A system and method for converting power notification text to 5G messages

Abstract

The present invention provides a system and method for converting power notification text into 5G messages. The event parsing module receives raw power event data from the power PMS and DMS systems, adopts rule parsing and key field extraction algorithms, and outputs a standardized structure; the summary generation module receives structured event data, combines the event type and the target user portrait, and generates a popular text summary; the graphic generation module maps the geographic coordinates and device number in the event structure to the GIS platform, and highlights the event area with a hot zone on the GIS map; the graphic and text linkage module anchors the summary text and the graphic, corresponds the event range in the text to the graphic hot zone, and binds the event time to the estimated recovery time countdown; the message encapsulation module encapsulates the text summary and graphic content into a 5G rich media message format, and finally sends it to the user terminal through the operator's 5G core network.

Description

Technical Field

[0001] The present invention relates to the field of 5G communication technologies, and in particular to a system and method for converting power notification text into 5G messages. Background Art

[0002] For a long time, power companies have mainly relied on traditional SMS to send notifications to users. The SMS format is mainly fixed templates and the content is simple text. It has many problems such as poor readability, weak content carrying capacity, and low user response rate. It can no longer meet the needs of modern power grid intelligent operation and maintenance and diversified user services. Against the background of rapid digitalization and intelligent development of power systems, power grid companies urgently need to use efficient and intelligent information transmission methods to promptly send notifications such as emergencies, planned maintenance, and load adjustments to the majority of user terminals to ensure that users perceive them in a timely manner and respond appropriately. 5G rich media communication RCS technology, with its advantages of high bandwidth, low latency, and multimodal transmission, provides basic support for building the next generation of power message push system.

[0003] With the widespread deployment of 5G networks, messaging services based on the rich media communication technology RCS are gradually maturing. This technology supports multiple modal content such as images, text, voice, interactive buttons, and location information, and has stronger expressiveness and interactivity. It is particularly suitable for the scenario-based presentation of complex events and personalized information push. However, the existing power system has not yet formed a mature mechanism to efficiently convert original text-based power notification SMS messages into 5G rich media messages. There are still large technical gaps in data structure standardization, modal generation, anchor point binding, encapsulation adaptation, and distribution control.

[0004] The present invention provides a system and method for converting power notification text into 5G messages, which automatically convert traditional power text messages into structured, multimodal 5G rich media messages, and realize the intelligent upgrade of power event notifications from pure text to 5G messages with a combination of text and pictures, voice assistance, and interactive feedback. It not only significantly improves the information expression and user perception experience, but also performs differentiated content matching and precise distribution based on the user's power grid portrait, with stronger reach efficiency, adaptability and service flexibility, and is particularly suitable for smart grid message push scenarios for large-scale terminals. Summary of the Invention

[0005] The purpose of the present invention is to provide a system and method for converting power notification text into 5G messages.

[0006] The problems to be solved by the present invention are: first, traditional power event notifications are still mainly plain text text messages, which have a single information carrying form and weak expression ability, and cannot effectively present complex key information such as power outage scope, fault location and restoration time; second, the existing message push mode lacks the ability to accurately adapt to the user's power grid portrait, and it is difficult to achieve differentiated content organization and personalized distribution according to user type, terminal capability and event urgency level, resulting in poor notification effect and low user response rate. For this reason, the present invention proposes a system and method that can automatically parse power event text into structured semantic information, and integrate multimodal content such as graphics, text, and voice to generate 5G rich media messages, thereby realizing the leap of power notification from "text notification" to "smart push".

[0007] A system for converting power notification text to 5G messages, comprising:

[0008] Event parsing module: Receives raw power event data from the power PMS and DMS systems, pre-processes the power event data, uses rule parsing and key field extraction algorithms, and outputs a standardized structure;

[0009] Summary generation module: Receives structured event data and models it into a multi-dimensional semantic graph structure based on the power grid event graph model. It then calls the semantic summary generation model to generate a popular text summary based on the event type and target user profile.

[0010] Graphics generation module: maps the geographic coordinates and device numbers in the event structure to the GIS platform, obtains the corresponding area layer, renders the event area, event location, and impact range boundary, and highlights the event area as a hot spot in the map;

[0011] Graphics and text linkage module: anchors the summary text to the graphics, uses field mapping strategies to map the text summary to the graphic hot zone ID, and binds the time to the estimated recovery time countdown timer;

[0012] Message encapsulation module: encapsulates text summaries and graphic content into 5G rich media message format, selects the appropriate message version based on the user terminal capabilities, and finally sends the message to the user terminal through the operator's 5G core network.

[0013] A method for converting power notification text into 5G messages is based on a system for converting power notification text into 5G messages, and the technical solution adopted is as follows:

[0014] S1: Receives raw power event data from the power PMS and DMS systems, preprocesses the event data, including field cleaning, device mapping, and time normalization, and converts the raw event data into structured event information using power event semantic segmentation and multi-channel parsing algorithms.

[0015] S2: Based on a preset power grid event graph model, semantic modeling is performed on the structured event information to construct a multi-dimensional semantic graph structure. The node information in the semantic graph includes event type, involved equipment, geographic location, expected recovery time, and affected area. Subsequently, the semantic graph vector is input into the semantic summary generation model. Combined with the event type and target user profile, a popular notification summary text expressed in natural language is generated.

[0016] S3: Extract the device number and geographic coordinate data from the event information body, connect to the GIS platform through the interface, locate the geographical area involved in the event on the map layer, generate the corresponding graphic representation, and highlight the hot zone of the relevant area according to the event scope to form a visual image of the event area;

[0017] S4: Mark the generated summary text with field anchors and establish mapping relationships with graphic elements. The mapping relationships include binding the place name and equipment name in the summary to the GIS hot zone ID, and binding the estimated recovery time field to the countdown module, to implement interactive logic that highlights the graphic when the user clicks the text anchor.

[0018] S5: Encapsulate the generated text summary and graphic content, generate an RCS message structure according to the 5G rich media communication standard format, and dynamically select the appropriate message version based on the user terminal capabilities and network capabilities, including plain text, graphic and text fusion, and voice-assisted versions. Finally, the encapsulated message is transmitted to the target user terminal through the 5G core network.

[0019] Furthermore, in step S1, the power event semantic segmentation and multi-channel parsing algorithm is used to convert the original event data into a structured event information body, including:

[0020] S31: Receives raw text data from the PMS and DMS systems. The raw text data includes semi-structured and free expression formats, performs character and time unit unification, punctuation completion, word error correction, and encoding compatibility processing. The processed text is marked with a unique event ID and encapsulated into an event raw object EventRawObject, completing basic cleaning and standardization of the raw input power grid event description text.

[0021] S32: Perform lexical segmentation and word segmentation on the preprocessed text. A self-trained DependencyParser-BiLSTM model is introduced to construct a syntactic dependency graph for the text. Combining a predefined event semantic dictionary with grammatical rule templates, it identifies candidate segments of event-centered words, device subject phrases, location complement phrases, and time adverb phrases in the sentence. Subsequently, a rule-weighted window sliding strategy is used to merge and disambiguate redundant and nested segments. Ultimately, the text is decomposed into four semantic segments: device segments, event segments, location segments, and time segments. Each segment is encapsulated into a standardized structure called SemanticChunk, annotated with its semantic category and character position information.

[0022] S33: Based on the semantic type field in the category label SemanticChunk of the semantic fragment, each semantic fragment is dynamically routed to the corresponding dedicated parsing channel for parallel processing. The scheduling of the parsing channel adopts an event-driven microservice architecture. Each type of semantic fragment is internally configured with an independent processing module, which is registered as a device identification service, a location extraction service, a time parsing service, and an event classification service in the form of a REST API. The structured semantic fragment is encapsulated in a standard JSON format through the internal message queue and attached with an event context ID, and then delivered to the message bus. Each parsing service module listens to the corresponding channel. After receiving the message, it automatically schedules the loading of the corresponding sub-model to complete the entity extraction, classification, and standardization processing of the semantic fragment. The scheduling center uses a thread pool + asynchronous callback mechanism to control the processing status of each channel and writes the results back to the event semantic cache pool EventSemanticCache for subsequent fusion module reading and calling;

[0023] S34: A named entity recognition model based on the BERT-CRF architecture is used to extract key information from the device fragment, including substation, line number, voltage level, and feeder name. The model takes the syntactic context as input and outputs the annotated device entity category and text location. The parsed results are then fed into the device hierarchy completion module, which automatically infers and fills in missing fields based on the trained device coding rules and dictionary system. The final output is a device information structure, DeviceStruct, which contains the device ID, voltage level, substation, type, and a confidence score.

[0024] S35: Use the BiGRU-CRF geographic entity recognition model, which combines word vector representation with prior knowledge of place names, to identify residential areas, landmarks, and community names in location segments, output the geographic entity names and their locations in the original text, and map the identified place names to standardized coordinate pairs using a place name-coordinate mapping table. For fuzzy matching place names, weighted similarity using geographic distance is used for fuzzy error correction to ensure positioning accuracy, and output a multi-location parallel structure, LocationStruct.

[0025] S36: Uses a regular expression-driven semantic pattern tree matching method to perform pattern recognition and type classification on time phrases. Then, based on the current system time context, it converts natural language time expressions into the ISO 8601 time format. The output time structure includes event start time, recovery time, and estimated duration fields.

[0026] S37: Perform intent recognition on the content of the event fragment to determine whether it belongs to the type of predefined event types such as tripping, planned maintenance, sudden power outage, and load abnormality. The model input is the text representation vector of the event fragment, and the output is the event type category and its probability distribution. The system presets the confidence threshold to 0.75 by default. Samples with a confidence greater than or equal to 0.75 are high-confidence samples, and samples with a confidence less than 0.75 are low-confidence samples. According to the set threshold, the event type with a confidence higher than a specific threshold is output. The low-confidence samples are sent to the manual review module. The result will be used as the upstream input for generating the summary template and the selection of the graphic template;

[0027] S38: After each channel completes the parsing, the system aggregates the parsing results of each SemanticChunk to generate a complete EventStruct object. For field overlaps and entity conflicts, the system uses a weighted voting mechanism to resolve conflicts. The weighted items include combined confidence, fragment position, and context consistency. For missing fields, the graph completion logic is called to finally generate a power event structure with complete structure and standardized fields for call by the summary generation and graphic generation modules.

[0028] Natural language time expressions are uniformly converted to ISO 8601 time format using the self-developed T-TimeParser time standardization engine. The specific implementation method is as follows:

[0029] Extracting candidate time phrases: First, the input power notification text is segmented, breaking the entire sentence into several consecutive word segments. Then, all segments between two and six words are matched sequentially against a predefined time expression template. The template is a set of rules based on common Chinese expressions. Whenever a segment meets the template rules, it is marked as a candidate time phrase, and its position range in the original text and its expression type are recorded for subsequent processing.

