Power distribution network intelligent scheduling method and system based on AI large model

Abstract

The invention relates to a power distribution network intelligent scheduling method and system based on an AI large model, and the method achieves the efficient intelligent scheduling through edge calculation preprocessing, historical data semantic compression, multi-task AI collaborative reasoning and multi-target optimization decision. According to context compression, three mechanisms of selection retention, abstract compression and sentence element extraction are adopted, historical operation data are compressed and uniformly coded to obtain a knowledge base, historical fragments matched with real-time data are screened respectively, semantic abstracts are generated through a large model, key information is extracted through an attention mechanism, redundancy is reduced, and the reasoning efficiency is improved. The AI large model reasoning module is a multi-task neural network architecture, calls a knowledge base, and fuses historical and real-time data to realize multi-task collaborative reasoning. Compared with the prior art, the scheduling strategy provided by the invention has high adaptability to new energy output fluctuation, and provides safe, efficient and adaptive scheduling support for the power distribution network with high-proportion new energy access.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent scheduling of power systems, in particular to a power distribution network intelligent scheduling method and system based on an AI large model. BACKGROUND

[0002] With the large-scale access of new energy and the diversification of power load patterns, the operating environment of modern power distribution networks is becoming increasingly complex, and the uncertainty and operating risk of the system are significantly increasing. The strong fluctuation of new energy output and the multi-time and space scale superposition of load changes make the traditional power distribution network scheduling system relying on static models and artificial rules gradually expose problems such as response lag, poor adaptability, and weak optimization capability.

[0003] In the existing power system operation and management framework, power dispatching is a core link for ensuring the safe and stable operation of the power grid and the optimal allocation of resources, and its dispatching decision-making capability directly affects the economy, reliability, and emergency disposal capability of the system. However, the current dispatching center generally relies on human experience or fixed rules for operation, and in the face of complex scenarios such as dynamic changes, nonlinear coupling, and fault bursts, it is difficult to achieve efficient global optimization control. Most traditional dispatching models are based on deterministic prediction and static power source configuration, and lack the ability to quickly feedback and roll back the actual operating state, resulting in the lagging and non-adaptive nature of the dispatching strategy.

[0004] In addition, although the power distribution network has deployed a large number of sensing devices and accumulated a wealth of operating data, due to the lack of efficient data processing mechanisms and intelligent analysis models, the current system generally suffers from low data utilization, poor reasoning efficiency, and lagging model updates. Control personnel still need to rely on a large amount of manual intervention and experience-based judgment when faced with massive heterogeneous information, and the repetitive "brain labor" is difficult to support the growing demand for operating complexity and control precision, and the overall automation and intelligence level of the system needs to be improved.

[0005] Therefore, in view of the low degree of automation, weak intelligent assistance, and inability to quickly respond to dynamic disturbances in the existing power distribution network scheduling method, a new type of scheduling system that integrates edge computing, context compression, and multi-task AI models is needed, which can make comprehensive decision-making and reasoning based on real-time sensing and historical data, realize an end-to-end intelligent closed loop from data acquisition, feature extraction, model prediction to scheduling execution, and provide solid technical support for the safe, efficient, and economic operation of the power distribution system. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art and provide a power distribution network intelligent scheduling method and system based on an AI large model.

[0007] The purpose of the present application can be achieved by the following technical solutions:

[0008] An AI large model-based power distribution network intelligent scheduling method, the method comprising:

[0009] Collecting power distribution network operation state data and preprocessing, based on the preprocessed power distribution network operation state data, realizing preliminary fault detection and fault location;

[0010] According to the preprocessed power distribution network operation state data and the current scheduling task, the historical operation data of the power distribution network is compressed and uniformly coded, and the knowledge base is stored and obtained, the compression processing includes selection reservation compression mechanism, abstract compression mechanism or sentence element extraction mechanism, the knowledge base includes preprocessed power distribution network operation state data, preliminary fault detection and fault location results and compressed power distribution network historical operation data;

[0011] According to the current scheduling task, the corresponding data in the knowledge base is called, based on the pre-established AI large model, a plurality of task results are obtained, the AI large model includes a plurality of task sub-networks for different scheduling tasks;

[0012] Using the output result of the AI large model, the optimal scheduling instruction of the power distribution network is generated based on the multi-objective optimization algorithm.

[0013] Further, the selection reservation compression mechanism comprises:

[0014] According to the data characteristics of the preprocessed power distribution network operation state data, the matching rules are set, the data in the power distribution network historical operation data that meet the matching rules are screened and obtained, and other irrelevant data are removed.

