Intelligent prediction method and system for whole life cycle cost of power grid equipment

Abstract

The invention provides an intelligent prediction method and system for the whole life cycle cost of power grid equipment, and belongs to the technical field of power grid asset management and artificial intelligence. The method comprises the following steps: constructing a multi-dimensional original data set; performing preprocessing and mechanism characteristic generation on the original data, and constructing a space-time atlas sequence; constructing an STA-STGCN dynamic deep learning model, and performing training and hyper-parameter optimization on the model by using a historical space-time atlas sequence; and inputting a space-time atlas sequence of to-be-predicted equipment into the trained STA-STGCN dynamic deep learning model, outputting a classification cost prediction value and a total prediction value of a specified period in the future, and generating a confidence interval and a risk level of a prediction result through a Monte Carlo Dropout technology. According to the method, prediction with higher precision, higher robustness and better decision value is realized, and lean management requirements of power grid assets are met.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent prediction of the full life cycle cost of power grid equipment, belonging to the fields of power grid asset management and artificial intelligence technology. Background Technology

[0002] In power grid companies, key equipment such as transformers and circuit breakers are core assets, making their life-cycle cost management crucial. Accurate LCC forecasting is the scientific basis for formulating optimal maintenance strategies and making decisions on technical upgrades and major overhaul projects. Existing LCC forecasting methods for power grid equipment mainly suffer from the following limitations: Limited data dimensions: Most rely on traditional financial and ledger data, failing to effectively integrate real-time and near-real-time equipment status data from production systems, online monitoring, and meteorological environments, easily leading to a disconnect between cost forecasting and changes in equipment status. Static and rigid models: Traditional models (such as parameter estimation methods and simple regression models) or some early machine learning models (such as BP neural networks) are fixed once established, unable to adaptively adjust to real-time degradation of equipment status or changes in the operating environment. Lack of uncertainty assessment: At the forecast output level: The forecast result is usually a single, definite value, unable to quantify the uncertainty of the forecast itself, making it difficult for decision-makers to assess forecast risks and formulate robust strategies. In recent years, although some studies have attempted to introduce intelligent algorithms into LCC prediction, such as using genetic algorithms to optimize support vector machines or employing fuzzy theory to handle data uncertainty, these methods have limited ability to handle the inherent spatiotemporal correlations of power grid equipment (e.g., the failure of one component can trigger stress changes and costs in related components) and long-sequence time dependencies. Therefore, this invention aims to propose a novel solution to overcome the above limitations.

[0003] As the core assets of the power system, the life-cycle cost management of power grid equipment directly affects the economic benefits and power supply reliability of power grid companies. Life-cycle cost (LCC) covers the costs of equipment from design and manufacturing, commissioning and operation to decommissioning and scrapping, including initial investment costs, operation and maintenance costs, repair costs, failure costs and energy consumption costs. Accurate forecasting is a key prerequisite for achieving lean management of power grid assets.

[0004] Existing power grid equipment forecasting methods have the following core limitations, making it difficult to meet the needs of power grid asset management under the new circumstances: (1) Fragmented data utilization: Existing methods rely heavily on static data such as financial ledgers and basic operation and maintenance records, and fail to effectively integrate dynamic data such as online monitoring (e.g., DGA dissolved gas analysis, vibration monitoring), meteorological environment (temperature, pollution level), and equipment health status (insulation aging degree). Furthermore, the data lacks structured correlation and cannot reflect the physical mechanism of "component failure-cost transmission", resulting in the prediction results being out of sync with the actual equipment status.

[0005] (2) Static and rigid model architecture: Traditional prediction models (such as parameter estimation and regression analysis) or early machine learning models (such as BP neural network and basic support vector machine) adopt a fixed architecture, which can only capture single-dimensional features (such as only time series or only spatial correlation). They cannot dynamically adjust model parameters according to equipment type (such as transformer and circuit breaker) and life cycle stage (initial operation and aging period). They have poor adaptability and the long-term prediction accuracy decays rapidly.

