Historical case retrieval and recommendation method for power distribution network dispatching auxiliary decision-making

Abstract

The invention belongs to the technical field of intelligent dispatching of power systems, and particularly relates to a power distribution network historical case retrieval and recommendation method based on multi-modal data fusion and depth similarity matching. The method comprises the following steps: constructing a multi-modal case feature library, and storing historical case data including a topological structure, a load curve and a scheduling text; receiving current power distribution network fault or scheduling demand information; calculating the topological similarity between the current scene and the historical case by adopting a graph neural network; calculating load distribution similarity by applying a dynamic time warping algorithm; calculating the semantic similarity of the scheduling text based on a fine-tuned BERT model of the electric power major; dynamically adjusting each feature weight according to the real-time new energy permeability, and generating a comprehensive similarity score; candidate cases conforming to power grid safe operation conditions are screened through a physical constraint verification layer; and outputting the recommendation case and the applicability analysis report thereof.

Description

Technical Field

[0001] This invention belongs to the field of intelligent dispatching technology for power systems, specifically relating to a method for retrieving and recommending historical cases of distribution networks based on multimodal data fusion and deep similarity matching. Background Technology

[0002] The current management of power distribution network dispatch cases faces the following technical bottlenecks: Low retrieval efficiency: Traditional databases rely on manual label classification (such as searching by fault type), which makes it difficult to support complex query needs such as "similar power grid status + similar handling targets"; Single matching dimension: Existing methods (such as keyword matching and collaborative filtering) only consider local features (such as faulty equipment type) and ignore key factors such as topology and spatiotemporal distribution of load; Insufficient dynamic adaptation: In scenarios with high penetration of new energy, the matching degree between historical cases and real-time operating status is significantly reduced.

[0003] Typical limitations of comparison techniques: Knowledge graph-based case retrieval does not solve the problem of feature extraction for unstructured data (such as scheduling recordings); Case similarity is calculated based on Euclidean distance, but no power grid physical constraints (such as ring network topology connectivity verification) are introduced.

[0004] In view of the shortcomings of existing technologies, it is necessary to design a new solution to meet the actual application needs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as single-dimensional case retrieval, poor dynamic adaptability, and lack of physical constraints, this invention proposes a multimodal fusion intelligent recommendation architecture as shown in the figure. Its core is to solve the adaptation problem between historical experience and real-time scenarios through a heterogeneous data collaboration engine and a physical-semantic dual verification mechanism. It is suitable for auxiliary decision-making in scenarios such as power grid fault handling, load transfer, and operation mode optimization.

[0006] The present invention is implemented using the following technical means.

[0007] A method for retrieving and recommending historical cases for distribution network dispatching auxiliary decision-making includes the following steps: Construct a multimodal case feature library to store historical case data containing topology, load curves, and scheduling text; Receive information on current power distribution network faults or dispatch requests; A graph neural network is used to calculate the topological similarity between the current scene and historical cases; The dynamic time warping algorithm is applied to calculate the load distribution similarity. Calculate the semantic similarity of scheduling texts using a BERT model fine-tuned based on power industry expertise; The weights of each feature are dynamically adjusted based on the real-time penetration rate of new energy sources to generate a comprehensive similarity score. Candidate cases that meet the conditions for safe operation of the power grid are screened through a physical constraint verification layer; Output recommended cases and their applicability analysis report.

[0008] The construction of the multimodal case feature library includes the following steps: Structured feature extraction: Obtaining equipment parameters, topology connection matrix, and protection action sequence from supervisory control and data acquisition; Unstructured feature processing: Speech recognition and sentiment analysis are performed on dispatch voice recordings, and convolutional neural networks are used to extract equipment defect features from on-site inspection images; Dynamic feature association: Associate real-time data such as meteorological data at the time of the fault and power generation curves of new energy sources.

[0009] The receipt of current distribution network fault or dispatch request information includes: Fault information acquisition: Real-time reception of fault alarm signals reported by the power distribution network monitoring system; Dispatch requirement analysis: Identify key operational objectives in the dispatch instruction text, including load transfer, renewable energy consumption, and network reconfiguration requirements; Multi-source data fusion: linking measurement data from monitoring and control and data acquisition systems, work order descriptions from production management systems, and meteorological and environmental information, and constructing structured input features.

