Power grid asset depreciation calculation and residual life prediction method based on large model

Abstract

The invention relates to the technical field of power grid asset management, and provides a power grid asset depreciation calculation and residual life prediction method based on a large model. According to the method, a unified data view is constructed by integrating multi-source heterogeneous data such as equipment ledgers, operation monitoring, defect records, environmental data and financial information; a pre-training large model architecture of time sequence and text parallel coding is adopted, and multi-modal feature alignment fusion is achieved in a shared semantic space; inputting the fused sequence into a pre-training model based on Transform, capturing a long-term dependency relationship by using an attention mechanism, and synchronously outputting residual life prediction and health state evaluation; and finally, realizing automatic identification and early warning of the depreciation state of the equipment by combining a depreciation standard and a threshold rule through a large language model subjected to fine adjustment of domain knowledge. According to the invention, a data-driven and knowledge-coordinated intelligent management model is established, the accuracy and scientificity of asset value evaluation are remarkably improved, and reliable technical support is provided for asset management of power grid enterprises.

Description

Technical Field

[0001] This invention belongs to the field of power grid asset management and evaluation technology, specifically relating to a method for calculating asset depreciation and predicting the remaining life of equipment for power grid enterprises. Background Technology

[0002] In the current digital economy era, big data, artificial intelligence, and other new-generation information technologies are profoundly changing the operation and management models of traditional industries. The large scale, high technical level, high load density, and high reliability requirements of power grid equipment highlight the shortcomings of traditional depreciation methods. Faced with massive amounts of heterogeneous data from multiple sources, including equipment ledgers, real-time operation data, defect and fault data, inspection and testing data, and environmental and meteorological data, leveraging cutting-edge technologies to extract data value and achieve digital transformation in asset value management has become a key path to enhancing a company's core competitiveness.

[0003] Currently, the research and practice of power grid asset depreciation calculation and management methods are evolving from traditional experience-based models to modern data-driven models. Traditional methods primarily focus on technical analysis, using equipment health status as a basis for reliability assessment to indirectly infer its remaining lifespan. While these methods are safety-oriented, they are essentially experience-based estimations, often neglecting key value factors such as the economic costs of equipment operation and maintenance and environmental performance. With the deepening of power system reform and increasing regulatory emphasis on investment efficiency and precise cost accounting, their extensive nature is becoming increasingly apparent, failing to meet the needs of scientific asset management. Therefore, there is an urgent need for an asset depreciation calculation and remaining life prediction method that can comprehensively consider influencing factors to improve the accuracy and reliability of assessment results. Summary of the Invention

[0004] This invention aims to overcome the limitations of existing power grid asset depreciation management methods, which rely on single-dimensional data and ignore the actual condition of equipment and the nonlinear decay of value. It proposes a method for calculating power grid asset depreciation and predicting remaining useful life based on a large model. This method systematically integrates multi-source heterogeneous data, including equipment ledgers, operational monitoring, defect records, environmental data, and financial information, to construct a unified data view. Through multimodal feature fusion and deep learning techniques, it achieves accurate prediction of asset remaining useful life and intelligent judgment of depreciation status.

[0005] The core of this invention lies in constructing a pre-trained large-scale model architecture that uses parallel encoding of time-series and textual data. This architecture extracts and aligns features from numerical time-series data and textual descriptive data, respectively. It utilizes the deep attention mechanism of the Transformer architecture to capture long-term dependencies in multimodal data, simultaneously outputting continuous remaining lifetime values ​​and health status assessment results for the equipment. Furthermore, through a large language model fine-tuned with domain knowledge, combined with the depreciation standards and threshold rules of power grid companies, it achieves automatic identification and early warning of equipment nearing or exceeding its service life.

[0006] The significant advantages of this invention are: by deeply fusing and modally aligning multi-source heterogeneous data, it breaks through the limitation of traditional depreciation methods that rely solely on the time dimension, significantly improving the accuracy and robustness of remaining useful life prediction; by introducing a pre-trained large model and a domain rule collaboration mechanism, it realizes a closed loop from data-driven prediction to business decision support, enhancing the applicability and interpretability of the model in actual asset management; the entire method has strong scalability and can provide reliable technical support for the lean asset management and digital transformation of power grid enterprises.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for calculating the depreciation and predicting the remaining useful life of power grid assets based on a large model includes the following steps:

[0009] S1: Construction and preprocessing of multi-source heterogeneous datasets: Systematically integrate multi-source heterogeneous data generated by typical power grid assets throughout their entire lifecycle, and perform data cleaning, alignment and normalization preprocessing to build a unified data view at the equipment level;

[0010] S2: Parallel Encoding and Fusion of Multimodal Features: Construct a pre-trained large model architecture for parallel encoding of time series and text, extract features from numerical time series data and text description data respectively, and align and fuse feature vectors of different modalities in a shared semantic space;

[0011] S3: Synchronous prediction of remaining lifetime based on fused features: The fused multimodal high-dimensional feature sequence is input into the Transformer-based pre-trained model for deep information interaction and feature extraction. The continuous remaining lifetime value of the device is predicted synchronously through the output layer and the health status is assessed.