[0030] Context-anchored parsing: Next, the candidate time phrase identified in the previous step is combined with the current system time to determine the specific calendar time it represents. If the phrase is ambiguous, it is mapped to a specific date and time value using the rule anchor dictionary. If the current system time is later than the time point, it is inferred to be the same time the next day. If the phrase expresses a relative time, the current time is added to the corresponding time interval to obtain a new specific time point.

[0031] Temporal semantic role recognition: Determine the temporal meaning of each temporal phrase based on its contextual position in the sentence and the accompanying verb, such as whether it represents the "start time of an event," "end time of an event," or "duration of an event."

[0032] Semantic completion and reasoning: For incomplete time expressions in power text, we use other contextual information to supplement the specific meaning of these expressions. This supplementary information includes common time patterns in the user's location and the presence of logical relationships between multiple time words in the text. Ultimately, the fuzzy expressions are completed into clearly structured time values.

[0033] Time standardization conversion: After completing the above analysis, each parsed time phrase is converted to a standard time format, which is the international standard ISO 8601. If a phrase represents a time period, the system will generate two time values, one for the start and one for the end of the period.

[0034] Fusion of structured output and confidence weighting: Finally, the system encapsulates the original content of each time phrase, the parsed standard time, the time role it represents, and a credibility score into a structured data entry. The credibility score reflects the reliability of the system's time parsing results. The scoring reference factors include: whether the time expression is standard, whether the context is clear, and the accuracy of this type of time expression in historical data. For time phrases with clear expression and standard structure, the system score will be higher, while for expressions with ambiguous semantics and unclear context, the score will be relatively low.

[0035] The power event semantic segmentation and multi-channel parsing algorithm aims to accurately identify the core semantic fragments of the event, including equipment, location, time, and event type, from the original power event text data obtained from the power dispatching system PMS and DMS. It also performs entity recognition and structuring on various fragments through a parallel multi-channel processing architecture, and finally constructs a standard structure that can be used for subsequent summary generation and graphics rendering. It has high scalability and domain adaptability, and can effectively deal with problems such as non-standard text, loose structure, and polysemy overlap. It is a basic capability module for building a power notification intelligent conversion system.

[0036] The specific operation steps of the power event semantic segmentation and multi-channel parsing algorithm are as follows:

[0037] Step 1: Text preprocessing: perform character normalization and semantic cleaning on the input text. The main processing methods include encoding standardization, Chinese punctuation restoration and sentence processing, and time unit unification. Example function: def normalize_text(raw_text);

[0038] Step 2: Semantic fragment segmentation, using dependency syntax graph + rule dictionary for semantic fragment identification; the model used is self-trained BiLSTM + DependencyParser, and the output types include device fragment, location fragment, event fragment, and time fragment. The structure of each fragment is SemanticChunk = {text, span, type}

[0039] Step 3: Multi-channel parallel parsing routes different types of fragments to different parsing channels for dedicated structured processing. The device channel includes entity recognition, voltage level completion, and encoding; the location channel includes place name recognition, coordinate mapping, and defuzzification; the time channel includes time word recognition and ISO standardization; and the event channel includes event intent recognition. Each channel is executed concurrently through a message queue and thread pool, and the results are written back to the cache after parsing is completed.

[0040] Step 4: Conflict resolution and result fusion: merge, disambiguate, and complete missing fields on the outputs from each channel.

[0041] The power event semantic segmentation and multi-channel parsing algorithm adopts error tolerance and belief propagation mechanism.

[0042] To prevent individual channel parsing failures from affecting the overall results, the system sets an information completeness threshold in the semantic cache pool. When certain fields are missing or conflicting, the semantic completion module is automatically triggered, including:

[0043] 1. Event graph completion;

[0044] 2. Default inference based on contextual time or place;

[0045] 3. Channel confidence propagation strategy: When multiple channels return conflicting entity results, the confidence propagation formula is used to determine the final result: ,in, For all candidate entities, is the set of participating channels, is the channel weight, which is adaptively updated based on the channel’s historical accuracy. For channel The confidence level of the output.

[0046] Each semantic parsing channel is registered with the system's event processing and dispatching center using a RESTful API service module. The interface uniformly adopts the following standard input and output formats:

[0047] Input structure: {chunk_id, chunk_type, raw_text, context_id};

[0048] Output structure: {chunk_id, semantic_type, extracted_entity, normalized_result, confidence_score};

[0049] This interface specification supports the independent evolution of different parsing modules while ensuring interoperability, and supports the addition of new channels such as sentiment analysis channel, semantic confidence filtering channel, etc., to achieve horizontal scalability of the system.

[0050] The power event semantic segmentation and multi-channel parsing algorithm relies on Python 3.9+ as the main language environment, Jieba as the Chinese word segmentation and syntax analysis library, PyTorch as the deep learning model inference framework, RabbitMQ as the message queue for inter-module parallelism and scheduling, Elasticsearch as the location fuzzy query, and FastAPI as the service interface framework for each channel.

[0051] Explicit segmentation of semantic structures improves parsing accuracy: This algorithm uses structured semantic fragment recognition technology based on a combination of syntactic dependencies and semantic rules to accurately decompose raw event text into four semantic blocks: "device, time, location, and event." This significantly reduces semantic ambiguity and contextual confusion, making it particularly suitable for processing text with complex sentence structures and overlapping entities in power business scenarios.

[0052] A multi-channel parallel parsing architecture significantly improves processing efficiency: Different types of semantic fragments are routed to dedicated parsing channels, enabling asynchronous concurrent processing within a microservices framework. The system boasts high throughput and low latency, making it particularly well-suited for the batch structuring of large-scale, real-time power event text.

[0053] Fusion of semantic rules and deep learning models, balancing generalization and precision: Device, location, and time channels are each combined with specialized models such as BERT-CRF, BiGRU, and regularized semantic tree matching. This fusion of business rules and model outputs, combined with a confidence voting mechanism for disambiguation, enables highly robust entity extraction and field standardization, with excellent performance for non-standard expressions and colloquial text.

[0054] Supports subsequent module linkage and complete semantic chain closed loop: the output structured result EventStruct retains all fragment location information, confidence, and semantic tags, supports direct reference of modules such as summary generation, graphics rendering, and rich media encapsulation, forming a semantic closed loop consistent with the front-end and back-end, effectively improving the consistency of graphics and text, interactive linkage, and message personalization.

[0055] Furthermore, in step S2, a summary generation model is used. The summary generation model uses a dynamic summary generation method based on an event semantic element weight matrix and user profile drive, including:

[0056] S41: Extracting a preset semantic element set from the structured power event data, the system first represents the structured event object as ,in is the event type, is the starting time, For recovery time, For the affected area, The name of the main device, As the responsible unit, each field is mapped into a semantic element vector , as the basic semantic representation input before summary generation;

[0057] S42: Use the target user's identity attributes, device capabilities, and behavioral preferences to build a user portrait vector ,in Whether you are an enterprise user, For elderly users, Whether there is a graphics terminal, Pay attention to the weight of time information in past clicks, The attention weight of the location information in the past clicks is finally obtained as a set of standardized user vectors ;

[0058] S43: Construct a weight matrix ,in Semantic elements In user portrait features The initial value of the matrix is ​​set by business experience, and is subsequently fine-tuned through feedback data. , calculate its weight sorting vector , the formula is ,in, Semantic elements The user's summary contribution score is sorted to obtain the semantic generation order;

[0059] S44: Select the top k semantic elements according to the descending sorting result of S and fill them into the prompt template Prompt. For missing fields and low weight fields, they will not be filled into Prompt. Prompt is encoded as an input vector and together with the semantic element content, it constitutes the final language model input structure LLMInput={prompt, event-emantics};

[0060] S45: The Prompt template and the event semantic vector are fed into a summary generation model. The model is a fine-tuned Transformer architecture that uses formatted methods as input and ultimately outputs a natural language summary.

[0061] S46: Finally, the system evaluates the quality of the generated summary text and uses it to dynamically update the weight matrix ,Quality evaluation indicators include semantic consistency score, click feedback index, and error ,correction feedback rate.

[0062] The generation model uses a dynamic summary generation method based on the event semantic element weight matrix and user profile drive. Its core functions include:

[0063] Achieve semantic element-level summary content customization: This method uses multiple semantic fields (such as event type, time, location, and equipment) in structured power event data as a basic semantic element set. Based on user profiles, it dynamically filters the information that different users are most concerned about and generates personalized summary text, improving the relevance and readability of the summary.

[0064] A user-profile-driven mechanism is introduced to optimize summary content: Based on user profile attributes such as terminal type, age level, and business preferences, the system automatically adjusts the presentation order of semantic elements in the summary, retains fields, and uses language style to make the message content more aligned with the recipient's cognitive habits and lower the threshold for information comprehension.

[0065] Dynamic calculation of semantic weights and summary generation control: By constructing a weight matrix of "semantic elements × user profile features," the importance of different semantic elements to different types of users is calculated, and the content priority and compression level during summary generation are determined. This allows for dynamic clipping of summary content and adaptive generation templates.

[0066] Supports continuous learning and optimization of summary generation quality: Utilizes multi-dimensional indicators such as user click feedback, error correction rate, and semantic matching to iteratively optimize the semantic weight matrix, achieve self-adjustment of semantic sorting weights and enhance corpus adaptability, and form a closed-loop learning mechanism.

[0067] The dynamic summary generation method based on the event semantic element weight matrix and user portrait driving includes the following steps:

[0068] Event semantic element vectorization: multiple core field event types in structured power event data , start time , Estimated recovery time of the incident ,Place , device name , responsible unit Mapped into a set of semantic elements , and construct a semantic representation vector for the initial semantic basis of summary construction;

[0069] User portrait vector construction: Extract the user's static attributes including whether they are corporate users , whether elderly users , terminal capability graphics support , focus on time weight , focus on location weight , forming a standardized user portrait vector ;

[0070] Semantic element weight matrix calculation: Constructing the semantic weight matrix , representing semantic elements In user portrait features The importance of the matrix is ​​calculated by matrix multiplication to calculate the comprehensive semantic score vector , sort the semantic score vectors in descending order to obtain the output priority of the semantic elements;

[0071] Prompt summary template construction: According to the sorted score vector, intercept the front The high-weight semantic elements are selected to construct a summary and generate a prompt template. Unselected low-weight fields will be ignored. The generated prompt and semantic content together constitute the language model input structure.

[0072] Summary generation and natural language output: The prompt template and semantic content are fed into the fine-tuned Transformer summary generation model, which outputs a popular, user-customized natural language notification summary text;

[0073] Summary evaluation and weight matrix fine-tuning: Use semantic consistency scoring, user click behavior feedback, error correction rate statistics and other indicators to comprehensively evaluate the quality of summary generation, and then adjust the weight matrix Fine-tune and optimize the corresponding parameters to form a continuous learning closed loop.