[0010] The process of calculating the topological similarity between the current scene and historical cases using a graph neural network includes the following steps: Topology modeling: The topology of the current distribution network scenario and historical cases are modeled as weighted directed graphs, where nodes represent substations, distribution transformers or load points, edges represent lines or switching equipment, and edge weights include line impedance, rated capacity and real-time operating status. Graph Neural Network Processing: A pre-trained graph neural network model is used to extract features from the topology graph. Node features include voltage level, load type, and equipment health status, while edge features include impedance parameters and current load rate. Similarity calculation: The cosine similarity between the graph embedding vectors of the current scene and historical cases is calculated, and the deviation caused by local topological differences is corrected by the subgraph matching algorithm. The output is a normalized similarity score in the range of [0,1].

[0011] The calculation of load similarity uses a dynamic time warping algorithm, which specifically includes: Normalize historical cases and real-time load curves; Find the optimal matching path between two curves using dynamic programming. Introduce a regional electricity consumption weighting coefficient to differentiate the matching contribution of important users from that of ordinary loads; By combining the analysis of the impact of meteorological factors on load patterns, the accuracy of curve matching can be improved.

[0012] The BERT model based on power industry fine-tuning calculates the semantic similarity of scheduling texts, including the following steps: Power sector pre-training: The basic BERT model is pre-trained using a corpus containing distribution network dispatching procedures, fault handling manuals, and operation ticket texts, with a professional terminology retention rate of no less than 95%. Similarity calculation framework construction: The input layer receives scheduling text pairs, including the current scheduling request text and historical case text; The embedding layer outputs a 768-dimensional semantic vector, in which electrical engineering terms are encoded using forced alignment. The similarity calculation layer generates a score in the range of 0-1 using the cosine similarity function.

[0013] The dynamic adjustment mechanisms for feature weights are as follows: When the penetration rate of new energy exceeds a preset threshold, the weight of the load curve similarity is increased. In severe weather scenarios, increase the weight of topological similarity; For important users with power supply needs, increase the weight of text semantic similarity.

[0014] The physical constraint verification layer includes: a topology connectivity verification module, a line capacity verification module, a voltage over-limit detection module, and a protection coordination verification module.

[0015] The applicability analysis report includes: Case similarity score, key measures comparison table, explanation of differences in constraints, and implementation risk warning.

[0016] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described above.

[0017] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any step of the method described above.

[0018] This invention has achieved several technological breakthroughs in the field of power distribution network dispatching auxiliary decision-making, and has significant innovation and practicality. Its core advantages are reflected in the following aspects: 1. Multimodal fusion improves retrieval accuracy

[0019] Traditional case retrieval methods typically rely on single features (such as fault type or equipment number), making it difficult to comprehensively reflect the complexity of the power grid status. This invention pioneers a multimodal fusion architecture combining "topology structure + load curve + dispatch text." It quantifies network connectivity similarity using graph neural networks (GNNs), analyzes load spatiotemporal distribution using dynamic time warping (DTW) algorithms, and parses semantic associations using a power industry-optimized BERT model, achieving multi-dimensional feature collaborative matching. Tests show that this architecture elevates case retrieval accuracy to an industry-leading level, maintaining stable matching performance, especially in scenarios with a high proportion of renewable energy integration. 2. Dynamic weighting mechanism enhances scenario adaptability

[0020] Traditional fixed-weight models struggle to adapt to dynamic scenarios such as fluctuating renewable energy output and extreme weather. This invention proposes a dynamic weight adjustment strategy based on real-time operational data. When photovoltaic / wind power penetration exceeds a threshold, the matching weight of load curve features is automatically increased; during severe weather events such as typhoons and snowstorms, the focus is on topological similarity assessment; and for critical users' power supply needs, textual semantic analysis is strengthened. This mechanism enables the system to intelligently adapt to different scenarios, significantly improving recommendation accuracy in complex operating environments. 3. Security of the physical constraint verification scheme

