Energy equipment data report generation method based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based energy equipment data report generation method, which comprises the following steps of: creating a dynamic data mirror image corresponding to physical energy equipment, receiving and carrying out isomorphic processing on a native data stream from a heterogeneous data source, and generating a standard data frame sequence with a unified time sequence; injecting the standard data frame sequence into a dynamic data mirror image to drive the dynamic data mirror image to simulate real-time operation evolution; capturing a mirror image state snapshot under a preset rule in an evolution process, and analyzing the mirror image state snapshot by utilizing a behavior understanding model to obtain equipment behavior semantic description; and according to the equipment behavior semantic description, the logic fragments are matched and instantiated and narrated from the report template knowledge base, and a structured data report is generated. The method has the beneficial effects that automatic and intelligent generation from multi-source heterogeneous data to a semantic coherence report can be realized, and the accuracy and practicability of an energy equipment data report are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for energy equipment, and more specifically, to a method for generating data reports for energy equipment based on artificial intelligence. Background Technology

[0002] Currently, the generation of energy equipment data reports for physical energy equipment mainly relies on manual aggregation of discrete data sources and analysis based on fixed rules. Conventional technical solutions typically directly interface with heterogeneous sources such as sensor databases and log systems, extracting data through pre-set query scripts and generating statistical reports containing charts and status codes based on simple logic such as threshold judgments. This approach treats data as independent, static points or sequences, limiting the content of energy equipment data reports to the listing of values, over-limit alarms, and trend curve displays, failing to reflect the complex coupling relationships and coherent behavioral logic between various parameters during the dynamic operation of physical energy equipment.

[0003] The shortcoming of existing technologies lies in their processing methods, which sever the intrinsic connection between data and the actual operational status of equipment. The raw data stream lacks deep integration and contextual reconstruction, preventing analytical models from understanding the overall, continuous behavior of the equipment. The generated energy equipment data reports are merely a collection of discrete events or indicators, lacking a coherent, semantic narrative of "what happened" and "why it happened." This stems from two fundamental problems: the lack of a unified dynamic model capable of real-time fusion of multi-source data and accurate mapping of the physical equipment's operational evolution as an analytical foundation; and the lack of advanced artificial intelligence analytical capabilities to understand complex operational states and translate them into human-readable behavioral descriptions.

[0004] Therefore, there is an urgent need for an automated and intelligent method to generate energy equipment data reports from multi-source heterogeneous data to produce semantically coherent reports, in order to improve the accuracy and practicality of energy equipment data reports. Summary of the Invention

[0005] The technical problem this invention aims to solve is how to achieve the automatic and intelligent generation of reports with semantic coherence from multi-source heterogeneous data, thereby improving the accuracy and practicality of energy equipment data reports. To overcome the shortcomings of the prior art (or related technologies), this invention provides an artificial intelligence-based method for generating energy equipment data reports.

[0006] This invention provides a method for generating energy equipment data reports based on artificial intelligence, comprising: Step S1: For the physical energy device to be analyzed, create a dynamic data mirror corresponding to the physical energy equipment, and receive the native data stream about the physical energy device from a heterogeneous data source; Step S2: The original data stream is homogenized and translated to generate a standard data frame sequence with a unified time sequence label, and the standard data frame sequence is injected into the dynamic data image to drive the dynamic data image to simulate the real-time operation and evolution of the physical energy device; Step S3: Capture a mirror state snapshot that conforms to preset rules during real-time operation and evolution, and submit the mirror state snapshot to the asynchronous analysis channel. Based on the pre-trained behavior understanding model, parse the mirror state snapshot to obtain a semantic description of equipment behavior. Step S4: Based on the semantic description of the equipment behavior, match and instantiate the corresponding narrative logic fragment in the report template knowledge base, and then aggregate multiple narrative logic fragments that are consecutive in time to generate a structured energy equipment data report.

[0007] In one possible implementation, the heterogeneous data source includes multiple data producers, namely a sensor network, an operation and maintenance log database, and a remote monitoring terminal. The process of generating the standard data frame sequence in step S2 includes: Assign a unique data source fingerprint to each of the data producers in the heterogeneous data source, receive raw data from each of the data producers and add an arrival timestamp to each of the raw data; Based on the fingerprint query of the data source, the conversion logic from a specific data format to a standard data model is obtained from the preset translation rule set; Based on the transformation logic obtained from the query, the raw data with the arrival timestamp stamped is converted into a standard data object containing data values, physical units, and quality labels; Based on the global time synchronization service, the arrival timestamp is calibrated as a global time stamp, and the standard data object carrying the global time stamp is encapsulated into a standard data frame. The standard data frames are then arranged in the order of the global time stamp to form the standard data frame sequence.

[0008] In one possible implementation, the process of driving the dynamic data mirror simulation to perform real-time evolution in step S2 includes: A set of state variables is predefined in the dynamic data mirror, and the state variables correspond to the key operating parameters of the physical energy equipment. Establish an injection queue and cache the standard data frames to be injected in the standard data frame sequence in order; Extract a standard data frame from the head of the injection queue, and locate the target state variable that needs to be updated in the dynamic data image based on the physical dimensions in the standard data frame; Assign the data value from the standard data frame to the target state variable and record the update time; After assigning values ​​to all the target state variables, the internal state transition function of the dynamic data mirror is triggered to calculate the derived state variable values ​​of the dynamic data mirror at the next moment, and the real-time running evolution of one simulation step is completed based on the derived state variable values.

[0009] In one possible implementation, the process of capturing the mirror state snapshot in step S3 includes: After each simulation step of the dynamic data mirror is completed, the values ​​of all current state variables are combined into a complete state vector, and the complete state vector is matched with the preset rule in the rule engine. The preset rule is a Boolean expression defined based on the logical combination of state variables. When the complete state vector satisfies any of the preset rules, the current simulation time point and the identifier of the preset rule being satisfied are recorded. Then, a fixed-duration time window is extracted both forward and backward, centered on the current simulation time point. Extract the complete state vector corresponding to all simulation steps within the time window from the historical state records of the dynamic data mirror, package them to generate the mirror state snapshot, and attach the preset rule identifier as metadata to the mirror state snapshot.

