Equipment state trend maintenance decision-making system and method for new energy power station

Abstract

The invention discloses an equipment state trend prediction and maintenance decision-making system and method for a new energy power station, and the system is deployed based on a cloud-side-end collaborative architecture, and a full-process closed-loop logic of data governance, trend prediction, uncertainty calibration, life evaluation, causal analysis, maintenance optimization, feedback execution, and safety audit is constructed. Accurate sensing, trend pre-judgment and dynamic maintenance decision making of the equipment operation state are achieved. According to the system, multi-source heterogeneous operation data and a physical mechanism model are fused, a cooperative work mechanism of data input-model operation-decision output-feedback iteration is formed by means of core technologies such as space-time diagram prediction, causal reasoning and risk optimization, all the modules are in close linkage with a logic dependency relationship through a standardized data interface, and the reliability of the system is improved. According to the method, the professionality of a single-module function is ensured, the collaboration of the whole scheme is also realized, the active maintenance, risk self-adaption and economical efficiency optimization of new energy power station equipment are finally achieved, and the operation safety and the power generation efficiency of the equipment are remarkably improved.

Description

Technical Field

[0001] This application belongs to the field of new energy power technology, and specifically relates to a new energy power plant equipment status trend maintenance decision system and method. Background Technology

[0002] As a crucial component of modern power systems, the operational status of new energy power plants directly impacts the safety and economy of the power grid. With the continuous expansion of installed capacity for wind power, photovoltaics, and energy storage devices, the number and types of equipment are becoming increasingly complex, and the operating environment is characterized by wide distribution, multiple data dimensions, and high real-time requirements. How to perceive the operational status of equipment through multi-source data and identify potential risks in advance has become an important research direction for new energy operation and maintenance management.

[0003] In existing technologies, equipment condition monitoring and prediction largely rely on machine learning models or empirical rules based on historical features. For example, some systems use SCADA data to perform statistical analysis or short-term trend prediction of equipment failures, but their prediction accuracy is limited when dealing with multi-source heterogeneous data, physical constraints, and nonlinear degradation characteristics. Another type of method uses traditional threshold diagnosis or life estimation models for health assessment, lacking comprehensive consideration of environmental changes, task planning, and risk costs. Furthermore, maintenance decisions are mostly executed in a static planning manner, making it difficult to dynamically adjust based on prediction results and operational status. This leads to over-maintenance or delayed maintenance of some equipment, thereby affecting power generation efficiency and safety reliability.

[0004] Based on the above situation, there is an urgent need to build a comprehensive system that integrates data-driven and mechanism-constrained approaches. By conducting multi-dimensional analysis and trend prediction of equipment operation data, the system can identify the laws of state evolution and health degradation trends. Combined with uncertainty assessment and causal reasoning mechanisms, it can form a quantifiable and interpretable basis for maintenance decisions, thereby solving the problems of insufficient prediction accuracy, delayed maintenance plans, and uncertainty in risk decisions for existing new energy power plant equipment. Summary of the Invention

[0005] The purpose of this application is to provide a maintenance decision-making system and method for equipment status trends in new energy power plants. This addresses the problems of insufficient equipment prediction accuracy, delayed maintenance plans, and uncertain risk decisions in the prior art, as mentioned in the background section.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a system for predicting equipment status trends and making maintenance decisions for new energy power plants is provided, including: The data governance and feature extraction module is used to collect and process multi-source data from SCADA, meteorological monitoring and power grid operation plans. It cleans, synchronizes, detects anomalies and extracts features from the data to form a unified structured feature sequence, providing standardized data input for subsequent modules. The spatiotemporal graph prediction module with physical prior constraints receives the structured feature sequence output by the data governance and feature extraction module, which is used to establish a heterogeneous spatiotemporal graph model containing equipment nodes, feeder nodes and meteorological nodes. It combines graph attention network and temporal convolutional network to realize the prediction of operating status trend, and introduces physical consistency conditions in model training to output the equipment operating status trend prediction results. The uncertainty quantification module receives the prediction results from the spatiotemporal graph prediction module with physical prior constraints, and uses them to generate confidence intervals for the prediction results based on deep integration of multiple sub-models and random dropout mechanism. After calibration, a credible health trend result is formed, providing a reliable basis with confidence for subsequent evaluation and decision-making. The life assessment and degradation analysis module receives the reliable health trend results output by the uncertainty quantification module, and uses them to calculate the degradation rate and remaining service life based on the degradation model of different equipment types, generate life distribution information, and clarify the equipment health status and maintenance priority. The causal reasoning and counterfactual analysis module receives the life distribution information from the life assessment and degradation analysis module, and uses it to construct a causal relationship diagram between key variables of the equipment, perform counterfactual simulation assessment of maintenance operations, quantify the impact of different actions on the evolution of health status, and provide causal support for maintenance strategy optimization. The opportunity window identification and maintenance optimization module receives the quantification results of the maintenance action impact from the causal reasoning and counterfactual analysis module, generates opportunity windows by combining external factors such as power generation load and meteorological conditions, and outputs a maintenance scheduling plan based on the risk optimization model. The execution and feedback module receives the maintenance scheduling plan from the opportunity window identification and maintenance optimization module, and uses it to issue maintenance tasks, monitor the execution progress, collect equipment operation data after maintenance and feed it back to the data governance and feature extraction module. At the same time, it performs model retraining and rolling calibration based on the input data distribution drift to achieve closed-loop iteration. The security and audit module runs through the entire operation of all modules and is used to implement system access control, data anonymization, logging, and abnormal operation alarms to ensure the security and traceability of data interaction and function execution of each module.