[0015] Further, the abstract compression mechanism comprises:

[0016] The screened historical data segment is input into the pre-trained semantic large model to generate a semantic abstract of the historical data; the screened historical data segment includes the power distribution network historical operation data processed by the selection reservation compression mechanism; the content of the semantic abstract retains the key information about the load change trend, the fault occurrence law and the power output fluctuation in the historical data.

[0017] Further, the sentence element extraction mechanism comprises:

[0018] Based on the attention score mechanism, the descriptive sentences in the power distribution network historical operation data are analyzed, the most relevant sentences to the task are extracted by calculating the correlation degree of each sentence to the current scheduling task, and the redundant description contents are removed.

[0019] Further, the pre-established AI large model adopts a multi-task neural network structure that shares an input encoding layer and sets task sub-networks at the output end to simultaneously perform load prediction, fault prediction, power flow optimization, and emergency scheduling strategy generation; the task sub-networks include a load prediction sub-network, a fault prediction sub-network, a power flow optimization sub-network, and an emergency scheduling strategy generation sub-network;

[0020] The load prediction sub-network predicts the load demand of a future period based on historical load data and real-time weather data;

[0021] The fault prediction sub-network models the distribution network topology through a graph neural network, analyzes the connection relationship and information interaction mode between nodes, and combines historical fault data of the nodes, preliminary fault detection and fault location results, and current operating state features to predict potential fault nodes;

[0022] The power flow optimization sub-network dynamically adjusts the power flow distribution according to the topology structure, load distribution, and power output of the current power grid, combines a reinforcement learning algorithm, aims to reduce network loss and improve power supply balance, and outputs the adjusted branch power flow distribution and device operating state;

[0023] The emergency scheduling strategy generation sub-network generates an optimal emergency scheduling strategy that meets safety constraints based on a multi-objective optimization algorithm, comprehensively considers safety, economy, and reliability safety constraint conditions, and according to the fault prediction results, the current power grid operating state, and the standby power source situation.

[0024] Further, the process of generating an optimal scheduling instruction for the distribution network based on a multi-objective optimization algorithm includes:

[0025] After obtaining the multiple task results output by the AI large model, a multi-objective optimization problem is constructed according to the operating safety constraints of the distribution network, the economic cost target, the power supply reliability requirement, the load balance target, and the new energy utilization rate;

[0026] According to the multi-objective optimization problem, a number of scheduling schemes that meet different combinations of targets are generated to form a Pareto frontier solution set;

[0027] According to the actual operating state and priority requirements of the current distribution network, the optimal scheduling strategy is selected from the Pareto frontier solution set to generate a scheduling instruction;

[0028] The generated scheduling instruction is refined, the operating constraints of the actual devices in the distribution network are considered, constraint verification and execution logic checking are performed, and the optimal scheduling instruction is output.

[0029] Further, the power distribution network operation state data is collected by installing sensors and IoT devices at key nodes of the power distribution network, including but not limited to voltage transformers, current transformers, smart meters, temperature sensors, humidity sensors, and light sensors; the power distribution network operation state data includes voltage values, current values, load power, power output, device operation states, and meteorological environment information.

[0030] Further, the preprocessing includes:

[0031] Data cleaning operation: based on the preset threshold range, the obviously abnormal values in the power distribution network operation state data are removed; interpolation algorithm is used to fill in the missing data in the power distribution network operation state data; filtering algorithm is used to remove noise data in the power distribution network operation state data;

[0032] Data feature extraction: based on the sliding time window mechanism, the power distribution network operation state data after the data cleaning operation is subjected to feature extraction, and the power distribution network operation state data is analyzed within the preset time window; in the time domain, the statistical features of various data in the unit time window are calculated, including mean, variance, maximum and minimum; in the frequency domain, the frequency components of voltage and current signals are analyzed by fast Fourier transform to extract the amplitude and phase information of the main frequency components; in the time-frequency domain, wavelet transform is used to obtain the features of the data in different time scales and frequency scales.

[0033] Further, the process of fault detection and fault location includes:

[0034] Using the data features extracted in the preprocessing, combined with the preset threshold, abnormal detection is performed one by one, if the data features are abnormal, it is preliminarily determined that there is a fault in the corresponding node;

[0035] Using the preset classification model to judge the fault type of the node preliminarily determined as fault, and according to the propagation law of fault features in the topology structure of the power distribution network, combined with the data information of adjacent nodes, preliminary fault location is performed to determine the possible area of fault occurrence, and alarm information is generated.