[0010] In one possible implementation, the process of parsing the semantic description of the equipment behavior in step S3 includes: Assign a unique analysis task identifier to the image state snapshot and store the image state snapshot in the analysis task queue; The behavior understanding model sequentially retrieves the image state snapshots and the attached preset rule identifiers from the analysis task queue. According to the trigger condition identifier, the corresponding attention weight configuration is loaded from the model parameter library, and the behavior understanding model assigns different attention to different state variables in the mirror state snapshot based on the attention weight configuration; The behavior understanding model is used to perform temporal convolution and feature extraction on the mirror state snapshot to generate a high-dimensional feature vector. The high-dimensional feature vector is mapped to a predefined equipment behavior semantic space, and a natural language text is output as the equipment behavior semantic description. The natural language text describes the operating behavior pattern of the physical energy equipment within the time window.

[0011] In one possible implementation, step S3, loading the corresponding attention weight configuration from the model parameter library according to the trigger condition identifier, and assigning different levels of attention to different state variables in the mirror state snapshot based on the attention weight configuration by the behavior understanding model, includes: A set of attention weight vectors is pre-configured for each of the preset rule identifiers defined in the rule engine and stored in the model parameter library; The attention weight vector corresponding to the preset rule identifier is retrieved from the model parameter library, and the retrieved attention weight vector is loaded into the attention mechanism layer of the behavior understanding model; The behavior understanding model uses the loaded attention weight vector to calculate the attention distribution at each time step in the state variable sequence; Based on the calculated attention distribution, the feature representations of different time steps and different variables in the state variable sequence are weighted and fused to obtain the attention of different state variables.

[0012] In one possible implementation, the report template knowledge base is organized in a directed graph structure, where nodes are atomic narrative units and edges represent logical connections. The process of matching and instantiating the narrative logic fragments in step S4 includes: The semantic skeleton is formed by extracting key entities and key actions from the semantic description of the equipment behavior. Using the semantic skeleton as the query condition, a multi-level search is performed in the report template knowledge base to locate the subgraph path with the highest matching degree with the semantic skeleton in the directed graph structure. The subgraph path is composed of a series of ordered atomic narrative units. Following the order of the subgraph paths, the original text template of each atomic narrative unit is obtained sequentially; The names, data values, and time information of the key entities extracted from the equipment behavior semantic description are filled into the placeholders in the original text template to generate the corresponding text sentences. The multiple text sentences generated in the order of the subgraph path are combined to form a coherent narrative logic segment.

[0013] In one possible implementation, step S4, which involves extracting the process for forming the semantic skeleton, includes: The natural language text describing the equipment behavior semantics is segmented and tagged with parts of speech. Based on a predefined entity dictionary in the field of energy equipment, dictionary matching and named entity recognition technologies are used to identify the key entities representing physical energy equipment components, operating parameters, and fault modes from the word segmentation results. Using dependency parsing techniques, the grammatical modification relationships between words in the natural language text are analyzed, and the core verbs related to the key entities are identified as the key actions. The identified key entities and key actions are combined according to their subject-verb-object or subject-verb relationship in the original sentence to form triples with entities and actions as nodes. All extracted triples are deduplicated and merged to remove redundant information. The deduplicated and merged triples are then sorted according to their logical importance to form the semantic skeleton that describes the core content of the semantic description of the equipment behavior.

[0014] In one possible implementation, the process of generating the energy equipment data report in step S4 includes: Record the corresponding simulation time window range for each generated narrative logic fragment, and sort all the narrative logic fragments to be aggregated in ascending order according to the start time of the simulation time window; Check whether there are gaps or overlaps in time between adjacent narrative logic segments: If a time gap exists, a neutral transition text will be generated and inserted into the gap. If there is time overlap, the overlapping narrative logic segments are trimmed and merged based on the time length ratio of the overlapping parts to generate new merged segments. All the logical segments arranged in sequence and processed continuously are placed into a preset hierarchical report structure, associated and indexed to generate the original mirror state snapshot on which the narrative logical segments are based, and the energy equipment data report is output.

[0015] In one possible implementation, the pre-training process of the behavior understanding model includes: Collect historical snapshots of the operational status of historical energy equipment and corresponding expert-annotated behavioral description texts to construct a training sample set. Design an encoder network structure, which includes multiple temporal convolutional layers and long short-term memory network layers, to process the sequence of state variables in the historical mirror state snapshots; Design a decoder network structure, which is a recurrent neural network based on an attention mechanism, for generating natural language descriptions; The encoder network structure and the decoder network structure are connected end-to-end to form a sequence-to-sequence model framework; The model framework is trained using the training sample set, and the model parameters are optimized by minimizing the difference between the generated text and the expert-annotated behavioral description text. The performance of the trained model framework is evaluated on a validation set. When the accuracy of the generated text reaches a preset threshold, the trained behavior understanding model is obtained.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By constructing a dynamic data mirror that can be driven by native data streams, a unified, continuous, and high-fidelity analytical foundation is provided for the generation of energy equipment data reports. This dynamic data mirror integrates data from multiple heterogeneous data sources in real time and transforms it into a digital evolution state with complete contextual temporal correlation. This changes the problem of data being separated from equipment operation scenarios in traditional methods, transforming the analysis object from a discrete data set into a virtual equipment behavior trajectory that can be simulated and deduced. All subsequent analyses are based on this coherent real-time operation evolution process, ensuring the inherent consistency and dynamic authenticity of the energy equipment data report analysis basis. 2. A pre-trained behavior understanding model is used to parse the mirror state snapshot, realizing the direct conversion from complex numerical states to high-level operational semantics. The behavior understanding model can identify specific operating modes and abnormal behavior sequences characterized by multi-parameter coupling changes, and output equipment behavior semantic descriptions that can be directly used for narration. This replaces the traditional method that relies on manual feature extraction and hard-coded rule judgment, and solves the semantic gap between low-level data and high-level cognition. The production of energy equipment data report content is thus transformed into the direct description and organization of identified behavioral events with clear engineering significance, enhancing the logic and interpretability of energy equipment data reports. 3. By matching and assembling narrative logic fragments in the report template knowledge base based on the semantic description of equipment behavior, dynamic adaptive generation of energy equipment data report structure is achieved. The content and structure of energy equipment data report are determined by the actual behavior sequence of physical energy equipment, rather than simple filling of static templates. The temporally continuous behavioral semantics are aggregated into a complete narrative with inherent causal and temporal logic, enabling the generated energy equipment data report to clearly present the behavioral evolution of physical energy equipment within a specific time period, forming a coherent technical narrative, thereby improving the information density and decision support value of energy equipment data report. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a matching analysis diagram of the wind power equipment state triggering conditions during the mirror state capture phase of the present invention. Figure 3 This is a diagram showing the coordinated change of operating parameters of the physical energy equipment of the present invention. Detailed Implementation