[0007] In one possible implementation, the data governance and feature extraction module detects outliers by jointly using the median absolute deviation algorithm and the isolated forest algorithm, repairs short-period missing data by time interpolation, and completes long-period missing data by adjacent device regression. Based on a sliding time window, features in the time domain, frequency domain, and energy domain are extracted, and a multi-dimensional feature vector containing operational state statistics and vibration spectrum features is output. The feature vector is synchronized to the spatiotemporal graph prediction module with physical prior constraints through a standardized interface, serving as the core input data for model training and inference.

[0008] In one possible implementation, the node types of the spatiotemporal map prediction module include wind turbine nodes, photovoltaic inverter nodes, energy storage cluster nodes, and meteorological nodes. The edge weights consist of physical coupling weights, statistical correlation weights, and planning correlation weights, with the weight ratios of the three being 0.6, 0.3, and 0.1, respectively. Among them, the physical coupling weight is determined based on the electrical topology and geographical layout, the statistical correlation weight is calculated based on mutual information and Granger causality, and the planned correlation weight is generated based on grid connection permits and maintenance plans. The physical prior constraints include wind turbine drivetrain efficiency constraints, photovoltaic inverter thermal balance constraints, and energy storage unit equivalent circuit constraints. During the training phase, the prediction model incorporates physical constraint errors as a penalty term into the loss function. Through the constraint learning mechanism, the prediction results are kept consistent with the actual operating conditions of the equipment. The output trend prediction results are directly used as the input data source for the uncertainty quantification module.

[0009] In one possible implementation, the uncertainty quantification module consists of an integrated system of deep learning sub-models initialized with multiple different parameters. During the inference phase, multiple prediction distributions are generated using random dropout technology, and confidence intervals are calibrated using an order-preserving regression algorithm to ensure that the prediction interval coverage is not less than 0.9 and the interval width is controlled within a set range. The reliable health trend results output by the module include predicted values ​​and corresponding confidence intervals, which are synchronized to the life assessment and degradation analysis module to provide input data with risk boundaries for equipment degradation rate calculation and remaining life assessment.

[0010] In one possible implementation, the life assessment and degradation analysis module includes a multi-model evaluation mechanism for different equipment types, wherein the wind turbine gearbox uses the Gamma degradation process model to calculate the degradation rate, the photovoltaic inverter power module estimates fatigue life based on temperature cycle statistics, and the energy storage battery cluster uses a parameter-adaptive unscented Kalman filter method to achieve dynamic life prediction. The health index and remaining lifespan distribution information output by the module serve as the core basis for the causal reasoning and counterfactual analysis module to identify maintenance priorities and quantify maintenance effectiveness.

[0011] In one possible implementation, the causal reasoning and counterfactual analysis module strengthens the constraints of true causal paths and reduces spurious correlation interference during the training phase by using causal consistency regularization terms. The system uses structured causal graphs to identify key causal chains between maintenance operations, environmental factors, and equipment health indicators, and generates health change prediction results under different maintenance actions through counterfactual deduction during the simulation phase. The maintenance action impact quantification ranking results output by the module are directly input into the opportunity window identification and maintenance optimization module, providing interpretable support for the screening and optimization of maintenance strategies.

[0012] In one possible implementation, the opportunity window identification and maintenance optimization module comprehensively considers multiple factors such as power generation load level, meteorological conditions, traffic accessibility, grid connection permits and weather risks to form an opportunity window scoring model. When the comprehensive score reaches a set threshold and meets the safety constraints, the system determines that time period as an effective maintenance window. During the maintenance and optimization phase, the system takes expected cost, power generation loss and risk cost as objectives, and human resources, spare parts inventory and safety red line as constraints. It achieves dynamic multi-objective scheduling based on a two-stage solution strategy that combines reinforcement learning and mixed integer programming. The upper-level reinforcement learning model is used to generate resource allocation and time planning strategies, and the lower-level mixed integer programming model is used to determine the specific task sequence and execution time. The maintenance scheduling plan output by the module is synchronized to the execution and feedback module, serving as the direct basis for issuing maintenance tasks.

[0013] In one possible implementation, the execution and feedback module monitors the progress, collects status data, and learns feedback during the maintenance task execution process. The collected post-maintenance equipment operation data is transmitted back to the data governance and feature extraction module through a standardized interface, forming a data closed loop. When a significant shift in the distribution of input data is detected, the module automatically triggers a small-step retraining mechanism to update the model parameters of the spatiotemporal graph prediction module with physical prior constraints, the uncertainty quantification module, and the lifetime assessment and degradation analysis module, and performs gray-scale release verification. If the prediction confidence of the new model is lower than the set threshold during the verification phase, the system automatically reverts to the previous stable version to ensure the stability and security of the operation and maintenance strategy.

[0014] In one possible implementation, the security and auditing module controls the access permissions of each module through a hierarchical permission mechanism. It performs full anonymization and log recording on the raw data of the data governance and feature extraction module, the intermediate data of the interaction between modules, and the maintenance records of the execution and feedback module. It also provides real-time alarms for abnormal operations, ensuring the security and traceability of the entire system's data flow and function execution.