[0074] The accuracy of summary content is significantly improved, effectively reducing redundancy and ambiguity: This method vectorizes multiple fields in structured power event data (such as event type, time, location, equipment, etc.) and inputs them into a weight matrix model. It dynamically assigns semantic weights based on user profiles, thereby screening the most noteworthy summary elements for presentation. Experimental results show that, based on comparisons with traditional fixed template summaries, the summaries generated by this method improve the ROUGE-2 index (an accuracy index measuring semantic consistency) from 0.67 to 0.81, and the entity hit rate increases from 74.2% to 89.6%, significantly improving the semantic accuracy and expression clarity of the summaries.

[0075] Improved user click-through rate and interactive participation rate: The summary content is more closely aligned with the reading habits and information needs of different user groups, significantly increasing users' willingness to actively respond to power notifications. In an A / B test of 800,000 real power notification messages, the summary version generated by this method saw its click-through rate increase from 12.3% to 17.9% among residential users and from 9.8% to 14.1% among corporate high-voltage users, with an average click-through rate increase of over 45%, effectively improving notification communication effectiveness and interactive response rate.

[0076] Effectively compresses summary length to meet the needs of 5G rich media packaging: During RCS rich media message packaging, content length and compression ratio directly impact the packaging success rate and display aesthetics. This method uses a weight-driven summary control mechanism to precisely trim redundant information and achieve condensed information expression. Test data shows that the average length of generated summaries is compressed from 112 words (using traditional methods) to 82 words, achieving a compression ratio of 26.8%, while retaining at least 95% information coverage, perfectly adapting to the constraints of mixed text and image messages.

[0077] Improved summary generation efficiency: Since the summary generation process only constructs the input structure based on a small number of high-weight semantic fields, it can significantly reduce the language model inference burden and improve real-time response capabilities. Tests in server-side deployment show that the generation time of each summary is shortened from 147 milliseconds for traditional structure input to 93 milliseconds, and the inference delay is reduced by 36.7%. It is particularly suitable for high-concurrency event push scenarios and edge device deployment scenarios, and has good system scalability and stability.

[0078] Furthermore, in step S3, a method for generating electric GIS graphics based on event-region coupling and dynamic layer clipping is used, including:

[0079] S51: The system first extracts all fields with geographic attributes from the structured event data, including the list of affected cells, the main device number, the substation to which the device belongs, and the voltage level. It then searches the device geographic mapping table containing the list of affected cells, the main device number, the substation to which the device belongs, and the voltage level to generate the event impact area point set P and the device set D for subsequent spatial calculations.

[0080] S52: Using event-region coupling function The coupling function combines three factors: spatial distance effect, the straight-line distance between the event center and the regional center; topological correlation, the spatial intersection ratio between the main device service area and the user location; user density weight, the number of registered users per unit area on the GIS grid map; the coupling function is defined as ,in, is the center point of the event, Candidate area The center point, For equipment service area, For the region The boundary polygon of is a weight parameter, which is adjusted to adapt to user differences. The dist function represents the spatial straight-line distance between the event center and the area center, normalized to , calculated as ,in are the latitude and longitude coordinates of the event center point, is the latitude and longitude of the center point of the region, The maximum visual range distance set by the platform is used for normalization. The closer the distance, the closer the dist function value is to 1, and the higher the coupling degree. The farther the dist function distance is, the closer the value is to 0. The IOU function is the spatial intersection degree between the service range of the event master device and the candidate area, which measures the direct geographical coverage relationship. The specific calculation method is: ,in is the polygonal area formed by the service range of the event master device. is the polygon corresponding to the candidate region r, Area is the area calculation function in the map system, and the IOU function value is in the range of [0,1]. The closer it is to 1, the more completely the region is covered by the event, indicating a high degree of coupling. The function is used to calculate the number of registered electricity users per unit area. The specific calculation method is: ,in is the total number of registered users in region r, obtained from the main grid data, is the area of ​​region r, The maximum user density per unit area in the entire network area is used for normalization and is selected after the final calculation. The area before the score 𝑁 enters the candidate set of graph cutting;

[0081] S53: Based on the set of regions with the highest coupling degree, the system constructs a minimum coverage rectangle and defines a cropping boundary area 𝐵. This area includes all hotspots and controls the total map size to not exceed the set image resolution. The geographic base layer, power-specific layer, and user interest layer are loaded inside the cropping area. The layer overlay adopts a hierarchical transparency + raster priority mechanism.

[0082] S54: Within the cropped area, the system analyzes all device point sets D using a spatial clustering algorithm to generate multiple event hotspots H. The system then adds attributes to each hotspot, including the hotspot ID, event level, and clickable range. Finally, each hotspot is bound to the corresponding field in the text summary, forming a complete graphic-text linkage model.

[0083] S55: Encode the rendered graphic cropping area into a format supported by RCS, automatically select a version based on the user terminal capabilities, and finally encapsulate it into a 5G message data packet and send it together with the text summary and countdown linkage information.

[0084] The event-region coupling degree scoring in the power GIS graphics generation method based on event-region coupling degree and dynamic layer clipping not only adopts a static weight combination including spatial distance, topological association, and user density, but also introduces a dynamic weight adjustment mechanism driven by event type; the system automatically adjusts the coupling degree function according to the business type of the event, such as planned maintenance, fault tripping, and load warning. Weight parameters; for example, in the event of a sudden failure, the system will increase the spatial distance weight to quickly mark nearby areas, while in planned maintenance notifications, it will place more emphasis on topological coverage and affected user density, making the graphical display more consistent with business characteristics.

[0085] To make graphic information richer and less crowded, a layer-level configurable overlay mechanism was designed based on the power GIS graphic generation method of event-region coupling and dynamic layer clipping. The layers are divided into: basic geographic layers, power-specific layers, and user interest layers. The system assigns different transparency and priority to each layer, controls the display effect through layer synthesis logic Z-index sorting + inter-layer filtering, and adapts the layer complexity according to the user terminal capabilities. For example, when the terminal performance is low, the system only retains the main hot zone and equipment layers and ignores non-essential interest layers to ensure rendering performance.

[0086] Hotspot is not just a simple geographical area selection, but also needs to have semantic perception capabilities; using density-spatial weighted clustering algorithm, in the geographical point set Analyze and form several hot blocks Each hot zone is attached with attributes such as a unique ID, event level, estimated recovery time, etc.; the ID of each hot zone is then bound to the anchor field in the summary text through the graphic-text linkage module to achieve high-light linkage when clicking on the graphic and text; when generating the anchor, the system automatically extracts content such as place names and device names based on the entity recognition position field and establishes a mapping with the hot zone to ensure that when the user clicks on the text, the graphic area can be accurately highlighted and positioned.

[0087] To meet the standardized requirements for RCS rich media message encapsulation, a structured graphic encapsulation model is used. This model includes image content, hot zone definition JSON, layer configuration description, and anchor binding data. An example of the encapsulation format design is {"image": "base64 image or URL", "hotzones": {"id": "HZ001", "name": "Dongfang Community", "polygon": [...], "type": "residential area", "bind": "text segment #3"}, "layers": ["map background", "device layer", "user interest layer"], "interaction": {"anchor_map": {"Dongfang Community": "HZ001", "Fuxing Station": "HZ004"}}}.

[0088] Significantly improve the matching degree between the graphic content and the areas of user concern: By introducing a coupling function to fuse spatial distance, device topological relationship, and user density factors, the system can accurately identify the areas that users are most concerned about for graphic rendering; compared with the traditional static cropping method of "main device position point + radius expansion", the coverage rate of the target user concern area in the generated graphic has increased from 72.8% to 93.5%, significantly enhancing the fit between the graphic and text content and the actual needs, and effectively avoiding irrelevant redundancy of graphic information;

[0089] Effectively control the graphic cropping range and reduce the terminal transmission and rendering burden: Based on the minimum covering rectangle construction strategy, only the area layers strongly associated with the event are retained, which can effectively compress the graphic size and complexity; in the test environment, under the same event, the amount of graphic data for rendering has been reduced from an average of 1.4MB to 0.87MB, a decrease of about 38%. On mid- to low-end user terminals, the image loading time has dropped from 2.6 seconds to 1.5 seconds, improving the rendering efficiency and user experience;

[0090] The accuracy of hot zone click linkage has been significantly improved, enhancing the user interaction experience: The hot zone generation adopts a spatial density + device clustering logic, combined with the anchor annotation in the text summary to achieve graphic-text linkage; in the test of 8000 real notifications, the corresponding rate of user clicks on the hot zone and the correct graphic anchor has increased from 76.1% to 91.4%, and the misclick rate has decreased by more than 50%, which can significantly improve the user's understanding of graphic elements and the accuracy of operation feedback;

[0091] The layer adaptation mechanism enhances compatibility and adapts to the differentiated display of multiple terminals: it has the ability to dynamically adjust the layer transparency and map complexity, and can automatically control the number of layers and content granularity according to the terminal performance information in the user portrait; in the three types of terminal adaptation tests, the system's average encapsulation success rate increased from 82.7% to 97.2%, significantly improving the rendering success rate of 5G rich media messages in heterogeneous terminals, which is particularly suitable for State Grid's tiered user group coverage strategy.

[0092] Furthermore, in step S5, a method for differentiated and accurate delivery of 5G messages based on user power grid profiles is used, including:

[0093] S61: Based on the grid master data and historical interaction data, a grid profile of each user is constructed, including the nature of electricity consumption, sensitivity level, historical behavior pattern, terminal capabilities, and location information. The above attributes are integrated into a vectorized representation to form a user profile vector. , as a comprehensive characterization of users’ receiving capabilities and needs;

[0094] S62: Establish a content level system for power notifications, classifying messages into multiple levels: L1 summary reminder, L2 text and image fusion, and L3 text and image + voice. Each level corresponds to a different information granularity and modal combination structure. Preset a structural template for each level, including whether to enable a graphic area, whether to include a recovery countdown, whether to add voice assistance, and whether to enable text and image anchor linkage content elements.