[0036] An AI large model-based power distribution network intelligent scheduling system, the system includes:

[0037] A data acquisition module for acquiring power distribution network operation state data, including voltage, current, load, power output, device state, and meteorological environment;

[0038] An edge computing module for local preprocessing, data cleaning, and feature extraction of the data collected by the data acquisition module, and realizing preliminary fault detection and fault location;

[0039] A context compression module is configured to compress historical operation data of the power distribution network;

[0040] An AI large model inference module is deployed on a cloud server or a local server, configured to train and optimize a multi-task neural network based on historical operation data and real-time collected data of the power distribution network, and to implement load prediction, fault prediction, power flow optimization and emergency dispatch strategy generation by using the multi-task neural network;

[0041] A dispatch decision module is configured to receive output results of the AI large model inference module and generate optimal dispatch instructions based on a set optimization target;

[0042] A human-computer interaction interface is configured to allow an operation and maintenance personnel to view AI large model recommended schemes, execute dispatch operations and backtrack abnormal events.

[0043] Compared with the prior art, the present application has the following advantages:

[0044] 1. The present application realizes a closed-loop intelligent dispatch process of intelligent sensing, autonomous judgment, efficient decision-making and visual collaboration by organically integrating edge intelligent processing, context semantic compression, multi-task learning and multi-target dispatch optimization, and builds an end-to-end intelligent dispatch system for the power distribution network, thereby solving the problems of traditional dispatch relying on manual experience, delayed response and weak adaptability to dynamic scenarios, and overall realizing comprehensive sensing of the operation state of the power distribution network, rapid prediction and positioning of faults and intelligent optimization of dispatch strategies, thereby significantly improving the operation efficiency, power supply reliability, new energy consumption capacity and adaptability to complex scenarios of the power distribution network, and enhancing the scientificity and feasibility of dispatch decisions.

[0045] 2. In the present application, the selection and retention compression mechanism selects and retains historical data related to the current state by matching rules, ensures data relevance, discards other irrelevant data, and also reduces the amount of data for subsequent processing; the abstract compression mechanism retains key information in the historical data and presents it in a concise form, further reducing the data dimension; the sentence element extraction mechanism extracts the most relevant sentences to the task by calculating the relevance of each sentence to the current dispatch task, and eliminates redundant description content; the combination of the selection and retention compression mechanism, the abstract compression mechanism and the sentence element extraction mechanism realizes the elimination of redundant data and the reduction of data dimension, reduces the consumption of storage and computing resources, retains the core information most valuable to dispatch decisions, and directly improves the inference efficiency of the AI large model; tests show that the present application can improve the inference efficiency of the AI model by more than 40% without increasing the computing resources.

[0046] 3. The application will store the preprocessed real-time operation data, preliminary fault detection results and compressed historical data in a centralized manner, forming a structured knowledge base, which solves the problem of traditional data dispersion and heterogeneous format, and facilitates unified calling and management; through unified coding processing, the multi-source data format is standardized, reducing the AI model data preprocessing overhead; and the related data in the knowledge base can be accurately matched and called according to the current scheduling task, avoiding invalid data loading and shortening the model inference response time.

[0047] 4. The AI large model of the application is a multi-task neural network structure composed of multiple task sub-networks, which can complete multiple key tasks such as load prediction, fault prediction, power flow optimization and emergency scheduling strategy generation in parallel, improve the generalization ability of the model and the overall processing efficiency of the system, meet the actual needs of multi-objective and high complexity in the distribution network scheduling scene, and the test shows that the fault prediction accuracy reaches 95%, the load prediction error is less than 3%, and the scheduling strategy has high adaptability to new energy output fluctuation.

[0048] 5. The application introduces a multi-objective optimization algorithm to construct a scheduling decision module, which considers multiple optimization objectives such as power balance, line safety and economy when generating scheduling instructions, effectively improving the scientificity and rationality of scheduling decisions, providing more valuable decision support for distribution network operation, and solving the limitations of traditional single-objective optimization, improving the global optimality and safety of the decision.

[0049] 6. In the intelligent scheduling system of the application, through the introduction of the edge computing module and the context compression module cooperative working mechanism, data preprocessing and feature extraction are completed on the terminal side, while through semantic compression and time window selection technologies, redundant information is reduced, effectively reducing the data transmission overhead, improving the processing capacity and real-time response performance of the system to large-scale data flow.