[0015] Secondly, a method for predicting equipment status trends and making maintenance decisions in new energy power plants is provided. This method utilizes the system described in the first aspect. The method forms a closed-loop process through logical progression and data flow between steps, including the following steps: S1: Data Governance and Feature Extraction. The data governance and feature extraction module collects multi-source data from SCADA, meteorological monitoring, power grid operation plans, and equipment operation and maintenance records. After cleaning, synchronization, anomaly detection, and feature extraction, a unified structured feature sequence is formed, providing standardized input for subsequent trend prediction. S2: Spatiotemporal trend prediction under physical constraints. Based on the structured feature sequence obtained in step S1, a heterogeneous spatiotemporal graph model is constructed through the spatiotemporal graph prediction module with physical prior constraints. Combining graph attention network and temporal convolutional network, physical consistency constraints are introduced to realize the trend prediction of equipment operating status and output the trend prediction results. S3: Uncertainty calibration. The trend prediction results from step S2 are input into the uncertainty quantification module. A prediction distribution is generated based on the deep integration of multiple sub-models and the random discarding mechanism. After calibration by the order-preserving regression algorithm, a reliable health trend result with confidence interval is output. S4: Equipment life assessment and degradation analysis. Input the reliable health trend results from step S3 into the life assessment and degradation analysis module, and use the dedicated degradation model for the corresponding equipment type to calculate the degradation rate and remaining service life, and generate life distribution information. S5: Causal analysis and quantification of maintenance actions. Input the life distribution information from step S4 into the causal reasoning and counterfactual analysis module to construct a causal relationship diagram of key equipment variables. Evaluate the effects of different maintenance operations through counterfactual simulation and quantify the impact of maintenance actions on the evolution of health status. S6: Opportunity window identification and maintenance optimization. Based on the quantitative results of maintenance actions in step S5, the opportunity window identification and maintenance optimization module generates an effective maintenance window by integrating multiple factors and outputting a maintenance scheduling plan in combination with the risk optimization model. S7: Maintenance execution and closed-loop feedback. The maintenance scheduling plan of step S6 is issued and executed through the execution and feedback module. The maintenance progress is monitored and the equipment operation data after maintenance is collected. The operation data is fed back to step S1. At the same time, the models corresponding to steps S2, S3 and S4 are retrained and rolled calibrated according to the data distribution drift. S8: Full-process security and audit monitoring. Through the security and audit module, the entire process of data flow, model calculation and operation execution in steps S1-S7 is subject to access control, data anonymization, logging and anomaly alarms to ensure the security and traceability of the method execution.

[0016] Compared with the prior art, this application has the following beneficial effects: This application provides a new energy power plant equipment status trend maintenance decision system. By constructing a cloud-edge-device collaborative architecture, it realizes unified governance and feature extraction of the operation data of wind power, photovoltaic and energy storage equipment, and can maintain data consistency and efficient calculation under multi-source data conditions.

[0017] By introducing physical prior constraints and uncertainty quantification mechanisms into the spatiotemporal graph prediction model, the system takes into account both physical rationality and credibility when predicting equipment operating trends, significantly improving the accuracy and stability of health status prediction. At the same time, the combined application of life assessment, causal reasoning and opportunity window identification modules enables maintenance plans to be dynamically optimized based on the actual degradation patterns of equipment and the external environment.

[0018] The system structure of this application is complete and the data flow is clear. It can form a closed-loop operation mechanism of prediction-evaluation-decision-feedback, realize active maintenance and risk adaptive control of equipment, thereby reducing unplanned downtime rate and improving the operation safety and comprehensive economic benefits of new energy power plants.

[0019] This application provides a maintenance decision-making system and method for equipment status trends in new energy power plants. Through a cloud-edge-device collaborative architecture, it achieves intelligent management of the entire process from data acquisition to maintenance execution. Logically, the system constructs a complete closed-loop working mechanism, and methodologically, it defines clear execution steps. In particular, by deeply integrating the weight structure of the spatiotemporal graph model—combining physical, statistical, and planning weights at a ratio of 0.6:0.3:0.1—with the prior constraints of key physical equipment, the accuracy of status prediction and physical consistency are significantly improved, laying a reliable foundation for subsequent intelligent decision-making. The modules are interdependent, with bidirectional information and control flow, forming a comprehensive operation and maintenance platform with self-learning, self-adaptation, and self-evolution capabilities. Attached Figure Description

[0020] Figure 1 This application provides a schematic diagram of the modules of a new energy power plant equipment status trend maintenance decision system. Detailed Implementation

[0021] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0022] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly defined. The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0024] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a communication connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0025] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] like Figure 1 As shown in the figure, this embodiment provides a new energy power plant equipment status trend maintenance decision system. The system is deployed based on a cloud-edge-device collaborative architecture and is used to realize trend prediction, health assessment and maintenance decision optimization of the operating status of equipment such as wind turbines, photovoltaic inverters and energy storage systems.

[0028] The system forms a complete closed loop from data acquisition, status perception, trend prediction, life assessment, maintenance decision-making, and execution feedback, aiming to solve the problems of low accuracy in equipment health prediction, delayed maintenance plans, and uncertain risk decisions in existing technologies.