[0012] S4: Equipment depreciation status identification based on rule and model collaboration: The remaining life predicted by the model is input into a large language model that has been fine-tuned by domain knowledge, so that it can automatically identify and warn of equipment that is approaching or has reached its service life by combining the threshold rules set by the fixed asset classification and depreciation period standards of the power grid enterprise.

[0013] The construction and preprocessing of the multi-source heterogeneous dataset in step S1 specifically includes:

[0014] Based on Internet of Things (IoT) technology, the gradual establishment of China's power IoT has enabled power grid companies to achieve basic data connectivity. Because the power system supplies electricity to the entire society, its power equipment is complex and diverse, the system operates in real-time, almost 24 / 7 without interruption, and its coverage is extremely broad, encompassing a wide range of entities. It can acquire massive amounts of monitoring data, power generation equipment operation data, user load data, and electricity carbon emission data, among others. With the construction of my country's smart grid, the data volume obtained from smart power terminal equipment is becoming increasingly rich, and the data quality is improving due to the development of various data platforms, while the data structure is becoming increasingly complex. Simultaneously, with the in-depth advancement of the ubiquitous power IoT construction, the sharing of power big data brings unprecedented value to government departments, electricity users, and all sectors of society.

[0015] Typical power asset data forms the cornerstone of this research model. This section will systematically analyze the multi-dimensional characteristics and internal structure of data related to typical power grid assets, laying a solid data foundation for the subsequent construction of a multimodal fusion prediction model. We take core assets such as transformers and power cables as research objects, and conduct in-depth analysis of the massive, multi-source, and heterogeneous data generated throughout their entire life cycle.

[0016] In step S2, a pre-trained large model architecture for parallel temporal and textual encoding is constructed, and the output feature vectors are fused, specifically as follows:

[0017] This invention proposes a remaining lifetime prediction method based on pre-trained GPT. The method first fuses temporal and textual features from multi-source heterogeneous data and injects location information to construct a multimodal sequence. This sequence is then input into a GPT architecture composed of multiple Transformer modules for deep feature interaction and information extraction. Finally, an accurate prediction of the remaining lifetime is achieved through a linear output layer.

[0018] S2.1: Temporal coding path: For various numerical sequences, a channel-independent strategy is adopted to process each variable separately. The input is optimized by block division and reversible instance normalization techniques. Then, each sequence block is mapped to a high-dimensional feature vector through a linear embedding layer.

[0019] S2.2: Text Encoding Path: For maintenance records and defect description texts, use the BERT encoder to extract text features, and then project them through a linear layer and the ReLU activation function to a vector space with the same dimension as the temporal features to achieve modality alignment;

[0020] S2.3: Feature Fusion: The feature vectors output by the two paths are added element by element and fused together, and position embedding information is injected to form a fused multimodal sequence.

[0021] In step S3, Sobol global sensitivity analysis is used to quantify the contribution of each factor to the carbon asset assessment results, and key influencing factors are selected from them, specifically:

[0022] This invention inputs a fused multimodal high-dimensional feature sequence into a pre-trained model based on the Transformer architecture. This model, through its internal multi-layered self-attention mechanism, performs deep information interaction and hierarchical feature extraction within the sequence. Based on this powerful feature representation capability, the model can simultaneously complete two key tasks: accurately predicting the continuous remaining lifetime of equipment and comprehensively assessing its macroscopic health status. This process fully utilizes the superior sequence modeling and contextual understanding capabilities of large models, enabling them to effectively extract the inherent laws, micro-patterns, and long-term evolution trends characterizing equipment performance degradation from complex high-dimensional fused features. Ultimately, this method achieves high-precision inference of the remaining economic and technological life of assets from complex data containing multi-dimensional and non-linear correlations, thus fundamentally overcoming the limitations of traditional prediction methods that rely on a single data dimension or simplified linear assumptions.

