions until the maximum transfer is completed.

[0112] After the optimal power transfer strategy is executed, the master station receives the local power transfer results uploaded by each distribution terminal. Based on the global network topology, power margin, and all received local power transfer results, the master station performs global multi-terminal coordination optimization to detect and eliminate conflicts between different terminal power transfer schemes in terms of tie switch status and reserve capacity allocation. The master station sends the global coordination command generated after coordination optimization to the corresponding distribution terminal to correct or confirm the local power transfer strategy of each distribution terminal.

[0113] In another aspect, this invention proposes a full-cycle fault handling system for power distribution terminals based on NPU collaboration. The power distribution terminal adopts a heterogeneous multi-core architecture, which includes a real-time core and an NPU core. The real-time core collects electrical quantity data from the distribution network and records waveforms. The real-time core transmits the waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. The NPU core determines whether a fault has occurred based on the waveform data. The waveform recording of the real-time core and the fault identification of the NPU core are performed simultaneously. When a fault is detected, the NPU core sends a fault warning to the real-time core. When the real-time core receives a fault warning, it pre-locates the fault section. The NPU core pre-calculates the level and power of the power loss load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power loss area as the optimization objective, and finally locates the fault section and performs physical isolation based on the optimal isolation boundary. The NPU core aims to maximize the total restored load, minimize distribution network transfer losses, and minimize the number of switching operations as fault recovery objectives. Under distribution network constraints, it solves for these objectives to obtain the optimal transfer strategy. This optimal strategy is then distributed to the real-time core in the form of a circuit breaker operation sequence, which includes the operation order and the time interval between adjacent operations. After executing one operation, the real-time core transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it uses real-time global topology and power flow data of the distribution network to solve for the fault recovery objectives under distribution network constraints to obtain a supplementary transfer operation, which the real-time core executes. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

[0114] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0115] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0116] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0117] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for full-cycle fault handling of distribution terminals based on NPU collaboration, characterized in that, The power distribution terminal adopts a heterogeneous multi-core architecture, which includes a real-time core and an NPU core; the method includes: The real-time core collects electrical quantity data from the distribution network and records waveforms. The real-time core transmits the waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. The NPU core determines whether a fault has occurred based on the waveform data. The waveform recording of the real-time core and the fault identification of the NPU core are performed simultaneously. When a fault is detected, the NPU core sends a fault warning to the real-time core. When the real-time core receives a fault warning, it pre-locates the fault section. The NPU core pre-calculates the level and power of the power loss load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power loss area as the optimization objective, and finally locates the fault section and performs physical isolation based on the optimal isolation boundary. The NPU core aims to maximize the total restored load, minimize distribution network transfer losses, and minimize the number of switching operations as fault recovery objectives. Under distribution network constraints, it solves for these objectives to obtain the optimal transfer strategy. This optimal strategy is then distributed to the real-time core in the form of a circuit breaker operation sequence, which includes the operation order and the time interval between adjacent operations. After executing one operation, the real-time core transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it uses real-time global topology and power flow data of the distribution network to solve for the fault recovery objectives under distribution network constraints to obtain a supplementary transfer operation, which the real-time core executes. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

2. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 1, characterized in that, After the optimal power transfer strategy is executed, the master station receives the local power transfer results uploaded by each power distribution terminal; The main station performs global multi-terminal coordination optimization based on the global network topology, power margin, and all received local power transfer results in order to detect and eliminate conflicts between different terminal power transfer schemes in terms of contact switch status and backup capacity allocation. The master station will send the globally coordinated instructions generated after coordination and optimization to the corresponding power distribution terminals to correct or confirm the local power transfer strategies of each power distribution terminal.

