Lower guide shoe temperature adaptive prediction method based on machine learning

Abstract

The invention discloses a lower guide shoe temperature adaptive prediction method based on machine learning, and relates to the technical field of hydroelectric generating set operation and maintenance monitoring and intelligent monitoring, and the method comprises the steps: collecting and synchronizing multi-source data in real time, and screening feature subsets through XGBoost feature importance; dividing the operation state into a plurality of working condition subclasses, and training an XGBoost submodel for each working condition subclass; constructing an integrator, outputting a multi-step predicted value, and generating a confidence interval and a confidence index based on residual distribution or integration difference; judging an abnormal risk according to a multi-level alarm rule; and deploying a sliding window residual error and concept drift detection strategy to detect distribution drift or residual error abnormity. According to the method, sliding window residual monitoring, concept drift detection and integrated uncertainty evaluation are introduced, so that the stability and reliability of temperature prediction of the hydroelectric generating set are improved. And the accuracy of temperature early warning is improved based on the confidence interval and the multi-level warning rule.

Description

A Machine Learning-Based Adaptive Prediction Method for Lower Conductor Tile Temperature Technical Field

[0001] This invention relates to the field of hydropower unit operation and maintenance monitoring and intelligent monitoring technology, and in particular to a machine learning-based adaptive prediction method for the temperature of the lower guide tile. Background Technology

[0002] As the digitalization of the power industry accelerates, the operation monitoring systems for hydropower units, wind turbines, and various key equipment are continuously accumulating large amounts of real-time data. Temperature prediction, vibration prediction, and health status assessment models built using machine learning and deep learning methods are gradually becoming an important foundation for intelligent operation and maintenance. However, during the long-term operation of power equipment, its operating conditions, environmental conditions, load structure, and equipment aging status constantly change, causing the statistical characteristics of the model input data to drift over time, resulting in the concept drift problem.

[0003] When data distribution changes, models trained offline struggle to maintain stable performance over the long term. Prediction residuals accumulate and exhibit abnormal fluctuations, ultimately leading to excessive prediction bias, decreased warning reliability, and compromised equipment safety and stability. Therefore, a key challenge for current intelligent operations and maintenance technologies is how to detect residual change trends in real time after a model is deployed, identify data distribution drift, and automatically select appropriate adaptive update strategies based on the degree of drift.

[0004] Existing methods typically rely on periodic offline retraining or simple threshold alarms, which cannot provide fine-grained discrimination of drift levels and lack the ability to differentiate between different anomaly types. When residuals fluctuate in the short term, operating conditions change slightly, or the model degrades slightly, direct retraining will waste resources; while untimely response to severe drift may lead to inaccurate prediction models or even missing key early warning signals.

[0005] Therefore, there is an urgent need for a model dynamic adaptation method that can statistically analyze residual changes based on sliding time windows and combine concept drift detection mechanism to achieve real-time identification of residual anomalies and data distribution drift, and automatically select update strategies such as incremental fine-tuning, sample reweighting, or offline retraining according to the drift type, so as to ensure the stability and reliability of the equipment status prediction model in long-term operation. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the problem to be solved by this invention is to address the common issues in existing hydropower unit temperature prediction and intelligent operation and maintenance models during long-term online operation, such as residual accumulation instability, inability to identify concept drift in a timely manner, insufficient model adaptability, and decreased early warning reliability.

[0008] To address the aforementioned technical problems, the present invention provides the following technical solution: Firstly, embodiments of the present invention provide a machine learning-based adaptive prediction method for lower guide tile temperature, comprising: real-time acquisition and synchronization of multi-source data; fusion of multi-source data; and selection of feature subsets through XGBoost feature importance; identification of operating conditions based on the selected feature subsets, dividing the operating state into multiple operating condition subclasses, and training an XGBoost sub-model for each operating condition subclass; constructing an integrator, outputting multi-step predicted values, and generating confidence intervals and confidence indices based on residual distribution or integrator differences; determining abnormal risks according to multi-level alarm rules, and automatically providing causes and maintenance suggestions when an alarm is triggered; deploying a sliding window residual and concept drift detection strategy, and prioritizing incremental fine-tuning, sample reweighting, or triggering offline retraining if distribution drift or residual abnormalities are detected.

[0009] As a preferred embodiment of the machine learning-based adaptive prediction method for lower guide tile temperature described in this invention, the method includes: real-time acquisition and synchronization of multi-source data; fusion of the multi-source data; and selection of a feature subset using XGBoost feature importance evaluation. This includes: performing time synchronization, temporal alignment, and unit normalization on the real-time acquired multi-source data, fusing it into a unified format multi-dimensional temporal feature set; performing missing value imputation, outlier detection, and noise filtering on the fused multi-dimensional temporal feature set; constructing extended features based on lubrication theory and hydroelectric generator operation mechanism; and using the feature importance evaluation mechanism of the XGBoost model to select the extended features, thereby obtaining a feature subset characterizing the factors influencing the temperature change of the lower guide tile.

[0010] As a preferred embodiment of the machine learning-based adaptive prediction method for lower guide tile temperature described in this invention, the missing value imputation adopts a hybrid interpolation strategy based on temporal interpolation and weighted similar working conditions, and the outlier detection adopts a joint identification method of statistical threshold judgment and isolated forest model.

[0011] As a preferred embodiment of the machine learning-based adaptive temperature prediction method for lower guide tiles described in this invention, the following steps are taken: Operating conditions are identified based on a selected feature subset of multi-source data, dividing the operating state into multiple operating condition subclasses. For each operating condition subclass, an XGBoost sub-model is trained, including: operating condition subclasses such as low load-high cooling condition, rated stable condition, overload-low cooling condition, and insufficient lubrication condition; for each operating condition subclass, corresponding feature distributions, feature weights, and typical temperature rise patterns are extracted to construct operating condition labels and operating condition feature templates; based on the divided operating condition subclasses, corresponding XGBoost sub-models are trained, with the training process including sample balancing, hyperparameter search, and cross-validation.