[0018] First, those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0020] See Figure 1 This invention discloses a method for generating energy equipment data reports based on artificial intelligence, comprising: Step S1: For the physical energy device to be analyzed, create a dynamic data mirror corresponding to the physical energy equipment, and receive the native data stream about the physical energy device from a heterogeneous data source; Step S2: The original data stream is homogenized and translated to generate a standard data frame sequence with a unified time sequence label, and the standard data frame sequence is injected into the dynamic data image to drive the dynamic data image to simulate the real-time operation and evolution of the physical energy device; Step S3: Capture a mirror state snapshot that conforms to preset rules during real-time operation and evolution, and submit the mirror state snapshot to the asynchronous analysis channel. Based on the pre-trained behavior understanding model, parse the mirror state snapshot to obtain a semantic description of equipment behavior. Step S4: Based on the semantic description of the equipment behavior, match and instantiate the corresponding narrative logic fragment in the report template knowledge base, and then aggregate multiple narrative logic fragments that are consecutive in time to generate a structured energy equipment data report.

[0021] In this embodiment of the invention, the method involves creating a dynamic data mirror corresponding to physical energy equipment and processing raw data streams from heterogeneous data sources. For a gas turbine generator set, the implementation process involves the standardized integration of multi-source data and the driven update of the dynamic data mirror. A unique data source fingerprint is assigned to each independent data producer in the heterogeneous data source. For example, the raw data stream from the vibration sensor is assigned the data source fingerprint "VS-001", the raw data stream from the temperature sensor is assigned the data source fingerprint "TS-100", and the raw data stream from the distributed control system log is assigned the data source fingerprint "DCS-LOG-01". A preset translation rule set is queried according to the data source fingerprint. The translation rule set defines the conversion logic from a specific data format to a standard data model. For the vibration sensor corresponding to the data source fingerprint "VS-001", the translation rule set specifies dividing the original integer value by 1000 to convert it into a floating-point number with "m / s²" as the physical unit, and adding the quality label "normal". For the log entry corresponding to the data source fingerprint "DCS-LOG-01", the translation rule set specifies using a natural language processing model to parse the text, extract the operation entity "valve V001" and the operation action "open", and map them into a standard data object with the physical unit of "state" and a value of "1". Applying the queried translation rule set, the raw data unit with the arrival timestamp is converted into a standard data object containing data value, physical unit, and quality label. Based on the global time-series synchronization service, the arrival timestamps of each independent heterogeneous data source are calibrated to a unified global time-series clock to generate a global time-series marker. The standard data object carrying the global time-series marker is encapsulated into a standard data frame. The standard data frame contains fields such as frame header, global time-series marker, data value, physical unit, and quality label, and is arranged in strict order according to the global time-series marker to form a standard data frame sequence.

[0022] In this embodiment of the invention, a set of state variables are predefined in the dynamic data mirror, corresponding to key operating parameters of the gas turbine. An injection queue is established, and standard data frames to be injected are cached sequentially. A standard data frame is extracted from the head of the injection queue, and the target state variable "bearing vibration amplitude" that needs to be updated in the dynamic data mirror is located based on the physical dimension identifier "vibration amplitude" in the standard data frame. The data value "0.123" in the standard data frame is assigned to the target state variable "bearing vibration amplitude", and the update time "GTS-102727143005123" is recorded. After completing the assignment of values ​​to all target state variables, the internal state transition function of the dynamic data mirror is triggered. The internal state transition function is a set of differential equations or state machines based on physical principles and operating rules, used to simulate the mutual influence between variables.

[0023] In this embodiment of the invention, the execution of the internal state transition function can be expressed as a process of calculating derived state variable values. The internal state transition function calculates the derived state variable values ​​for the next simulation time step based on the values ​​of all state variables at the current time. For example, based on current state variables such as "turbine speed," "fuel flow rate," and "compressor efficiency," the predicted value of "power generation" for the next time step is calculated using the internal state transition function. Executing the internal state transition function calculates the derived state variable values ​​of the dynamic data mirror at the next time step, thus completing the evolution of one simulation step. After the simulation step is completed, the state vector of the dynamic data mirror is updated. This state vector includes all directly injected measured state variable values ​​and the derived state variable values ​​calculated by the internal state transition function, providing a basis for subsequent rule matching and snapshot capture. It can be understood that this process is continuous, with standard data frame sequences being continuously injected, and the state of the dynamic data mirror continuously evolving accordingly, forming a synchronous high-fidelity simulation of the operating state of the physical energy equipment.