[0050] 7. The application is based on time domain, frequency domain and time-frequency domain feature extraction of sliding time window, which effectively captures the dynamic characteristics and key patterns of data, provides high-quality features for fault detection and AI model input, and through threshold discrimination and classification model, the application also realizes preliminary identification and regional positioning of faults, quickly generates alarm information, significantly shortens the fault response delay, and reduces the computing pressure of the central system, which saves time for subsequent accurate analysis and emergency scheduling. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of an AI large model-based distribution network intelligent scheduling method according to the application is shown in the figure;

[0052] Figure 2 A structure block diagram of the context compression module in the application is shown in the figure;

[0053] Figure 3 Flow chart for generating scheduling decision in the present application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the protection scope of the present application.

[0055] Embodiment 1

[0056] The embodiment discloses an intelligent scheduling method for power distribution network based on AI large model, and the method specifically comprises steps S1-S5 as shown in the figure, and specific descriptions of steps S1-S5 are as follows: Figure 1

[0057] Step S1, collecting power distribution network operation state data.

[0058] The power distribution network operation state data is collected by installing sensors and IoT devices at key nodes of the power distribution network, and the sensors and IoT devices include but are not limited to voltage transformers, current transformers, smart meters, temperature sensors, humidity sensors and illumination sensors; the power distribution network operation state data includes voltage values, current values, load powers, power output, device operation states and meteorological environment information.

[0059] Step S2, preprocessing the collected power distribution network operation state data, and realizing preliminary fault detection and fault positioning based on the preprocessed power distribution network operation state data.

[0060] The specific steps of data preprocessing include:

[0061] Data cleaning operation: by setting a reasonable threshold range, the obviously abnormal values in the power distribution network operation state data are eliminated; interpolation algorithm is used to fill in the missing data in the power distribution network operation state data; filtering algorithm is used to remove noise data in the power distribution network operation state data;

[0062] Data feature extraction: based on the sliding time window mechanism, the power distribution network operation state data after the data cleaning operation is subjected to feature extraction, and the power distribution network operation state data is analyzed within the preset time window; in the time domain, the statistical characteristics of various data in the unit time window are calculated, and the statistical characteristics include mean, variance, maximum and minimum; in the frequency domain, the frequency components of voltage and current signals are analyzed by fast Fourier transform, and the amplitude and phase information of main frequency components are extracted; in the time-frequency domain, wavelet transform is used to obtain the characteristics of data in different time scales and frequency scales. ​

[0063] The process of fault detection and fault location includes:

[0064] Using each data feature extracted in the preprocessing, combined with the preset threshold, abnormal detection is carried out one by one, if the data feature is abnormal, it is preliminarily judged that the corresponding node has a fault;

[0065] Using the preset classification model to judge the fault type of the node preliminarily judged as fault, and according to the propagation rule of fault characteristics in the topology structure of the power distribution network, combined with the data information of the adjacent nodes, the preliminary fault location is carried out, the possible fault area is determined, and the alarm information is generated.

[0066] Step S3, according to the preprocessed power distribution network operation state data and the current scheduling task, the power distribution network historical operation data is compressed and uniformly coded, and the knowledge base is stored and acquired.

[0067] The compression processing includes selection reservation compression mechanism, abstract compression mechanism or sentence element extraction mechanism. In actual application, according to the task requirement and the type of actual collected data, one or more compression processing methods are selected for compression processing.

[0068] Selection reservation compression mechanism: according to the data characteristics of the preprocessed power distribution network operation state data, the matching rules are set, the data in the power distribution network historical operation data that meets the matching rules is screened and obtained, and other irrelevant data is removed. For example, if the current load is in the peak period and mainly industrial load, then the historical time segment with high industrial load proportion in the peak period is screened out from the historical data, and other irrelevant time segments are discarded, so as to reduce the data amount of subsequent processing.

[0069] Abstract compression mechanism: the screened historical data segment is input into the pre-trained semantic large model to generate the semantic abstract of the historical data; the screened historical data segment includes the power distribution network historical operation data processed by the selection reservation compression mechanism; the content of the semantic abstract retains the key information about the load change trend, fault occurrence rule and power output fluctuation in the historical data, and presents in a simple form, further reducing the data dimension.