[0006] (3) Lack of uncertainty quantification: Existing methods mostly output a single definite value. Even if some studies introduce fuzzy theory and genetic algorithms to deal with uncertainty, they only address the noise of the input data and do not consider the prediction risks caused by model generalization error and sudden changes in equipment state (such as short circuit impact). Decision-makers cannot assess the reliability of the prediction results and find it difficult to formulate a sound asset management strategy.

[0007] (4) Insufficient interpretability and decision support: Most existing intelligent algorithms are "black box models" that cannot locate cost-driven key components (such as increased maintenance costs due to bushing aging) and key time nodes (such as a surge in fault costs after a short circuit event). The output results can only be used for cost estimation and cannot directly guide refined decision-making such as component-level maintenance and targeted cost control.

[0008] (5) Industry adaptability limitations: Due to the lack of long-term historical data, existing models cannot quickly achieve accurate prediction for newly commissioned equipment. At the same time, existing methods mostly focus on single equipment prediction and ignore the correlation between equipment in the power grid system (such as transformer failure leading to a surge in load and increased costs of associated circuit breakers), thus limiting the applicable scenarios.

[0009] In recent years, although some studies have attempted to introduce deep learning technology into prediction, the integrated problems of "cooperative extraction of spatiotemporal features," "dynamic adaptation," "risk quantification," and "cross-scenario transfer" have not yet been solved. Therefore, this invention aims to propose an intelligent prediction scheme that integrates multi-source data structured representation, dynamic deep learning models, and closed-loop system evolution capabilities, thereby breaking through the bottlenecks of existing technologies. Summary of the Invention

[0010] The purpose of this invention is to provide an intelligent prediction method and system for the full life cycle cost of power grid equipment, overcome the shortcomings of existing power grid equipment prediction methods, achieve higher accuracy, stronger robustness, and better decision value prediction, and meet the needs of lean management of power grid assets.

[0011] To achieve the above objectives, the present invention employs the following technical solution: A method for intelligent prediction of the life-cycle cost of power grid equipment includes the following steps: Collect basic attribute data, dynamic status data, cost and operation and maintenance data, and environmental and system-related data of power grid equipment to construct a multi-dimensional raw dataset; The raw data is preprocessed and mechanistic features are generated to construct a three-level spatiotemporal map sequence of equipment, components, and systems; A dynamic deep learning model STA-STGCN, comprising an STGC module, an Attention-LSTM module, and a prediction and uncertainty quantification layer, is constructed. The model is then trained and its hyperparameters are optimized using historical spatiotemporal graph sequences. The spatiotemporal map sequence of the device to be predicted is input into the trained STA-STGCN dynamic deep learning model, which outputs the categorical cost prediction value and the total prediction value for a specified future period, and generates the confidence interval and risk level of the prediction results through Monte Carlo Dropout technology.

[0012] Preferably, the basic attribute data is obtained from the PMS system, including equipment model, capacity, manufacturer, commissioning time, factory test report, and initial investment cost; The dynamic status data is obtained from the online monitoring system and PMS system, including load current, top oil temperature, winding hot spot temperature, vibration frequency, DGA characteristic gas concentration, SF6 gas pressure, and insulation resistance. The cost and maintenance data are obtained from the ERP system and PMS system, including historical maintenance records, repair work orders, fault types, repair or replacement parts records, and the corresponding maintenance costs, repair costs, fault costs, and energy consumption costs. The environmental and system-related data includes ambient temperature, humidity, pollution level, and the operating status and fault records of related equipment.

[0013] Preferably, the preprocessing includes using interpolation to complete missing data and removing outliers based on the 3σ criterion; The generation of the mechanism features includes calculating the equipment health score, fault transmission coefficient, absolute gas production rate of characteristic gas, and annual cumulative heavy load hours.

[0014] Preferably, the multidimensional spatiotemporal map is represented as follows: Each time slice ; in, It is a set of nodes, including device component nodes and associated device nodes, and each node carries a feature vector; This is a set of edges, representing the physical connections and cost transmission relationships between nodes; It is a set of edge weights, dynamically assigned based on device runtime, failure frequency, and state change rate.