[0024] In this embodiment of the invention, the process of capturing and analyzing mirror state snapshots that conform to preset rules is illustrated by the operation of a steam turbine generator set. After each simulation step of the dynamic data mirroring is completed, the values ​​of all current state variables form a complete state vector. The state variables include "main steam pressure", "turbine speed", "bearing temperature", and "output power". An example of a complete state vector is [main steam pressure = 8.5 MPa, turbine speed = 3000 rpm, bearing temperature = 85 degrees Celsius, output power = 100 MW]. The complete state vector is matched with preset rules in the rule engine. The preset rules are Boolean expressions defined based on the logical combination of state variables. For example, the Boolean expression corresponding to the preset rule identifier "TC-ALERT-002" is "bearing temperature > 90 degrees Celsius AND bearing temperature change rate > 5 degrees Celsius per minute", and the Boolean expression corresponding to the preset rule identifier "TC-PERF-001" is "output power fluctuation standard deviation > 10% of the baseline output power fluctuation standard deviation". When a complete state vector satisfies any preset rule, the system records the current simulation time point, such as "Global Time Marker GTS-110415093020456", and the satisfied preset rule identifier, such as "TC-ALERT-002". Centered on the current simulation time point, a fixed-length time window is extracted both forward and backward. The time window length is set to 10 simulation steps, so the extracted time range is from Global Time Marker GTS-110415093010456 to Global Time Marker GTS-110415093030456. From the historical state records of the dynamic data mirror, the complete state vectors corresponding to all simulation steps within the time window are extracted. For example, 20 state vectors are extracted. Each state vector contains the values ​​of the same state variables at the corresponding simulation steps. These are packaged to generate a mirror state snapshot. The data structure of the mirror state snapshot is a multi-dimensional array with the array dimensions being [number of time steps, number of state variables]. The preset rule identifier "TC-ALERT-002" is appended to the mirror state snapshot as metadata.

[0025] In this embodiment of the invention, the preset rules in the rule engine cover various physical energy equipment types and operating events. For wind turbine generators, the preset rule identifier "TC-WIND-005" corresponds to the Boolean expression "blade vibration frequency > 2 Hz AND gearbox oil temperature > 75 degrees Celsius". For lithium-ion battery energy storage systems, the preset rule identifier "TC-BATT-003" corresponds to the Boolean expression "single cell voltage range > 0.3 volts AND system temperature difference > 10 degrees Celsius". It can be understood that the matching process of the preset rules is a step-by-step calculation of the numerical values ​​in the state vector and the logical operators of the Boolean expression. When the calculation result is true, snapshot capture is triggered. A unique analysis task identifier is assigned to the mirror state snapshot, for example, the analysis task identifier "ATASK-20231104-150934", and the mirror state snapshot is stored in the analysis task queue. The analysis task queue adopts a priority scheduling mechanism, assigning higher dequeue priority to mirror state snapshots with high severity level identifiers.

[0026] In this embodiment of the invention, the behavior understanding model sequentially obtains mirror state snapshots and additional preset rule identifiers from the analysis task queue. For example, it obtains mirror state snapshot data and metadata "TC-ALERT-002". Based on the preset rule identifier, it loads the corresponding attention weight configuration from the model parameter library. The model parameter library stores key-value pairs, where the key is the preset rule identifier and the value is the corresponding attention weight vector. The behavior understanding model assigns different attention levels to different state variables in the mirror state snapshot according to the attention weight configuration. For example, for the preset rule identifier "TC-ALERT-002", the loaded attention weight vector specifies that the attention weight of the state variable "bearing temperature" is 0.6, the attention weight of the state variable "bearing temperature change rate" is 0.3, the attention weight of the state variable "output power" is 0.1, and the attention weight of the state variable "main steam pressure" is 0.0. The behavior understanding model performs temporal convolution and feature extraction on the mirror state snapshot to generate a high-dimensional feature vector. The sequence of state variables in the mirror state snapshot is arranged in the time dimension to form a matrix of shape [20,4], where 20 is the number of time steps and 4 is the number of state variables.

[0027] In this embodiment of the invention, the temporal convolution and feature extraction process involves a multi-layer one-dimensional convolutional neural network. Each layer includes convolution, activation, and pooling operations. The generation of high-dimensional feature vectors can be expressed by the following formula: in: This represents the generated high-dimensional feature vector. Represents a non-linear activation function. This represents the intermediate feature matrix after multiple convolutional processing. This indicates a global average pooling operation. This represents the feature weighted sum matrix after the attention weight configuration is adjusted. The term represents the bias term of the fully connected layer. All symbols in the formula are defined within the scope of this embodiment and do not repeat the formula symbols in other embodiments. The high-dimensional feature vector is mapped to a predefined equipment behavior semantic space, and a natural language text is output as a semantic description of the equipment behavior. For example, the output natural language text is "The bearing temperature rises continuously from 82 degrees Celsius to 94 degrees Celsius within the time window, and the temperature change rate reaches a peak of 6.2 degrees Celsius per minute in the sixth minute. During the same period, the output power fluctuates between 98 MW and 102 MW."

[0028] In this embodiment of the invention, the behavior understanding model mapping process uses an encoder-decoder architecture. The encoder network structure compresses high-dimensional feature vectors into context vectors, and the decoder network structure generates word sequences based on autoregression of the context vectors until the end symbol is generated. It can be understood that the asynchronous nature of the asynchronous analysis channel is reflected in the decoupling of the submission of the mirror state snapshot from the behavior understanding model's analysis. The analysis task queue buffers the submitted mirror state snapshots, and the behavior understanding model consumes tasks from the queue as an independent process. Optionally, the behavior understanding model is deployed on a dedicated computing node, supporting GPU acceleration to process high-dimensional time-series data. The model parameter library is stored in a distributed database such as Redis for fast retrieval and loading by multiple instance behavior understanding models. The output format of the equipment behavior semantic description is a JSON object, containing a behavior summary, key data points, and time range fields, for use by the subsequent report generation module.