[0095] S63: Build a hierarchical profile matching rule library , each rule Expressed in the form of a conditional tuple, ,in, For user types, such as residential, commercial, and medical institutions, enumeration codes are used; The user importance level is 0 for low, 1 for medium, and 2 for high; Terminal capability level, scored based on device performance, with low-end being 0, medium-end being 1, and high-end being 2; User historical interaction activity, quantified based on the user's message click-through rate over the past 30 days, with 1 being active and 0 being inactive; The emergency level of the event, with values ​​ranging from 0 for normal, 1 for important, and 2 for urgent; Message level templates can be L1 plain text, L2 text and image, or L3 text and image + voice. The matching process is as follows:

[0096] Conditional judgment, the system generates a target user portrait vector and the urgency of the current incident Perform field-level matching;

[0097] Rule filtering, traverse the rule base, and filter all conditions that meet , , , , A set of rules equal to the corresponding rule value ;

[0098] Prioritize the rule set By message level priority Sorting: the higher the message level, the higher the priority;

[0099] Fuzzy matching, some fields in the rule conditions use wildcards;

[0100] Finally, select the message level corresponding to the rule with the highest ranking As the message push level of the user's current event;

[0101] Dynamic adjustment, combined with user click feedback and message sending results, adjustment 、 Weight, real-time update of rule matching priority, and optimization of matching results;

[0102] S64: After the event summary, graphic area, and voice content are generated, the system trims and packages the rich media content based on the matching message level, including selecting an appropriate summary compression level, filtering graphic hot zone layers, adding and omitting voice content, and preserving the original anchor binding relationship;

[0103] S65: Before sending a message to the 5G core network, the distribution control strategy is set according to the user profile, including distribution time window control, priority control, and frequency suppression. The push strategy can be dynamically adjusted, and a lighter message type can be switched if the user has not clicked on the message for a long time.

[0104] A differentiated and precise 5G message delivery method based on user power grid profiles accurately profiles users by integrating the following five dimensions:

[0105] Nature of electricity consumption: residential, industrial and commercial, special industries (such as hospitals and schools);

[0106] Business sensitivity level: whether the customer is a key security client (such as a government agency or a key project);

[0107] Terminal capability model: determines whether the device supports rich media such as graphics, text, and voice.

[0108] User behavior preferences: Dynamic modeling based on historical click-through rates, reading time, and content type preferences;

[0109] Geographic and substation access information: used to identify whether events are highly relevant.

[0110] The core mechanism of a differentiated and precise 5G message delivery method based on user power grid profiles is as follows:

[0111] Use quintuples to express rule conditions, combining user profiles and event types;

[0112] Each rule is associated with an RCS message template (L1, L2, L3). The system automatically matches the optimal template based on the event level and user profile.

[0113] Supports wildcard rules to adapt to situations where user portraits are incomplete or features are ambiguous;

[0114] Dynamic feedback mechanism: Iteratively optimize rule priorities and content structure based on users' actual click behavior and response results.

[0115] A differentiated and precise 5G message delivery method based on user power grid profiles introduces rich media structure clipping and format adaptation strategies before the message is finally delivered:

[0116] Light users receive the plain text version (L1);

[0117] Graphics support users receive a fusion version of graphics and text (L2);

[0118] A voice-assisted version (L3) is provided for elderly or special users;

[0119] Each version maintains a unified semantic anchor structure to ensure that different terminals are equally interactive and understandable.

[0120] Beneficial effects of the invention: The present invention proposes a system and method for converting power notification text into 5G messages, which has high scalability, intelligence level and adaptability, and is significantly better than the existing power grid notification methods based on traditional text messages, static graphics or manual templates; First, through the power event semantic segmentation and multi-channel parsing algorithm, the present invention introduces a multi-dimensional semantic fragment recognition strategy and a microservice-driven parallel processing mechanism in the power business text parsing for the first time, which effectively responds to the challenges of heterogeneous power grid event data sources, non-standard language expression, and more redundant information; Through lexical-syntactic joint analysis, self-training dependency model and rule-based window merging strategy, the system can accurately divide and identify Four key semantic fragments, including devices, events, locations and time; then, an independently scheduled multi-channel parsing service is used to perform parallel entity recognition, standardization and structure generation on the fragments, which greatly shortens the parsing delay; secondly, through a dynamic summary generation method based on the event semantic element weight matrix and user portrait driving, the present invention breaks through the limitations of traditional fixed templates, redundant or over-simplified summary content, and dynamically screens the event semantic elements most relevant to the user for combined expression based on the semantic vector-portrait weight joint driving mechanism. The core of the method is to construct a weight matrix to quantify the degree of attention of different user groups to various semantic elements, and realize the personalized sorting of content generation order and focus, which not only retains the relevant Key information, and avoid information overload, effectively improve the acceptance and understanding efficiency of notifications; in terms of graphic visualization presentation, the power GIS graphic generation method based on event-region coupling and layer dynamic clipping proposed in this invention introduces a three-factor coupling calculation mechanism of spatial distance, topological association and user density, and combines the minimum coverage strategy of the clipping area, multi-layer overlay and transparency control to achieve refined clipping, regional highlighting and multimodal fusion of notification graphics. Compared with traditional static map display, this method generates more focused graphic areas, prominent hot events and clear boundaries, which greatly enhances the readability and interactivity of graphic information; combined with text anchor binding technology, it also realizes the ability to click on summary text The linkage function of automatic positioning of characters in graphic areas improves the interactive experience and spatial perception ability; the present invention effectively optimizes the matching between rich media notification content and user capabilities and preferences through a differentiated and precise 5G message delivery method based on the user's power grid profile; the system integrates user historical behavior through power master data to construct a five-dimensional user profile vector, dynamically selects the optimal RCS message hierarchy structure based on the urgency of the event and the terminal capability, and realizes automatic decision-making and version tailoring through rule base matching. While ensuring the integrity of the core elements of the message, personalized delivery is achieved, and the integrity of the message reception and interaction logic is maintained even in network fluctuation scenarios, greatly ensuring the information reach rate and service continuity;This invention efficiently extracts key information from complex, unstructured power event text data, generates structured, highly semantically relevant, and text-linked 5G rich media messages, and adapts them to end-user terminals on demand, significantly improving the efficiency, accuracy, and user experience of power information transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0121] Figure 1 A system module diagram for converting power notification text to 5G messages;

[0122] Figure 2 This is a flow chart of a method for converting power notification text into 5G messages. DETAILED DESCRIPTION

[0123] The present invention is further clearly and completely described below, but the protection scope of the present invention is not limited thereto.

[0124] A system for converting power notification text to 5G messages, comprising:

[0125] Event parsing module: Receives raw power event data from the power PMS and DMS systems, pre-processes the power event data, uses rule parsing and key field extraction algorithms, and outputs a standardized structure;

[0126] Summary generation module: Receives structured event data and models it into a multi-dimensional semantic graph structure based on the power grid event graph model. It then calls the semantic summary generation model to generate a popular text summary based on the event type and target user profile.

[0127] Graphics generation module: maps the geographic coordinates and device numbers in the event structure to the GIS platform, obtains the corresponding area layer, renders the event area, event location, and impact range boundary, and highlights the event area as a hot spot in the map;

[0128] Graphics and text linkage module: anchors the summary text to the graphics, uses field mapping strategies to map the text summary to the graphic hot zone ID, and binds the time to the estimated recovery time countdown timer;

[0129] Message encapsulation module: encapsulates text summaries and graphic content into 5G rich media message format, selects the appropriate message version based on the user terminal capabilities, and finally sends the message to the user terminal through the operator's 5G core network.

[0130] like Figure 2 The figure shows a flow chart of a method for converting power notification text into 5G messages.

[0131] A method for converting power notification text into 5G messages is based on the above-mentioned system for converting power notification text into 5G messages, and the technical solution adopted is as follows:

[0132] S1: Receives raw power event data from the power PMS and DMS systems, preprocesses the event data, including field cleaning, device mapping, and time normalization, and converts the raw event data into structured event information using power event semantic segmentation and multi-channel parsing algorithms.

[0133] S2: Based on a preset power grid event graph model, semantic modeling is performed on the structured event information to construct a multi-dimensional semantic graph structure. The node information in the semantic graph includes event type, involved equipment, geographic location, expected recovery time, and affected area. Subsequently, the semantic graph vector is input into the semantic summary generation model. Combined with the event type and target user profile, a popular notification summary text expressed in natural language is generated.

[0134] S3: Extract the device number and geographic coordinate data from the event information body, connect to the GIS platform through the interface, locate the geographical area involved in the event on the map layer, generate the corresponding graphic representation, and highlight the hot zone of the relevant area according to the event scope to form a visual image of the event area;

[0135] S4: Mark the generated summary text with field anchors and establish mapping relationships with graphic elements. The mapping relationships include binding the place name and equipment name in the summary to the GIS hot zone ID, and binding the estimated recovery time field to the countdown module, to implement interactive logic that highlights the graphic when the user clicks the text anchor.

[0136] S5: Encapsulate the generated text summary and graphic content, generate an RCS message structure according to the 5G rich media communication standard format, and dynamically select the appropriate message version based on the user terminal capabilities and network capabilities, including plain text, graphic and text fusion, and voice-assisted versions. Finally, the encapsulated message is transmitted to the target user terminal through the 5G core network.

[0137] Furthermore, in step S1, the power event semantic segmentation and multi-channel parsing algorithm is used to convert the original event data into a structured event information body, including:

[0138] S31: Receives raw text data from the PMS and DMS systems. The raw text data includes semi-structured and free expression formats, performs character and time unit unification, punctuation completion, word error correction, and encoding compatibility processing. The processed text is marked with a unique event ID and encapsulated into an event raw object EventRawObject, completing basic cleaning and standardization of the raw input power grid event description text.

[0139] S32: Perform lexical segmentation and word segmentation on the preprocessed text. A self-trained DependencyParser-BiLSTM model is introduced to construct a syntactic dependency graph for the text. Combining a predefined event semantic dictionary with grammatical rule templates, it identifies candidate segments of event-centered words, device subject phrases, location complement phrases, and time adverb phrases in the sentence. Subsequently, a rule-weighted window sliding strategy is used to merge and disambiguate redundant and nested segments. Ultimately, the text is decomposed into four semantic segments: device segments, event segments, location segments, and time segments. Each segment is encapsulated into a standardized structure called SemanticChunk, annotated with its semantic category and character position information.

[0140] S33: Based on the semantic type field in the category label SemanticChunk of the semantic fragment, each semantic fragment is dynamically routed to the corresponding dedicated parsing channel for parallel processing. The scheduling of the parsing channel adopts an event-driven microservice architecture. Each type of semantic fragment is internally configured with an independent processing module, which is registered as a device identification service, a location extraction service, a time parsing service, and an event classification service in the form of a REST API. The structured semantic fragment is encapsulated in a standard JSON format through the internal message queue and attached with an event context ID, and then delivered to the message bus. Each parsing service module listens to the corresponding channel. After receiving the message, it automatically schedules the loading of the corresponding sub-model to complete the entity extraction, classification, and standardization processing of the semantic fragment. The scheduling center uses a thread pool + asynchronous callback mechanism to control the processing status of each channel and writes the results back to the event semantic cache pool EventSemanticCache for subsequent fusion module reading and calling;

[0141] S34: A named entity recognition model based on the BERT-CRF architecture is used to extract key information from the device fragment, including substation, line number, voltage level, and feeder name. The model takes the syntactic context as input and outputs the annotated device entity category and text location. The parsed results are then fed into the device hierarchy completion module, which automatically infers and fills in missing fields based on the trained device coding rules and dictionary system. The final output is a device information structure, DeviceStruct, which contains the device ID, voltage level, substation, type, and a confidence score.