[0029] The system mainly includes the following functional modules: 1. Data Governance and Feature Extraction Module: Used to synchronize, clean, and extract features from multi-source data collected by SCADA, meteorological, and environmental monitoring devices, and output structured datasets; 2. Spatiotemporal graph prediction module with physical prior constraints: Based on the equipment topology and physical constraints, a heterogeneous spatiotemporal graph model is established to predict the trends of key operating indicators; 3. Uncertainty Quantification Module: Employs deep integration and quantile regression to generate predictive confidence intervals and output reliable state trend results; 4. Life Assessment and Degradation Analysis Module: Utilizes degradation models and thermal cycling fatigue models to assess the remaining life (RUL) of the equipment. 5. Causal Reasoning and Counterfactual Analysis Module: Establish a causal relationship model for equipment and quantify the impact of different maintenance actions on the evolution of health status; 6. Opportunity Window Identification and Maintenance Optimization Module: Generates opportunity windows by integrating meteorological data, power generation plans, grid connection permits, and safety constraints, and outputs maintenance strategies based on a risk optimization model; 7. Execution and Feedback Module: Issues maintenance tasks, monitors execution status, and sends back results; supports model self-calibration and rolling policy updates. 8. Security and Audit Module: Implements data anonymization, access control, and operation tracking to ensure the security and traceability of system operation.

[0030] This system can be applied to wind farms, photovoltaic power plants, and hybrid power plants with energy storage, enabling accurate prediction of equipment operating status and proactive maintenance decisions, thereby improving the safety and economy of operation and maintenance.

[0031] The system in this application adopts a modular and distributed design in its structure. The modules interact with each other through standardized interfaces to exchange data and commands, thereby realizing multi-layer collaborative equipment status prediction and maintenance decision-making functions. The following section provides a detailed description of the system's technical solution in conjunction with the specific functions of each module.

[0032] The system first uses a data governance and feature extraction module to clean, synchronize, and structure operational data from wind turbines, photovoltaic inverters, energy storage units, and meteorological monitoring systems. The raw data includes torque, speed, output power, temperature, current, voltage, vibration, and meteorological variables. Data is synchronized in real-time at edge nodes before entering the system to eliminate time deviations caused by different sampling frequencies. For abnormal or missing data, the system employs a hybrid detection strategy based on median absolute deviation and the isolated forest algorithm for identification and repair, ensuring the stability of the input data. The final generated feature data includes statistical characteristics of time, frequency, and energy domains for subsequent modeling and analysis.

[0033] After data preprocessing, the system enters the spatiotemporal graph prediction module stage with physical prior constraints. This module describes the energy, spatial, and control coupling relationships between different devices by establishing a heterogeneous graph structure including wind turbines, inverters, energy storage clusters, and meteorological units. Nodes in the graph structure represent device units, and edge weights are composed of physical coupling weights, statistical correlation weights, and planning association weights, thus reflecting the inherent dependencies in device operation. The system uses a graph attention network and a temporal convolutional network to jointly construct the spatiotemporal prediction model, capturing spatial correlation features while maintaining the continuity of temporal evolution. To ensure that the model output conforms to actual operating conditions, a physical consistency constraint term is introduced into the loss function. For example, the output power of the wind turbine drivetrain should satisfy:

[0034] in, For mechanical transmission efficiency, Main spindle torque, The main axis angular velocity.

[0035] Furthermore, photovoltaic inverters meet the following requirements under thermal stability conditions.

[0036] in, For the junction temperature, For ambient temperature, and These are thermal resistance and thermal capacity constant, respectively.

[0037] By introducing the above constraints, the model training process is forced to follow the physical laws of the device, thereby achieving a balance between prediction accuracy and interpretability.

[0038] After completing trend prediction, the system enters the uncertainty quantification stage. To address the lack of confidence expression in traditional deep learning model predictions, this system employs a combination of deep ensemble of multiple sub-models and Monte Carlo Dropout to sample the prediction distribution. Each inference iteration outputs multiple quantile results, which are then post-calibrated using ordinal-preserving regression to obtain prediction results with interpretable confidence intervals. This method not only provides the expected value of health indicators but also outputs upper and lower confidence boundaries, giving the predictions statistically significant confidence coverage. Through this mechanism, the system can explicitly measure the uncertainty of predictions when assessing equipment health status, thereby enhancing the risk perception capability of subsequent maintenance decisions.

[0039] During the life assessment and degradation analysis phase, the system establishes corresponding degradation models based on different equipment types. The degradation behavior of the wind turbine gearbox is described by a Gamma process, whose state variables monotonically increase with time and follow a non-stationary distribution; the photovoltaic inverter IGBT module adopts a thermal cycling damage model, based on the temperature difference change... Rainflow cycle counting is performed to calculate cumulative fatigue damage; the energy storage battery cells are estimated in real time based on unscented Kalman filtering, combined with an empirical attenuation model. Predict lifetime evolution trends. The remaining useful life (RUL) output by each model is stored in the database in the form of a probability distribution, providing a quantifiable time reference for the maintenance decision module.

[0040] In the causal reasoning and counterfactual analysis module, the system models the causal relationship between equipment operating status and maintenance operations. By constructing a local causal graph, variables such as lubrication status, friction coefficient, vibration amplitude, temperature rise, and failure rate are formed into a directed acyclic graph structure. During model training, a causal consistency regularization term is introduced to maximize the relevance of true causal paths and minimize the interference of spurious correlation terms. Based on the established causal structure, the system can perform counterfactual simulations of different maintenance actions (such as lubricant replacement or fan blade cleaning) to calculate the change in state after maintenance. This allows for the assessment of the impact of different strategies on equipment health evolution, enabling interpretable optimization of maintenance strategies.

[0041] The opportunity window identification and maintenance optimization module transforms prediction and evaluation results into actionable maintenance strategies. The system comprehensively considers factors such as power generation load, meteorological conditions, traffic accessibility, grid connection permits, and weather risks, constructing a multi-index scoring model. Its overall score function is defined as...