[0023] S3.1: Deep Feature Extraction of Multimodal Sequences: The fused multimodal sequences are input into a pre-trained model based on the Transformer architecture. This model performs deep encoding and feature transformation on the input sequences through its multiple Transformer Blocks. Each Transformer Block sequentially performs multi-head self-attention calculation, residual connection and layer normalization, feedforward neural network processing, and residual connection and layer normalization again. Through this stacked deep structure, the high-level semantics and complex patterns in the input information are gradually abstracted and refined.

[0024] S3.2: Long-term dependency capture based on attention mechanism: The multi-head self-attention mechanism is the core of the model to realize deep information interaction. This mechanism first linearly transforms the input sequence into three matrices: query, key, and value. Then, in each attention head, the attention weight distribution is obtained by calculating the dot product of the query vector and all key vectors, and normalizing it by scaling and the Softmax function. Finally, the value vector is weighted and summed using this weight distribution to obtain the output of the attention head. By computing multiple attention heads in parallel and concatenating and linearly transforming their outputs, the model can capture complex, long-term dependencies and dynamic interaction patterns between arbitrary positions within the sequence from different subspaces.

[0025] S3.3: Synchronous Prediction Output of Remaining Lifetime and Health Status: The final sequence representation after processing through multiple Transformer Blocks aggregates the deep features of the input multimodal data. The model decodes this sequence representation through a specific output layer (usually a linear layer), mapping it to a continuous scalar value, which is the predicted remaining lifetime of the device. At the same time, the model architecture is designed to be scalable, adding a parallel classification output head to classify and evaluate the health status of the device based on the same deep features, outputting discrete status labels such as "healthy", "attention", "abnormal", and "severe", thereby achieving simultaneous output of quantitative prediction of remaining lifetime and qualitative assessment of health status.

[0026] In step S4, a mechanism for identifying equipment depreciation status is constructed based on the collaboration of rules and models, specifically as follows:

[0027] Based on a rule-and-model collaborative mechanism, this invention aims to achieve accurate identification and early warning of equipment depreciation status. Specifically, it inputs the remaining useful life predicted by the model into a large language model fine-tuned with professional knowledge of the power grid domain. This large language model deeply integrates the explicit threshold rules set in the fixed asset classification standards and depreciation period regulations of power grid enterprises, analyzes and judges the input information, and automatically identifies equipment that is approaching or has reached the prescribed service life, issuing timely warnings. This key step successfully combines data-driven prediction results with rigorous domain knowledge rules. Its output is no longer limited to a single quantitative remaining useful life indicator, but is further transformed into depreciation status judgments and specific decision-making suggestions that conform to the actual management processes of enterprises and industry regulatory requirements, ultimately constructing a complete closed loop from bottom-level data perception to top-level intelligent decision-making.

[0028] S4.1: Structuring the input of prediction results and domain knowledge: The remaining life of the equipment predicted in step S3, the basic information of the equipment (including equipment classification code, asset name, commissioning date, and design life), and the relevant depreciation policy provisions are combined to form a structured natural language prompt; this prompt template aims to clearly present the predicted facts and the question to be judged to the large language model.

[0029] S4.2: Depreciation status rule matching based on fine-tuned large language model: The above structured prompts are input into a large language model that has been fine-tuned in depth with knowledge of power grid depreciation. The model encodes the complex depreciation rules and judgment logic defined in industry standard documents in its internal parameters. After receiving the input, the model starts its embedded reasoning mechanism, compares the predicted remaining life with the statutory depreciation period or economic life corresponding to the equipment category, and makes a status judgment based on the preset, configurable multi-level threshold rules.

[0030] S4.3: Depreciation Status Determination and Structured Early Warning Report Generation: The threshold rules include at least two levels: The first level is the early warning threshold, where the model determines the equipment to be in a "nearing the end of its service life" state when the remaining useful life of the equipment is less than or equal to a certain percentage (e.g., 20%, which is adjustable) of the statutory depreciation period; the second level is the critical threshold, where the model determines the equipment to be in a "reached or exceeded service life" state when the remaining useful life of the equipment is less than or equal to a smaller percentage (e.g., 5%, which is adjustable) or has reached zero / negative values. Based on the matching results of these rules, the model automatically identifies a list of equipment that meets the conditions and generates a structured early warning report containing key information such as the equipment's unique identifier, current predicted remaining useful life, determined depreciation status, and recommended measures (e.g., strengthening monitoring, arranging assessments, and planning replacement), for asset management personnel to make decisions.

