Power distribution network collaborative management method and system based on artificial intelligence

Abstract

The invention discloses a power distribution network collaborative management method and system based on artificial intelligence, and relates to the technical field of electrochemical detection, and the method comprises the steps: constructing a distributed edge computing node network, deploying nodes at key positions of a power distribution network, achieving the collection and preprocessing of local power data, and reducing the cross-regional data transmission pressure; an AI real-time communication scheduling model is established based on the preprocessed data, communication resources are dynamically allocated according to the operation state of the power distribution network, and fault data transmission is guaranteed preferentially; seamless interaction of multi-protocol equipment is realized through a self-adaptive protocol conversion mechanism containing protocol identification, format conversion and data verification; training a fault diagnosis model by using a federated learning framework, and enabling edge nodes to only upload parameters to a coordination center for aggregation and updating, so as to balance model precision and data privacy; when a fault is detected, a millisecond response mechanism is started, and a processing strategy is generated and executed in combination with edge local decision and central global optimization.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electrochemical detection, in particular to a power distribution network collaborative management method and system based on artificial intelligence. BACKGROUND

[0002] With large-scale access of renewable energy, the operation characteristics of the power distribution network have fundamentally changed. For example, it is estimated that by 2030, the installed capacity of distributed photovoltaic power in China will reach 1 billion kilowatts, and the number of electric vehicles will reach 110 million. The strong intermittency and volatility of these resources pose a serious challenge to the stability of the grid voltage and frequency. The traditional centralized control mode is difficult to cope with the real-time changes in power flow distribution, resulting in high curtailment rate of wind and light and increased risk of user power outage.

[0003] The power distribution network needs to change from the one-way power supply mode of "source following load" to the multi-agent collaborative interaction of "source grid load storage". For example, the output fluctuation of distributed power needs to be smoothed by energy storage, and user-side adjustable load (such as electric vehicles and smart buildings) needs to participate in peak shaving. However, dispersed resources are difficult to integrate in a unified space-time dimension, resulting in insufficient grid regulation capacity.

[0004] AI technology (such as machine learning and reinforcement learning) provides a bottom support for solving the above problems. High-precision weather prediction and new energy power prediction improve the control of wind and light output; based on the dynamic prediction of load and resource distribution of regional society and natural attributes; driving complex grid simulation and power market transaction prediction, optimizing dispatching decisions.

[0005] Therefore, there is an urgent need for a power distribution network collaborative management method and system based on artificial intelligence. In the face of data real-time and communication bottlenecks, including the need for millisecond-level response of the power distribution network, but the traditional communication technology has high latency and insufficient bandwidth. For example, under multi-line fault conditions, traditional methods may cause protection misoperation or refusal due to communication delay; although 5G URLLC (Ultra-Reliable Low-Latency Communication) and Time-Sensitive Network (TSN) can reduce the latency to within 10ms, actual deployment still faces challenges such as protocol compatibility and device heterogeneity. SUMMARY

[0006] To solve the above technical problems, a power distribution network collaborative management method and system based on artificial intelligence are provided, which solves the problem of data real-time and communication bottlenecks.

[0007] To achieve the above purposes, the technical scheme adopted by the application is as follows: A power distribution network collaborative management method based on artificial intelligence, comprising: A distributed edge computing node network is constructed, and the edge computing nodes are deployed at key nodes of the power distribution network for collecting and preprocessing local power data; Based on local power data, an artificial intelligence-based real-time communication scheduling model is established, which dynamically allocates communication resources according to the operation state of the distribution network, and prioritizes the transmission of fault handling related data; An adaptive protocol conversion mechanism is adopted to realize data interaction between devices with different communication protocols, which includes three sub-processes of protocol identification, format conversion and data verification. A distribution network fault diagnosis model is trained through a federated learning framework. The model only uploads model parameters after local training on each edge computing node, and is updated in the coordination center. When a fault in the distribution network is detected, a millisecond-level response mechanism is started, and a fault handling strategy is generated and executed through a combination of local decision-making by edge computing nodes and global optimization by the coordination center.

[0008] Preferably, the construction of the distributed edge computing node network specifically includes: Based on the topology of the distribution network and the load density, the edge computing nodes are divided into three levels. The first level nodes are deployed at feeder section switches and tie switches, responsible for regional data aggregation and collaborative control. The second level nodes are deployed at distributed power grid connection points and energy storage system interfaces, dedicated to high-frequency collection of new energy output data. The third level nodes are deployed at user-side smart meters, charging piles and adjustable load control terminals, in the form of embedded micro nodes, realizing real-time sensing of user-side data. Each edge computing node adopts a heterogeneous computing architecture, including a main processing unit, an industrial-grade CPU responsible for task scheduling and regular data processing, an acceleration unit integrating a GPU for local AI model inference and an FPGA dedicated to high-precision time-frequency analysis. The data acquisition interface is configured with a multi-protocol acquisition module, supporting analog, digital and high-frequency pulse quantity acquisition, with dynamic adjustable sampling frequency. The communication module is equipped with a TSN chip and a 5G URLLC module, supporting wired and wireless dual-link redundant communication.

[0009] Preferably, the edge computing node for collecting and preprocessing local power data specifically includes: Local data collection is divided into timed collection and triggered collection. Timed collection collects basic data such as voltage, current and power at a preset period during normal operation. Triggered collection automatically triggers high-frequency collection to record transient waveform characteristics before and after the fault when voltage sudden change or current out-of-limit anomaly is detected. Both collected data are attached with time stamps and node IDs to ensure spatiotemporal consistency. The data preprocessing includes: the cleaning layer uses isolated forest to detect abnormal values; the feature extraction layer extracts transient and steady-state features through wavelet transform and Fourier transform to generate compressed feature vectors; the standardization layer converts the data to a certain interval and dynamically updates the mean and variance through a sliding window; and the preprocessed data is compressed through sparse coding to reduce transmission volume and retain key features. Each edge node constructs a self-organizing network through a distributed hash table, a new node automatically joins and obtains a topological relationship; when the CPU occupancy rate of a node exceeds a certain proportion, non-real-time tasks are automatically migrated to a low-load adjacent node; when the communication of a certain node is interrupted, an adjacent node takes over its data acquisition task and temporarily stores data through a local cache, and after the communication is restored, the data is supplemented.