[0042] in, Indicates the power generation load level; Indicates the meteorological condition coefficient; Indicates the accessibility score; Indicates the weight of the grid connection permit; Indicates weather risk indicators; when Furthermore, a maintenance window is considered valid when safety constraints are met. Based on this window information, the system further establishes a risk-sensitive maintenance optimization model, with the objective function being to minimize the overall expected cost and risk loss.

[0043] in, : Expectation operator, used to represent the overall expected cost; Maintenance execution costs include direct operating expenses such as spare parts replacement, labor hours, transportation and tool consumption; Power generation loss cost represents the loss of power generation revenue due to equipment downtime during maintenance, calculated based on the equipment power output curve and planned downtime. Risk-added costs are used to measure the potential losses caused by delayed maintenance or sudden failures. Their value is dynamically estimated based on the equipment health index and failure probability. Conditional Value at Risk (VaR) represents the extreme expected loss of the overall cost distribution at a confidence level of 0.95, and is used to reflect the high-risk state faced by the system under uncertainty. Risk trade-off coefficient, a non-negative real number, is used to adjust the balance between expected cost and extreme risk. The larger the value, the more emphasis is placed on risk avoidance during the optimization process.

[0044] Model constraints include personnel and vehicle capacity, spare parts inventory, grid connection permits, and operational safety limits. The upper layer uses reinforcement learning to generate maintenance strategies, while the lower layer uses mixed-integer programming to sort tasks and allocate resources, ultimately outputting a multi-objective balanced optimization solution.

[0045] The execution and feedback module is responsible for converting maintenance plans into work orders and monitoring the execution process in real time. When the system detects a significant shift in the input distribution (statistical test p-value ≤ 0.01), it will trigger a small-step retraining mechanism with a learning rate set to 1×10⁻⁶. -5The system iterates 500 times. If the confidence interval coverage of the new model drops by more than 0.03 during the online validation phase, it automatically reverts to the previous stable version. The system adopts a canary release mechanism, initially covering only 10% of devices, and gradually expanding to all applications after 72 hours of stable operation to ensure the security and continuity of policy updates.

[0046] The security and auditing module is integrated throughout the entire system operation, responsible for data security, identity authentication, and operation tracking. Data anonymization and hash encryption are performed at the edge, while access levels are restricted in the cloud through a role-based access control system. All critical operational events are recorded in the audit log using a hash chain to ensure traceability and immutability. Upon detecting unauthorized work assignments or unauthorized operations, the system will immediately block the operation and trigger a security alert to prevent potential security risks from escalating.

[0047] Through the synergistic effect of the aforementioned modules, this application achieves intelligent management of the entire process of new energy power plant equipment, from condition prediction to maintenance execution. Compared with traditional maintenance methods based on thresholds or experience-based judgments, this system introduces a spatiotemporal prediction model with physical constraints, an uncertainty confidence mechanism, and a causal-driven decision optimization method. This makes equipment health prediction more accurate, maintenance scheduling more economical, and the decision-making process more interpretable, thereby significantly improving the safety and economic benefits of operation and maintenance in complex new energy scenarios.

[0048] In summary, the equipment status trend maintenance decision-making system for new energy power plants proposed in this application achieves intelligent management of the entire process from data collection to maintenance execution through a cloud-edge-device collaborative architecture. Logically, the system constructs a closed-loop working mechanism encompassing multi-source data governance, status trend prediction, health life assessment, causal reasoning analysis, opportunity window identification, risk-sensitive optimization decision-making, and execution and feedback calibration. The modules are interdependent, with bidirectional information and control flows, forming a comprehensive operation and maintenance platform with self-learning, self-adaptive, and self-evolving capabilities.

[0049] In terms of operational principles, the system first achieves real-time cleaning and feature extraction of high-frequency operational data through edge nodes. Then, it utilizes a heterogeneous spatiotemporal graph prediction model based on physical constraints to predict the trends of multi-dimensional indicators, and outputs a health status range with confidence through an uncertainty quantification module. Next, the lifespan assessment module calculates the remaining useful life (RUL) based on degradation and fatigue models, providing a time basis for maintenance planning. The causal reasoning module then introduces structured causal relationship analysis to achieve counterfactual inference of the effects of maintenance actions, thereby supporting intelligent judgment at the decision-making level. Finally, the maintenance optimization module outputs the optimal scheduling scheme by integrating opportunity windows and risk indicators. The system continuously performs model drift detection and online calibration during execution and feedback to ensure long-term stability and robustness.

[0050] In one possible embodiment, this application takes the monthly maintenance decision of the hybrid renewable energy power station as an example to describe in detail the complete execution steps of the equipment status trend prediction and maintenance decision method for renewable energy power stations: S1: Multi-source data acquisition and structured processing (1 day) The data governance and feature extraction module collects SCADA data, meteorological data, power grid operation plans, and equipment maintenance records, including historical maintenance times, types of replaced spare parts, and fault handling results from multiple sources. After synchronization, anomaly detection and repair, and feature extraction, a unified structured feature sequence is formed. For example, 64-dimensional feature vectors of 30 wind turbines over the past month, totaling 900 sample data points, are extracted for subsequent modeling.

[0051] S2: Equipment operating status trend prediction (2 hours) A heterogeneous spatiotemporal graph model is constructed through a spatiotemporal graph prediction module with physical prior constraints. The structured feature sequence obtained by inputting S1 is combined with a graph attention network and a temporal convolutional network. Under the physical consistency constraints of wind turbine drive chain efficiency, photovoltaic inverter thermal balance, and energy storage unit equivalent circuit, the model predicts the equipment operation status trend for the next 30 days and outputs the predicted values ​​of key indicators for each day, such as wind turbine output power, inverter junction temperature, and battery SOC.