[0015] Preferably, the STA-STGCN dynamic deep learning model includes an input layer, an STGC module, an Attention-LSTM module, and a prediction and uncertainty quantification layer; The STGC module is composed of graph convolutional layers and multi-scale temporal convolutional layers stacked alternately to collaboratively extract the spatial correlation features and multi-scale temporal evolution features of device components. The Attention-LSTM module includes a component attention branch and a time attention branch, which are used to focus on key components and key historical time nodes that have a significant impact on cost, respectively. The prediction and uncertainty quantification layer outputs categorized cost predictions and total predictions for a specified future period based on the fully connected layer, and outputs the probability distribution and confidence interval of the cost predictions based on Monte Carlo Dropout technology.

[0016] Preferably, the component attention branch maps the component features output by the STGC module to the same dimension, calculates the Pearson correlation coefficient between each component feature and the cost label, and uses it as the component attention weight. : , in, Let V be the feature vector of component V. Retain weight for cost tags >0.1 core components, Representing variables variance express and covariance; The temporal attention branch inputs the filtered core component features into the LSTM network to obtain the hidden state at each time step. Calculate the attention weights over time The calculation formula is as follows: , in, , Here are the trainable parameters, and T is the total number of time steps. Let be the feature vector of the k-th object; The context vector is obtained by weighted summation based on the hidden state and temporal attention weights. The key historical time nodes are identified by focusing on the time steps corresponding to key events based on the context vector.

[0017] Preferably, the STA-STGCN dynamic deep learning model is trained using HuberLoss as the loss function, trained through the AdamW optimizer, and its hyperparameters are tuned using Bayesian optimization; the specific details of HuberLoss are as follows: , in, This is the true cost value. To predict costs, This is a constant value, set to 1000.

[0018] Preferably, the prediction and uncertainty quantification layer specifically includes: performing Monte Carlo Dropout random forward propagation, calculating the mean and standard deviation of the prediction results, and determining the confidence interval; Risk levels are determined by the ratio of the standard deviation to the mean: a ratio of less than 5% indicates low risk; a ratio of 5% or greater but less than or equal to 15% indicates medium risk; and a ratio greater than 15% indicates high risk.

[0019] Preferably, the STA-STGCN dynamic deep learning model is trained using an incremental fine-tuning strategy, freezing the graph convolutional layer weights of the STGC module based on new detection data, and updating the parameters of the Attention-LSTM module and the fully connected layer.

[0020] A smart prediction system for the entire life cycle cost of power grid equipment, the system comprising: The data acquisition module is used to collect multi-source data from power grid equipment; The spatiotemporal mapping construction module is used to process data and construct a three-level spatiotemporal mapping sequence of devices, components, and systems. The model training and optimization module is used to build, train, and optimize the STA-STGCN dynamic deep learning model. The forecasting and risk quantification module is used to perform cost forecasting and output the forecast results and their uncertainty quantification information. The online learning and decision-making module is used to achieve closed-loop incremental learning of the model and generate decision support information; The storage module is used to store various types of data and model parameters involved in the system.

[0021] The advantages of this invention are: Significantly improved prediction accuracy and adaptability: By fusing multi-source mechanism data through spatiotemporal mapping, the STA-STGCN model captures spatiotemporal collaborative features and key focuses, reducing MAPE in type-specific cost prediction and improving accuracy compared to traditional support vector machines and basic GCN models; at the same time, the model's dynamic adaptation mechanism and transfer learning capabilities can cover multiple types of equipment such as transformers and circuit breakers, shortening the prediction and deployment cycle of new equipment.

[0022] Outstanding risk quantification and decision support capabilities: By using Monte Carlo Dropout to output predicted values, confidence intervals, risk levels, and sources of uncertainty, decision-makers can clearly assess predicted risks; combined with two-level attention weights, it can pinpoint cost-driven key components and time points, directly guiding component-level maintenance and targeted cost control.