[0010] Preferably, the establishment of the real-time communication scheduling model based on artificial intelligence specifically includes: The local power data preprocessed by the edge node is taken as input to construct a joint perception mechanism of the power distribution network and the communication network state; The power grid parameter features are fused through a time series convolution network to generate a power grid operation state vector; the channel occupancy rate, time delay and packet loss rate of the TSN / 5GURLLC network are collected in real time, and a network health score is calculated through a sliding window; an attention mechanism is used to establish an association model between the power grid and the communication network, and the communication network hard constraints are automatically associated during power grid failure; A three-level priority system is constructed based on the power grid state and data types: the first level is the hard real-time data of fault feature data and protection control instructions, which are bound to exclusive channels; the second level is the soft real-time data of distributed power regulation instructions and energy storage control signals; and the third level is the non-real-time data of conventional measurement data and equipment state data; during power grid failure, the relevant second-level data in the fault area is temporarily upgraded to the first level through a rule engine, and falls back within 100ms after recovery.

[0011] Preferably, the establishment of the real-time communication scheduling model based on artificial intelligence further includes: A resource allocation sub-model based on deep reinforcement learning is used to construct an intelligent decision-making framework for dynamic allocation of communication resources, the state space includes the power grid operation state vector, the communication network health score, the data backlog of each priority queue, and the available channel bandwidth; the action space defines three types of actions: channel allocation, modulation mode adjustment, and transmission cycle compression; a multi-level reward mechanism is set for the reward function; a double-delay deep deterministic policy gradient algorithm is used to train the agent, and the agent is locally inferred on the edge node to update the resource allocation strategy periodically; The communication load prediction and pre-allocation sub-model realizes short-time prediction of communication flow by combining historical data and real-time state, and reserves key resources in advance; the input features include communication flow time series data, current power grid state, meteorological data, and user power consumption mode label; the prediction model uses an LSTM network enhanced by an attention mechanism to predict the peak value of communication flow and focuses on identifying flow characteristics before failure; the heterogeneous network adaptive switching submodule constructs an intelligent switching mechanism in view of the protocol differences and performance complementarity of TSN and 5G URLLC.

[0012] Preferably, the mechanism includes three sub-processes of protocol identification, format conversion and data verification, which specifically include: The protocol identification sub-process parses the received data packet in multiple layers to extract 18-dimensional features of the physical layer, data link layer and application layer; a two-level identification architecture of deep CNN and expert rules is adopted, the first level outputs the protocol type probability distribution through lightweight CNN classification, and the second level verifies the low probability result through the expert rule library; when a new protocol device is accessed, the edge node automatically collects samples and updates the model through federated transfer learning; The format conversion sub-process converts heterogeneous data into a unified intermediate format with extended JSON-LD as the carrier, which includes core and extended fields; a protocol-intermediate format mapping table is constructed to realize field mapping, data type conversion and time sequence alignment; the conversion time is optimized through reinforcement learning, FPGA acceleration channels are enabled for high-priority data, and compression conversion is adopted for non-real-time data; The data verification sub-process generates SHA-256 hash values for data packets before and after conversion and stores them in the local blockchain, checks the integrity through chain hash verification, and verifies large data packets using Merkle tree; the data rationality is detected based on the Isolation Forest model, and abnormal values are removed through weighted voting of multi-source data; the data timestamp timeout condition is checked and retransmission is triggered, the interval continuity of time sequence data is verified, and interpolation correction is automatically performed when the deviation exceeds 1ms.

[0013] Preferably, the power distribution network fault diagnosis model trained through the federated learning framework specifically includes: The framework topology is constructed using a two-level architecture of a coordination center and edge nodes, the coordination center is deployed in the power distribution network control master station and is responsible for global model aggregation and parameter distribution; each edge computing node acts as a federated learning client and trains based on locally collected fault data, and the nodes interact through encrypted communication links; The basic model uses a lightweight hybrid neural network as the initial model, including a feature extraction layer, a time series modeling layer, and a classification output layer; the coordination center pre-trains the initial model based on the historical global fault data set, compresses and packages the model parameters, and then distributes them to each edge node to ensure that the initial model has basic fault identification capability.

[0014] Preferably, the model only uploads model parameters after local training at each edge computing node, and the specific update aggregation at the coordination center includes: Each edge node constructs a training set based on historical fault data and labels the fault labels; a timing and event-driven training trigger mechanism is adopted, and only when the local model performance is better than the last round, the parameters are prepared for uploading; The edge node extracts the update parameters, compares only the threshold change amount with the last round of global parameters, and uses homomorphic encryption after L1 regularization sparsification; After the coordination center receives the parameters, it detects abnormal parameters through range checking, cosine similarity calculation and contribution evaluation; An improved federated average algorithm is used to calculate the weight according to the sample size, local model accuracy and historical contribution, and the decrypted parameters are weighted and aggregated, and after verification, the accuracy is improved to confirm the effectiveness, otherwise trigger the second aggregation.

[0015] Preferably, by combining local decision-making of edge computing nodes with global optimization of the coordination center, a fault handling strategy is generated and executed, which specifically includes: After the coordination center receives the fault information, it integrates global data including real-time load distribution of non-fault areas; current output and predicted value of distributed power; remaining capacity and charge / discharge rate limit of energy storage system; available state of tie switches and standby power; The recovery strategy is generated, a deep deterministic policy gradient reinforcement learning algorithm is used, a multi-objective function is generated, a load transfer path is generated, a distributed power output adjustment instruction is generated, a charge / discharge strategy of an energy storage system is generated, and a switch operation sequence is generated.