[0031] The significant features of this invention compared to the prior art are as follows:

[0032] (1) Based on the data perception and representation capabilities of multimodal fusion, unlike traditional methods that rely on only a single type of time series data or static ledgers, this invention first constructs a time series and text parallel coding architecture for power grid asset depreciation; (2) The decision-making paradigm of pre-trained large model and domain knowledge collaboration, unlike directly training shallow models or simply using rule engines, this invention innovatively adopts a two-stage intelligent decision-making process of "pre-trained large model fine-tuning + domain large language model (LLM) rule reasoning"; (3) This invention dynamically captures the personalized value decay trajectory of equipment caused by factors such as load, environment, and maintenance history through a data-driven approach, realizing a fundamental shift in asset depreciation from "static time dimension" to "dynamic state dimension". Attached Figure Description

[0033] Figure 1 This is a flowchart of the multimodal fusion prediction architecture method based on a pre-trained large model for power grid enterprises, which is the subject of this invention. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, a multimodal fusion prediction architecture method based on a pre-trained large model for power grid enterprises includes the following steps:

[0036] S1: Construction and Preprocessing of Multi-Source Heterogeneous Datasets

[0037] Power grid asset data exhibits significant big data characteristics, and its complexity and unique features place higher demands on data analysis methods. Multi-source heterogeneity is the most prominent feature of power grid asset data. This data originates from various stages throughout the asset's lifecycle, including planning and design, equipment procurement, engineering construction, operation and maintenance, and decommissioning. From the perspective of the systems generating the data, it is mainly dispersed across multiple information platforms such as Production Management System (PMS), Energy Management System (EMS), Distribution Automation System (DAS), meteorological monitoring system, and financial management system. This multi-source heterogeneity leads to significant differences in data format, standards, and quality, requiring specialized cleaning, transformation, and integration processing to provide a unified and standardized data foundation for subsequent analysis and modeling.

[0038] S1.1: Taking typical power grid assets such as transformers and cables as research objects, this paper constructs a large-scale model for depreciation calculation and remaining life prediction. By integrating multi-source heterogeneous data such as asset ledgers, operational data, maintenance records, and asset depreciation records, a pre-trained large-scale model architecture with parallel time-series and text encoding is designed to achieve deep fusion of multimodal features and inject knowledge from the power grid depreciation domain, ultimately achieving accurate prediction of equipment remaining life and health status assessment.

[0039] Equipment static data acquisition: including basic attribute data such as equipment model, years of operation, original value, and technical parameters;

[0040] Dynamic operation data acquisition: including time-series monitoring data such as load rate, operating temperature, vibration data, and fault records;

[0041] Maintenance record text data processing: Key information extraction and standardized summary generation are performed using a domain-fine-tuned large model;

[0042] Financial depreciation data integration: including historical depreciation records, maintenance costs, energy efficiency indicators, and other economic data;

[0043] S1.2: Data preprocessing. For numerical time-series data, missing value imputation, outlier detection and correction are performed, and parameters of different dimensions are normalized to scale them to the range of 0 to 1. For text data, a domain-adjusted large model is used to extract key information and generate summaries, transforming unstructured maintenance reports into standardized, time-series text descriptions.

[0044] The development of multi-objective optimization algorithms has evolved from traditional scalarization methods to modern metaheuristic algorithms. Early research mainly transformed multi-objective problems into single-objective problems through weighted summation and ε-constraints, but these had limitations such as the inability to fully capture the Pareto front and reliance on prior knowledge. In the early 21st century, multi-objective evolutionary algorithms, represented by NSGA-II, SPEA2, and MOEA / D, achieved breakthroughs. Through mechanisms such as non-dominated sorting, crowding calculation, and decomposition-aggregation, they achieved the acquisition of a well-distributed approximate Pareto solution set in a single run, becoming the mainstream research direction. In the power grid asset depreciation calculation model proposed in this patent, although a multi-objective optimization framework is not explicitly adopted, its core design concept embodies the idea of ​​weight-based multi-objective optimization. The model training process essentially balances three key objectives: maximizing prediction accuracy (ensuring the accuracy of remaining lifetime prediction), modal fusion effectiveness (ensuring the alignment quality of text and time-series features), and compliance with business rules (ensuring that the output conforms to depreciation accounting standards). These objectives achieve implicit weight allocation through the careful design of the loss function and balanced sampling of the training data. The final model is the optimal balance solution obtained among these mutually constraining objectives, which satisfies the accuracy requirements at the technical level and ensures the practical value at the business level.