[0029] In this embodiment of the invention, based on the semantic description of equipment behavior, the corresponding narrative logic fragments are matched and instantiated in the report template knowledge base. Taking the data analysis of a photovoltaic power generation system as an example, the input of the semantic description of equipment behavior is a natural language text: "The DC bus voltage continuously decreased from 650 volts to 580 volts between 14:25 and 14:35. The photovoltaic inverter issued a DC undervoltage alarm at 14:33, while the ambient irradiance remained stable during the same period." The natural language text of the semantic description of equipment behavior is processed by word segmentation and part-of-speech tagging. The part-of-speech tagging marks each word with noun, verb, time word, and other part-of-speech categories. Based on a predefined entity dictionary in the energy equipment field, dictionary matching and named entity recognition technologies are used to identify key entities representing physical energy equipment components, operating parameters, and fault modes from the word segmentation results. The identified key entities include "DC bus voltage," "photovoltaic inverter," "DC undervoltage alarm," and "ambient irradiance." Dependency parsing technology is used to analyze the grammatical modification relationships between words in the natural language text, and the core verbs related to the key entities are identified as key actions. The analysis shows that the key actions include "decreases," "issues," and "maintains." The identified key entities and key actions are combined according to their subject-verb-object or subject-verb relationship in the original sentence to form triples with entities and actions as nodes.

[0030] In this embodiment of the invention, the entity dictionary for the energy equipment field includes standard terms from multiple subfields such as photovoltaics, wind power, and energy storage. Dictionary matching employs a prefix tree data structure for efficient lookup, and named entity recognition technology uses a neural network model based on an attention mechanism. Dependency parsing technology uses a transformation-based parser or a graph-based neural network parser to establish dominance relationships between words to accurately extract core predicates. All extracted triples are deduplicated and merged to remove redundant information; for example, triples describing the same event but with slightly different time signatures are merged into one. The deduplicated and merged triple set is then sorted according to logical importance, forming a semantic skeleton that describes the core content of the equipment behavior semantics. The logical importance ranking is based on a comprehensive determination of the core importance of key entities in physical energy equipment, the abnormality level of key actions, and the recentity of the event.

[0031] In this embodiment of the invention, a multi-level retrieval is performed in the report template knowledge base using a semantic skeleton as the query condition. The report template knowledge base is organized in a directed graph structure, where nodes are atomic narrative units. An atomic narrative unit is the smallest unit of a text template describing a single event or state, and edges represent logical connections, including chronological order, causal relationships, and transitional relationships. The subgraph path with the highest matching degree to the semantic skeleton is located in the directed graph structure. The matching degree calculation considers the coverage of key entities, key actions, and preset labels of atomic narrative units in the triplet, as well as the logical coherence of the subgraph path. The subgraph path consists of a series of ordered atomic narrative units. For example, the located subgraph path contains three atomic narrative units in the order of "Device X reports alarm Y at time T1", "Associated parameter Z shows a changing trend in time period T2", and "Parameter W remains stable during the same period". Following the order of the subgraph path, the original text template of each atomic narrative unit is obtained sequentially. The original text template of the atomic narrative unit "Device X reports alarm Y at time T1" is "[Device] reports [alarm type] alarm at [time]". The key entity names, data values, and time information extracted from the equipment behavior semantic description are filled into placeholders in the original text template to generate specific text sentences. The placeholders are set as follows: [Equipment] = "Photovoltaic Inverter", [Time] = "14:33", and [Alarm Type] = "DC Undervoltage". The generated sentence is "The photovoltaic inverter reported a DC undervoltage alarm at 14:33". Multiple specific text sentences generated according to the subgraph path sequence are combined to form a coherent narrative logic segment.

[0032] In this embodiment of the invention, the report template knowledge base is constructed based on historical expert reports and domain knowledge, and the directed graph structure allows for flexible combination of various narrative logics. In some embodiments, the atomic narrative unit template contains strictly defined placeholder types and format constraints, and the filling process performs data type validation, such as time format and numerical units. Optionally, the subgraph path search with the highest matching degree uses a graph search algorithm based on dynamic programming to ensure efficient path location even when the knowledge base is large. Multiple specific text sentences generated according to the subgraph path order are combined to form a coherent narrative logic fragment. The combination process inserts preset connecting words or phrases between sentences to ensure paragraph fluency. An example of the generated narrative logic fragment is: "The photovoltaic inverter reported a DC undervoltage alarm at 14:33. The associated parameter, DC bus voltage, continuously decreased from 650 volts to 580 volts between 14:25 and 14:35, while the environmental parameter, irradiance, remained stable during the same period." Optionally, the quantitative calculation of the subgraph path matching degree can be performed using the following formula: in: Indicates the matching score. Indicates the number of triples in the semantic skeleton. This represents the entity matching weight coefficient. Indicates the first Entities in a triplet With atomic narrative unit label The similarity function, This represents the action matching weight coefficient. Indicates the first Actions in a triplet With atomic narrative unit label The similarity function, This indicates the path of the candidate subgraph. The logical coherence scoring function. It can be understood that multi-level retrieval in a knowledge base first performs coarse-grained tag filtering, followed by fine-grained semantic similarity calculation and path scoring.

[0033] In this embodiment of the invention, multiple sequential narrative logic segments are aggregated and assembled to generate a structured energy equipment data report. Taking the operation data analysis of a doubly-fed wind turbine generator set over 24 hours as an example scenario, the system has generated three independent narrative logic segments. The simulation time window range corresponding to each generated narrative logic segment is recorded. The simulation time window ranges for the three narrative logic segments are "2023-11-05 10:15:00 to 2023-11-05 10:30:00", "2023-11-05 10:28:00 to 2023-11-05 10:45:00", and "2023-11-05 11:05:00 to 2023-11-05 11:20:00". According to the start time of the simulation time window, all narrative logic segments to be aggregated are sorted in ascending order. The sorted order is: first narrative logic segment, second narrative logic segment, and third narrative logic segment. See Table 1 below for details. Table 1. Status Table of Temporal Aggregation Processing for Narrative Logic Fragments Check whether adjacent narrative logic segments are consecutive or overlapping in time. Compare the first and second narrative logic segments. The first segment ends at "10:30:00," and the second segment begins at "10:28:00," indicating an overlap from "10:28:00" to "10:30:00." Compare the second and third narrative logic segments. The second segment ends at "10:45:00," and the third segment begins at "11:05:00," indicating a 20-minute gap. If a time gap exists, a neutral transition text is generated and inserted to maintain narrative continuity. For the 20-minute gap from "10:45:00" to "11:05:00," the generated transition text is: "From then until 11:05, the unit operating parameters were within the normal fluctuation range." If time overlap exists, the overlapping narrative segments are trimmed and merged based on the proportion of their overlapping time lengths to generate a new merged segment. The proportion of time overlap is calculated by dividing the overlap duration by the total segment duration. The logic for trimming and merging overlapping segments can be expressed as follows: in: This represents the content of the generated fused fragment. Represents the fusion function. This represents the content of the first overlapping narrative logical segment. This represents the content of the second overlapping narrative logical segment. This represents the weighting coefficient for the overlapping portion of the two segments, calculated based on a time window. All symbols in this formula are defined and used only within the context of this embodiment. Based on this logic, the system merges the description of the first narrative segment regarding the period from "10:28:00 to 10:30:00" with the description of the second narrative segment during the same period, removes duplicate time information, and generates a new fused segment describing the concurrent abnormality of generator bearing vibration and temperature during the period from "10:15:00 to 10:45:00".