[0016] Further, an artificial intelligence-based power distribution network collaborative management system includes: A distributed edge computing module is deployed at a key node of the power distribution network and is used to collect and preprocess local power data; An intelligent communication scheduling module is in communication connection with the distributed edge computing module and is used to dynamically allocate communication resources based on the operating state of the power distribution network, the module includes a communication load prediction unit, a resource dynamic allocation unit, and a priority management unit, wherein the priority management unit sets the fault handling related data as the highest transmission priority; An adaptive protocol conversion module is connected with the distributed edge computing module and the intelligent communication scheduling module, respectively, and the module includes a protocol identification unit, a format conversion unit, and a data verification unit, wherein the protocol identification unit uses a hybrid identification architecture that combines deep learning and rule engine; A federated learning fault diagnosis module includes an edge node local training unit that trains a fault diagnosis model based on local power data and only uploads model parameters, and a coordination center aggregation unit that performs weighted aggregation on the uploaded parameters to update the global model; A millisecond-level fault response module is configured to generate and execute a fault processing strategy by combining local decision-making of an edge computing node with global optimization of a coordination center when a power distribution network fault is detected, and the module includes a fault rapid detection unit, a local rapid isolation unit, a global optimization decision unit, and a strategy execution unit. A coordination decision center is configured to receive information uploaded by each module and perform global coordination, and the coordination decision center communicates with the distributed edge computing module through a TSN and 5GURLLC fusion network.

[0017] Compared with the prior art, the present application has the following advantages: A distributed edge computing node network is hierarchically deployed according to key nodes of a power distribution network, localizes data acquisition and preprocessing, reduces cross-regional data transmission volume, and realizes millisecond-level data processing through a heterogeneous computing architecture (CPU+GPU+FPGA) to provide a low-latency, high-quality data source for subsequent AI models, thereby solving the problems of data accumulation and high latency in traditional centralized data acquisition.

[0018] An AI-based real-time communication scheduling model perceives power grid states through a timing convolution network, predicts communication load through LSTM, dynamically allocates TSN / 5GURLLC resources through deep reinforcement learning, sets fault data as "hard real-time" through a three-level priority system, and binds exclusive channels to ensure that the transmission latency of fault data is ≤5ms, thereby avoiding the waste of traditional static communication resources and the risk of protection misoperation caused by lag in fault data transmission.

[0019] An adaptive protocol conversion mechanism uses a two-level identification of deep CNN and expert rules to unify heterogeneous data such as IEC61850 and Modbus into an extended JSON-LD format, and combines blockchain hash verification and anomaly detection to solve the problems of low efficiency and poor compatibility of traditional manual protocol adaptation, and to ensure data integrity and lay a foundation for unified data interaction in multi-agent coordination.

[0020] A federated learning framework allows edge nodes to upload only model parameters, and the coordination center optimizes a global fault diagnosis model through weighted aggregation, thereby avoiding privacy leakage caused by cross-node transmission of original fault data, and improving the cross-region generalization ability of the model through knowledge distillation, thereby breaking through the limitations of high privacy risk and poor generalization of traditional centralized training. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of a power distribution network collaborative management method based on artificial intelligence; Figure 2 An internal framework diagram of a power distribution network collaborative management system based on artificial intelligence. DETAILED DESCRIPTION

[0022] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be made by those skilled in the art.

[0023] Referring to Figure 1 As shown in the figure, an artificial intelligence-based power distribution network collaborative management method comprises: A distributed edge computing node network is constructed, and the edge computing nodes are deployed at key nodes of the power distribution network for collecting and preprocessing local power data; Based on the local power data, an artificial intelligence-based real-time communication scheduling model is established, which dynamically allocates communication resources according to the operation state of the power distribution network, and prioritizes the transmission of fault handling related data; An adaptive protocol conversion mechanism is adopted to realize data interaction between devices with different communication protocols, which includes three sub-processes of protocol identification, format conversion and data verification; A power distribution network fault diagnosis model is trained through a federated learning framework, which only uploads model parameters after local training at each edge computing node, and is updated at the coordination center; When a fault in the power distribution network is detected, a millisecond-level response mechanism is started, and a fault handling strategy is generated and executed through a combination of local decision-making by the edge computing nodes and global optimization by the coordination center.

[0024] It should be noted that the distributed node network realizes millisecond-level collection and filtering processing of high-frequency data such as topological end current, voltage and harmonic, significantly reducing the bandwidth pressure of uploading original data; The adaptive protocol conversion mechanism realizes cross-vendor device data interconnection through protocol fingerprint library matching and data semantic analysis, with compatibility covering more than 95% of mainstream devices; The federated learning framework generates gradient parameters through local training of edge nodes under the premise of ensuring data privacy, and the coordination center generates a global diagnosis model through weighted average aggregation, solving the problems of data island and centralized training computing power bottleneck.

[0025] The distributed edge computing node network specifically comprises: Based on the topology structure and load density of the power distribution network, the edge computing nodes are divided into three levels of deployment, the first level nodes are deployed at feeder sectionalizing switches and tie switches, responsible for regional level data aggregation and collaborative control; the second level nodes are deployed at distributed power grid connection points and energy storage system interface ends, dedicated to high-frequency collection of new energy output data; the third level nodes are deployed at user-side smart meters, charging piles and adjustable load control terminals, in the form of embedded micro nodes, realizing real-time sensing of user-side data; Each edge computing node adopts a heterogeneous computing architecture, including a main processing unit, an industrial-grade CPU responsible for task scheduling and regular data processing, and an acceleration unit integrating a GPU for local AI model inference and an FPGA dedicated to high-precision time-frequency analysis. The data acquisition interface is configured with a multi-protocol acquisition module that supports analog, digital, and high-frequency pulse quantity acquisition, with a dynamic adjustable sampling frequency. The communication module is built-in with a TSN chip and a 5G URLLC module, supporting wired and wireless dual-link redundant communication.