[0142] S35: Use the BiGRU-CRF geographic entity recognition model, which combines word vector representation with prior knowledge of place names, to identify residential areas, landmarks, and community names in location segments, output the geographic entity names and their locations in the original text, and map the identified place names to standardized coordinate pairs using a place name-coordinate mapping table. For fuzzy matching place names, weighted similarity using geographic distance is used for fuzzy error correction to ensure positioning accuracy, and output a multi-location parallel structure, LocationStruct.

[0143] S36: Uses a regular expression-driven semantic pattern tree matching method to perform pattern recognition and type classification on time phrases. Then, based on the current system time context, it converts natural language time expressions into the ISO 8601 time format. The output time structure includes event start time, recovery time, and estimated duration fields.

[0144] S37: Perform intent recognition on the content of the event fragment to determine whether it belongs to the type of predefined event types such as tripping, planned maintenance, sudden power outage, and load abnormality. The model input is the text representation vector of the event fragment, and the output is the event type category and its probability distribution. The system presets the confidence threshold to 0.75 by default. Samples with a confidence greater than or equal to 0.75 are high-confidence samples, and samples with a confidence less than 0.75 are low-confidence samples. According to the set threshold, the event type with a confidence higher than a specific threshold is output. The low-confidence samples are sent to the manual review module. The result will be used as the upstream input for generating the summary template and the selection of the graphic template;

[0145] S38: After each channel completes the parsing, the system aggregates the parsing results of each SemanticChunk to generate a complete EventStruct object. For field overlaps and entity conflicts, the system uses a weighted voting mechanism to resolve conflicts. The weighted items include combined confidence, fragment position, and context consistency. For missing fields, the graph completion logic is called to finally generate a power event structure with complete structure and standardized fields for call by the summary generation and graphic generation modules.

[0146] Natural language time expressions are uniformly converted to ISO 8601 time format using the self-developed T-TimeParser time standardization engine. The specific implementation method is as follows:

[0147] Extracting candidate time phrases: First, the input power notification text is segmented, breaking the entire sentence into several consecutive word segments. Then, all segments between two and six words are matched sequentially against a predefined time expression template. The template is a set of rules based on common Chinese expressions. Whenever a segment meets the template rules, it is marked as a candidate time phrase, and its position range in the original text and its expression type are recorded for subsequent processing.

[0148] Context-anchored parsing: Next, the candidate time phrase identified in the previous step is combined with the current system time to determine the specific calendar time it represents. If the phrase is ambiguous, it is mapped to a specific date and time value using the rule anchor dictionary. If the current system time is later than the time point, it is inferred to be the same time the next day. If the phrase expresses a relative time, the current time is added to the corresponding time interval to obtain a new specific time point.

[0149] Temporal semantic role recognition: Determine the temporal meaning of each temporal phrase based on its contextual position in the sentence and the accompanying verb, such as whether it represents the "start time of an event," "end time of an event," or "duration of an event."

[0150] Semantic completion and reasoning: For incomplete time expressions in power text, we use other contextual information to supplement the specific meaning of these expressions. This supplementary information includes common time patterns in the user's location and the presence of logical relationships between multiple time words in the text. Ultimately, the fuzzy expressions are completed into clearly structured time values.

[0151] Time standardization conversion: After completing the above analysis, each parsed time phrase is converted to a standard time format, which is the international standard ISO 8601. If a phrase represents a time period, the system will generate two time values, one for the start and one for the end of the period.

[0152] Fusion of structured output and confidence weighting: Finally, the system encapsulates the original content of each time phrase, the parsed standard time, the time role it represents, and a credibility score into a structured data entry. The credibility score reflects the reliability of the system's time parsing results. The scoring reference factors include: whether the time expression is standard, whether the context is clear, and the accuracy of this type of time expression in historical data. For time phrases with clear expression and standard structure, the system score will be higher, while for expressions with ambiguous semantics and unclear context, the score will be relatively low.

[0153] The power event semantic segmentation and multi-channel parsing algorithm aims to accurately identify the core semantic fragments of events, including equipment, location, time, event type, etc., from the original power event text data obtained from the power dispatching system PMS and DMS, and perform entity recognition and structuring of various fragments through a parallel multi-channel processing architecture, and finally construct a standard structure that can be used for subsequent summary generation and graphics rendering. It has high scalability and domain adaptability, and can effectively deal with problems such as non-standard text, loose structure, and polysemy overlap. It is a basic capability module for building a power notification intelligent conversion system.

[0154] The specific operation steps of the power event semantic segmentation and multi-channel parsing algorithm are as follows:

[0155] Step 1: Text preprocessing: perform character normalization and semantic cleaning on the input text. The main processing methods include encoding standardization, Chinese punctuation restoration and sentence processing, and time unit unification. Example function: def normalize_text(raw_text);

[0156] Step 2: Semantic fragment segmentation, using dependency syntax graph + rule dictionary for semantic fragment identification; the model used is self-trained BiLSTM + DependencyParser, and the output types include device fragment, location fragment, event fragment, and time fragment. The structure of each fragment is SemanticChunk = {text, span, type}

[0157] Step 3: Multi-channel parallel parsing routes different types of fragments to different parsing channels for dedicated structured processing. The device channel includes entity recognition, voltage level completion, and encoding; the location channel includes place name recognition, coordinate mapping, and defuzzification; the time channel includes time word recognition and ISO standardization; and the event channel includes event intent recognition. Each channel is executed concurrently through a message queue and thread pool, and the results are written back to the cache after parsing is completed.

[0158] Step 4: Conflict resolution and result fusion: merge, disambiguate, and complete missing fields on the outputs from each channel.

[0159] The power event semantic segmentation and multi-channel parsing algorithm adopts error tolerance and belief propagation mechanism.

[0160] To prevent individual channel parsing failures from affecting the overall results, the system sets an information completeness threshold in the semantic cache pool. When certain fields are missing or conflicting, the semantic completion module is automatically triggered, including:

[0161] 1. Event graph completion;

[0162] 2. Default inference based on contextual time or place;

[0163] 3. Channel confidence propagation strategy: When multiple channels return conflicting entity results, the confidence propagation formula is used to determine the final result: ,in, For all candidate entities, is the set of participating channels, is the channel weight, which is adaptively updated based on the channel’s historical accuracy. For channel The confidence level of the output.

[0164] Furthermore, each semantic parsing channel is registered with the system's event processing and dispatching center using a RESTful API service module. The interface uniformly adopts the following standard input and output formats:

[0165] Input structure: {chunk_id, chunk_type, raw_text, context_id};

[0166] Output structure: {chunk_id, semantic_type, extracted_entity, normalized_result, confidence_score};

[0167] This interface specification supports the independent evolution of different parsing modules while ensuring interoperability, and supports the addition of new channels such as sentiment analysis channel, semantic confidence filtering channel, etc., to achieve horizontal scalability of the system.

[0168] The power event semantic segmentation and multi-channel parsing algorithm relies on Python 3.9+ as the main language environment, Jieba as the Chinese word segmentation and syntax analysis library, PyTorch as the deep learning model inference framework, RabbitMQ as the message queue for inter-module parallelism and scheduling, Elasticsearch as the location fuzzy query, and FastAPI as the service interface framework for each channel.

[0169] Explicit segmentation of semantic structures improves parsing accuracy: This algorithm uses structured semantic fragment recognition technology based on a combination of syntactic dependencies and semantic rules to accurately decompose raw event text into four semantic blocks: "device, time, location, and event." This significantly reduces semantic ambiguity and contextual confusion, making it particularly suitable for processing text with complex sentence structures and overlapping entities in power business scenarios.

[0170] A multi-channel parallel parsing architecture significantly improves processing efficiency: Different types of semantic fragments are routed to dedicated parsing channels, enabling asynchronous concurrent processing within a microservices framework. The system boasts high throughput and low latency, making it particularly well-suited for the batch structuring of large-scale, real-time power event text.

[0171] Fusion of semantic rules and deep learning models, balancing generalization and precision: Device, location, and time channels are each combined with specialized models such as BERT-CRF, BiGRU, and regularized semantic tree matching. This fusion of business rules and model outputs, combined with a confidence voting mechanism for disambiguation, enables highly robust entity extraction and field standardization, with excellent performance for non-standard expressions and colloquial text.

[0172] Supports subsequent module linkage and complete semantic chain closed loop: the output structured result EventStruct retains all fragment location information, confidence, and semantic tags, supports direct reference of modules such as summary generation, graphics rendering, and rich media encapsulation, forming a semantic closed loop consistent with the front-end and back-end, effectively improving the consistency of graphics and text, interactive linkage, and message personalization.

[0173] Furthermore, in step S2, a summary generation model is used. The summary generation model uses a dynamic summary generation method based on an event semantic element weight matrix and user profile drive, including:

[0174] S41: Extracting a preset semantic element set from the structured power event data, the system first represents the structured event object as ,in is the event type, is the starting time, For recovery time, For the affected area, The name of the main device, As the responsible unit, each field is mapped into a semantic element vector , as the basic semantic representation input before summary generation;

[0175] S42: Use the target user's identity attributes, device capabilities, and behavioral preferences to build a user portrait vector ,in Whether you are an enterprise user, For elderly users, Whether there is a graphics terminal, Pay attention to the weight of time information in past clicks, The attention weight of the location information in the past clicks is finally obtained as a set of standardized user vectors ;

[0176] S43: Construct a weight matrix ,in Semantic elements In user portrait features The initial value of the matrix is ​​set by business experience, and is subsequently fine-tuned through feedback data. , calculate its weight sorting vector , the formula is ,in, Semantic elements The user's summary contribution score is sorted to obtain the semantic generation order;

[0177] S44: Select the top k semantic elements according to the descending sorting result of S and fill them into the prompt template Prompt. For missing fields and low weight fields, they will not be filled into Prompt. Prompt is encoded as an input vector and together with the semantic element content, it constitutes the final language model input structure LLMInput={prompt, event-emantics};

[0178] S45: The Prompt template and the event semantic vector are fed into a summary generation model. The model is a fine-tuned Transformer architecture that uses formatted methods as input and ultimately outputs a natural language summary.

[0179] S46: Finally, the system evaluates the quality of the generated summary text and uses it to dynamically update the weight matrix ,Quality evaluation indicators include semantic consistency score, click feedback index, and error ,correction feedback rate.

[0180] The generation model uses a dynamic summary generation method based on the event semantic element weight matrix and user profile drive. Its core functions include:

[0181] Achieve semantic element-level summary content customization: This method uses multiple semantic fields (such as event type, time, location, and equipment) in structured power event data as a basic semantic element set. Based on user profiles, it dynamically filters the information that different users are most concerned about and generates personalized summary text, improving the relevance and readability of the summary.