3. The method for full-cycle fault handling of power distribution terminals based on NPU collaboration according to claim 1, characterized in that, The steps for real-time acquisition of electrical quantity data of the power distribution network and waveform recording include: The real-time kernel has multiple built-in abnormal condition triggering mechanisms, which will trigger waveform recording when any of the following conditions are met: 1) The rate of change of current is greater than the set threshold for the rate of change of current; 2) The voltage level after the drop is less than the set voltage threshold; 3) The harmonic content is greater than the set content threshold; 4) The zero-sequence current is greater than the set zero-sequence current threshold; Among them, various thresholds are adaptively adjusted based on historical statistical data, and the adjustment method for each threshold is the same, as shown in the following formula: In the formula, For the updated threshold, The threshold before the update. This is a static threshold. This is the smoothing coefficient.

4. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 1, characterized in that, The steps by which the real-time core transmits waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism include: A circular buffer is divided in the shared memory. The real-time core writes the waveform data into the buffer through direct memory access, and the NPU core reads the data by maintaining read and write pointers through atomic operations. The real-time core configures the NPU core's inference priority register to high-throughput mode and starts the NPU core's parallel inference task through the cooperative control instruction set; Based on this configuration, the NPU core executes multiple device status assessment models in parallel and processes the waveform recording data of multiple power distribution terminals concurrently.

5. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 4, characterized in that, The steps by which the NPU core determines whether a fault has occurred based on the recorded waveform data include: The NPU core uses a pre-trained first neural network model to extract features and recognize patterns from the waveform data, and outputs a probability distribution of fault types; the first neural network model is a fusion of a one-dimensional convolutional neural network and a long short-term memory network. When the probability distribution of the fault type meets the preset warning conditions, a fault warning message is generated and fed back to the real-time kernel; the fault warning message includes the fault type, probability value and expected occurrence time window.

6. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 1, characterized in that, The step of pre-locating the faulty section when the real-time core receives a fault warning includes: When a fault is confirmed, the fault characteristic information of the distribution terminal is extracted in real time, and the status information of the adjacent distribution terminals is obtained through horizontal peer-to-peer communication. The fault characteristic information includes fault current, current direction, voltage amplitude and other characteristic information, and the status information includes line switch status, load distribution and network topology. Real-time fault location is determined based on the fault characteristics of the current power distribution terminal and the status information of adjacent power distribution terminals.

7. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 6, characterized in that, When the real-time core detects a fault and determines that it has entered the isolation phase, the real-time core immediately updates the NPU core's inference priority register to the lowest latency mode. After completing the register update, the real-time core sends a high-priority preemptive interrupt to the NPU core through a bidirectional deterministic interrupt mechanism. After receiving an interrupt, the NPU core saves the complete context of the currently executing non-urgent task, loads the emergency inference model related to the current fault location, and begins to execute the single-path fault location inference task.

8. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 7, characterized in that, The NPU core pre-calculates the level and power of the power outage load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power outage area as the optimization objective. The steps of finally locating the fault section and performing physical isolation based on the optimal isolation boundary include: In the lowest latency mode, the NPU core uses known global topology and real-time power flow data to pre-calculate the total power outage load and the impact assessment of important loads under different potential isolation boundaries in parallel; the pre-calculation results are quickly sent to the real-time core in the form of "suggested isolation cut-off points"; After obtaining the "suggested isolation cut-off point" in real time, the optimal isolation boundary is selected with the minimum load power loss area as the optimization objective, and the fault section isolation is performed. The real-time core synchronously transmits the status data after fault isolation to the NPU core. The status data includes the fault segment number, the isolated network topology matrix, the circuit breaker status, and the load distribution data. The status data is transmitted in a structured data serialization format.

9. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 1, characterized in that, The NPU core aims to maximize the total restored load, minimize distribution network transfer losses, and minimize the number of switching operations as fault recovery objectives. The steps for solving these objectives to obtain the optimal transfer strategy under distribution network constraints include: The NPU core obtains status data, load priority, and short-term source-load forecast data from the real-time core and local cache. Under distribution network constraints, a multi-objective power transfer optimization model is established based on the fault recovery objective. The objective function of the model includes: 1) Maximize the total weighted recovery load, where the weighting coefficient is set according to the load priority; 2) Minimize power transfer losses in the distribution network; 3) Minimize the number of switching actions; The NPU core uses a first optimization algorithm to perform a preliminary solution to the multi-objective power transfer optimization model and generate a set of candidate power transfer schemes; the first optimization algorithm includes a heuristic swarm optimization algorithm. The NPU core uses a pre-trained second neural network model to finely optimize the set of candidate transfer schemes and generate the optimal transfer strategy; the second neural network model includes a deep reinforcement learning model.

10. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 9, characterized in that, Objective function to maximize the total weighted recovery load As shown in the following formula: In the formula, For load nodes The recovery power; For load nodes Priority weights; For load nodes The decision variable has a value of 1 indicating that power has been restored and a value of 0 indicating that power has not been restored. This represents the number of load nodes.

11. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 9, characterized in that, Objective function to minimize distribution network transfer losses As shown in the following formula: In the formula, , Branch roads The current and resistance, E represents the duration of the transfer, and E represents the set of branches.

12. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 9, characterized in that, Objective function: Minimize the number of switching actions As shown in the following formula: In the formula, For switch The state variable is denoted by 0, which indicates that the switch is open, and 1, which indicates that the switch is closed; S is the set of switches.

13. The method for full-cycle fault handling of distribution terminals based on NPU collaboration according to claim 9, characterized in that, The optimal transfer strategy is issued to the real-time core in the form of a circuit breaker operation sequence. The circuit breaker operation sequence includes the operation order and the time interval between two adjacent operations, and includes the following steps: The NPU core performs a security check on the optimal transfer strategy, and after the check passes, it sends the strategy to the real-time core for execution. Security checks include: 1) Calculate the three-phase / single-phase short-circuit current of each node after the power transfer, and ensure that it does not exceed the switching capacity of the switch; 2) Verify the coordination of protection settings between upper and lower levels to ensure that the time difference is not less than the set value; The time interval between two consecutive operations shall not be less than 2 seconds.

14. A distribution terminal fault full-cycle handling system based on NPU collaboration, used to implement the distribution terminal fault full-cycle handling method based on NPU collaboration as described in any one of claims 1 to 13, characterized in that, The power distribution terminal adopts a heterogeneous multi-core architecture, which includes a real-time core and an NPU core; The real-time core collects electrical quantity data from the distribution network and records waveforms. The real-time core transmits the waveform data to the NPU core via an on-chip high-speed bus and a shared memory mechanism. The NPU core determines whether a fault has occurred based on the waveform data. The waveform recording of the real-time core and the fault identification of the NPU core are performed simultaneously. When a fault is detected, the NPU core sends a fault warning to the real-time core. When the real-time core receives a fault warning, it pre-locates the fault section. The NPU core pre-calculates the level and power of the power loss load under different potential isolation boundaries in parallel based on the real-time global topology and power flow data of the distribution network. The real-time core selects the optimal isolation boundary with the minimum load power loss area as the optimization objective, and finally locates the fault section and performs physical isolation based on the optimal isolation boundary. The NPU core aims to maximize the total load recovery, minimize the distribution network transfer loss, and minimize the number of switching operations as fault recovery objectives. Under the constraints of the distribution network, it solves the fault recovery objectives to obtain the optimal transfer strategy. The optimal transfer strategy is issued to the real-time core in the form of a circuit breaker operation sequence, which includes the operation order and the time interval between two adjacent operations. After the real-time core completes an operation, it transmits the waveform data to the NPU core. If the NPU core determines that the total restored load is insufficient, it solves the fault recovery target based on the real-time global topology and power flow data of the distribution network under the constraints of the distribution network to obtain the power transfer supplement operation, and the real-time core executes the power transfer supplement operation. If the NPU core determines that the total restored load is sufficient, the real-time core executes the next operation.

15. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-13.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-13.

Images