[0045] S1.2.1: Normalization:

[0046]

[0047] Where x is the original data, min(X) and max(X) are the minimum and maximum values ​​of the parameter sequence, respectively, and x′ is the normalized value;

[0048] S1.2.2: Objective function:

[0049] min F(x)=w1f1(x)+w2f2(x)+w3f3(x)

[0050] Where f1(x) is the technical objective function, f2(x) is the economic objective function, f3(x) is the environmental objective function, and w1, w2, and w3 are weighting coefficients.

[0051] S2: Parallel Encoding and Fusion of Multimodal Features

[0052] The output feature vectors of the temporal encoding path and the text encoding path are fused based on multi-source heterogeneous datasets, and positional information is injected. The fused multimodal sequence is input into the pre-trained GPT architecture, which contains multiple TransformerBlocks. After deep feature interaction and information extraction, the remaining lifetime is predicted through the output linear layer.

[0053] S2.1: Temporal coding path: For various numerical sequences, a channel-independent strategy is adopted to process each variable separately. The input is optimized by block division and reversible instance normalization techniques. Then, each sequence block is mapped to a high-dimensional feature vector through a linear embedding layer.

[0054] S2.2: Text Encoding Path: For maintenance records and defect description texts, use the BERT encoder to extract text features, and then project them through a linear layer and the ReLU activation function to a vector space with the same dimension as the temporal features to achieve modality alignment;

[0055] S2.3: Feature Fusion: The feature vectors output by the two paths are added element by element and fused together, and position embedding information is injected to form a fused multimodal sequence.

[0056] S3: Synchronous prediction of remaining lifetime based on fusion features

[0057] The fused multimodal high-dimensional feature sequence is input into a Transformer-based pre-trained model for deep information interaction and feature extraction. The output layer synchronously predicts the continuous remaining lifetime value of the equipment and performs health status assessment. This step fully utilizes the powerful sequence modeling and contextual understanding capabilities of large models to mine the inherent patterns and long-term trends of equipment performance degradation from the fused features. It enables accurate inference of the remaining economic and technical lifetime of assets from multi-dimensional, non-linear correlated data, overcoming the limitations of traditional methods that rely on single-dimensional or linear assumptions.

[0058] S3.1: Deep Feature Extraction of Multimodal Sequences: The fused multimodal sequences are input into a pre-trained model based on the Transformer architecture. This model performs deep encoding and feature transformation on the input sequences through its multiple Transformer Blocks. Each Transformer Block sequentially performs multi-head self-attention calculation, residual connection and layer normalization, feedforward neural network processing, and residual connection and layer normalization again. Through this stacked deep structure, the high-level semantics and complex patterns in the input information are gradually abstracted and refined.

[0059] S3.2: Long-term dependency capture based on attention mechanism: The multi-head self-attention mechanism is the core of the model to realize deep information interaction. This mechanism first linearly transforms the input sequence into three matrices: query, key, and value. Then, in each attention head, the attention weight distribution is obtained by calculating the dot product of the query vector and all key vectors, and normalizing it by scaling and the Softmax function. Finally, the value vector is weighted and summed using this weight distribution to obtain the output of the attention head. By computing multiple attention heads in parallel and concatenating and linearly transforming their outputs, the model can capture complex, long-term dependencies and dynamic interaction patterns between arbitrary positions within the sequence from different subspaces.

[0060] S3.3: Synchronous Prediction Output of Remaining Lifetime and Health Status: The final sequence representation after processing through multiple Transformer Blocks aggregates the deep features of the input multimodal data. The model decodes this sequence representation through a specific output layer (usually a linear layer), mapping it to a continuous scalar value, which is the predicted remaining lifetime of the device. At the same time, the model architecture is designed to be scalable, adding a parallel classification output head to classify and evaluate the health status of the device based on the same deep features, outputting discrete status labels such as "healthy", "attention", "abnormal", and "severe", thereby achieving simultaneous output of quantitative prediction of remaining lifetime and qualitative assessment of health status.

[0061] S4: Equipment Depreciation Status Identification Based on Rule-Based and Model-Based Collaboration

[0062] Equipment depreciation status identification based on rule and model collaboration: The remaining useful life predicted by the model is input into a large language model that has been fine-tuned with domain knowledge. This model, combined with the threshold rules set by the power grid company's fixed asset classification and depreciation period standards, automatically identifies and warns of equipment that is nearing or has reached its service life. This step achieves an organic combination of data-driven prediction and domain knowledge rules, not only providing quantitative remaining useful life indicators, but also further transforming them into depreciation status judgments and decision-making suggestions that conform to enterprise management practices and regulatory requirements, thus completing a closed loop from data perception to intelligent decision-making.