[0052] S3: Calibration of prediction uncertainty (1 hour) Through the uncertainty quantification module, based on the deep integration of 5 sub-models and a random discarding mechanism, 50 prediction distributions are generated from the prediction results of S2. After calibration by the ordinal-preserving regression algorithm, a reliable health trend result with a 90% confidence interval is output. For example, the average output power of a wind turbine in the next 10 days is predicted to be 1.7MW, with a confidence interval of [1.55MW, 1.85MW].

[0053] S4: Equipment Remaining Life Assessment (3 hours) Through the lifespan assessment and degradation analysis module, specific degradation models are employed for different equipment types: the Gamma degradation process model is applied to wind turbine gearboxes, the temperature cycling statistics are used for photovoltaic inverter power modules, and the parameter-adaptive unscented Kalman filtering method is used for energy storage battery clusters. Combined with the reliable prediction results of S3, the degradation rate and remaining service life are calculated to generate lifespan distribution information. For example, the assessment found that 3 out of 10 wind turbine gearboxes have a remaining service life of less than 180 days, requiring close monitoring.

[0054] S5: Quantitative Analysis of Maintenance Action Effectiveness (2 hours) Through the causal reasoning and counterfactual analysis module, a causal relationship graph of key equipment variables is constructed. During the training phase, a causal consistency regularization term is introduced to strengthen the real causal path. Counterfactual simulations are performed on different maintenance actions, such as lubricant replacement, gear overhaul, and inverter power module cleaning, to quantify the impact of each action on the evolution of the health status and rank them to obtain the optimal combination of maintenance actions.

[0055] S6: Maintenance Window Identification and Scheduling Plan Generation (4 hours) Through the opportunity window identification and maintenance optimization module, based on factors such as the power generation load level, meteorological conditions, traffic accessibility, grid connection permits, and weather risks for the next 30 days, a scoring model is applied to identify effective maintenance windows, such as 00:00-08:00 on days 10-12. Then, based on a two-stage risk-sensitive optimization model that combines reinforcement learning and mixed integer programming, with the goal of minimizing expected cost, power generation loss, and risk cost, and in conjunction with personnel, spare parts, and safety constraints, a maintenance scheduling plan is output, which clarifies the maintenance time, responsible team, and required spare parts for each piece of equipment.

[0056] S7: Maintenance execution and model rolling calibration (15 days) Maintenance work orders are issued through the execution and feedback module, and the execution progress is monitored in real time. Post-maintenance equipment status data and feedback information are collected. During maintenance execution, the distribution of input data is continuously monitored for drift. If the distribution of operating data for a certain device drifts significantly (p-value = 0.006), a small-step retraining mechanism is triggered to update the model parameters. After gray-scale release verification (covering 2 devices for 72 hours), if the new model prediction confidence is 0.88 (higher than the threshold of 0.85) and the confidence interval coverage is 0.91 (meeting the requirements), it will be gradually promoted to all devices.

[0057] S8: System Security Assurance and Operational Traceability (Full Process) Through the security and auditing modules, access control, data anonymization, and log recording are implemented throughout the process. All operations during maintenance are stored using a hash chain. If any abnormal operation occurs, such as unauthorized personnel attempting to modify the scheduling plan, the system will immediately block and issue an alarm, ensuring system security and traceability of operations.

[0058] By fully implementing the above eight steps, a closed-loop management system for the entire process from equipment status prediction to maintenance execution was achieved. After the implementation of this monthly maintenance decision-making scheme, the unplanned downtime rate of the plant was reduced by 25%, the operation and maintenance cost was reduced by 20%, and the power generation efficiency was increased by 8%, significantly improving operational safety and overall economic benefits.

[0059] In another possible embodiment, it is described as follows: 1. System applications in extreme weather scenarios When the station faces a typhoon warning, with a typhoon expected to make landfall within the next 72 hours and wind speeds reaching level 12 or higher, the system will implement the following special adaptation: Opportunity window identification and maintenance optimization module: Adjust the weights of the opportunity window scoring model; increase the weight of the weather risk index W_r to 0.3; and adjust the power generation load level L. p The weight is reduced to 0.2, and priority is given to identifying the 24 hours before the arrival of the typhoon as the emergency maintenance window, such as 06:00-14:00 3 days before the typhoon makes landfall. Priority is given to scheduling maintenance tasks such as wind turbine blade reinforcement, photovoltaic module fastening, and waterproof inspection of energy storage systems. The spatiotemporal graph prediction module with physical prior constraints updates the efficiency constraints of the wind turbine drivetrain, considering extreme loads during typhoons, η. m The value range was adjusted to 0.88-0.92 to ensure that the prediction results conform to the physical laws under extreme working conditions; Life assessment and degradation analysis module: In response to potential equipment damage caused by typhoons, the degradation model parameters are temporarily adjusted, and the β value of the Gamma degradation model for wind turbine gearboxes is increased to 0.002, accelerating the calculation of degradation rate and allowing for early assessment of remaining life after potential damage.