[0021] Existing recommendation systems often overlook the physical limitations of power grid operation, rendering some solutions infeasible. This invention innovatively introduces a four-layer verification mechanism, including topology connectivity verification, line capacity check, voltage over-limit detection, and protection coordination analysis, eliminating illegal solutions at the source. All recommended cases are accompanied by a "Safety Compliance Assessment Report," clearly indicating the satisfaction status of key constraints, significantly reducing dispatch decision-making risks. 4. Improved efficiency and usability

[0022] Through optimizations in parallel computing and feature preprocessing, the system reduces case retrieval time from several minutes using traditional methods to seconds, meeting the timeliness requirements for emergency fault handling. Simultaneously, the availability of recommended solutions has been improved to industry-leading levels, effectively reducing the need for manual intervention. 5. Significant industrial application value

[0023] This technology can be widely applied in scenarios such as power grid dispatching, fault handling, and operation mode optimization. Its implementation will bring the following benefits: Improve power supply reliability: By accurately matching cases, shorten fault handling time and reduce power outage losses for users; Reduce operational costs: reduce reliance on manual experience and optimize resource allocation. Supporting energy transition: Adapting to the complex operating environment of high-proportion renewable energy integration and contributing to the construction of new power systems. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings. Figure 1 This is a schematic diagram of the architecture design of the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention.

[0025] Figure 2 This is a schematic diagram of the multimodal fusion recommendation in the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention.

[0026] Figure 3 This is a schematic diagram illustrating the construction of a multimodal case feature library in the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention.

[0027] Figure 4 This is a schematic diagram of the dispatch voice recording processing in the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention.

[0028] Figure 5 This is a schematic diagram illustrating the similarity calculation in the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention.

[0029] Figure 6 This is a schematic diagram of dynamic weight fusion in the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention.

[0030] Figure 7 This is a schematic diagram of physical constraint verification in the historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making of the present invention. Detailed Implementation

[0031] See attached document Figure 1-7 A method for retrieving and recommending historical cases for auxiliary decision-making in power distribution network dispatching includes the following steps: Construct a multimodal case feature library to store historical case data containing topology, load curves, and scheduling text; Receive information on current power distribution network faults or dispatch requests; A graph neural network is used to calculate the topological similarity between the current scene and historical cases; The dynamic time warping algorithm is applied to calculate the load distribution similarity. Calculate the semantic similarity of scheduling texts using a BERT model fine-tuned based on power industry expertise; The weights of each feature are dynamically adjusted based on the real-time penetration rate of new energy sources to generate a comprehensive similarity score. Candidate cases that meet the conditions for safe operation of the power grid are screened through a physical constraint verification layer; Output recommended cases and their applicability analysis report.

[0032] The construction of the multimodal case feature library includes the following steps: Structured feature extraction: Obtaining equipment parameters, topology connection matrix, and protection action sequence from the supervisory control and data acquisition system; Unstructured feature processing: Speech recognition and sentiment analysis are performed on dispatch voice recordings, and convolutional neural networks are used to extract equipment defect features from on-site inspection images; Dynamic feature association: Associate real-time data such as meteorological data at the time of the fault and power generation curves of new energy sources.

[0033] The receipt of current distribution network fault or dispatch request information includes: Fault information acquisition: Real-time reception of fault alarm signals reported by the power distribution network monitoring system; Dispatch requirement analysis: Identify key operational objectives in the dispatch instruction text, including load transfer, renewable energy consumption, and network reconfiguration requirements; Multi-source data fusion: linking measurement data from monitoring and control and data acquisition systems, work order descriptions from production management systems, and meteorological and environmental information, and constructing structured input features.

[0034] The process of calculating the topological similarity between the current scene and historical cases using a graph neural network includes the following steps: Topology modeling: The topology of the current distribution network scenario and historical cases are modeled as weighted directed graphs, where nodes represent substations, distribution transformers or load points, edges represent lines or switching equipment, and edge weights include line impedance, rated capacity and real-time operating status. Graph Neural Network Processing: A pre-trained graph neural network model is used to extract features from the topology graph. Node features include voltage level, load type, and equipment health status, while edge features include impedance parameters and current load rate. Similarity calculation: The cosine similarity between the graph embedding vectors of the current scene and historical cases is calculated, and the deviation caused by local topological differences is corrected by the subgraph matching algorithm. The output is a normalized similarity score in the range of [0,1].