[0070] Sentence element extraction mechanism: based on the attention score mechanism, the descriptive sentences in the power distribution network historical operation data are analyzed, the most relevant sentences to the task are extracted by calculating the correlation degree of each sentence to the current scheduling task, and the redundant description contents are removed. For example, when performing fault prediction, the sentences related to the device fault reason and the pre-fault signs in the historical data are focused on, and the sentences irrelevant to the normal operation and maintenance of the device are ignored.

[0071] The unified coding process is for the pre-processed power distribution network operation state data, preliminary fault detection and fault location results and compressed power distribution network historical operation data. For numerical data, after standardization processing, the data is input into the model. For time series data, it is converted into a sequence format suitable for Transformer processing. For power distribution network topology structure data, node embedding technology of graph neural network is used to embed it into a low-dimensional space as one of the inputs of the subsequent AI large model. After unified coding processing, all data is stored in the knowledge base which can be used for subsequent scheduling tasks. The data in the knowledge base is stored according to the time sequence and relationship of each data, and indexed according to data type, timestamp and scheduling task type, which facilitates subsequent scheduling of knowledge base data according to current scheduling task requirements.

[0072] The power distribution network historical operation data is real-time updated data. The power distribution network operation state data collected in this scheduling task will also be stored in the power distribution network historical operation data after use. In the next scheduling task, it is compressed and processed as power distribution network historical operation data to obtain a new knowledge base.

[0073] Step S4, according to the current scheduling task, calling the corresponding data in the knowledge base, based on the pre-established AI large model, obtaining multiple task results.

[0074] The pre-established AI large model adopts a multi-task neural network structure, which shares an input coding layer and sets task sub-networks at the output end to simultaneously perform load prediction, fault prediction, power flow optimization and emergency scheduling strategy generation. The task sub-networks include load prediction sub-network, fault prediction sub-network, power flow optimization sub-network and emergency scheduling strategy generation sub-network.

[0075] The specific steps of model architecture building are as follows:

[0076] A multi-task neural network based on Transformer structure is adopted as the core architecture. The Transformer structure can effectively process sequence data and is suitable for processing time series data in the power distribution network. At the same time, a graph neural network module is introduced to model the topology of the power distribution network and capture the connection relationship and information propagation path between nodes.

[0077] The input coding layer of the network calls the data required by the current scheduling task in the knowledge base and transmits it to the output end.

[0078] Multiple task sub-networks are set at the output end, corresponding to different scheduling tasks respectively:

[0079] The load prediction sub-network predicts the load demand in the future period based on historical load data and real-time weather data; for example, the input includes historical load data and real-time weather data, the complex relationship between load changes and weather factors is learned through a multi-layer Transformer structure, and the load prediction curve for the next 1 hour is output.

[0080] The fault prediction sub-network models the distribution network topology through a graph neural network, analyzes the connection relationship and information interaction mode between nodes, and combines the historical fault data of the nodes, the preliminary fault detection and fault location results, and the current operating state features to predict potential fault nodes; for example, by calculating the abnormal feature propagation path and probability of the nodes, it determines which nodes are more likely to fail in the future period.

[0081] The power flow optimization sub-network dynamically adjusts the power flow distribution based on the current topology of the power grid, load distribution, and power output, and combines a reinforcement learning algorithm to reduce network loss and improve power supply balance. Through interaction with the environment, the optimal power flow adjustment strategy is learned, and the adjusted branch power flow distribution and device operating state are output.

[0082] The emergency dispatch strategy generation sub-network generates an optimal emergency dispatch strategy that meets the safety constraints based on a multi-objective optimization algorithm, considering safety, economy, and reliability safety constraints, according to the fault prediction results, current power grid operating state, and standby power source conditions.

[0083] The specific steps of model training and inference are as follows:

[0084] The PyTorch framework is used for model training and inference. A large amount of historical operation data of the distribution network is collected as the training data set, and the data set is divided into training set, validation set, and test set.

[0085] An end-to-end training method is used, and the model is trained using the backpropagation algorithm by defining a suitable loss function. During training, the model's hyperparameters are adjusted based on the performance of the validation set to ensure good generalization ability. After the model is trained, it is deployed to the cloud or a local high-performance server. In actual operation, according to the current dispatch task, real-time data and compressed historical data in the knowledge base are called to perform fast inference, generate load prediction results, fault prediction results, power flow optimization schemes, and emergency dispatch strategies, and these results are used for subsequent dispatch decisions.

[0086] Step S5, using the output results of the AI large model, generating optimal dispatch instructions for the distribution network based on a multi-objective optimization algorithm.