[0023] The system exhibits strong robustness and evolutionary capabilities: the online learning closed-loop mechanism enables the model to continuously iterate as the equipment status changes, reducing the long-term prediction accuracy decay rate; at the same time, the lightweight design of the model can be adapted to the computing resources of power grid companies of different sizes, and can be efficiently deployed from small distribution companies to large State Grid companies. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0025] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0026] Figure 2 This is a schematic diagram of the implementation process of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 like Figures 1 to 2 As shown, a method for intelligent prediction of the life-cycle cost of power grid equipment includes the following steps: S1: Collect basic attribute data, dynamic status data, cost and operation and maintenance data, and environmental and system-related data of power grid equipment to construct a multi-dimensional raw dataset; S2: Preprocess the raw data and generate mechanism features to construct a three-level spatiotemporal map sequence of equipment-component-system; S3: Construct a dynamic deep learning model STA-STGCN that includes an STGC module, an Attention-LSTM module, and a prediction and uncertainty quantification layer. Train the model and optimize its hyperparameters using historical spatiotemporal graph sequences. S4: Input the spatiotemporal map sequence of the device to be predicted into the trained STA-STGCN dynamic deep learning model, output the categorical cost prediction value and total prediction value for a specified future period, and generate the confidence interval and risk level of the prediction results through Monte Carlo Dropout technology.

[0029] As a refinement of the above embodiments, the basic attribute data is obtained from the PMS system, including equipment model, capacity, manufacturer, commissioning time, factory test report, and initial investment cost. The dynamic status data is obtained from the online monitoring system and the PMS system, including load current, top oil temperature, winding hot spot temperature, vibration frequency, DGA characteristic gas concentration, SF6 gas pressure, and insulation resistance.

[0030] The cost and maintenance data are obtained from the ERP system and PMS system, including historical maintenance records, repair work orders, fault types, repair or replacement parts records, and the corresponding maintenance costs, repair costs, fault costs, and energy consumption costs.

[0031] The environmental and system-related data includes ambient temperature, humidity, pollution level, and the operating status and fault records of related equipment.

[0032] As a refinement of the above embodiments, the preprocessing includes using interpolation to complete missing data and removing outliers based on the 3σ criterion.

[0033] The generation of the mechanism features includes calculating the equipment health score, fault transmission coefficient, absolute gas production rate of characteristic gas, and annual cumulative heavy load hours.

[0034] As a refinement of the above embodiments, the multi-dimensional spatiotemporal map is represented as follows: Each time slice .

[0035] in, It is a set of nodes, with equipment components as the basic nodes (such as the windings, cores, and bushings of a transformer) and extending to related equipment nodes (such as circuit breakers that are matched with transformers). Each node carries a feature vector Xvt=[basic attribute features, dynamic state features, mechanism features, and cost features].

[0036] The edge set represents the physical connections and cost transmission relationships between nodes, including electrical connections (such as windings and bushings), mechanical linkages (such as circuit breaker operating mechanisms and arc-extinguishing chambers), and system connections (such as transformers and circuit breakers).

[0037] The edge weight set is adjusted based on equipment runtime, fault frequency, and state change rate. For example, after the bushing ages, its "fault transmission edge weight" with the winding increases from 0.3 to 0.8, reflecting the change in the correlation strength between components in actual operation.

[0038] As a refinement of the above embodiments, the STA-STGCN dynamic deep learning model includes an input layer, an STGC module, an Attention-LSTM module, and a prediction and uncertainty quantification layer.

[0039] (1) The input layer receives the above-mentioned spatiotemporal map sequence as the original input data of the model.

[0040] (2) The STGC module is composed of graph convolutional layers and multi-scale temporal convolutional layers stacked alternately to extract the spatial correlation features and multi-scale temporal evolution features of device components.

[0041] Graph convolutional layer: at each time slice Above, based on adjacency matrix and edge weights Aggregate the features of neighboring nodes and update the features of the current node, using the following formula: , in, Let I be the normalized self-adjacency matrix, and D be the identity matrix and degree matrix. For the node features of the l-th layer, For trainable weight matrix, This is the ReLU activation function.