[0034] In this embodiment of the invention, all narrative logical segments, arranged sequentially and processed continuously, are placed into a preset hierarchical report structure. The hierarchical report structure includes a summary layer, an event detail layer, and a data support layer. The summary layer generates a report summary, extracting key events, times, and conclusive descriptions from all narrative logical segments for condensation, for example, generating "During the reporting period, the main records include one generator bearing abnormality event and one power adjustment due to a decrease in wind speed." The event detail layer arranges the processed narrative logical segments and transitional text in chronological order to form the main narrative content of the report. In the data support layer, the original mirror state snapshot on which the narrative logical segments are based is associated and indexed, providing a traceable data source for each statement in the event detail layer. The index information includes the storage path, timestamp, and preset rule identifier of the mirror state snapshot.

[0035] In this embodiment of the invention, the trimming and fusion of overlapping time segments requires semantic understanding. The system compares the descriptions of the same time point in two overlapping narrative segments. If the descriptions are consistent, one is retained; if the descriptions are complementary, they are merged; if the descriptions contradict each other, they are selected based on data quality labels or rule priorities. In some embodiments, the simulation time window range of the fused new segment is adjusted to range from the earliest start time to the latest end time. Finally, a complete energy equipment data report is output, presented in a standard document format, containing a complete hierarchical report structure, text content, and data index. Optionally, the data index in the data support layer can be in the form of hyperlinks, allowing direct access to the corresponding original mirror state snapshot data within the electronic report.

[0036] In this embodiment of the invention, see Figure 2 This is a chart analyzing the triggering conditions of wind power equipment during the mirror state capture phase. As shown in the chart, between 10:25 and 10:40, the effective value of vibration acceleration (red line) exceeded the 7.0 m / s² threshold, and the bearing temperature simultaneously exceeded the 45℃ threshold. This interval is marked as the "triggering state snapshot interval." Vibration and temperature show a positive correlation (temperature rises synchronously with vibration), which is consistent with the physical logic of "intensified vibration → increased friction temperature" when the bearing is abnormal. This type of chart is used to identify abnormal operating periods of equipment. By matching parameters with thresholds, the capture of mirror state snapshots is automatically triggered, providing key event data support for subsequent semantic analysis and report generation. It is the core link of "state perception → event extraction" in the generation of energy equipment data reports.

[0037] In this embodiment of the invention, the construction of the behavior understanding model includes the following steps. Taking the construction of a behavior understanding model for abnormal pattern recognition of a gas turbine control system as an example, historical mirror state snapshots generated during the operation of historical energy equipment and corresponding expert-annotated behavior description texts are collected to construct a training sample set. The historical mirror state snapshots are derived from thousands of state sequence snapshots captured by the rule engine during the operation of the gas turbine in the past three years. Each mirror state snapshot is attached with a preset rule identifier. The corresponding expert-annotated behavior description texts are written by domain experts based on the snapshot data. The overspeed protection logic is triggered, and the intake valve opening is synchronously closed. An encoder network structure is designed. The encoder network structure includes multiple layers of temporal convolutional layers and long short-term memory network layers to process the state variable sequences in the mirror state snapshots. The temporal convolutional layers are used to extract the interaction features of state variables at the local time scale, and the long short-term memory network layers are used to capture the long-term temporal dependencies of the state variable sequences.

[0038] In this embodiment of the invention, the decoder network structure is a recurrent neural network based on an attention mechanism, used to generate natural language descriptions. At each time step, the decoder network dynamically focuses on different parts of the encoder network output based on the context vector output by the encoder network and its own hidden state at the previous time step, in order to predict the word output at the current time step. The encoder and decoder networks are connected end-to-end to form a sequence-to-sequence model framework. The encoder network maps the variable-length mirror state snapshot sequence to a fixed-dimensional context vector representation. The decoder network uses the context vector as initial input and autoregressively generates a variable-length natural language description sequence. The sequence-to-sequence model framework is trained using a training sample set. The model parameters are optimized by minimizing the difference between the generated text and the expert-annotated behavioral description text. The difference is typically measured using the cross-entropy loss function. The optimization process employs the backpropagation algorithm and a gradient descent optimizer such as Adam. The performance of the trained behavior understanding model is evaluated on the validation set, which is a snapshot and annotation of a portion of the historical data that was not used in training. Evaluation metrics include the BLEU score, ROUGE score, and human readability score of the generated text. When the accuracy of the generated text reaches a preset threshold, the trained behavior understanding model is obtained. The preset threshold is set according to the actual application requirements, such as requiring a BLEU-4 score greater than 0.65.