[0063] S4.1: Structuring the input of prediction results and domain knowledge: The remaining life of the equipment predicted in step S3, the basic information of the equipment (including equipment classification code, asset name, commissioning date, and design life), and the relevant depreciation policy provisions are combined to form a structured natural language prompt; this prompt template aims to clearly present the predicted facts and the question to be judged to the large language model.

[0064] S4.2: Depreciation status rule matching based on fine-tuned large language model: The above structured prompts are input into a large language model that has been fine-tuned in depth with knowledge of power grid depreciation. The model encodes the complex depreciation rules and judgment logic defined in industry standard documents in its internal parameters. After receiving the input, the model starts its embedded reasoning mechanism, compares the predicted remaining life with the statutory depreciation period or economic life corresponding to the equipment category, and makes a status judgment based on the preset, configurable multi-level threshold rules.

[0065] S4.3: Depreciation Status Determination and Structured Early Warning Report Generation: The threshold rules include at least two levels: The first level is the early warning threshold, where the model determines the equipment to be in a "nearing the end of its service life" state when the remaining useful life of the equipment is less than or equal to a certain percentage (e.g., 20%, which is adjustable) of the statutory depreciation period; the second level is the critical threshold, where the model determines the equipment to be in a "reached or exceeded service life" state when the remaining useful life of the equipment is less than or equal to a smaller percentage (e.g., 5%, which is adjustable) or has reached zero / negative values. Based on the matching results of these rules, the model automatically identifies a list of equipment that meets the conditions and generates a structured early warning report containing key information such as the equipment's unique identifier, current predicted remaining useful life, determined depreciation status, and recommended measures (e.g., strengthening monitoring, arranging assessments, and planning replacement), for asset management personnel to make decisions.

[0066] This invention proposes a large-scale model-based method for calculating the depreciation and remaining useful life of power grid assets, aiming to address the problems of traditional depreciation methods relying on a single time dimension and neglecting the actual condition of equipment and the nonlinear decay of value. This method integrates multi-source heterogeneous data, including equipment ledgers, operational monitoring, defect records, environmental data, and financial information, to construct a unified data view. It employs a pre-trained large-scale model architecture with parallel temporal and textual encoding to achieve deep fusion and alignment of multimodal features. A deep attention mechanism based on Transformer captures the long-term dependencies in equipment state evolution, simultaneously outputting remaining useful life predictions and health status assessments. Finally, a domain-knowledge-tuned large language model, combined with depreciation rules and threshold settings, enables automatic identification and early warning of equipment depreciation status.

[0067] This invention establishes an intelligent depreciation management model that combines data-driven and domain knowledge collaboration, significantly improving the accuracy of remaining useful life prediction and the scientific nature of depreciation status judgment, and providing reliable technical support for power grid companies' lean asset management, accurate cost accounting, and investment decisions.

[0068] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent substitutions. As long as they meet the purpose of the invention and do not deviate from the technical principle and inventive concept of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A method for calculating the depreciation and predicting the remaining useful life of power grid assets based on a large model, characterized in that, Includes the following steps: S1: Construction and preprocessing of multi-source heterogeneous datasets: Systematically integrate multi-source heterogeneous data generated by typical power grid assets throughout their entire lifecycle, and perform data cleaning, alignment and normalization preprocessing to build a unified data view at the equipment level; S2: Parallel Encoding and Fusion of Multimodal Features: Construct a pre-trained large model architecture for parallel encoding of time series and text, extract features from numerical time series data and text description data respectively, and align and fuse feature vectors of different modalities in a shared semantic space; S3: Synchronous prediction of remaining lifetime based on fused features: The fused multimodal high-dimensional feature sequence is input into the Transformer-based pre-trained model for deep information interaction and feature extraction. The continuous remaining lifetime value of the device is predicted synchronously through the output layer and the health status is assessed. S4: Equipment depreciation status identification based on rule and model collaboration: The remaining life predicted by the model is input into a large language model that has been fine-tuned by domain knowledge, so that it can automatically identify and warn of equipment that is approaching or has reached its service life by combining the threshold rules set by the fixed asset classification and depreciation period standards of the power grid enterprise.