[0035] The calculation of load similarity uses a dynamic time warping algorithm, which specifically includes: Normalize historical cases and real-time load curves; Find the optimal matching path between two curves using dynamic programming. Introduce a regional electricity consumption weighting coefficient to differentiate the matching contribution of important users from that of ordinary loads; By combining the analysis of the impact of meteorological factors on load patterns, the accuracy of curve matching can be improved.

[0036] The BERT model based on power industry fine-tuning calculates the semantic similarity of scheduling texts, including the following steps: Power sector pre-training: The basic BERT model is pre-trained using a corpus containing distribution network dispatching procedures, fault handling manuals, and operation ticket texts, with a professional terminology retention rate of no less than 95%. Similarity calculation framework construction: The input layer receives scheduling text pairs, including the current scheduling request text and historical case text; The embedding layer outputs a 768-dimensional semantic vector, in which electrical engineering terms are encoded using forced alignment. The similarity calculation layer generates a score in the range of 0-1 using the cosine similarity function.

[0037] The dynamic adjustment mechanisms for feature weights are as follows: When the penetration rate of new energy exceeds a preset threshold, the weight of the load curve similarity is increased. In severe weather scenarios, increase the weight of topological similarity; For important users with power supply needs, increase the weight of text semantic similarity.

[0038] The physical constraint verification layer includes: a topology connectivity verification module, a line capacity verification module, a voltage over-limit detection module, and a protection coordination verification module.

[0039] The applicability analysis report includes: Case similarity score, key measures comparison table, explanation of differences in constraints, and implementation risk warning.

[0040] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described above.

[0041] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any step of the method described above.

[0042] Example: Building a multimodal case feature library Reference Figure 3 A feature hierarchical extraction framework is constructed to extract three types of features: structured features, unstructured features, and dynamic features.

[0043] Structured Features: Standardized extraction of structured features of the distribution network is achieved through multi-source data fusion technology. Equipment parameters, including circuit breaker rated current and transformer impedance percentage, are acquired in real time from the monitoring, control, and data acquisition system and converted into a normalized vector format to ensure uniform processing of parameters with different dimensions.

[0044] Topological connectivity is established through the adjacency matrix A n×nIt is represented that, where matrix element a ij =1 indicates that nodes i and j are electrically connected, a ij =0 indicates no connection, and the adjacency matrix can be directly input into a graph neural network (GNN) for topological similarity calculation. Protection action sequences use timestamps to record fault events, for example, "t0: CB1 trips, t..." 0+200ms "CB2 locking" forms a temporal logic chain to verify the rationality of the cooperation. All structured features are time-aligned through a feature association engine, ultimately generating case feature vectors with uniform dimensions, providing standardized input for subsequent hybrid similarity calculations.

[0045] Unstructured features: See Appendix Figure 4 The system transforms raw, unstructured data into quantifiable feature vectors. For dispatch voice recordings, the system first converts the audio signal into text information using an Automatic Speech Recognition (ASR) engine, and then performs semantic encoding using a natural language processing model fine-tuned with power industry corpus. This allows for accurate identification of key elements in dispatch instructions, such as parsing "prioritize power supply to important users" into a structured semantic representation. The semantic encoding process employs an attention mechanism to capture the correlations between power dispatch terms, generating feature vectors with clear power industry semantics.

[0046] Receive current distribution network fault or dispatching request information Analyze information on power distribution network faults or dispatching needs, and extract the relevant structured, unstructured, and dynamic features from the system for future reference.