[0039] In this embodiment of the invention, the calculation process of the attention mechanism can be described as follows: in: This represents the attention query vector at the current decoding moment. This represents the query generating function, which is typically a linear transformation. This represents the hidden state of the decoder network structure at the previous time step. This represents the set of hidden state vectors at all time steps of the encoder network structure's output. It can be understood that the training process employs a teacher-forced strategy, using the actual words from the previous time step as input during decoder network training to accelerate convergence. A dedicated set of attention weight vectors is pre-configured for each preset rule identifier defined in the rule engine and stored in the model parameter library. The preset rule identifier "TC-GT-OVERSPEED" corresponds to a dedicated set of attention weight vectors, which are used in the model's attention mechanism to emphasize state variable features related to rotational speed and acceleration. The preset rule identifier "TC-GT-TEMP-RISE" corresponds to another set of dedicated attention weight vectors, used to emphasize state variable features related to temperature, temperature difference, and temperature rise rate.

[0040] In this embodiment of the invention, the corresponding attention weight configuration is loaded from the model parameter library according to the preset rule identifier. The behavior understanding model assigns different levels of attention to different state variables in the mirrored state snapshot based on the attention weight configuration. When the behavior understanding model obtains an additional preset rule identifier such as "TC-GT-OVERSPEED", it queries the model parameter library using the preset rule identifier "TC-GT-OVERSPEED" as the index key. The model parameter library can be a key-value storage database or a distributed cache. The dedicated attention weight vector corresponding to the preset rule identifier "TC-GT-OVERSPEED" is retrieved from the model parameter library and loaded into the attention mechanism layer of the behavior understanding model. The parameters of the attention mechanism layer are dynamically replaced or weighted by this dedicated vector during inference. When processing the mirrored state snapshot, the encoder network structure of the behavior understanding model uses the loaded dedicated attention weight vector to calculate the attention distribution of each time step in the state variable sequence. The attention distribution determines which time steps and variables in the state variable sequence the encoder network structure should focus on when generating different parts of the descriptive text.

[0041] In this embodiment of the invention, the configuration of the dedicated attention weight vector is based on domain knowledge. For each preset rule identifier, an expert specifies a set of key state variables and their relative importance, which are then converted into initial weights in vector form. In some embodiments, the dedicated attention weight vector in the model parameter library can also be fine-tuned and optimized in the later stages of model training through multi-task learning or meta-learning. Based on the calculated attention distribution, the feature representations of different time steps and different variables in the state variable sequence are weighted and fused, thereby achieving focused attention on key state variables. For example, when generating text describing an overspeeding event, the attention distribution will significantly point to the time steps with higher feature values ​​for "speed" and "speed change rate" in the state variable sequence. It can be understood that this mechanism of dynamically loading attention weights based on preset rule identifiers enables the same behavior understanding model to adaptively adjust its analysis focus according to different analysis scenarios, improving the accuracy and relevance of the behavior understanding model in describing different types of running events.

[0042] In this embodiment of the invention, see Figure 3 This is a chart analyzing the coordinated changes in operating parameters of energy equipment. Figure 3 As can be seen, when the temperature rises, the pressure decreases synchronously; when the temperature falls, the pressure rises synchronously. The temperature reaches its peak (≈575℃) and the pressure reaches its trough (≈465℃) in approximately 20 seconds; the temperature reaches its trough (≈420℃) and the pressure reaches its peak (≈415℃) in approximately 60 seconds. Both parameters exhibit periodic oscillations, consistent with the equipment's operational logic of "operating condition cycle → parameter alternation." These types of charts are used to analyze the synergistic patterns of equipment parameters and provide fundamental data support for the "state evolution → behavioral semantic analysis" process in energy equipment data report generation. They can help identify the equipment's normal operating mode or the risk of parameter mismatch under abnormal operating conditions.

[0043] In the description of this invention, the references to "one embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An artificial intelligence-based energy equipment data report generation method, characterized by, Includes the following steps: Step S1: For the physical energy device to be analyzed, create a dynamic data mirror corresponding to the physical energy equipment, and receive the native data stream about the physical energy device from a heterogeneous data source; Step S2: The original data stream is homogenized and translated to generate a standard data frame sequence with a unified time sequence mark, and the standard data frame sequence is injected into the dynamic data image to drive the dynamic data image to simulate the real-time operation and evolution of the physical energy device; Step S3: Capture a mirror state snapshot that conforms to preset rules during real-time operation and evolution, and submit the mirror state snapshot to the asynchronous analysis channel. Based on the pre-trained behavior understanding model, parse the mirror state snapshot to obtain a semantic description of equipment behavior. Step S4: Based on the semantic description of the equipment behavior, match and instantiate the corresponding narrative logic fragment in the report template knowledge base, and then aggregate multiple narrative logic fragments that are consecutive in time to generate a structured energy equipment data report.

2. The energy equipment data report generation method of claim 1, wherein, The heterogeneous data source includes multiple data producers, namely a sensor network, an operation and maintenance log database, and a remote monitoring terminal. The process of generating the standard data frame sequence in step S2 includes: Assign a unique data source fingerprint to each of the data producers in the heterogeneous data source, receive raw data from each of the data producers and add an arrival timestamp to each of the raw data; Based on the fingerprint query of the data source, the conversion logic from a specific data format to a standard data model is obtained from the preset translation rule set; Based on the transformation logic obtained from the query, the raw data with the arrival timestamp stamped is converted into a standard data object containing data values, physical units, and quality labels; Based on the global time synchronization service, the arrival timestamp is calibrated as a global time stamp, and the standard data object carrying the global time stamp is encapsulated into a standard data frame. The standard data frames are then arranged in the order of the global time stamp to form the standard data frame sequence.

3. The energy equipment data report generation method of claim 2, wherein, The process of driving the dynamic data mirror simulation to run and evolve in real time in step S2 includes: A set of state variables is predefined in the dynamic data mirror, and the state variables correspond to the key operating parameters of the physical energy equipment. Establish an injection queue and cache the standard data frames to be injected in the standard data frame sequence in order; Extract a standard data frame from the head of the injection queue, and locate the target state variable that needs to be updated in the dynamic data image based on the physical dimensions in the standard data frame; Assign the data value from the standard data frame to the target state variable and record the update time; After assigning values ​​to all the target state variables, the internal state transition function of the dynamic data mirror is triggered to calculate the derived state variable values ​​of the dynamic data mirror at the next moment, and the real-time running evolution of one simulation step is completed based on the derived state variable values.