[0182] A user-profile-driven mechanism is introduced to optimize summary content: Based on user profile attributes such as terminal type, age level, and business preferences, the system automatically adjusts the presentation order of semantic elements in the summary, retains fields, and uses language style to make the message content more aligned with the recipient's cognitive habits and lower the threshold for information comprehension.

[0183] Dynamic calculation of semantic weights and summary generation control: By constructing a weight matrix of "semantic elements × user profile features," the importance of different semantic elements to different types of users is calculated, and the content priority and compression level during summary generation are determined. This allows for dynamic clipping of summary content and adaptive generation templates.

[0184] Supports continuous learning and optimization of summary generation quality: Utilizes multi-dimensional indicators such as user click feedback, error correction rate, and semantic matching to iteratively optimize the semantic weight matrix, achieve self-adjustment of semantic sorting weights and enhance corpus adaptability, and form a closed-loop learning mechanism.

[0185] The dynamic summary generation method based on the event semantic element weight matrix and user portrait driving includes the following steps:

[0186] Event semantic element vectorization: multiple core field event types in structured power event data , start time , Estimated recovery time of the incident ,Place , device name , responsible unit Mapped into a set of semantic elements , and construct a semantic representation vector for the initial semantic basis of summary construction;

[0187] User portrait vector construction: Extract the user's static attributes including whether they are corporate users , whether elderly users , terminal capability graphics support , focus on time weight , focus on location weight , forming a standardized user portrait vector ;

[0188] Semantic element weight matrix calculation: Constructing the semantic weight matrix , representing semantic elements In user portrait features The importance of the matrix is ​​calculated by matrix multiplication to calculate the comprehensive semantic score vector , sort the semantic score vectors in descending order to obtain the output priority of the semantic elements;

[0189] Prompt summary template construction: According to the sorted score vector, intercept the front The high-weight semantic elements are selected to construct a summary and generate a prompt template. Unselected low-weight fields will be ignored. The generated prompt and semantic content together constitute the language model input structure.

[0190] Summary generation and natural language output: The prompt template and semantic content are fed into the fine-tuned Transformer summary generation model, which outputs a popular, user-customized natural language notification summary text;

[0191] Summary evaluation and weight matrix fine-tuning: Use semantic consistency scoring, user click behavior feedback, error correction rate statistics and other indicators to comprehensively evaluate the quality of summary generation, and then adjust the weight matrix Fine-tune and optimize the corresponding parameters to form a continuous learning closed loop.

[0192] The accuracy of summary content is significantly improved, effectively reducing redundancy and ambiguity: This method vectorizes multiple fields in structured power event data (such as event type, time, location, equipment, etc.) and inputs them into a weight matrix model. It dynamically assigns semantic weights based on user profiles, thereby screening the most noteworthy summary elements for presentation. Experimental results show that, based on comparisons with traditional fixed template summaries, the summaries generated by this method improve the ROUGE-2 index (an accuracy index measuring semantic consistency) from 0.67 to 0.81, and the entity hit rate increases from 74.2% to 89.6%, significantly improving the semantic accuracy and expression clarity of the summaries.

[0193] Improved user click-through rate and interactive participation rate: The summary content is more closely aligned with the reading habits and information needs of different user groups, significantly increasing users' willingness to actively respond to power notifications. In an A / B test of 800,000 real power notification messages, the summary version generated by this method saw its click-through rate increase from 12.3% to 17.9% among residential users and from 9.8% to 14.1% among corporate high-voltage users, with an average click-through rate increase of over 45%, effectively improving notification communication effectiveness and interactive response rate.

[0194] Effectively compresses summary length to meet the needs of 5G rich media packaging: During RCS rich media message packaging, content length and compression ratio directly impact the packaging success rate and display aesthetics. This method uses a weight-driven summary control mechanism to precisely trim redundant information and achieve condensed information expression. Test data shows that the average length of generated summaries is compressed from 112 words (using traditional methods) to 82 words, achieving a compression ratio of 26.8%, while retaining at least 95% information coverage, perfectly adapting to the constraints of mixed text and image messages.

[0195] Improved summary generation efficiency: Since the summary generation process only constructs the input structure based on a small number of high-weight semantic fields, it can significantly reduce the language model inference burden and improve real-time response capabilities. Tests in server-side deployment show that the generation time of each summary is shortened from 147 milliseconds for traditional structure input to 93 milliseconds, and the inference delay is reduced by 36.7%. It is particularly suitable for high-concurrency event push scenarios and edge device deployment scenarios, and has good system scalability and stability.

[0196] Furthermore, in step S3, a method for generating electric GIS graphics based on event-region coupling and dynamic layer clipping is used, including:

[0197] S51: The system first extracts all fields with geographic attributes from the structured event data, including the list of affected cells, the main device number, the substation to which the device belongs, and the voltage level. The system then searches the device geographic mapping table containing the list of affected cells, the main device number, the substation to which the device belongs, and the voltage level to generate the event impact area point set P and the device set D for subsequent spatial calculations.

[0198] S52: Using event-region coupling function The coupling function combines three factors: spatial distance effect, the straight-line distance between the event center and the regional center; topological correlation, the spatial intersection ratio between the main device service area and the user location; user density weight, the number of registered users per unit area on the GIS grid map; the coupling function is defined as ,in, is the center point of the event, Candidate area The center point, For equipment service area, For the region The boundary polygon of is a weight parameter, which is adjusted to adapt to user differences. The dist function represents the spatial straight-line distance between the event center and the area center, normalized to , calculated as ,in are the latitude and longitude coordinates of the event center point, is the latitude and longitude of the center point of the region, The maximum visual range distance set by the platform is used for normalization. The closer the distance, the closer the dist function value is to 1, and the higher the coupling degree. The farther the dist function distance is, the closer the value is to 0. The IOU function is the spatial intersection degree between the service range of the event master device and the candidate area, which measures the direct geographical coverage relationship. The specific calculation method is: ,in is the polygonal area formed by the service range of the event master device. is the polygon corresponding to the candidate region r, Area is the area calculation function in the map system, and the IOU function value is in the range of [0,1]. The closer it is to 1, the more completely the region is covered by the event, indicating a high degree of coupling. The function is used to calculate the number of registered electricity users per unit area. The specific calculation method is: ,in is the total number of registered users in region r, obtained from the main grid data, is the area of ​​region r, The maximum user density per unit area in the entire network area is used for normalization and is selected after the final calculation. The area before the score 𝑁 enters the candidate set of graph cutting;

[0199] S53: Based on the set of regions with the highest coupling degree, the system constructs a minimum coverage rectangle and defines a cropping boundary area 𝐵. This area includes all hotspots and controls the total map size to not exceed the set image resolution. The geographic base layer, power-specific layer, and user interest layer are loaded inside the cropping area. The layer overlay adopts a hierarchical transparency + raster priority mechanism.

[0200] S54: Within the cropped area, the system analyzes all device point sets D using a spatial clustering algorithm to generate multiple event hotspots H. The system then adds attributes to each hotspot, including the hotspot ID, event level, and clickable range. Finally, each hotspot is bound to the corresponding field in the text summary, forming a complete graphic-text linkage model.

[0201] S55: Encode the rendered graphic cropping area into a format supported by RCS, automatically select a version based on the user terminal capabilities, and finally encapsulate it into a 5G message data packet and send it together with the text summary and countdown linkage information.

[0202] The event-region coupling degree scoring in the power GIS graphics generation method based on event-region coupling degree and dynamic layer clipping not only adopts a static weight combination including spatial distance, topological association, and user density, but also introduces a dynamic weight adjustment mechanism driven by event type; the system automatically adjusts the coupling degree function according to the business type of the event, such as planned maintenance, fault tripping, and load warning. Weight parameters; for example, in the event of a sudden failure, the system will increase the spatial distance weight to quickly mark nearby areas, while in planned maintenance notifications, it will place more emphasis on topological coverage and affected user density, making the graphical display more consistent with business characteristics.

[0203] To make graphic information richer and less crowded, a layer-level configurable overlay mechanism was designed based on the power GIS graphic generation method of event-region coupling and dynamic layer clipping. The layers are divided into: basic geographic layers, power-specific layers, and user interest layers. The system assigns different transparency and priority to each layer, controls the display effect through layer synthesis logic Z-index sorting + inter-layer filtering, and adapts the layer complexity according to the user terminal capabilities. For example, when the terminal performance is low, the system only retains the main hot zone and equipment layers and ignores non-essential interest layers to ensure rendering performance.

[0204] Hotspot is not just a simple geographical area selection, but also needs to have semantic perception capabilities; using density-spatial weighted clustering algorithm, in the geographical point set Analyze and form several hot blocks Each hot zone is attached with attributes such as a unique ID, event level, and estimated recovery time; the ID of each hot zone is then bound to the anchor field in the summary text through the graphic-text linkage module to achieve highlighter linkage for graphic-text clicks; when generating the anchor, the system automatically extracts content such as place names and device names based on the entity recognition position field and establishes a mapping with the hot zone to ensure that the graphic area can be accurately highlighted and located when the user clicks on the text.

[0205] To meet the standardized requirements of RCS rich media message encapsulation, a structured graphic encapsulation model is used, which includes image content, hot zone definition JSON, layer configuration description, and anchor binding data. An example of the encapsulation format design is {"image": "base64 image or URL", "hotzones": {"id": "HZ001", "name": "Dongfang Community", "polygon": [...], "type": "residential area", "bind": "text segment #3"}, "layers": ["map background", "device layer", "user interest layer"], "interaction": {"anchor_map": {"Dongfang Community": "HZ001", "Fuxing Station": "HZ004"}}}.

[0206] Significantly improve the matching degree between the graphic content and the areas of user concern: by introducing a coupling function to fuse spatial distance, device topological relationship, and user density factors, the system can accurately identify the areas that users are most concerned about for graphic rendering; compared with the traditional static cropping method of "main device location point + radius expansion", the coverage rate of the target user concern area in the generated graphics has increased from 72.8% to 93.5%, significantly enhancing the fit between the graphic-text content and the actual needs, and effectively avoiding irrelevant redundancy of graphic information.

[0207] Effectively control the graphic cropping range and reduce the terminal transmission and rendering burden: based on the minimum covering rectangle construction strategy, only the area layers strongly associated with the event are retained, which can effectively compress the graphic size and complexity; in the test environment, the amount of graphic data rendered under the same event has been reduced from an average of 1.4MB to 0.87MB, a decrease of about 38%. On mid- to low-end user terminals, the image loading time has dropped from 2.6 seconds to 1.5 seconds, improving the rendering efficiency and user experience.