[0026] The edge computing node is used to collect and preprocess local power data, specifically including: Local data acquisition is divided into timed acquisition and triggered acquisition. Timed acquisition is the acquisition of basic data such as voltage, current, and power at a preset period during normal operation. Triggered acquisition is the automatic triggering of high-frequency acquisition to record transient waveform characteristics before and after the fault when voltage sudden change or current out-of-limit anomaly is detected. Both acquisition data are attached with timestamps and node IDs to ensure spatiotemporal consistency. Data preprocessing includes: the cleaning layer uses isolation forest to detect outliers; the feature extraction layer extracts transient and steady-state features through wavelet transform and Fourier transform to generate compressed feature vectors; the standardization layer converts data to a specified interval and dynamically updates mean and variance through a sliding window; after preprocessing, data is compressed through sparse coding to reduce transmission volume and retain key features. Each edge node constructs a self-organizing network through a distributed hash table, and new nodes automatically join and obtain topological relationships. When the CPU occupancy rate of a node exceeds a certain proportion, non-real-time tasks are automatically migrated to a low-load neighboring node. When a node's communication is interrupted, neighboring nodes take over its data acquisition tasks and temporarily store data through local caching. After communication is restored, the data is transmitted.

[0027] It should be noted that the three-level deployment architecture: the first-level node (regional control layer) is deployed at key switch points of the feeder and automatically switches topology by real-time monitoring of short-circuit current direction. The second-level node (energy interface layer) is directly embedded in the internal controller of the photovoltaic inverter and energy storage PCS, capturing millisecond-level power fluctuations of new energy at a sampling frequency of 10 kHz, solving the distortion of harmonic analysis caused by the second-level acquisition of traditional SCADA systems. The third-level node (user perception layer) is a micro node deeply integrated into the smart meter chip set, identifying electric vehicle charging behavior through load current high-frequency harmonic features. Computing power decreases exponentially: the first-level node has a computing power of 15 TOPS (AI inference), and the third-level node only needs 0.5 TOPS (edge feature extraction). Communication redundancy difference: the first-level node has a dual-link bandwidth of >1 Gbps, and the third-level node uses lightweight NB-IoT (uplink 50 kbps).

[0028] Heterogeneous computing, main processing unit (control flow core), real-time kernel (such as VxWorks RTOS), task scheduling delay <10 μs; dynamic load balancing algorithm, when CPU occupancy >75%, automatically migrate data cleaning tasks to adjacent nodes (migration decision time <5 ms); Acceleration unit (data flow engine), GPU parallel inference: deploy the quantized fault diagnosis model (ResNet18 compressed to 3MB), single-frame inference time <8 ms; FPGA time-frequency analysis, 128-channel synchronous acquisition, 1024-point FFT completed by CORDIC algorithm in 0.15 ms; Dynamic sampling technology, normal state, basic power sampling rate 1 kHz (satisfies 0.5S level measurement accuracy); transient state trigger, fault instant switches to 100 kHz sampling (captures 5 μs level arc characteristics).

[0029] Data preprocessing: abnormal cleaning layer, improved isolated forest algorithm, introduces the spatiotemporal correlation of power data; Feature extraction layer, wavelet transform uses db6 wavelet basis to extract transient traveling wave characteristics (time resolution reaches 10 μs); frequency domain enhancement, suppresses spectral leakage by FFT superimposed black window function, harmonic detection accuracy is improved to 0.1%; Sparse coding compression, train power-specific dictionary: generate over-complete basis based on K-SVD algorithm, feature compression ratio reaches 20:1; key feature preservation mechanism, scalar quantization (SQ) coding protection is performed on fault waveform peak points.

[0030] Self-organizing network, topology adaptation, distributed hash table (DHT) constructs logical overlay network, new node joins within 300 ms to complete topology registration; dynamic routing protocol, based on link quality to select the optimal transmission path in real time.

[0031] The real-time communication scheduling model based on artificial intelligence specifically includes: Taking the local power data preprocessed by the edge node as input, a joint perception mechanism of the power distribution network and the communication network state is constructed; Through a time series convolution network, power grid parameter features are fused to generate a power grid operating state vector; the channel occupancy rate, delay, and packet loss rate of the TSN / 5G URLLC network are collected in real time, and the network health score is calculated combined with a sliding window; an attention mechanism is used to establish an association model between the two, and the communication network hard constraints are automatically associated when the power grid fails; A three-level priority system is constructed based on the grid state and data type: the first level is the hard real-time data of fault feature data and protection control instructions, which is bound to a dedicated channel; the second level is the soft real-time data of distributed power regulation instructions and energy storage control signals; the third level is the non-real-time data of conventional measurement data and equipment state data; when the power grid fails, the relevant second-level data in the fault area is temporarily upgraded to the first level by the rule engine and falls back within 100 ms after recovery.

[0032] The real-time communication scheduling model based on artificial intelligence further includes: A resource allocation sub-model based on deep reinforcement learning constructs an intelligent decision-making framework for dynamic allocation of communication resources. The state space includes the grid operation state vector, the communication network health score, the data backlog of each priority queue, and the available channel bandwidth. The action space defines three types of actions: channel allocation, modulation mode adjustment, and transmission cycle compression. A multi-level reward mechanism is set for the reward function. The agent is trained using a double-delay deep deterministic policy gradient algorithm, and the lightweight inference is performed locally on the edge node to update the resource allocation strategy periodically. A communication load prediction and pre-allocation sub-model combines historical data and real-time state to achieve short-term prediction of communication traffic and reserve key resources in advance. The input features include communication traffic time series data, current grid state, weather data, and user power consumption pattern labels. The prediction model uses an LSTM network enhanced with an attention mechanism to predict communication traffic peaks, focusing on identifying traffic characteristics before a fault. A heterogeneous network adaptive switching submodule is constructed to address the protocol differences and performance complementarity of TSN and 5G URLLC, establishing an intelligent switching mechanism.