[0042] Multi-scale temporal convolutional layer: Three convolutional kernels of different scales (window sizes 3, 5, and 7) are used to convolve the temporal features of each node, capturing the evolution patterns of short-term time series (such as monthly load fluctuations) and medium-to-long-term time series (such as quarterly insulation aging). By splicing and fusing multi-scale information through feature concatenation, the adaptability of the model to cost changes over different time spans is improved.

[0043] (3) The Attention-LSTM module includes a component attention branch and a time attention branch, which are used to focus on key components and key historical time nodes that have a significant impact on cost, respectively.

[0044] Component attention branch: Maps the features of each component output by STGC to the same dimension, calculates the Pearson correlation coefficient between each component feature and the cost label, and uses it as the component attention weight. The formula is as follows: , in, Let V be the feature vector of component V. Retain weight for cost tags Core components with a value >0.1 (such as transformer windings and bushings) are included in subsequent calculations; Representing variables variance express and The covariance.

[0045] Temporal attention branch: Input the filtered core component features into the LSTM network to obtain the hidden state at each time step. Calculate the attention weights over time The calculation formula is as follows: , in, , Here are the trainable parameters, and T is the total number of time steps. Let be the feature vector of the k-th object; The context vector is obtained by weighted summation. The focus is on the time steps corresponding to key events (such as short-circuit impacts and component replacements).

[0046] (4) The prediction and uncertainty quantification layer outputs the categorized cost prediction value and the total prediction value for a specified future period based on the fully connected layer, and outputs the probability distribution and confidence interval of the cost prediction value based on the Monte Carlo Dropout technique.

[0047] Predicted output: Input the context vector C into the fully connected layer to output the "type-specific cost prediction value" (operation and maintenance cost, maintenance cost, failure cost, energy consumption cost) and "total prediction value" for the next 12-36 months; Uncertainty Quantification: Based on Monte Carlo Dropout technology, Dropout is kept active during the prediction phase, and 1000 random forward propagations are performed to generate the probability distribution for cost prediction. Calculate the predicted mean ( (where is the predicted value from the i-th sampling), and is used as the final predicted value; Calculate the standard deviation ,by As a 95% confidence interval; Based on the confidence interval width and equipment status, risk levels are classified (low risk). Medium risk High risk And locate the sources of uncertainty (data noise / model generalization / state mutation).

[0048] As a refinement of the above embodiment, the STA-STGCN dynamic deep learning model is trained using HuberLoss as the loss function, trained through the AdamW optimizer, and its hyperparameters are tuned using Bayesian optimization; the specific details of HuberLoss are as follows: , in, This is the true cost value. To predict costs, This is a constant value, set to 1000.

[0049] Optimizer: The AdamW optimizer (weight decay coefficient λ=0.01) is used to avoid model overfitting.

[0050] Hyperparameter tuning: Bayesian optimization method is used to automatically optimize the number of graph convolutional layers (2-4 layers), the number of LSTM hidden units (64-256), the learning rate (1e-4-1e-3), and the kernel size.

[0051] Dynamic adaptation mechanism: Based on the equipment type (transformer / circuit breaker / interchange transformer), the initial range of hyperparameters is preset, and the model automatically adjusts the architecture without the need for manual redesign.

[0052] As a refinement of the above embodiments, the STA-STGCN dynamic deep learning model also employs an incremental fine-tuning strategy for training, freezing the graph convolutional layer weights of the STGC module based on new detection data, and updating the parameters of the Attention-LSTM module and the fully connected layer.

[0053] The incremental fine-tuning strategy is as follows: Triggering conditions: New data is detected weekly (such as DGA data, maintenance work orders, and fault records). Online learning is triggered when the amount of data reaches the time slice threshold (such as 1 month's data) or when the equipment status changes abruptly (such as a 20% surge in DGA gas concentration).

[0054] Incremental fine-tuning: The strategy of freezing the bottom layer and updating the top layer is adopted. The weights of the graph convolutional layer of the STGC module are frozen (to preserve the spatiotemporal feature extraction capability), and only the parameters of the Attention-LSTM module and the fully connected layer are updated. The new data is used to fine-tune with a low learning rate (1e-5) to reduce the computational cost.