[0047] Calculate similarity See attached document Figure 5 This paper constructs a hybrid similarity calculation model for distribution network dispatching scenarios by integrating multi-dimensional features such as power grid topology, load spatiotemporal distribution, and dispatching text semantics. The method first uses a graph neural network to quantify topological connectivity similarity to ensure network structure matching; secondly, it applies a dynamic time warping algorithm to analyze the spatiotemporal characteristics of load curves and capture electricity consumption pattern similarity; simultaneously, it uses a BERT model fine-tuned based on power industry corpus to parse semantic associations in dispatching text. For scenarios with high penetration of renewable energy, a dynamic weight adjustment mechanism is introduced to automatically optimize the weight ratio of each feature based on real-time operational data (such as photovoltaic output ratio and meteorological conditions), ultimately generating a comprehensive similarity score. This method innovatively embeds power system physical constraints (such as topological connectivity and voltage limits) into the entire similarity calculation process, ensuring that recommended cases meet both historical experience matching requirements and current power grid safety operation requirements.

[0048] This application utilizes an intelligent similarity calculation system based on multi-dimensional feature fusion, employing a three-layer computing architecture to achieve accurate matching of power grid cases. The system employs feature enhancement technology to deeply embed the unique physical constraints and operational characteristics of the power system into the entire similarity calculation process.

[0049] The topological similarity between the current scenario and historical cases is calculated using a graph neural network (GNN): A topological similarity calculation model based on a graph neural network (GNN) is adopted, which quantifies the degree of matching of the structural features of the power distribution network through deep learning methods. The topological similarity calculation model abstracts the power grid topology into a weighted graph structure, where nodes represent electrical equipment such as substations and switching stations, and edges represent the connection relationships between equipment and include electrical parameters such as impedance and capacity.

[0050] The calculation process first performs graph embedding processing on the current operating network and historical cases, aggregating neighborhood features through a multi-layer message passing mechanism to generate node vector representations with power system characteristics. Then, a graph-level matching algorithm is used to calculate the similarity scores of two topologies in terms of equipment connection methods, electrical parameter distribution, etc. Specifically, the model introduces a ring network identification module and an electrical island detection mechanism to ensure that the topologies of recommended cases not only satisfy morphological similarity but also comply with the physical constraints of distribution network operation. For renewable energy access scenarios, the model automatically enhances the matching weight of the network structure surrounding the photovoltaic / wind power grid connection point, improving the recommendation accuracy under high penetration conditions. This calculation method overcomes the shortcomings of traditional topology comparison, which only considers connectivity relationships and ignores electrical characteristics, achieving deep semantic matching of power grid structural features.

[0051] Load distribution similarity is calculated using a Dynamic Time Warping (DTW) algorithm: This load curve matching method, with DTW as its core, achieves accurate comparison of electricity consumption patterns through spatiotemporal alignment technology. The calculation model first normalizes historical cases and real-time load curves to eliminate the influence of dimensional differences. Then, it uses dynamic programming to find the optimal matching path between the two curves, effectively overcoming the time-series offset problem caused by load fluctuations and asynchronous data sampling. The algorithm specifically considers the spatiotemporal distribution characteristics of the distribution network load, introducing a regional electricity consumption weighting coefficient in the similarity calculation to distinguish the matching contribution of important users from ordinary loads. For scenarios with high penetration of renewable energy, the system automatically identifies the coupling characteristics between photovoltaic / wind power output and load curves, enhancing the similarity assessment of net load curves during peak-valley matching. The calculation process simultaneously integrates the analysis of the impact of meteorological factors (such as temperature and humidity) on load patterns, ensuring the accuracy of curve matching under different environmental conditions. This scheme overcomes the limitations of the traditional Euclidean distance method in rigid time-series alignment, effectively capturing the morphological characteristics and temporal patterns of load changes, providing a more reliable load similarity assessment for dispatching decisions.

[0052] This invention employs a BERT model fine-tuned based on power industry expertise to calculate the semantic similarity of dispatch text: It utilizes natural language processing technology enhanced with power industry knowledge to achieve deep semantic parsing of dispatch text. Through a language model pre-trained in a professional domain, the system can accurately understand key information contained in dispatch records, such as equipment operation instructions, fault handling logic, and power supply requirements. During the training phase, the model integrates professional corpora including power dispatch regulations, equipment operation manuals, and typical fault cases to establish a semantic association network between power terminology. When processing text input, the system first identifies and standardizes key elements such as equipment numbers and operation instructions, and then analyzes the implicit priority of handling and operational logic relationships in the text through an attention mechanism. Specifically, the model develops an intent recognition module that can distinguish between different types of dispatch intents, such as "emergency isolation" and "load transfer," and maps text descriptions to a standardized set of handling measures. For dispatch instructions from renewable energy power plants, the system also integrates the ability to understand professional terms such as photovoltaic power prediction and wind turbine start-up and shutdown. This text processing scheme effectively solves the shortcomings of traditional keyword matching methods in terms of semantic understanding depth, achieving accurate conversion from free text to structured dispatch logic, and providing a reliable basis for semantic similarity evaluation for case matching.