[0033] It should be noted that the spatio-temporal fusion perception engine extracts 128-dimensional features (time resolution 1 ms) such as voltage phase jump and zero sequence current mutation using a time convolution network (TCN), generating a dynamic weight grid state vector. The communication state evaluation calculates the network health index (NHS) based on a sliding window and introduces a jitter sensitivity coefficient. The cross-domain attention model automatically upgrades the communication resource guarantee level to the highest when detecting a short-circuit current rise rate di / dt > 5 kA / ms (response delay < 8 ms). The fault correlation prediction mechanism establishes a mapping rule library of "electrical characteristics-communication demand" (e.g., when arc features are detected, 2 MHz bandwidth is forcibly reserved for traveling wave positioning data transmission). Extreme scenario warning: typhoon weather data triggers pre-allocation mode, and disaster prevention control instructions are upgraded to the highest priority level in advance.

[0034] Hierarchical logic essence:

[0035] Evolutionary ability of rule engine Upgrade decision based on causal reasoning, when the risk of photovoltaic inverter island operation is detected, its state data is automatically upgraded to level two; Time window optimization of fallback mechanism, exponential decay fallback algorithm is adopted (degrade by 3 steps in 100ms after recovery).

[0036] Deep reinforcement learning, resource allocation model realizes Pareto optimality under high-dimensional constraints; Decision space, state space dimension reduction, 128-dimensional power grid vector is compressed to 32-dimensional through autoencoder (information retention rate > 99%); Action space expansion, channel allocation, dynamic division of 5G URLLC Grant-Free resource block; Modulation adaptation, QPSK / 16QAM switching (error rate > 10 -5 Time backoff); Period compression, non-real-time data sampling period is extended from 1s to 5s; Double delay DDPG improvement, introduce priority experience replay (PER) mechanism, fault scenario sample weight is improved by 3 times; Edge-cloud collaborative training, agent performs inference on edge node (inference time ≤3ms), parameters are uploaded to cloud for retraining every month.

[0037] Triple protection of prediction-preallocation-switch, prediction model integrates multi-source data including power characteristics, load change rate dP / dt; Environmental factors, rainfall attenuation model (affecting 5G signal); Behavior label, industrial user lunchtime production peak mode; LSTM-ATT network architecture, attention mechanism focuses on 5-minute data window before failure (prediction accuracy reaches 94%); Seamless handover of heterogeneous network:

[0038] The mechanism includes three sub-processes of protocol identification, format conversion and data verification, which specifically include: Protocol identification sub-process, multi-layer parsing of received data packets, extracting 18-dimensional features of physical layer, data link layer and application layer; Adopting two-level identification architecture of deep CNN and expert rules, the first level classifies and outputs protocol type probability distribution through lightweight CNN, and the second level verifies the low probability results through expert rule base; When accessing new protocol devices, edge nodes automatically collect samples and update models through federated transfer learning; Format conversion sub-process, convert heterogeneous data into unified intermediate format with extended JSON-LD as carrier, including core and extended fields; Build protocol-intermediate format mapping table to realize field mapping, data type conversion and time alignment; Optimize conversion time through reinforcement learning, enable FPGA acceleration channel for high-priority data, and use compression conversion for non-real-time data; Data verification sub-process, SHA-256 hash value is generated before and after conversion and stored in the local blockchain, the integrity is verified by chain hash, large data packets are verified by Merkle tree; Based on the isolation forest model to detect data rationality, multiple source data is removed by weighted voting; Check the timeout of the data timestamp and trigger the retransmission, verify the interval continuity of the time series data, and automatically interpolate and correct when the deviation is more than 1ms.

[0039] It should be noted that the multi-layer feature deconstruction technology: Physical layer, extract modulation mode, symbol rate, carrier frequency offset and other 6-dimensional features (accuracy up to 0.1ppm); Data link layer, parse MAC frame structure, CRC check mode, traffic mode (such as Modbus polling period feature); Application layer, deep deconstruction of IEC61850 MMS service mapping, DNP3 object header structure (supporting 200+ protocol fingerprint library); Lightweight CNN model, using deep separable convolution to build MobileNet variant (model volume <500KB); Output protocol probability vector; Expert rule base bottom-up mechanism, start rule verification for results with confidence less than 85% (such as DNP3 0x0564 start symbol detection), built-in protocol conflict arbitration algorithm (when multiple protocol features coexist, decide by device model tracing); New protocol adaptation process, new device access → automatically capture 1000+ samples → locally fine-tune CNN bottom convolution kernel → upload gradient to coordination center → aggregate to generate global protocol identification model.

[0040] Blockchain storage mechanism, lightweight chain architecture, each node only stores its own data hash (storage overhead <1MB / year), using improved PBFT consensus (verification delay <10ms); Merkle tree optimization, million-level data packet verification time <50ms, supports dynamic segment-based verification; Intelligent rationality verification, isolation forest enhanced model, introduces power grid topology constraints, such as substation outgoing line power conservation law, adds weather correlation factor to photovoltaic output data (false negative rate <0.01%); Multi-source voting mechanism, the same measurement value is weighted voted by adjacent 3 nodes (weight = signal strength × device reputation value), automatically removes ±3σ data and marks the credibility label; Timeout retransmission strategy, hard real-time data, retransmitted within 200ms after loss (up to 3 times); Non-real-time data, using forward prediction compensation (ARIMA model filling); Continuity correction technology, Kalman filter smoothing is used for PMU phasor data, and virtual sampling points are automatically inserted to maintain waveform integrity in voltage sag events.