[0087] The process of generating optimal dispatch instructions for the distribution network based on a multi-objective optimization algorithm includes:

[0088] After obtaining the multi-task results output by the AI large model, a multi-objective optimization problem is constructed according to the operation safety constraints, economic cost targets, power supply reliability targets, load balancing targets and new energy utilization rates of the power distribution network;

[0089] According to the multi-objective optimization problem, a plurality of scheduling schemes satisfying different target combinations are generated to form a Pareto frontier solution set;

[0090] According to the actual operation state and priority requirements of the current power distribution network, an optimal scheduling strategy is selected from the Pareto frontier solution set to generate a scheduling instruction;

[0091] The generated scheduling instruction is refined, considering the operation constraints of the actual devices in the power distribution network, and constraint verification and execution logic checking are performed to output an optimal scheduling instruction;

[0092] The optimal scheduling instruction is sent to the control system of the power distribution network to guide the actual operation of the power distribution network.

[0093] Embodiment 2

[0094] This embodiment is based on the above-mentioned embodiment 1, and discloses a power distribution network intelligent scheduling system based on an AI large model, which comprises:

[0095] A data acquisition module is configured to acquire power distribution network operation state data, including voltage, current, load, power output, device state and meteorological environment;

[0096] An edge computing module is configured to perform local preprocessing, data cleaning and feature extraction on the data collected by the data acquisition module, and to implement preliminary fault detection and fault location;

[0097] A context compression module is configured to compress the historical operation data of the power distribution network;

[0098] An AI large model inference module is deployed on a cloud server or a local server, and is configured to train and optimize a multi-task neural network based on the historical operation data and real-time collected data of the power distribution network, and to use the multi-task neural network to implement load prediction, fault prediction, power flow optimization and emergency scheduling strategy generation;

[0099] A scheduling decision module is configured to receive the output results of the AI large model inference module, and to generate an optimal scheduling instruction based on the set optimization target;

[0100] A human-computer interaction interface is configured to allow operation and maintenance personnel to view AI large model recommended schemes, execute scheduling operations and backtrack abnormal events.

[0101] The system first collects multi-dimensional data such as voltage, current, load, power output, equipment status, and meteorological environment in real time through distributed sensors and IoT devices deployed at key nodes of the power distribution network, ensuring comprehensive coverage and real-time updating of operation information. The collected data is preprocessed locally by the edge computing module, including operations such as outlier removal, noise cleaning, and missing value filling. Key features are extracted through time-domain, frequency-domain, or time-frequency domain analysis methods combined with a sliding time window mechanism. Meanwhile, the module introduces a local fault detection mechanism based on threshold discrimination or lightweight classification models, enabling preliminary anomaly identification and alarm, significantly reducing response time and alleviating central computing pressure.

[0102] To further improve the reasoning efficiency and processing capacity of AI models when facing large-scale historical data, the system is equipped with a context compression module that performs semantic compression and redundancy removal on historical operation data and stores real-time data, historical data, and preliminary fault localization results centrally to obtain a knowledge base. The module integrates three compression mechanisms: a selection retention compression mechanism that filters out historical time segments matching the current operating state features; an abstract compression mechanism that generates semantic summaries of historical data based on large models; and a sentence element extraction mechanism that extracts key sentences and descriptions through an attention mechanism and automatically discards irrelevant content, effectively reducing data dimensionality and storage requirements while retaining context information most valuable for dispatch tasks.

[0103] The context compression module is specifically as shown in Figure 2 The actual operation process and role of each unit can be understood by referring to the related descriptions and effects in Embodiment 1. In actual dispatch tasks, appropriate units are selected for data compression processing based on actual needs and task requirements.

[0104] The AI large model reasoning module is the core computing unit of the system, deployed on a cloud or local high-performance server, and uses a multi-task neural network architecture based on Transformer structure and graph neural networks. The model processes heterogeneous input data (such as numerical features and time series data) from the knowledge base through a shared input encoding layer, and sets multiple task sub-networks at the output end to handle tasks such as load prediction, fault prediction, power flow optimization, and emergency dispatch strategy generation. The load prediction sub-network combines historical load data and weather information to accurately predict future load curves; the fault prediction sub-network builds a graph model based on power grid topology to identify potential high-risk nodes; the power flow optimization sub-network uses reinforcement learning algorithms to intelligently adjust power flow distribution, reducing network loss and improving power supply balance; and the emergency dispatch sub-network generates feasible dispatch strategies that meet various operating constraints based on multi-objective optimization algorithms. Model training and reasoning are based on the PyTorch framework, with end-to-end training capabilities and the ability to implement online fine-tuning and adaptive learning.