[0053] The weights of each feature are dynamically adjusted based on the real-time penetration rate of new energy sources to generate a comprehensive similarity score: see attached table. Figure 6 Based on a dynamic weight fusion algorithm with multi-dimensional environmental awareness, this system achieves intelligent optimization and adjustment of feature weights during case matching. This mechanism constructs a dynamic mapping relationship of "scenario-feature-weight" by monitoring the power grid's operating status in real time, automatically adjusting the contribution ratio of each feature dimension according to changes in the system's operating environment. The system continuously collects operating parameters such as renewable energy penetration rate, meteorological conditions, and load characteristics, and performs multi-factor collaborative analysis through a pre-set weight adjustment rule base. When the proportion of photovoltaic or wind power output exceeds a threshold, the algorithm automatically increases the matching weight of load curve features; under severe weather conditions such as typhoons and rainstorms, it emphasizes enhancing the matching priority of topology features; for scheduling needs involving power supply for important users, the system correspondingly increases the weight ratio of textual semantic features. This mechanism features a specially designed weight smoothing transition function to avoid oscillations in recommendation results caused by parameter mutations. All weight adjustment rules are verified through a power expert knowledge base to ensure compliance with the actual needs of power grid operation. Through this dynamic fusion approach, the system can flexibly adapt to various complex operating scenarios while maintaining the stability of the core matching logic, significantly improving the environmental adaptability and decision accuracy of case recommendations.

[0054] Candidate cases meeting the conditions for safe operation of the power grid are screened through a physical constraint verification layer. A four-tiered progressive physical constraint verification system was designed, such as... Figure 7As shown, multi-layered security verification ensures that recommended cases comply with power grid operation specifications. This verification layer serves as the final quality checkpoint for case recommendations, employing a "screening before evaluation" process to conduct a systematic security assessment of candidate cases.

[0055] Topology connectivity verification: A depth-first search algorithm is used to check whether the recommended solution will lead to grid disconnection, ensuring that all important load nodes maintain electrical connectivity. The system specifically considers the network structure changes after the ring network is broken to avoid the formation of unnecessary electrical islands.

[0056] Line capacity verification: Real-time power flow calculations are used to verify the load rate of each line. When the load on any line exceeds 90% of its rated capacity, a case elimination mechanism is automatically triggered. The verification process considers the N-1 safety criterion to ensure that the recommended solution has the necessary operational margin.

[0057] Voltage over-limit detection: Through node voltage scanning analysis, potential voltage violations are identified. The system has a built-in dynamic voltage limit standard that can automatically adjust the allowable voltage fluctuation range according to different operating modes (such as light load at night, photovoltaic backfeeding).

[0058] Protection Coordination Verification: Analyze the timing of protection device actions in the recommended scheme to verify their selective coordination relationships. The system establishes a protection action time difference model, requiring a difference of ≥200ms between adjacent protection levels to prevent the risk of cascading tripping.

[0059] Output recommended cases and their applicability analysis report The intelligent recommendation engine outputs the optimal set of historical cases and automatically generates a multi-dimensional applicability analysis report. Specifically, this includes: a visual display of the degree of fit between each recommended case and the current scenario; a feasibility assessment, calculating the success probability of the solution based on real-time grid conditions and highlighting potential violations of physical constraints (such as voltage limits exceeding limits, line overload, etc.); a difference comparison engine that automatically extracts key differences between the recommended cases and the current scenario in terms of topology, load levels, and environmental factors; and a dynamic adjustment suggestion generation module that provides parameter correction schemes based on real-time grid operation modes (such as high-penetration renewable energy modes). The analysis report uses a "one case, one report" visualization template, including interactive elements such as trend comparison curves, highlighted topology differences, and an operational risk matrix, enabling dispatchers to complete solution decisions within 30 seconds.