[0041] The power distribution network fault diagnosis model is trained through a federal learning framework, and specifically includes: The framework topology is constructed using a two-level architecture of a coordination center and edge nodes. The coordination center is deployed in the power distribution network control master station and is responsible for global model aggregation and parameter distribution. Each edge computing node serves as a federal learning client and carries out training based on locally collected fault data. The nodes interact with each other through an encrypted communication link to exchange parameters. The basic model uses a lightweight hybrid neural network as the initial model, which includes a feature extraction layer, a time series modeling layer, and a classification output layer. The coordination center pre-trains the initial model based on historical global fault data sets, compresses and packages the model parameters, and then distributes them to each edge node to ensure that the initial model has basic fault identification capabilities.

[0042] After the model is trained locally on each edge computing node, only the model parameters are uploaded, and the aggregation and update at the coordination center specifically include: Each edge node constructs a training set based on historical fault data and labels the fault labels. A timing and event-driven training trigger mechanism is used, and only when the local model performance is better than the previous round is the parameter uploaded. The edge node extracts the updated parameters, uploads only the threshold value changes compared to the previous round of global parameters, and uses homomorphic encryption after L1 regularization and sparsification. After the coordination center receives the parameters, it detects abnormal parameters through range checking, cosine similarity calculation, and contribution evaluation. An improved federal average algorithm is used to calculate the weights based on sample size, local model accuracy, and historical contribution, and the decrypted parameters are aggregated with weights. If the verification accuracy is improved, it is confirmed to be effective, otherwise a second aggregation is triggered.

[0043] It should be noted that incremental differential encoding only uploads the L1 norm difference (Δθ) with the previous round of global parameters, reducing the data volume by 72%:

[0044] In the formula, is the L1 norm difference with the previous round of global parameters, i.e., the change in incremental differential encoding; is the local parameter; is the global parameter; is the threshold value used to determine which parameter changes need to be uploaded; homomorphic encryption is enhanced using the CKKS scheme to encrypt the parameters, and the coordination center can directly aggregate in the ciphertext state (encryption and decryption delay <15ms).

[0045] The improved federal average algorithm has an aggregation weight calculation formula:

[0046] In the formula, is the aggregation weight of the i-th node; is the data volume of the i-th node; is the total data volume of all nodes; is the accuracy of the i-th node; is the historical contribution of the i-th node to the k-th task; T is the total number of tasks; A dynamic weight mechanism automatically increases the weight of regional nodes during a high-failure period (e.g., coastal node weight x 1.5 during a typhoon); secondary aggregation fault tolerance, when the first round of aggregation does not reach the accuracy improvement target, recall the Top 30% high-quality nodes for re-aggregation.

[0047] Through the combination of local decision-making by edge computing nodes and global optimization by the coordination center, a fault handling strategy is generated and executed, including: After receiving the fault information, the coordination center integrates global data, including the real-time load distribution of non-fault areas; the current output and predicted value of distributed power sources; the remaining capacity and charge / discharge rate limit of energy storage systems; the available state of tie switches and backup power sources; Recovery strategy generation, using a deep deterministic policy gradient reinforcement learning algorithm, with multiple objective functions, to generate load transfer paths, distributed power output adjustment instructions, energy storage system charge / discharge strategies, and switch operation sequences.

[0048] It should be noted that multi-objective reinforcement learning: state space modeling, 48-dimensional vector fusion of power grid physical characteristics: topology binary encoding (32-bit switch state), load thermal distribution (8-dimensional regional density), new energy output deviation rate (current / predicted value ratio), and energy storage state-of-charge-temperature coupling coefficient; Action space definition: Discrete action: 256 load transfer paths (based on graph theory to generate minimum cut sets) Continuous action: distributed power output adjustment coefficient (-100% ~ +100%), energy storage charge / discharge rate (0 ~ 2C interval); Mixed action: switch operation sequence (including operation timing dependency constraints); Multi-objective reward function:

[0049] In the formula, R is the total reward value; is the reduced power loss; is the number of operations, i.e., the number of switch operations; is the energy storage gain, i.e., the health condition gain of the battery; is the voltage change; Risk penalty term is introduced: when the voltage exceeds the limit, an additional -0.3 times the penalty coefficient is added.

[0050] Referring to Figure 2 An artificial intelligence-based power distribution network collaborative management system, comprising: A distributed edge computing module deployed at key nodes of the power distribution network for collecting and preprocessing local power data; An intelligent communication scheduling module in communication connection with the distributed edge computing module for dynamically allocating communication resources based on the operation state of the power distribution network, the module comprising a communication load prediction unit, a resource dynamic allocation unit, and a priority management unit, wherein the priority management unit sets fault processing related data as the highest transmission priority; An adaptive protocol conversion module connected with the distributed edge computing module and the intelligent communication scheduling module, the module comprising a protocol identification unit, a format conversion unit, and a data verification unit, wherein the protocol identification unit adopts a hybrid identification architecture combining deep learning and rule engine; A federated learning fault diagnosis module comprising an edge node local training unit for training a fault diagnosis model based on local power data and uploading only model parameters, and a collaborative center aggregation unit for weighting and aggregating the uploaded parameters to update the global model; A millisecond-level fault response module for generating and executing fault processing strategies through the combination of local decision-making by edge computing nodes and global optimization by the collaborative center when detecting a fault in the power distribution network, the module comprising a fault rapid detection unit, a local rapid isolation unit, a global optimization decision-making unit, and a strategy execution unit; A collaborative decision-making center for receiving information uploaded by each module and performing global coordination, the collaborative decision-making center being in communication with the distributed edge computing module through a TSN and 5G URLLC fusion network.