[0105] The scheduling decision module is based on AI model output, as shown in Figure 3 The scheduling optimal solution is generated by a multi-objective optimization algorithm in combination with operation safety, economic cost, power supply reliability, load balancing and new energy consumption of the power distribution system, and the optimal scheduling strategy under the current system constraint is selected in the Pareto frontier solution set. The generated scheduling instruction can be refined, constraint verified and execution logic checked according to the operation requirements, to ensure the landing and safety of the strategy.

[0106] The man-machine interface provides a visual operation platform, supports the operation and maintenance personnel to view the power grid state diagram, scheduling suggestion, model prediction result and historical operation trend. The system also has functions of scheduling log recording, manual correction of scheduling instruction, abnormal event backtracking analysis, etc., and shows the key events and system response process through the time axis mode, to enhance the traceability of operation and maintenance operation and the explainability of AI recommendation. The front-end interface can integrate multiple graphic engines to show the power grid topology and load distribution, and all operations and abnormal information are recorded in the database, to facilitate subsequent data mining and behavior optimization.

[0107] In summary, the application realizes the closed-loop intelligent scheduling process of intelligent perception, autonomous judgment, efficient decision and visual collaboration by organically integrating edge intelligent processing, context semantic compression, multi-task learning and multi-objective scheduling optimization. Actual test shows that the system can improve the AI model inference efficiency by more than 40% without increasing computing resources, the fault prediction accuracy rate reaches 95%, the load prediction error is less than 3%, and the scheduling strategy has high adaptability to new energy output fluctuation. The overall architecture of the system has good portability and expandability in the new power distribution system scenario, and provides a solid technical support for building a safe, efficient, green and intelligent new power system.

[0108] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the application, which should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. An AI large model-based power distribution network intelligent scheduling method, characterized in that, The method comprises: Collecting power distribution network operation state data and preprocessing, based on the preprocessed power distribution network operation state data to realize preliminary fault detection and fault location; According to the preprocessed power distribution network operation state data and the current scheduling task, the historical operation data of the power distribution network is compressed and uniformly coded, and the knowledge base is stored and obtained, the compression processing includes selection reservation compression mechanism, abstract compression mechanism or sentence element extraction mechanism, the knowledge base includes preprocessed power distribution network operation state data, preliminary fault detection and fault location results and compressed power distribution network historical operation data; According to the current scheduling task, the corresponding data in the knowledge base is called, based on the pre-established AI large model, a variety of task results are obtained, the AI large model includes a plurality of task subnetworks for different scheduling tasks; Using the output result of AI large model, based on multi-objective optimization algorithm to generate the optimal scheduling instruction of power distribution network.

2. The power distribution network intelligent scheduling method based on an AI large model according to claim 1, characterized in that, The selection reservation compression mechanism comprises: According to the data characteristics of the preprocessed power distribution network operation state data, the matching rules are set, the data in the power distribution network historical operation data that meets the matching rules is screened and obtained, and other irrelevant data is removed.

3. The power distribution network intelligent scheduling method based on an AI large model according to claim 2, characterized in that, The abstract compression mechanism comprises: The screened historical data segment is input into the pre-trained semantic large model to generate a semantic abstract of the historical data; The screened historical data segment includes the power distribution network historical operation data processed by the selection reservation compression mechanism; The content of the semantic abstract retains the key information about the load change trend, fault occurrence law and power output fluctuation in the historical data.

4. The power distribution network intelligent scheduling method based on an AI large model according to claim 1, characterized in that, The sentence element extraction mechanism comprises: Based on the attention score mechanism, the descriptive sentences in the power distribution network historical operation data are analyzed, the most relevant sentences to the task are extracted by calculating the correlation degree of each sentence to the current scheduling task, and the redundant description content is removed.