2. The method according to claim 1, characterized in that, The construction and preprocessing of the multi-source heterogeneous dataset in step S1 specifically includes: Power grid asset data exhibits significant big data characteristics, and its complexity and unique features place higher demands on data analysis methods. Multi-source heterogeneity is the most prominent feature of power grid asset data. This data originates from various stages throughout the asset's lifecycle, including planning and design, equipment procurement, engineering construction, operation and maintenance, and decommissioning. From the perspective of the systems generating the data, it is mainly dispersed across multiple information platforms such as Production Management System (PMS), Energy Management System (EMS), Distribution Automation System (DAS), meteorological monitoring system, and financial management system. This multi-source heterogeneity leads to significant differences in data format, standards, and quality, requiring specialized cleaning, transformation, and integration processing to provide a unified and standardized data foundation for subsequent analysis and modeling. S1.1: Taking typical power grid assets such as transformers and cables as research objects, this paper constructs a large-scale model for depreciation calculation and remaining life prediction. By integrating multi-source heterogeneous data such as asset ledgers, operational data, maintenance records, and asset depreciation records, a pre-trained large-scale model architecture with parallel time-series and text encoding is designed to achieve deep fusion of multimodal features and inject knowledge from the power grid depreciation domain, ultimately achieving accurate prediction of equipment remaining life and health status assessment. Equipment static data acquisition: including basic attribute data such as equipment model, years of operation, original value, and technical parameters; Dynamic operation data acquisition: including time-series monitoring data such as load rate, operating temperature, vibration data, and fault records; Maintenance record text data processing: Key information extraction and standardized summary generation are performed using a domain-fine-tuned large model; Financial depreciation data integration: including historical depreciation records, maintenance costs, energy efficiency indicators, and other economic data; S1.2: Data preprocessing. For numerical time-series data, missing value imputation, outlier detection and correction are performed, and parameters of different dimensions are normalized to scale them to the range of 0 to 1. For text data, a domain-adjusted large model is used to extract key information and generate summaries, transforming unstructured maintenance reports into standardized, time-series text descriptions. The development of multi-objective optimization algorithms has evolved from traditional scalarization methods to modern metaheuristic algorithms. Early research mainly transformed multi-objective problems into single-objective problems through weighted summation and ε-constraints, but these had limitations such as the inability to fully capture the Pareto front and reliance on prior knowledge. Breakthroughs have been achieved with multi-objective evolutionary algorithms represented by NSGA-II, SPEA2, and MOEA / D. Through mechanisms such as non-dominated sorting, crowding calculation, and decomposition-aggregation, they have achieved the acquisition of a well-distributed approximate Pareto solution set in a single run, becoming the mainstream research direction. In the power grid asset depreciation calculation model proposed in this patent, although a multi-objective optimization framework is not explicitly adopted, its core design concept embodies the idea of ​​weight-based multi-objective optimization. The model training process essentially balances three key objectives: maximizing prediction accuracy (ensuring the accuracy of remaining lifetime prediction), modal fusion effectiveness (ensuring the alignment quality of text and time-series features), and compliance with business rules (ensuring the output conforms to depreciation accounting standards). These objectives achieve implicit weight allocation through the careful design of the loss function and balanced sampling of the training data. The final model is the optimal balance solution obtained among these mutually constraining objectives, which satisfies the accuracy requirements at the technical level and ensures the practical value at the business level. S1.2.1: Normalization: Where x is the original data, min(X) and max(X) are the minimum and maximum values ​​of the parameter sequence, respectively, and x′ is the normalized value; S1.2.2: Objective function: min F(x)=w1f1(x)+w2f2(x)+w3f3(x) Where f1(x) is the technical objective function, f2(x) is the economic objective function, f3(x) is the environmental objective function, and w1, w2, and w3 are weighting coefficients.

3. The method according to claim 1, characterized in that, The parallel encoding and fusion of multimodal features in step S2 is specifically implemented as follows: The output feature vectors of the temporal encoding path and the text encoding path are fused based on multi-source heterogeneous datasets, and positional information is injected. The fused multimodal sequence is input into the pre-trained GPT architecture, which contains multiple TransformerBlocks. After deep feature interaction and information extraction, the remaining lifetime is predicted through the output linear layer. S2.1: Temporal coding path: For various numerical sequences, a channel-independent strategy is adopted to process each variable separately. The input is optimized by block division and reversible instance normalization techniques. Then, each sequence block is mapped to a high-dimensional feature vector through a linear embedding layer. S2.2: Text Encoding Path: For maintenance records and defect description texts, use the BERT encoder to extract text features, and then project them through a linear layer and the ReLU activation function to a vector space with the same dimension as the temporal features to achieve modality alignment; S2.3: Feature Fusion: The feature vectors output by the two paths are added element by element and fused together, and position embedding information is injected to form a fused multimodal sequence.