[0208] The accuracy of hot zone click linkage has been significantly improved, enhancing the user interaction experience: the hot zone generation adopts the spatial density + device clustering logic, combined with the anchor annotation in the text summary, to achieve graphic-text linkage; in the test of 8000 real notifications, the correspondence rate between the user's click on the hot zone and the correct graphic-text anchor has increased from 76.1% to 91.4%, and the misclick rate has decreased by more than 50%, which can significantly improve the user's understanding of graphic elements and the accuracy of operation feedback.

[0209] The layer adaptation mechanism enhances compatibility and adapts to the differentiated display of multiple terminals: it has the ability to dynamically adjust the layer transparency and map complexity, and can automatically control the number of layers and content granularity according to the terminal performance information in the user portrait; in the three types of terminal adaptation tests, the system's average encapsulation success rate increased from 82.7% to 97.2%, significantly improving the rendering success rate of 5G rich media messages in heterogeneous terminals, which is particularly suitable for State Grid's tiered user group coverage strategy.

[0210] Furthermore, in step S5, a method for differentiated and accurate delivery of 5G messages based on user power grid profiles is used, including:

[0211] S61: Based on the grid master data and historical interaction data, a grid profile of each user is constructed, including the nature of electricity consumption, sensitivity level, historical behavior pattern, terminal capabilities, and location information. The above attributes are integrated into a vectorized representation to form a user profile vector. , as a comprehensive characterization of users’ receiving capabilities and needs;

[0212] S62: Establish a content level system for power notifications, classifying messages into multiple levels: L1 summary reminder, L2 text and image fusion, and L3 text and image + voice. Each level corresponds to a different information granularity and modal combination structure. Preset a structural template for each level, including whether to enable a graphic area, whether to include a recovery countdown, whether to add voice assistance, and whether to enable text and image anchor linkage content elements.

[0213] S63: Build a hierarchical profile matching rule library , each rule Expressed in the form of a conditional tuple, ,in, For user types, such as residential, commercial, and medical institutions, enumeration codes are used; The user importance level is 0 for low, 1 for medium, and 2 for high; Terminal capability level, scored based on device performance, with low-end being 0, medium-end being 1, and high-end being 2; User historical interaction activity, quantified based on the user's message click-through rate over the past 30 days, with 1 being active and 0 being inactive; The emergency level of the event, with values ​​ranging from 0 for normal, 1 for important, and 2 for urgent; Message level templates can be L1 plain text, L2 text and image, or L3 text and image + voice. The matching process is as follows:

[0214] Conditional judgment, the system generates a target user portrait vector and the urgency of the current incident Perform field-level matching;

[0215] Rule filtering, traverse the rule base, and filter all conditions that meet , , , , A set of rules equal to the corresponding rule value ;

[0216] Prioritize the rule set By message level priority Sorting: the higher the message level, the higher the priority;

[0217] Fuzzy matching, some fields in the rule conditions use wildcards;

[0218] Finally, select the message level corresponding to the rule with the highest ranking As the message push level of the user's current event;

[0219] Dynamic adjustment, combined with user click feedback and message sending results, adjustment 、 Weight, real-time update of rule matching priority, and optimization of matching results;

[0220] S64: After the event summary, graphic area, and voice content are generated, the system trims and packages the rich media content based on the matching message level, including selecting an appropriate summary compression level, filtering graphic hot zone layers, adding and omitting voice content, and preserving the original anchor binding relationship;

[0221] S65: Before sending a message to the 5G core network, the distribution control strategy is set according to the user profile, including distribution time window control, priority control, and frequency suppression. The push strategy can be dynamically adjusted, and a lighter message type can be switched if the user has not clicked on the message for a long time.

[0222] A differentiated and precise 5G message delivery method based on user power grid profiles accurately profiles users by integrating the following five dimensions:

[0223] Nature of electricity consumption: residential, industrial and commercial, special industries (such as hospitals and schools);

[0224] Business sensitivity level: whether the customer is a key security client (such as a government agency or a key project);

[0225] Terminal capability model: determines whether the device supports rich media such as graphics, text, and voice.

[0226] User behavior preferences: Dynamic modeling based on historical click-through rates, reading time, and content type preferences;

[0227] Geographic and substation access information: used to identify whether events are highly relevant.

[0228] The core mechanism of a differentiated and precise 5G message delivery method based on user power grid profiles is as follows:

[0229] Use quintuples to express rule conditions, combining user profiles and event types;

[0230] Each rule is associated with an RCS message template (L1, L2, L3). The system automatically matches the optimal template based on the event level and user profile.

[0231] Supports wildcard rules to adapt to situations where user portraits are incomplete or features are ambiguous;

[0232] Dynamic feedback mechanism: Iteratively optimize rule priorities and content structure based on users' actual click behavior and response results.

[0233] A differentiated and precise 5G message delivery method based on user power grid profiles introduces rich media structure clipping and format adaptation strategies before the message is finally delivered:

[0234] Light users receive the plain text version (L1);

[0235] Graphics support users receive a fusion version of graphics and text (L2);

[0236] A voice-assisted version (L3) is provided for elderly or special users;

[0237] Each version maintains a unified semantic anchor structure to ensure that different terminals are equally interactive and understandable.

[0238] The embodiments disclosed in the present invention are preferred embodiments, but are not limited to them. Ordinary technicians in this field can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. As long as they do not deviate from the spirit of the present invention, they are all within the scope of protection of the present invention.

[0239] The present invention proposes a system and method for converting power notification text into 5G messages, which has high scalability, intelligence and adaptability, and is significantly superior to the existing power grid notification methods based on traditional text messages, static graphics or manual templates. First, through the semantic segmentation and multi-channel parsing algorithm of power events, the present invention introduces a multi-dimensional semantic fragment recognition strategy and a microservice-driven parallel processing mechanism in the power business text parsing for the first time, effectively responding to the challenges of heterogeneous power grid event data sources, non-standard language expression, and more redundant information. Through lexical-syntactic joint analysis, self-training dependency model and rule-based window merging strategy, the system can accurately divide and identify devices, events, and locations. Then, an independently scheduled multi-channel parsing service is used to perform parallel entity recognition, standardization and structure generation on the fragments, which greatly shortens the parsing delay. Secondly, through a dynamic summary generation method based on the event semantic element weight matrix and user portrait driving, the present invention breaks through the limitations of traditional fixed templates, redundant or over-simplified summary content, and dynamically screens the event semantic elements most relevant to the user for combined expression based on the semantic vector-portrait weight joint driving mechanism. The core of the method is to construct a weight matrix to quantify the degree of attention of different user groups to various semantic elements, and realize personalized sorting of content generation order and focus, which not only retains key information but also Avoiding information overload effectively improves the acceptance and comprehension efficiency of notifications; in terms of graphic visualization, the method of generating electric GIS graphics based on event-region coupling and dynamic layer clipping proposed in this invention introduces a three-factor coupling calculation mechanism of spatial distance, topological association and user density, and combines the minimum coverage strategy of the clipping area, multi-layer overlay and transparency control to achieve refined clipping, regional highlighting and multimodal fusion of notification graphics. Compared with traditional static map display, the graphic area generated by this method is more focused, hot events are highlighted, and boundaries are clear, which greatly enhances the readability and interactivity of graphic information; combined with text anchor binding technology, it also realizes automatic display of summary text upon clicking on it. The linkage function of the positioning graphic area improves the interactive experience and spatial perception ability. The present invention effectively optimizes the matching between rich media notification content and user capabilities and preferences through a differentiated and precise 5G message delivery method based on the user's power grid profile. The system integrates user historical behavior through power master data to construct a five-dimensional user profile vector. Combining the urgency of the event and the terminal capabilities, it dynamically selects the optimal RCS message hierarchy structure and realizes automatic decision-making and version tailoring through rule base matching. While ensuring the integrity of the core elements of the message, it achieves personalized delivery and maintains the integrity of message reception and interaction logic even in network fluctuation scenarios, greatly ensuring information reach and service continuity.This invention efficiently extracts key information from complex, unstructured power event text data, generates structured, highly semantically relevant, and text-linked 5G rich media messages, and adapts them to end-user terminals on demand, significantly improving the efficiency, accuracy, and user experience of power information transmission.

Claims

1. A method for converting power notification text to 5G message, characterized in that: include: S1: Receive raw power event data from the power PMS and DMS systems, preprocess the event data, including field cleaning, device mapping, and time normalization operations, and convert the raw event data into structured event information using the power event semantic segmentation and multi-channel parsing algorithm. The specific implementation of the power event semantic segmentation and multi-channel parsing algorithm is as follows: Receives raw text data from PMS and DMS systems, including semi-structured and free-form formats, and performs character and time unit unification, punctuation completion, word error correction, and encoding compatibility processing. The processed text is annotated with a unique event ID and encapsulated into an EventRawObject, completing basic cleaning and standardization of the raw input power grid event description text. The preprocessed text is lexically segmented and word-partitioned. A self-trained DependencyParser-BiLSTM model is introduced to construct a syntactic dependency graph for the text. Combining a predefined event semantic dictionary with grammatical rule templates, candidate segments of event-centered words, device subject phrases, location complement phrases, and time adverb phrases are identified in the sentence. Subsequently, a rule-weighted window sliding strategy is used to merge and disambiguate redundant and nested segments. Ultimately, the text is decomposed into four semantic segments: device segments, event segments, location segments, and time segments. Each segment is encapsulated into a standardized structure called a SemanticChunk, annotated with its semantic category and character position information. Based on the semantic type field in the semantic fragment's category label, SemanticChunk, each semantic fragment is dynamically routed to the corresponding dedicated parsing channel for parallel processing. The parsing channel scheduling adopts an event-driven microservice architecture. Each type of semantic fragment is configured with an independent processing module internally, which is registered as a device identification service, location extraction service, time parsing service, and event classification service in the form of a REST API. The structured semantic fragment is encapsulated in standard JSON format through the internal message queue and attached with an event context ID and delivered to the message bus. Each parsing service module listens to the corresponding channel and, upon receiving the message, automatically schedules the loading of the corresponding sub-model to complete the entity extraction, classification, and standardization of the semantic fragment. The scheduling center uses a thread pool + asynchronous callback mechanism to control the processing status of each channel and writes the results back to the event semantic cache pool EventSemanticCache for subsequent fusion module reading and calling. A named entity recognition model based on the BERT-CRF architecture is used to extract key information from device snippets, including substation, line number, voltage level, and feeder name. The model takes syntactic context as input and outputs the labeled device entity category and text location. The parsed results are then fed into the device hierarchy completion module, which automatically infers and fills in missing fields based on trained device coding rules and a dictionary system. The final output is a device information structure (DeviceStruct) containing the device ID, voltage level, substation, type, and a confidence score. The BiGRU-CRF geographic entity recognition model, which combines word vector representations with prior knowledge of place names, identifies residential areas, landmarks, and neighborhood names in location fragments, outputs the geographic entity names and their locations in the original text, and maps the identified place names to standardized coordinate pairs using a place name-coordinate mapping table. For fuzzy matching place names, weighted similarity using geographic distance is used for fuzzy error correction to ensure positioning accuracy, and outputs a multi-location parallel structure, LocationStruct. A regular expression-driven semantic pattern tree matching method is used to perform pattern recognition and type classification on time phrases. Then, combined with the current system time context, natural language time expressions are uniformly converted to the ISO 8601 time format. The output time structure includes event start time, recovery time, and estimated duration fields. The intent of the event fragment content is recognized to determine whether it belongs to the predefined event type of tripping, planned maintenance, sudden power outage, or load anomaly. The model input is the text representation vector of the event fragment, and the output is the event type category and its probability distribution. The system presets the confidence threshold to 0.75 by default. Samples with a confidence greater than or equal to 0.75 are high-confidence samples, and samples with a confidence less than 0.75 are low-confidence samples. According to the set threshold, the event type with a confidence higher than a specific threshold is output. The low-confidence samples are sent to the manual review module. The result will serve as the upstream input for generating summary templates and selecting graphic templates. After parsing each channel, the system aggregates the parsing results of each SemanticChunk to generate a complete EventStruct object. For field overlaps and entity conflicts, the system uses a weighted voting mechanism to resolve conflicts. The weighting factors include confidence, fragment position, and context consistency. For missing fields, the graph completion logic is invoked to ultimately generate a complete power event structure with standardized fields for use by the summary generation and image and text generation modules. S2: Based on a preset power grid event graph model, semantic modeling is performed on the structured event information to construct a multi-dimensional semantic graph structure. The node information in the semantic graph includes event type, involved equipment, geographic location, expected recovery time, and affected area. Subsequently, the semantic graph vector is input into the semantic summary generation model. Combined with the event type and target user profile, a popular notification summary text expressed in natural language is generated. S3: Extract the device number and geographic coordinate data from the event information body, connect to the GIS platform through the interface, locate the geographical area involved in the event on the map layer, generate the corresponding graphic representation, and highlight the hot zone of the relevant area according to the event scope to form a visual image of the event area; S4: Mark the generated summary text with field anchors and establish mapping relationships with graphic elements. The mapping relationships include binding the place name and equipment name in the summary to the GIS hot zone ID, and binding the estimated recovery time field to the countdown module, to implement interactive logic that highlights the graphic when the user clicks the text anchor. S5: Encapsulate the generated text summary and graphic content, generate an RCS message structure according to the 5G rich media communication standard format, and dynamically select the appropriate message version based on the user terminal capabilities and network capabilities, including plain text, graphic and text fusion, and voice-assisted versions. Finally, the encapsulated message is transmitted to the target user terminal through the 5G core network.