[0060] This invention adopts a microservice architecture design, achieving end-to-end processing of case retrieval and recommendation through layered collaboration of a data acquisition layer, feature processing layer, similarity calculation layer, constraint verification layer, and recommendation service layer. The system interfaces with the monitoring and control data acquisition system via a dedicated power data bus to ensure the real-time performance and security of data acquisition. All services are deployed on a cloud computing platform that meets the requirements of information security standards, and containerization technology is used to achieve elastic scaling.

[0061] Data preparation and governance: The system construction includes a rigorous data governance process, extracting historical cases from scheduling logs and establishing a quality scoring system, real-time access to monitoring and control and data acquisition systems, meteorological and new energy data, and transforming heterogeneous data such as topology, load and text into standardized feature vectors through feature engineering, providing a unified data foundation for subsequent analysis.

[0062] For topology similarity calculation, a graph neural network is used to construct a topology matching model. Power grid equipment is abstracted as nodes and electrical connections are abstracted as edges. The network structure features are learned through multi-layer message passing, and ring network identification and island detection mechanisms are introduced to ensure that the recommended solution meets the physical constraints of the power grid while satisfying structural similarity.

[0063] Load curve matching is achieved by calculating the similarity of load curves based on the dynamic time warping algorithm. The sampling offset effect is eliminated by time alignment. Combined with the regional power consumption characteristics and the coupling relationship of new energy output, the degree of load pattern matching under different scenarios is accurately evaluated.

[0064] Text semantic analysis utilizes a finely tuned natural language processing model based on power industry corpus to parse scheduling text, establishes a mapping relationship from free text to structured scheduling logic, and captures key operational intentions and equipment associations through an attention mechanism to improve the accuracy of semantic understanding.

[0065] Dynamic weight fusion is implemented by designing an environment-aware weight adjustment strategy. The weight ratio of each feature is automatically optimized based on the real-time operating status. In scenarios with high penetration of new energy sources, the weight of load features is increased. In severe weather, topology matching is emphasized. During important power supply tasks, text analysis is enhanced to achieve intelligent feature fusion.

[0066] Physical constraint verification is conducted, and a four-fold safety verification system is constructed. The system sequentially checks topological connectivity, line capacity, voltage limits, and protection coordination. A "screening before evaluation" mechanism is adopted to ensure that the recommended scheme fully complies with the power grid safety operation specifications and generates a detailed safety compliance assessment report.

[0067] System integration testing includes functional testing, performance testing, and extreme scenario verification. It integrates with the existing scheduling and control system through standard interfaces and establishes a sound operation and maintenance mechanism, including health checks, feature drift monitoring, and regular model updates.

[0068] In a typical application scenario, taking typhoon weather fault handling as an example, the system integrates environmental data such as wind speed and precipitation, matches similar handling strategies from historical cases, and outputs recommended solutions that include safety assessments, significantly improving decision-making efficiency in complex scenarios.

[0069] To ensure effective implementation, data quality control, algorithm performance optimization, and multi-layered security protection are implemented to ensure reliable system operation, compatibility with existing power standards and specifications, and comprehensive improvement in case retrieval efficiency and recommendation quality. Example: Recommended Case Studies of Typhoon Weather Fault Handling

[0070] Real-time data input Current status: 10kV line #23 is broken, wind speed is 15m / s, and photovoltaic output has dropped to 30%. Dispatch objective: Prioritize ensuring power supply to the hospital, with a time limit of 1 hour for restoration.

[0071] Case search results:

[0072] This invention has achieved several technological breakthroughs in the field of power distribution network dispatching auxiliary decision-making, and has the following significant innovations and practical applications.

[0073] 1. Multimodal fusion improves retrieval accuracy; 2. Dynamic weighting mechanism enhances scenario adaptability; 3. Physical constraint verification ensures the security of the solution; 4. Improved efficiency and usability; 5. Significant industrial application value.