5. The power distribution network intelligent scheduling method based on an AI large model according to claim 1, characterized in that, The pre-established AI large model adopts a multi-task neural network structure, which shares an input encoding layer and sets task subnetworks at the output end to simultaneously perform load prediction, fault prediction, power flow optimization and emergency scheduling strategy generation; The task subnetworks include load prediction subnetwork, fault prediction subnetwork, power flow optimization subnetwork and emergency scheduling strategy generation subnetwork; The load prediction subnetwork predicts the load demand in the future period based on historical load data and real-time weather data; The fault prediction subnetwork models the power distribution network topology through the graph neural network, analyzes the connection relationship and information interaction mode between nodes, combines the historical fault data of nodes, the preliminary fault detection and fault location results and the current operation state characteristics, and predicts the potential fault nodes; The power flow optimization subnetwork dynamically adjusts the power flow distribution according to the topology structure, load distribution and power output of the current power grid, combines the reinforcement learning algorithm, takes reducing network loss and improving power supply balance as the goal, and outputs the adjusted branch power flow distribution and device operation state; The emergency dispatching strategy generation sub-network is based on a multi-objective optimization algorithm, comprehensively considers safety, economy and reliability safety constraint conditions, generates an optimal emergency dispatching strategy meeting safety constraints according to fault prediction results, current power grid operation states and standby power supply conditions.

6. The power distribution network intelligent scheduling method based on an AI large model according to claim 1, characterized in that, The process of generating optimal dispatching instructions of the distribution network based on the multi-objective optimization algorithm includes: After obtaining the multiple task results output by the AI large model, the multi-objective optimization problem is constructed according to the operation safety constraints, economic cost targets, power supply reliability targets, load balancing targets and new energy utilization rates of the distribution network; According to the multi-objective optimization problem, a plurality of dispatching schemes meeting different target combinations are generated to form a Pareto frontier solution set; According to the actual operation state and priority requirements of the current distribution network, the optimal dispatching strategy is selected from the Pareto frontier solution set to generate dispatching instructions. The generated dispatching instructions are refined, the operation constraints of the actual devices in the distribution network are considered, constraint verification and execution logic checking are performed, and the optimal dispatching instructions are output.

7. The power distribution network intelligent scheduling method based on an AI large model according to claim 1, characterized in that, The distribution network operation state data is collected by installing sensors and IoT devices at key nodes of the distribution network, including but not limited to voltage transformers, current transformers, smart meters, temperature sensors, humidity sensors and light sensors; the distribution network operation state data includes voltage values, current values, load power, power output, device operation states and meteorological environment information.

8. The power distribution network intelligent scheduling method based on an AI large model according to claim 1, characterized in that, The preprocessing includes: Data cleaning operation: based on the preset threshold range, the obviously abnormal values in the distribution network operation state data are removed; interpolation algorithms are used to fill in the missing data in the distribution network operation state data; filtering algorithms are used to remove noise data in the distribution network operation state data; Data feature extraction: based on the sliding time window mechanism, the distribution network operation state data after the data cleaning operation is subjected to feature extraction, and the distribution network operation state data is analyzed within the preset time window; in the time domain, the statistical characteristics of various data in the unit time window are calculated, including mean, variance, maximum and minimum; in the frequency domain, the frequency components of the voltage and current signals are analyzed by fast Fourier transform to extract the amplitude and phase information of the main frequency components; in the time-frequency domain, wavelet transform is used to obtain the features of the data in different time scales and frequency scales.

9. The power distribution network intelligent scheduling method based on an AI large model according to claim 8, characterized in that, The process of fault detection and fault location includes: Using each data feature extracted in the preprocessing, combined with the preset threshold, abnormal detection is performed one by one, if the data feature is abnormal, it is preliminarily determined that the corresponding node has a fault; Using a preset classification model to determine the fault type of the node preliminarily determined as faulty, and according to the propagation law of fault features in the topology structure of the distribution network, combined with the data information of adjacent nodes, preliminary fault location is performed to determine the possible area of the fault, and alarm information is generated.

10. An AI large model-based power distribution network intelligent scheduling system, characterized in that, The system includes: A data acquisition module for acquiring distribution network operation state data, including voltage, current, load, power output, device state and meteorological environment; An edge computing module is configured to perform local preprocessing, data cleaning and feature extraction on the data collected by the data collection module, and to implement preliminary fault detection and fault location. A context compression module is configured to compress the historical operation data of the power distribution network. An AI large model inference module is deployed on a cloud server or a local server, and is configured to train and optimize a multi-task neural network based on the historical operation data and the real-time collected data of the power distribution network, and to implement load prediction, fault prediction, power flow optimization and emergency dispatch strategy generation by using the multi-task neural network. A dispatch decision module is configured to receive the output result of the AI large model inference module, and to generate optimal dispatch instructions based on a set optimization target. A human-computer interaction interface is configured to allow an operation and maintenance personnel to view the AI large model recommended scheme, perform dispatch operations and backtrack abnormal events.

Images