[0060] 2. System Expansion in Large-Scale Station Cluster Scenarios When the system is applied to a cluster containing 5 hybrid renewable energy power stations, the following expansion is implemented: Cloud-edge-device collaborative architecture optimization: Edge nodes are deployed according to the site, responsible for local data preprocessing and real-time monitoring, while the cloud is responsible for global model training, cross-site maintenance resource scheduling and overall optimization; The spatiotemporal graph prediction module has been expanded: the node types have been expanded to include station-level convergence line nodes and regional meteorological nodes, and the edge weights have been expanded to include cross-station power grid coupling weights, enabling cluster-level state trend prediction. Opportunity window identification and maintenance optimization module: Considering cross-site resource sharing, such as spare parts allocation and cross-site team support, cross-site resource transportation time and cost are added to the constraints to optimize the overall maintenance schedule of the cluster and reduce the total operation and maintenance cost.

[0061] This application achieves accurate prediction and closed-loop control of equipment status in new energy power plants through multi-source data fusion, physical consistency modeling, uncertainty calibration, and causal-driven maintenance optimization. The system has a reasonable structure, clear logic, and adjustable parameters, enabling stable operation in various new energy scenarios such as wind power, photovoltaics, and energy storage. It provides proactive, safe, and efficient intelligent operation and maintenance support for large-scale new energy power plants.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A system for equipment state trend prediction and maintenance decision of a new energy power plant, characterized in that, include: The data governance and feature extraction module is used to collect and process multi-source data from SCADA, meteorological monitoring and power grid operation plans. It cleans, synchronizes, detects anomalies and extracts features from the data to form a unified structured feature sequence, providing standardized data input for subsequent modules. The spatiotemporal graph prediction module with physical prior constraints receives the structured feature sequence output by the data governance and feature extraction module, which is used to establish a heterogeneous spatiotemporal graph model containing equipment nodes, feeder nodes and meteorological nodes. It combines graph attention network and temporal convolutional network to realize the prediction of operating status trend, and introduces physical consistency conditions in model training to output the equipment operating status trend prediction results. The uncertainty quantification module receives the prediction results from the spatiotemporal graph prediction module with physical prior constraints, and uses them to generate confidence intervals for the prediction results based on deep integration of multiple sub-models and random dropout mechanism. After calibration, a credible health trend result is formed, providing a reliable basis with confidence for subsequent evaluation and decision-making. The life assessment and degradation analysis module receives the reliable health trend results output by the uncertainty quantification module, and uses them to calculate the degradation rate and remaining service life based on the degradation model of different equipment types, generate life distribution information, and clarify the equipment health status and maintenance priority. The causal reasoning and counterfactual analysis module receives the life distribution information from the life assessment and degradation analysis module, and uses it to construct a causal relationship diagram between key variables of the equipment, perform counterfactual simulation assessment of maintenance operations, quantify the impact of different actions on the evolution of health status, and provide causal support for maintenance strategy optimization. The opportunity window identification and maintenance optimization module receives the quantification results of the maintenance action impact from the causal reasoning and counterfactual analysis module, generates opportunity windows by combining external factors such as power generation load and meteorological conditions, and outputs a maintenance scheduling plan based on the risk optimization model. The execution and feedback module receives the maintenance scheduling plan from the opportunity window identification and maintenance optimization module, and uses it to issue maintenance tasks, monitor the execution progress, collect equipment operation data after maintenance and feed it back to the data governance and feature extraction module. At the same time, it performs model retraining and rolling calibration based on the input data distribution drift to achieve closed-loop iteration. The security and audit module runs through the entire operation of all modules and is used to implement system access control, data anonymization, logging, and abnormal operation alarms to ensure the security and traceability of data interaction and function execution of each module.

2. The system of claim 1, wherein, The data governance and feature extraction module uses the median absolute deviation algorithm and the isolated forest algorithm to jointly detect outliers, uses time interpolation to repair short-period missing data, and uses adjacent device regression to complete long-period missing data. Based on a sliding time window, features in the time domain, frequency domain, and energy domain are extracted, and a multi-dimensional feature vector containing operational state statistics and vibration spectrum features is output. The feature vector is synchronized to the spatiotemporal graph prediction module with physical prior constraints through a standardized interface, serving as the core input data for model training and inference.

3. The system of claim 1, wherein, The node types of the spatiotemporal map prediction module include wind turbine nodes, photovoltaic inverter nodes, energy storage cluster nodes, and meteorological nodes. The edge weights consist of physical coupling weights, statistical correlation weights, and planning correlation weights, with weight ratios of 0.6, 0.3, and 0.1, respectively. Among them, the physical coupling weight is determined based on the electrical topology and geographical layout, the statistical correlation weight is calculated based on mutual information and Granger causality, and the planned correlation weight is generated based on grid connection permits and maintenance plans. The physical prior constraints include wind turbine drivetrain efficiency constraints, photovoltaic inverter thermal balance constraints, and energy storage unit equivalent circuit constraints. During the training phase, the prediction model incorporates physical constraint errors as a penalty term into the loss function. Through the constraint learning mechanism, the prediction results are kept consistent with the actual operating conditions of the equipment. The output trend prediction results are directly used as the input data source for the uncertainty quantification module.

4. The system of claim 1, wherein, The uncertainty quantification module consists of an integrated system of deep learning sub-models initialized with different parameters. During the inference phase, it uses random dropout technology to generate multiple prediction distributions and uses an order-preserving regression algorithm to calibrate the confidence interval, so that the prediction interval coverage is not less than 0.9 and the interval width is controlled within a set range. The reliable health trend results output by the module include predicted values ​​and corresponding confidence intervals, which are synchronized to the life assessment and degradation analysis module to provide input data with risk boundaries for equipment degradation rate calculation and remaining life assessment.