Claims

1. A method for retrieving and recommending historical cases for distribution network dispatching auxiliary decision-making, characterized in that, Includes the following steps: Construct a multimodal case feature library to store historical case data containing topology, load curves, and scheduling text; Receive information on current power distribution network faults or dispatch requests; A graph neural network is used to calculate the topological similarity between the current scene and historical cases; The dynamic time warping algorithm is applied to calculate the load distribution similarity. Calculate the semantic similarity of scheduling texts using a BERT model fine-tuned based on power industry expertise; The weights of each feature are dynamically adjusted based on the real-time penetration rate of new energy sources to generate a comprehensive similarity score. Candidate cases that meet the conditions for safe operation of the power grid are screened through a physical constraint verification layer; Output recommended cases and their applicability analysis report.

2. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The construction of the multimodal case feature library includes the following steps: Structured feature extraction: Obtaining equipment parameters, topology connection matrix, and protection action sequence from the supervisory control and data acquisition system; Unstructured feature processing: Speech recognition and sentiment analysis are performed on dispatch voice recordings, and convolutional neural networks are used to extract equipment defect features from on-site inspection images; Dynamic feature association: real-time data of meteorological data at the time of the fault and the power generation curve of new energy sources.

3. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The receipt of current distribution network fault or dispatch request information includes: Fault information acquisition: Real-time reception of fault alarm signals reported by the power distribution network monitoring system; Dispatch requirement analysis: Identify key operational objectives in the dispatch instruction text, including load transfer, renewable energy consumption, and network reconfiguration requirements; Multi-source data fusion: linking measurement data from monitoring and control and data acquisition systems, work order descriptions from production management systems, and meteorological and environmental information, and constructing structured input features.

4. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The process of calculating the topological similarity between the current scene and historical cases using a graph neural network includes the following steps: Topology modeling: The topology of the current distribution network scenario and historical cases are modeled as weighted directed graphs, where nodes represent substations, distribution transformers or load points, edges represent lines or switching equipment, and edge weights include line impedance, rated capacity and real-time operating status. Graph Neural Network Processing: A pre-trained graph neural network model is used to extract features from the topology graph. Node features include voltage level, load type, and equipment health status, while edge features include impedance parameters and current load rate. Similarity calculation: The cosine similarity between the graph embedding vectors of the current scene and historical cases is calculated, and the deviation caused by local topological differences is corrected by the subgraph matching algorithm. The output is a normalized similarity score in the range of [0,1].

5. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The calculation of load similarity uses a dynamic time warping algorithm, which specifically includes: Normalize historical cases and real-time load curves; Find the optimal matching path between two curves using dynamic programming. Introduce a regional electricity consumption weighting coefficient to differentiate the matching contribution of important users from that of ordinary loads; By combining the analysis of the impact of meteorological factors on load patterns, the accuracy of curve matching can be improved.

6. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The BERT model based on power industry fine-tuning calculates the semantic similarity of scheduling texts, including the following steps: Power sector pre-training: The basic BERT model is pre-trained using a corpus containing distribution network dispatching procedures, fault handling manuals, and operation ticket texts, with a professional terminology retention rate of no less than 95%. Similarity calculation framework construction: The input layer receives scheduling text pairs, including the current scheduling request text and historical case text; The embedding layer outputs a 768-dimensional semantic vector, in which electrical engineering terms are encoded using forced alignment. The similarity calculation layer generates a score in the range of 0-1 using the cosine similarity function.

7. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The dynamic adjustment mechanisms for feature weights are as follows: When the penetration rate of new energy exceeds a preset threshold, the weight of the load curve similarity is increased. In severe weather scenarios, increase the weight of topological similarity; For important users with power supply needs, increase the weight of text semantic similarity.

8. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The physical constraint verification layer includes: a topology connectivity verification module, a line capacity verification module, a voltage over-limit detection module, and a protection coordination verification module.

9. The historical case retrieval and recommendation method for distribution network dispatch auxiliary decision-making according to claim 1, characterized in that, The applicability analysis report includes: Case similarity score, key measures comparison table, explanation of differences in constraints, and implementation risk warning.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-9.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-9.

Images