4. The method according to claim 3, characterized in that, Step S3 specifically includes: The fused multimodal high-dimensional feature sequence is input into a Transformer-based pre-trained model for deep information interaction and feature extraction. The output layer synchronously predicts the continuous remaining lifetime value of the equipment and performs health status assessment. This step fully utilizes the powerful sequence modeling and contextual understanding capabilities of large models to mine the inherent patterns and long-term trends of equipment performance degradation from the fused features. It enables accurate inference of the remaining economic and technical lifetime of assets from multi-dimensional, non-linear correlated data, overcoming the limitations of traditional methods that rely on single-dimensional or linear assumptions. S3.1: Deep Feature Extraction of Multimodal Sequences: The fused multimodal sequences are input into a pre-trained model based on the Transformer architecture. This model performs deep encoding and feature transformation on the input sequences through its multiple Transformer Blocks. Each Transformer Block sequentially performs multi-head self-attention calculation, residual connection and layer normalization, feedforward neural network processing, and residual connection and layer normalization again. Through this stacked deep structure, the high-level semantics and complex patterns in the input information are gradually abstracted and refined. S3.2: Long-term dependency capture based on attention mechanism: The multi-head self-attention mechanism is the core of the model to realize deep information interaction. This mechanism first linearly transforms the input sequence into three matrices: query, key, and value. Then, in each attention head, the attention weight distribution is obtained by calculating the dot product of the query vector and all key vectors, and normalizing it by scaling and the Softmax function. Finally, the value vector is weighted and summed using this weight distribution to obtain the output of the attention head. By computing multiple attention heads in parallel and concatenating and linearly transforming their outputs, the model can capture complex, long-term dependencies and dynamic interaction patterns between arbitrary positions within the sequence from different subspaces. S3.3: Synchronous Prediction Output of Remaining Lifetime and Health Status: The final sequence representation after processing through multiple Transformer Blocks aggregates the deep features of the input multimodal data. The model decodes this sequence representation through a specific output layer (usually a linear layer), mapping it to a continuous scalar value, which is the predicted remaining lifetime of the device. At the same time, the model architecture is designed to be scalable, adding a parallel classification output head to classify and evaluate the health status of the device based on the same deep features, outputting discrete status labels such as "healthy", "attention", "abnormal", and "severe", thereby achieving simultaneous output of quantitative prediction of remaining lifetime and qualitative assessment of health status.

5. The method according to claim 1, characterized in that, Step S4 specifically includes: Equipment depreciation status identification based on rule and model collaboration: The remaining useful life predicted by the model is input into a large language model that has been fine-tuned with domain knowledge. This model, combined with the threshold rules set by the power grid company's fixed asset classification and depreciation period standards, automatically identifies and warns of equipment that is nearing or has reached its service life. This step achieves an organic combination of data-driven prediction and domain knowledge rules, not only providing quantitative remaining useful life indicators, but also further transforming them into depreciation status judgments and decision-making suggestions that conform to enterprise management practices and regulatory requirements, thus completing a closed loop from data perception to intelligent decision-making. S4.1: Structuring the input of prediction results and domain knowledge: The remaining life of the equipment predicted in step S3, the basic information of the equipment (including equipment classification code, asset name, commissioning date, and design life), and the relevant depreciation policy provisions are combined to form a structured natural language prompt; this prompt template aims to clearly present the predicted facts and the question to be judged to the large language model. S4.2: Depreciation status rule matching based on fine-tuned large language model: The above structured prompts are input into a large language model that has been fine-tuned in depth with knowledge of power grid depreciation. The model encodes the complex depreciation rules and judgment logic defined in industry standard documents in its internal parameters. After receiving the input, the model starts its embedded reasoning mechanism, compares the predicted remaining life with the statutory depreciation period or economic life corresponding to the equipment category, and makes a status judgment based on the preset, configurable multi-level threshold rules. S4.3: Depreciation Status Determination and Structured Early Warning Report Generation: The threshold rules include at least two levels: the first level is the early warning threshold, when the remaining useful life of the equipment is less than or equal to a certain percentage of the statutory depreciation period, the model determines that the equipment is in the "nearing the end of its service life" state; the second level is the critical threshold, when the remaining useful life of the equipment is less than or equal to a smaller percentage or has reached zero / negative values, the model determines that the equipment is in the "reached or exceeded service life" state; based on the matching results of this rule, the model automatically identifies a list of equipment that meets the conditions and generates a structured early warning report containing key information such as the equipment's unique identifier, current predicted remaining useful life, determined depreciation status, and recommended measures (such as strengthening monitoring, arranging assessments, and planning replacement), for asset management personnel to make decisions.

Images