[0051] It should be noted that the distributed edge computing module constitutes a system distal perception network, is deployed at key nodes such as transformer substation areas / microgrid grid points, and realizes μs-level electrical quantity collection (voltage sag detection <100μs) with built-in lightweight preprocessing algorithms; The intelligent communication scheduling module builds an information highway, with TSN+5G URLLC dual-channel redundancy (main channel switching delay <2ms), and dynamic priority management realizes "zero congestion of fault data"; The adaptive protocol conversion module breaks through the "Tower of Babel dilemma" of the energy Internet of Things, supports intelligent interoperation of 137 kinds of protocols such as IEC61850 / DNP3 / Modbus, and reduces conversion delay from ms level to μs level (FPGA acceleration technology breakthrough); The collaborative decision-making center leads global optimization, integrates the output of the four intelligent modules, and builds a digital twin to realize strategy pre-rehearsal (pre-rehearsal accuracy >98%).

[0052] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only the principles of the present application. Various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for collaborative management of power distribution networks based on artificial intelligence, characterized in that, include: A distributed edge computing node network is constructed, wherein the edge computing nodes are deployed at key nodes of the power distribution network to collect and preprocess local power data; Based on local power data, an artificial intelligence-based real-time communication scheduling model is established. The model dynamically allocates communication resources according to the operating status of the distribution network, giving priority to ensuring the transmission of fault handling-related data. An adaptive protocol conversion mechanism is adopted to realize data interaction between devices with different communication protocols. The mechanism includes three sub-processes: protocol identification, format conversion, and data verification. A distribution network fault diagnosis model is trained using a federated learning framework. After the model is trained locally on each edge computing node, only the model parameters are uploaded and then aggregated and updated in the coordination center. When a distribution network fault is detected, a millisecond-level response mechanism is activated. The fault handling strategy is generated and executed by combining local decision-making at the edge computing node with global optimization at the collaboration center.

2. The distribution network collaborative management method based on artificial intelligence according to claim 1, characterized in that, The construction of the distributed edge computing node network specifically includes: Based on the distribution network topology and load density, edge computing nodes are divided into three levels of deployment. Level 1 nodes are deployed at feeder sectionalizing switches and tie switches, responsible for regional data aggregation and collaborative control. Level 2 nodes are deployed at distributed power grid connection points and energy storage system interfaces, dedicated to high-frequency acquisition of new energy output data. Level 3 nodes are deployed at user-side smart meters, charging piles, and adjustable load control terminals, using embedded micro nodes to achieve real-time perception of user-side data. Each edge computing node adopts a heterogeneous computing architecture, including a main processing unit, an industrial-grade CPU, responsible for task scheduling and routine data processing; and an acceleration unit, which integrates a GPU and an FPGA, where the GPU is used for local AI model inference and the FPGA is dedicated to high-precision time-frequency analysis. The data acquisition interface is equipped with a multi-protocol acquisition module, supporting the acquisition of analog, digital, and high-frequency pulse signals, with dynamically adjustable sampling frequency; the communication module has a built-in TSN chip and a 5G URLLC module, supporting wired and wireless dual-link redundant communication.

3. The distribution network collaborative management method based on artificial intelligence according to claim 2, characterized in that, The edge computing node is used to collect and preprocess local power data, specifically including: Local data acquisition is divided into timed acquisition and triggered acquisition. Timed acquisition is the collection of basic data, including voltage, current and power, according to a preset cycle during normal operation. Triggered acquisition is the automatic triggering of high-frequency acquisition to record transient waveform characteristics before and after the fault when a sudden voltage change or current exceeding the limit is detected. All acquired data is appended with a timestamp and node ID to ensure spatiotemporal consistency. Data preprocessing includes: a cleaning layer that uses an isolated forest to detect outliers; a feature extraction layer that extracts transient and steady-state features through wavelet transform and Fourier transform to generate compressed feature vectors; a standardization layer that transforms the data to a predetermined interval and dynamically updates the mean and variance through a sliding window; and sparse coding compression of the preprocessed data to reduce the amount of data transmitted while retaining key features. Each edge node builds a self-organizing network through a distributed hash table. New nodes automatically join and acquire topology relationships. When a node's CPU utilization exceeds a predetermined ratio, non-real-time tasks are automatically migrated to adjacent nodes with lower load. When a node's communication is interrupted, adjacent nodes take over its data acquisition tasks and temporarily store the data in a local cache. The data is then retransmitted after communication is restored.

4. The distribution network collaborative management method based on artificial intelligence according to claim 3, characterized in that, The establishment of a real-time communication scheduling model based on artificial intelligence specifically includes: Using preprocessed local power data from edge nodes as input, a joint sensing mechanism for the status of the power distribution network and communication network is constructed. A power grid operation status vector is generated by fusing power grid parameter features through a temporal convolutional network; channel occupancy, latency, and packet loss rate of TSN / 5GURLLC networks are collected in real time, and network health score is calculated by combining a sliding window; an attention mechanism is used to establish a correlation model between the two, and hard constraints of the communication network are automatically associated when a power grid fault occurs. A three-level priority system is constructed based on grid status and data types: Level 1 is hard real-time data of fault characteristic data and protection control commands, which is bound to a dedicated channel; Level 2 is soft real-time data of distributed power source regulation commands and energy storage control signals; Level 3 is non-real-time data of conventional measurement data and equipment status data; When a grid fault occurs, the relevant Level 2 data of the fault area is temporarily promoted to Level 1 through the rule engine, and will fall back to Level 1 within 100ms after recovery.

5. The distribution network collaborative management method based on artificial intelligence according to claim 4, characterized in that, The establishment of the artificial intelligence-based real-time communication scheduling model also includes: Based on a resource allocation sub-model using deep reinforcement learning, an intelligent decision-making framework for dynamic allocation of communication resources is constructed. The state space includes the power grid operation state vector, communication network health score, data backlog of each priority queue, and available channel bandwidth. The action space defines three types of actions: channel allocation, modulation mode adjustment, and transmission cycle compression. A multi-level reward mechanism is set up for the reward function. The agent is trained using a dual-delay deep deterministic policy gradient algorithm, and lightweight inference is performed locally at the edge node to achieve periodic updates of the resource allocation strategy. The communication load prediction and pre-allocation sub-model combines historical data and real-time status to achieve short-term prediction of communication traffic and reserve key resources in advance. Input features include communication traffic time-series data, current power grid status, meteorological data, and user electricity consumption pattern labels. The prediction model uses an attention-enhanced LSTM network to predict communication traffic peaks and focuses on identifying traffic characteristics before faults. The heterogeneous network adaptive switching sub-module constructs an intelligent switching mechanism to address the protocol differences and performance complementarity between TSN and 5GURLLC.