5. The system of claim 1, wherein, The life assessment and degradation analysis module includes a multi-model evaluation mechanism for different equipment types. The wind turbine gearbox uses the Gamma degradation process model to calculate the degradation rate, the photovoltaic inverter power module estimates fatigue life based on temperature cycle statistics, and the energy storage battery cluster uses a parameter-adaptive unscented Kalman filter method to achieve dynamic life prediction. The health index and remaining lifespan distribution information output by the module serve as the core basis for the causal reasoning and counterfactual analysis module to identify maintenance priorities and quantify maintenance effectiveness.

6. The system of claim 1, wherein, The causal reasoning and counterfactual analysis module strengthens the constraints of true causal paths and reduces spurious correlation interference during the training phase by using causal consistency regularization terms. The system uses structured causal graphs to identify key causal chains between maintenance operations, environmental factors, and equipment health indicators, and generates health change prediction results under different maintenance actions through counterfactual deduction during the simulation phase. The maintenance action impact quantification ranking results output by the module are directly input into the opportunity window identification and maintenance optimization module, providing interpretable support for the screening and optimization of maintenance strategies.

7. The system of claim 1, wherein, The opportunity window identification and maintenance optimization module comprehensively considers multiple factors such as power generation load level, meteorological conditions, transportation accessibility, grid connection permits and weather risks to form an opportunity window scoring model. When the comprehensive score reaches the set threshold and meets the safety constraints, the system determines that time period as an effective maintenance window. During the maintenance and optimization phase, the system takes expected cost, power generation loss and risk cost as objectives, and human resources, spare parts inventory and safety red line as constraints. It achieves dynamic multi-objective scheduling based on a two-stage solution strategy that combines reinforcement learning and mixed integer programming. The upper-level reinforcement learning model is used to generate resource allocation and time planning strategies, and the lower-level mixed integer programming model is used to determine the specific task sequence and execution time. The maintenance scheduling plan output by the module is synchronized to the execution and feedback module, serving as the direct basis for issuing maintenance tasks.

8. The system of claim 1, wherein, The execution and feedback module monitors the progress, collects status and learns feedback during the maintenance task execution process. The collected post-maintenance equipment operation data is sent back to the data governance and feature extraction module through a standardized interface, forming a data closed loop. When a significant shift in the distribution of input data is detected, the module automatically triggers a small-step retraining mechanism to update the model parameters of the spatiotemporal graph prediction module with physical prior constraints, the uncertainty quantification module, and the lifetime assessment and degradation analysis module, and performs gray-scale release verification. If the prediction confidence of the new model is lower than the set threshold during the verification phase, the system automatically reverts to the previous stable version to ensure the stability and security of the operation and maintenance strategy.

9. The system of claim 1, wherein, The security and auditing module controls the access permissions of each module through a hierarchical permission mechanism. It performs full anonymization and log recording on the raw data of the data governance and feature extraction module, the intermediate data of the interaction between modules, and the maintenance records of the execution and feedback module. It provides real-time alarms for abnormal operations to ensure the security and traceability of the entire system's data flow and function execution.

10. A method for equipment state trend prediction and maintenance decision of a new energy power station, characterized in that, The system described in any one of claims 1-9, wherein the method forms a closed-loop process through logical progression and data flow between steps, includes the following steps: S1: Data Governance and Feature Extraction. The data governance and feature extraction module collects multi-source data from SCADA, meteorological monitoring, power grid operation plans, and equipment operation and maintenance records. After cleaning, synchronization, anomaly detection, and feature extraction, a unified structured feature sequence is formed, providing standardized input for subsequent trend prediction. S2: Spatiotemporal trend prediction under physical constraints. Based on the structured feature sequence obtained in step S1, a heterogeneous spatiotemporal graph model is constructed through the spatiotemporal graph prediction module with physical prior constraints. Combining graph attention network and temporal convolutional network, physical consistency constraints are introduced to realize the trend prediction of equipment operating status and output the trend prediction results. S3: Uncertainty calibration. The trend prediction results from step S2 are input into the uncertainty quantification module. A prediction distribution is generated based on the deep integration of multiple sub-models and the random discarding mechanism. After calibration by the order-preserving regression algorithm, a reliable health trend result with confidence interval is output. S4: Equipment life assessment and degradation analysis. Input the reliable health trend results from step S3 into the life assessment and degradation analysis module, and use the dedicated degradation model for the corresponding equipment type to calculate the degradation rate and remaining service life, and generate life distribution information. S5: Causal analysis and quantification of maintenance actions. Input the life distribution information from step S4 into the causal reasoning and counterfactual analysis module to construct a causal relationship diagram of key equipment variables. Evaluate the effects of different maintenance operations through counterfactual simulation and quantify the impact of maintenance actions on the evolution of health status. S6: Opportunity window identification and maintenance optimization. Based on the quantitative results of maintenance actions in step S5, the opportunity window identification and maintenance optimization module generates an effective maintenance window by integrating multiple factors and outputting a maintenance scheduling plan in combination with the risk optimization model. S7: Maintenance execution and closed-loop feedback. The maintenance scheduling plan of step S6 is issued and executed through the execution and feedback module. The maintenance progress is monitored and the equipment operation data after maintenance is collected. The operation data is fed back to step S1. At the same time, the models corresponding to steps S2, S3 and S4 are retrained and rolled calibrated according to the data distribution drift. S8: Full-process security and audit monitoring. Through the security and audit module, the entire process of data flow, model calculation and operation execution in steps S1-S7 is subject to access control, data anonymization, log recording and anomaly alarms to ensure the security and traceability of method execution.

Images