[0055] Accuracy verification: After fine-tuning, the accuracy is verified by mean absolute percentage error (MAPE). If MAPE < 8%, the model is updated; otherwise, fine-tuning is retried and the learning rate is optimized.

[0056] To address the issue of insufficient data for new devices in cross-scenario transfer learning, the following strategy is adopted: Pre-training phase: Using historical data from 1000+ multi-type power grid equipment (500 transformers, 300 circuit breakers, and 200 instrument transformers), the basic STA-STGCN model is trained to learn the general laws governing the evolution of power grid equipment.

[0057] Fine-tuning phase: For newly commissioned equipment (such as 110kV circuit breakers), only 6 months of short-term data are needed. Freeze the STGC module of the basic model (retaining the ability to extract general spatiotemporal features), update the parameters of the attention layer and the fully connected layer, and achieve rapid adaptation.

[0058] Performance verification: After transfer learning, the model's MAPE improved by more than 25% compared to that without transfer learning, meeting the requirements for engineering applications (MAPE < 10%).

[0059] Example 2 A smart prediction system for the entire life cycle cost of power grid equipment, the system comprising: The data acquisition module is used to collect multi-source data from power grid equipment; The spatiotemporal mapping construction module is used to process data and construct a three-level spatiotemporal mapping sequence of devices, components, and systems. The model training and optimization module is used to build, train, and optimize the STA-STGCN dynamic deep learning model. The forecasting and risk quantification module is used to perform cost forecasting and output the forecast results and their uncertainty quantification information. The online learning and decision-making module is used to achieve closed-loop incremental learning of the model and generate decision support information; The storage module is used to store various types of data and model parameters involved in the system.

[0060] As a refinement of the above embodiments, the system also provides visual decision support. The system outputs a visual interface, including: Cost Forecast Dashboard: Displays the cost forecasts by category, the total forecast, and the 95% confidence interval for the next 12-36 months; Risk assessment dashboard: Displays risk level, sources of uncertainty and key risk points (such as "high maintenance cost risk due to aging of bushings"). Decision suggestion module: Based on the attention weight of key components, it automatically generates targeted suggestions (such as "prioritize the maintenance of bushings next quarter, which is expected to reduce maintenance costs by 30%").

[0061] Example 3 This disclosure also provides an intelligent prediction device for the life-cycle cost of power grid equipment, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the intelligent prediction method for the life-cycle cost of power grid equipment described in the above embodiments.

[0062] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0063] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes the program instructions / modules stored in the memory to perform functional applications and data processing, thereby realizing the intelligent prediction method for the full life-cycle cost of power grid equipment in the above embodiments.

[0064] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0065] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of the full life cycle cost of power grid equipment, characterized in that, Includes the following steps: Collect basic attribute data, dynamic status data, cost and operation and maintenance data, and environmental and system-related data of power grid equipment to construct a multi-dimensional raw dataset; The raw data is preprocessed and mechanistic features are generated to construct a three-level spatiotemporal map sequence of equipment, components, and systems; A dynamic deep learning model STA-STGCN, comprising an STGC module, an Attention-LSTM module, and a prediction and uncertainty quantification layer, is constructed. The model is then trained and its hyperparameters are optimized using historical spatiotemporal graph sequences. The spatiotemporal map sequence of the device to be predicted is input into the trained STA-STGCN dynamic deep learning model, which outputs the categorical cost prediction value and the total prediction value for a specified future period, and generates the confidence interval and risk level of the prediction results through Monte Carlo Dropout technology.

2. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 1, characterized in that, The basic attribute data is obtained from the PMS system, including equipment model, capacity, manufacturer, commissioning time, factory test report, and initial investment cost; The dynamic status data is obtained from the online monitoring system and PMS system, including load current, top oil temperature, winding hot spot temperature, vibration frequency, DGA characteristic gas concentration, SF6 gas pressure, and insulation resistance. The cost and maintenance data are obtained from the ERP system and PMS system, including historical maintenance records, repair work orders, fault types, repair or replacement parts records, and the corresponding maintenance costs, repair costs, fault costs, and energy consumption costs. The environmental and system-related data includes ambient temperature, humidity, pollution level, and the operating status and fault records of related equipment.

3. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 1, characterized in that, The preprocessing includes using interpolation to complete missing data and removing outliers based on the 3σ criterion; The generation of the mechanism features includes calculating the equipment health score, fault transmission coefficient, absolute gas production rate of characteristic gas, and annual cumulative heavy load hours.

4. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 1, characterized in that, The multidimensional spatiotemporal map is represented as follows: Each time slice ; in, For timestamps; It is a set of nodes, including device component nodes and associated device nodes, and each node carries a feature vector; This is a set of edges, representing the physical connections and cost transmission relationships between nodes; It is a set of edge weights, dynamically assigned based on device runtime, failure frequency, and state change rate.

5. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 1, characterized in that, The STA-STGCN dynamic deep learning model includes an input layer, an STGC module, an Attention-LSTM module, and a prediction and uncertainty quantification layer. The STGC module is composed of graph convolutional layers and multi-scale temporal convolutional layers stacked alternately to collaboratively extract the spatial correlation features and multi-scale temporal evolution features of device components. The Attention-LSTM module includes a component attention branch and a time attention branch, which are used to focus on key components and key historical time nodes that have a significant impact on cost, respectively. The prediction and uncertainty quantification layer outputs categorized cost predictions and total predictions for a specified future period based on the fully connected layer, and outputs the probability distribution and confidence interval of the cost predictions based on Monte Carlo Dropout technology.

6. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 5, characterized in that, The component attention branch maps the component features output by the STGC module to the same dimension, calculates the Pearson correlation coefficient between each component feature and the cost label, and uses it as the component attention weight. : ， in, Let V be the feature vector of component V. Retain weight for cost tags >0.1 core components, Representing variables variance express and covariance; The temporal attention branch inputs the filtered core component features into the LSTM network to obtain the hidden state at each time step. Calculate the attention weights over time The calculation formula is as follows: ， in, , Here are the trainable parameters, and T is the total number of time steps. Let be the feature vector of the k-th object; The context vector is obtained by weighted summation based on the hidden state and temporal attention weights. The key historical time nodes are identified by focusing on the time steps corresponding to key events based on the context vector.

7. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 5, characterized in that, The STA-STGCN dynamic deep learning model is trained using HuberLoss as the loss function, trained through the AdamW optimizer, and its hyperparameters are tuned using Bayesian optimization. The specific details of HuberLoss are as follows: ， in, This is the true cost value. To predict costs, This is a constant value, set to 1000.

8. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 5, characterized in that, The prediction and uncertainty quantification layer specifically includes: performing Monte Carlo Dropout random forward propagation, calculating the mean and standard deviation of the prediction results, and determining the confidence interval; Risk levels are determined by the ratio of the standard deviation to the mean: a ratio of less than 5% indicates low risk; a ratio of 5% or greater but less than or equal to 15% indicates medium risk; and a ratio greater than 15% indicates high risk.

9. The intelligent prediction method for the whole life cycle cost of power grid equipment according to claim 5, characterized in that, The STA-STGCN dynamic deep learning model also employs an incremental fine-tuning strategy for training, freezing the graph convolutional layer weights of the STGC module based on new detection data, and updating the parameters of the Attention-LSTM module and the fully connected layer.

10. A smart prediction system for the whole life cycle cost of power grid equipment, characterized in that, The system for implementing the method as described in any one of claims 1 to 9 comprises: The data acquisition module is used to collect multi-source data from power grid equipment; The spatiotemporal mapping construction module is used to process data and construct a three-level spatiotemporal mapping sequence of devices, components, and systems. The model training and optimization module is used to build, train, and optimize the STA-STGCN dynamic deep learning model. The forecasting and risk quantification module is used to perform cost forecasting and output the forecast results and their uncertainty quantification information. The online learning and decision-making module is used to achieve closed-loop incremental learning of the model and generate decision support information; The storage module is used to store various types of data and model parameters involved in the system.

Images