6. The distribution network collaborative management method based on artificial intelligence according to claim 5, characterized in that, The mechanism comprises three sub-processes: protocol identification, format conversion, and data verification. Specifically, it includes: The protocol identification subprocess performs multi-layer parsing of received data packets, extracting 18-dimensional features from the physical layer, data link layer, and application layer. It adopts a two-level identification architecture of deep CNN and expert rules. The first level outputs the probability distribution of protocol type through lightweight CNN classification, and the second level verifies low-probability results through an expert rule base. When a new protocol device is connected, the edge node automatically collects samples and updates the model through federated transfer learning. The format conversion subprocess converts heterogeneous data into a unified intermediate format based on extended JSON-LD, containing core and extended fields; it constructs a protocol-intermediate format mapping table to realize field mapping, data type conversion, and time sequence alignment; it optimizes conversion time through reinforcement learning, enables FPGA acceleration channels for high-priority data, and uses compression conversion for non-real-time data. The data verification subprocess generates SHA-256 hash values ​​for the data packets before and after conversion and stores them on the local blockchain. Integrity is verified through chain hashing, and large data packets are verified using a Merkle tree. The rationality of the data is detected based on the isolated forest model, and outliers are removed from multi-source data through weighted voting. Data timestamp timeout is checked and retransmission is triggered. The continuity of time-series data intervals is verified, and automatic interpolation correction is performed when the deviation exceeds 1ms.

7. The distribution network collaborative management method based on artificial intelligence according to claim 6, characterized in that, The training of the power grid fault diagnosis model using the federated learning framework specifically includes: The framework topology is constructed using a two-tier architecture of a collaborative center and edge nodes. The collaborative center is deployed at the distribution network control master station and is responsible for global model aggregation and parameter distribution. Each edge computing node acts as a federated learning client, conducting training based on locally collected fault data. Parameters are exchanged between nodes through encrypted communication links. The basic model uses a lightweight hybrid neural network as the initial model, including a feature extraction layer, a temporal modeling layer, and a classification output layer. The collaborative center pre-trains the initial model based on a historical global fault dataset, compresses and packages the model parameters, and distributes them to each edge node to ensure that the initial model has basic fault recognition capabilities.

8. The distribution network collaborative management method based on artificial intelligence according to claim 7, characterized in that, The model is trained locally on each edge computing node, and only the model parameters are uploaded. The aggregation and updating at the coordination center specifically includes: Each edge node constructs a training set based on historical fault data and labels it with faults; a timed and event-driven training triggering mechanism is adopted, and parameters are prepared to be uploaded only when the local model performance is better than the previous round; Edge nodes extract update parameters, and only upload the changes exceeding the threshold by comparing them with the global parameters of the previous round. After L1 regularization and sparsification, homomorphic encryption is used. After receiving the parameters, the collaboration center detects abnormal parameters through range checks, cosine similarity calculations, and contribution assessments. An improved federated averaging algorithm is used to calculate weights based on sample size, local model accuracy, and historical contribution. The decryption parameters are then weighted and aggregated. Once the accuracy is verified to be improved, the algorithm is confirmed to be effective; otherwise, a second aggregation is triggered.

9. The distribution network collaborative management method based on artificial intelligence according to claim 8, characterized in that, The method of generating and executing fault handling strategies by combining local decision-making at edge computing nodes with global optimization at the collaboration center specifically includes: After receiving fault information, the coordination center integrates global data, including real-time load distribution in non-faulty areas; current and predicted output of distributed power sources; remaining capacity and charge / discharge rate limits of energy storage systems; and availability status of interconnection switches and backup power supplies. The recovery strategy generation adopts a deep deterministic policy gradient reinforcement learning algorithm with multiple objective functions to generate load transfer paths, distributed power generation output adjustment commands, energy storage system charging and discharging strategies, and switching operation sequences.

10. A power distribution network collaborative management system based on artificial intelligence, characterized in that, include: Distributed edge computing modules are deployed at key nodes of the power distribution network to collect and preprocess local power data. The intelligent communication scheduling module is communicatively connected to the distributed edge computing module and is used to dynamically allocate communication resources based on the operating status of the power distribution network. The module includes a communication load prediction unit, a resource dynamic allocation unit, and a priority management unit, wherein the priority management unit sets fault handling-related data as the highest transmission priority. An adaptive protocol conversion module is connected to the distributed edge computing module and the intelligent communication scheduling module, respectively. The module includes a protocol identification unit, a format conversion unit and a data verification unit, wherein the protocol identification unit adopts a hybrid identification architecture that integrates deep learning and rule engine. The federated learning fault diagnosis module includes a local training unit at the edge node that trains a fault diagnosis model based on local power data and only uploads the model parameters, and a collaborative center aggregation unit that weights and aggregates the uploaded parameters to update the global model. A millisecond-level fault response module is used to generate and execute a fault handling strategy when a distribution network fault is detected by combining local decision-making at the edge computing node with global optimization at the collaborative center. The module includes a rapid fault detection unit, a local rapid isolation unit, a global optimization decision-making unit, and a strategy execution unit. The collaborative decision-making center is used to receive information uploaded by each module and perform global coordination. The collaborative decision-making center and the distributed edge computing module communicate through the TSN and 5G URLLC converged network.

Images