2. The method for converting a power notification text to a 5G message according to claim 1, characterized in that In step S2, a summary generation model is used. The summary generation model uses a dynamic summary generation method based on an event semantic element weight matrix and user profile drive, including: S41: Extracting a preset semantic element set from the structured power event data, the system first represents the structured event object as ,in is the event type, is the starting time, For recovery time, For the affected area, is the name of the main device, As the responsible unit, each field is mapped into a semantic element vector , as the basic semantic representation input before summary generation; S42: Use the target user's identity attributes, device capabilities, and behavioral preferences to build a user portrait vector ,in Whether you are an enterprise user, Is the user elderly? Whether there is a graphics terminal, The attention weight of the time information of past clicks, The attention weight of the location information in the past clicks is finally obtained as a set of standardized user vectors ; S43: Construct a weight matrix ,in Semantic elements The importance weight of the user portrait feature is determined by business experience, and the initial value of the matrix is ​​set by business experience, and then fine-tuned by feedback data. , calculate its weight sorting vector , the formula is ,in, Semantic elements The user's summary contribution score is sorted to obtain the semantic generation order; S44: Select the top k semantic elements according to the descending sorting result of S and fill them into the prompt template Prompt. For missing fields or low weight fields, Prompt is not filled in. Prompt is encoded as an input vector and together with the semantic element content, it constitutes the final language model input structure LLMInput={prompt, event-emantics}; S45: The prompt template and the event semantic vector are fed into a summary generation model. The model is a fine-tuned Transformer architecture that uses formatted information as input and outputs a natural language summary. S46: Finally, the system evaluates the quality of the generated summary text and uses it to dynamically update the weight matrix ,Quality evaluation indicators include semantic consistency score, click feedback index, and error ,correction feedback rate.

3. The method for converting a power notification text into a 5G message according to claim 1, characterized in that In step S3, a method for generating electric power GIS graphics based on event-region coupling and dynamic layer clipping is used, including: S51: The system first extracts all fields with geographic attributes from the structured event data, including the list of affected cells, the main device number, the substation to which the device belongs, and the voltage level. The system then searches the device geographic mapping table containing the list of affected cells, the main device number, the substation to which the device belongs, and the voltage level to generate the event impact area point set P and the device set D for subsequent spatial calculations. S52: Using event-region coupling function The coupling function combines three factors: spatial distance effect, the straight-line distance between the event center and the regional center; topological correlation, the spatial intersection ratio between the main device service area and the user location; user density weight, the number of registered users per unit area on the GIS grid map; the coupling function is defined as ,in, is the center point of the event, Candidate area The center point, For equipment service area, For the region The boundary polygon of is a weight parameter, which is adjusted to adapt to user differences. The dist function represents the spatial straight-line distance between the event center and the area center, normalized to , calculated as ,in are the latitude and longitude coordinates of the event center point, is the latitude and longitude of the center point of the region, The maximum visual range distance set by the platform is used for normalization. The closer the distance, the closer the dist function value is to 1, and the higher the coupling degree. The farther the dist function distance is, the closer the value is to 0. The IOU function is the spatial intersection degree between the service range of the event master device and the candidate area, which measures the direct geographical coverage relationship. The specific calculation method is: ,in is the polygonal area formed by the service range of the event master device. is the polygon corresponding to the candidate region r, Area is the area calculation function in the map system, and the IOU function value is in the range of [0,1]. The closer it is to 1, the more completely the region is covered by the event, indicating a high degree of coupling. The function is used to calculate the number of registered electricity users per unit area. The specific calculation method is: ,in is the total number of registered users in region r, obtained from the main grid data, is the area of ​​region r, The maximum user density per unit area in the entire network area is used for normalization and is selected after the final calculation. The area before scoring enters the candidate set of graphic cutting; S53: Based on the set of regions with the highest coupling degree, the system constructs a minimum coverage rectangle to define the cropping boundary area. This area includes all hotspots and controls the total map size to not exceed the set image resolution. The geographic base layer, power-specific layer, and user interest layer are loaded within the cropping area. The layer overlay adopts a hierarchical transparency + raster priority mechanism. S54: Within the cropped area, the system analyzes all device point sets D using a spatial clustering algorithm to generate multiple event hotspots H. The system then adds attributes to each hotspot, including the hotspot ID, event level, and clickable range. Finally, each hotspot is bound to the corresponding field in the text summary, forming a complete graphic-text linkage model. S55: Encode the rendered graphic cropping area into an RCS-supported format, automatically select a version based on the user terminal's capabilities, and finally encapsulate it into a 5G message data packet, which is sent together with the text summary and countdown linkage information.

4. The method for converting a power notification text to a 5G message according to claim 1, characterized in that In step S5, a method for accurately delivering 5G messages based on user power grid profiles is used, including: S61: Based on the grid master data and historical interaction data, a grid profile of each user is constructed, including the nature of electricity consumption, sensitivity level, historical behavior pattern, terminal capabilities, and location information. The above attributes are integrated into a vectorized representation to form a user profile vector. , as a comprehensive characterization of users’ receiving capabilities and needs; S62: Establish a content level system for power notifications, classifying messages into multiple levels: L1 summary reminder, L2 text and image fusion, and L3 text and image + voice. Each level corresponds to different information granularity and modal combination structure. Preset structural templates for each level include whether to enable the graphic area, whether to include a recovery countdown, whether to add voice assistance, and whether to enable the text and image anchor linkage content elements. S63: Build a hierarchical profile matching rule library , each rule Expressed in the form of a conditional tuple, ,in, For user types, such as residential, commercial, and medical institutions, enumeration codes are used; The user importance level is 0 for low, 1 for medium, and 2 for high; Terminal capability level, scored based on device performance, with low-end being 0, medium-end being 1, and high-end being 2; The user's historical interaction activity is quantified based on the user's message click-through rate in the past 30 days, with 1 being active and 0 being inactive; The emergency level of the event, with values ​​ranging from 0 for normal, 1 for important, and 2 for urgent; Message level templates can be L1 plain text, L2 text and image, or L3 text and image + voice. The matching process is as follows: Conditional judgment, the system generates a target user portrait vector and the urgency of the current incident Perform field-level matching; filter rules, traverse the rule base, and filter all conditions that meet the requirements , , , , A set of rules equal to the corresponding rule value ; Prioritize the rule set By message level priority Sorting: the higher the message level, the higher the priority; Fuzzy matching, some fields in the rule conditions use wildcards; Finally, select the message level corresponding to the rule with the highest ranking As the message push level of the user's current event; Dynamic adjustment, combined with user click feedback and message sending results, adjustment 、 Weight, real-time update of rule matching priority, and optimization of matching results; S64: After the event summary, graphic area, and voice content are generated, the system trims and packages the rich media content based on the matching message level, including selecting an appropriate summary compression level, filtering graphic hot zone layers, adding and omitting voice content, and preserving the original anchor binding relationship; S65: Before sending a message to the 5G core network, the distribution control strategy is set according to the user profile, including distribution time window control, priority control, and frequency suppression. The push strategy can be dynamically adjusted, and a lighter message type can be switched if the user has not clicked on the message for a long time.

5. A system for converting power notification text to 5G messages based on the method of claim 1, characterized in that: include: Event parsing module: Receives raw power event data from the power PMS and DMS systems, pre-processes the power event data, uses rule parsing and key field extraction algorithms, and outputs a standardized structure; Summary generation module: Receives structured event data and models it into a multi-dimensional semantic graph structure based on the power grid event graph model. It then calls the semantic summary generation model to generate a popular text summary based on the event type and target user profile. Graphics generation module: maps the geographic coordinates and device numbers in the event structure to the GIS platform, obtains the corresponding area layer, renders the event area, event location, and impact range boundary, and highlights the event area as a hot spot in the map; Graphics and text linkage module: anchors the summary text to the graphics, uses field mapping strategies to map the text summary to the graphic hot zone ID, and binds the time to the estimated recovery time countdown timer; Message encapsulation module: encapsulates text summaries and graphic content into 5G rich media message format, selects the appropriate message version based on the user terminal capabilities, and finally sends the message to the user terminal through the operator's 5G core network.

Images