4. The energy equipment data report generation method of claim 3, wherein, The process of capturing the mirror state snapshot in step S3 includes: After each simulation step of the dynamic data mirror is completed, the values ​​of all current state variables are combined into a complete state vector, and the complete state vector is matched with the preset rule in the rule engine. The preset rule is a Boolean expression defined based on the logical combination of state variables. When the complete state vector satisfies any of the preset rules, the current simulation time point and the identifier of the preset rule being satisfied are recorded. Then, a fixed-duration time window is extracted both forward and backward, centered on the current simulation time point. Extract the complete state vector corresponding to all simulation steps within the time window from the historical state records of the dynamic data mirror, package them to generate the mirror state snapshot, and attach the preset rule identifier as metadata to the mirror state snapshot.

5. The energy equipment data report generation method of claim 4, wherein, The process of parsing the semantic description of the equipment behavior in step S3 includes: Assign a unique analysis task identifier to the image state snapshot and store the image state snapshot in the analysis task queue; The behavior understanding model sequentially retrieves the image state snapshots and the attached preset rule identifiers from the analysis task queue. According to the trigger condition identifier, the corresponding attention weight configuration is loaded from the model parameter library, and the behavior understanding model assigns different attention to different state variables in the mirror state snapshot based on the attention weight configuration; The behavior understanding model is used to perform temporal convolution and feature extraction on the mirror state snapshot to generate a high-dimensional feature vector. The high-dimensional feature vector is mapped to a predefined equipment behavior semantic space, and a natural language text is output as the equipment behavior semantic description. The natural language text describes the operating behavior pattern of the physical energy equipment within the time window.

6. The energy equipment data report generation method of claim 5, wherein, In step S3, the process of loading the corresponding attention weight configuration from the model parameter library according to the trigger condition identifier, and assigning different levels of attention to different state variables in the mirror state snapshot based on the attention weight configuration by the behavior understanding model includes: A set of attention weight vectors is pre-configured for each of the preset rule identifiers defined in the rule engine and stored in the model parameter library; The attention weight vector corresponding to the preset rule identifier is retrieved from the model parameter library, and the retrieved attention weight vector is loaded into the attention mechanism layer of the behavior understanding model; The behavior understanding model uses the loaded attention weight vector to calculate the attention distribution at each time step in the state variable sequence; Based on the calculated attention distribution, the feature representations of different time steps and different variables in the state variable sequence are weighted and fused to obtain the attention of different state variables.

7. The energy equipment data report generation method of claim 5, wherein, The report template knowledge base is organized in a directed graph structure, where nodes are atomic narrative units and edges represent logical connections. The process of matching and instantiating the narrative logic fragment in step S4 includes: The semantic skeleton is formed by extracting key entities and key actions from the semantic description of the equipment behavior. Using the semantic skeleton as the query condition, a multi-level search is performed in the report template knowledge base to locate the subgraph path with the highest matching degree with the semantic skeleton in the directed graph structure. The subgraph path is composed of a series of ordered atomic narrative units. Following the order of the subgraph paths, the original text template of each atomic narrative unit is obtained sequentially; The names, data values, and time information of the key entities extracted from the equipment behavior semantic description are filled into the placeholders in the original text template to generate the corresponding text sentences. The multiple text sentences generated in the order of the subgraph path are combined to form a coherent narrative logic segment.

8. The method for generating energy equipment data reports according to claim 7, characterized in that, The process of extracting the semantic skeleton in step S4 includes: The natural language text describing the equipment behavior semantics is segmented and tagged with parts of speech. Based on a predefined entity dictionary in the field of energy equipment, dictionary matching and named entity recognition technologies are used to identify the key entities representing physical energy equipment components, operating parameters, and fault modes from the word segmentation results. Using dependency parsing techniques, the grammatical modification relationships between words in the natural language text are analyzed, and the core verbs related to the key entities are identified as the key actions. The identified key entities and key actions are combined according to their subject-verb-object or subject-verb relationship in the original sentence to form triples with entities and actions as nodes. All extracted triples are deduplicated and merged to remove redundant information. The deduplicated and merged triples are then sorted according to their logical importance to form the semantic skeleton that describes the core content of the semantic description of the equipment behavior.

9. The method for generating energy equipment data reports according to claim 7, characterized in that, The process of generating the energy equipment data report in step S4 includes: Record the corresponding simulation time window range for each generated narrative logic fragment, and sort all the narrative logic fragments to be aggregated in ascending order according to the start time of the simulation time window; Check whether there are gaps or overlaps in time between adjacent narrative logic segments: If a time gap exists, a neutral transition text will be generated and inserted into the gap. If there is time overlap, the overlapping narrative logic segments are trimmed and merged based on the time length ratio of the overlapping parts to generate new merged segments. All the logical segments arranged in sequence and processed continuously are placed into a preset hierarchical report structure, associated and indexed to generate the original mirror state snapshot on which the narrative logical segments are based, and the energy equipment data report is output.

10. The method for generating energy equipment data reports according to claim 1, characterized in that, The pre-training process of the behavior understanding model includes: Collect historical snapshots of the operational status of historical energy equipment and corresponding expert-annotated behavioral description texts to construct a training sample set. Design an encoder network structure, which includes multiple temporal convolutional layers and long short-term memory network layers, to process the sequence of state variables in the historical mirror state snapshots; Design a decoder network structure, which is a recurrent neural network based on an attention mechanism, for generating natural language descriptions; The encoder network structure and the decoder network structure are connected end-to-end to form a sequence-to-sequence model framework; The model framework is trained using the training sample set, and the model parameters are optimized by minimizing the difference between the generated text and the expert-annotated behavioral description text. The performance of the trained model framework is evaluated on a validation set. When the accuracy of the generated text reaches a preset threshold, the trained behavior understanding model is obtained.

Images