[0012] As a preferred embodiment of the machine learning-based adaptive prediction method for lower guide tile temperature described in this invention, the method includes: constructing an integrator to output multi-step predicted values ​​and generating confidence intervals and confidence indices based on residual distribution or integrator differences, comprising: using the XGBoost model and XGBoost sub-models corresponding to multiple operating condition subclasses as a candidate model set; calculating integrator weights based on the current identified operating condition, historical residuals, and model stability; fusing the integrator weights to obtain multi-step predicted values ​​for the lower guide tile temperature; generating prediction confidence intervals based on residual distribution and estimating prediction uncertainty using statistical results of near-real-time residuals; constructing an integrator uncertainty index based on model output differences, with the confidence intervals and confidence indices jointly serving as the prediction credibility output.

[0013] As a preferred embodiment of the machine learning-based adaptive prediction method for lower guide tile temperature described in this invention, the multi-step prediction value calculation formula is as follows: ;in, To predict the temperature of the lower conductor tile in the next k steps, Let be the predicted value of the i-th model. Let M be the ensemble weight of the i-th model, and M be the number of models; the confidence interval is calculated using the following formula: ;in, , is the confidence coefficient The standard deviation of the residuals within the sliding window; the confidence index is calculated using the following formula: ;in, This is the preset maximum acceptable variance.

[0014] As a preferred embodiment of the adaptive temperature prediction method for lower conductor tiles based on machine learning described in this invention, the abnormal risk is determined according to multi-level alarm rules, including: if the predicted value is within the confidence interval and the temperature rise trend is stable, it is determined to be normal; if the predicted value is close to the upper limit of the confidence interval or the short-term temperature rise rate increases, it is determined to be of concern; if the predicted value exceeds the upper limit of the confidence interval or the confidence level decreases at a faster rate, it is determined to be a warning; if the predicted value exceeds the confidence interval and the prediction deviation exceeds the threshold, it is determined to be serious.

[0015] As a preferred embodiment of the machine learning-based adaptive temperature prediction method for lower guide vanes described in this invention, the method includes: deploying a sliding window residual and concept drift detection strategy; if distribution drift or residual anomalies are detected, incremental fine-tuning, sample reweighting, or triggering offline retraining are prioritized. This includes: deploying a sliding window-based residual monitoring mechanism to perform dynamic statistical analysis on the online prediction residuals, calculating the residual mean, standard deviation, and skewness index; when the residuals continuously exceed the set confidence interval, they are determined to be residual anomalies; introducing a concept drift detection strategy to perform real-time detection of the input feature distribution and output distribution based on distribution comparison tests or drift detection algorithms; when feature distribution drift and changes in the frequency of operating condition categories are detected, they are determined to be concept drift events.

[0016] In a second aspect, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of a machine learning-based adaptive prediction method for lower guide tile temperature as described in the first aspect of the present invention.

[0017] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of a machine learning-based adaptive prediction method for lower guide tile temperature as described in the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows: By introducing sliding window residual monitoring, concept drift detection, multi-condition subclass modeling, and integrated uncertainty assessment, this invention achieves the model's adaptability to long-term online operating environments, significantly improving the stability and reliability of hydropower unit temperature prediction. Compared with traditional fixed models, this invention can promptly identify and trigger incremental fine-tuning or retraining when data distribution changes or residual anomalies occur, ensuring stable prediction accuracy across the entire operating range. Simultaneously, based on confidence intervals and multi-level alarm rules, it provides explainable causes of anomalies and maintenance suggestions, significantly improving the accuracy, real-time performance, and availability of temperature warnings, effectively supporting safe unit operation and intelligent maintenance decisions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 is a flowchart of the method of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Referring to Figure 1, Example 1 is the first embodiment of the present invention. This embodiment provides a machine learning-based adaptive prediction method for the temperature of the lower guide tile, including: S1: Real-time acquisition and synchronization of multi-source data, fusion of multi-source data, and selection of feature subsets through XGBoost feature importance.

[0025] Furthermore, time synchronization, time alignment, and unit normalization are performed on the real-time acquired multi-source data, and the data is fused into a unified format multi-dimensional time-series feature set. Missing value imputation, outlier detection, and noise filtering are performed on the fused multi-dimensional time-series feature set. Extended features are constructed based on lubrication theory and hydropower unit operation mechanism. The feature importance evaluation mechanism of the XGBoost model is used to screen the extended features to obtain a feature subset that characterizes the factors affecting the temperature change of the lower guide tile.

[0026] Furthermore, missing value imputation adopts a hybrid interpolation strategy based on temporal interpolation and weighted similar working conditions, and outlier detection adopts a joint identification method of statistical threshold judgment and isolated forest model.

[0027] For example, real-time data collection of unit operation and auxiliary equipment measurement points includes, but is not limited to: unit power P, speed n, head H, cooling water temperature Tc, cooling water flow rate Fc, lubricating oil temperature To, lubricating oil pressure Po, shaft vibration amplitude Va, and lower guide bearing temperature T_actual; each data source is from the upper-level monitoring system, PLC, online monitoring instruments, and vibration monitoring unit, with sampling periods ranging from 1s to 5s; all data are uniformly resampled to a 1s time resolution, and time alignment is performed using linear interpolation.

[0028] The aligned data are merged into a multi-dimensional time-series feature set in a unified format according to the timestamp. Then, the feature set is processed using a hybrid strategy of "time-series interpolation + weighted similar operating conditions": when the missing time period is short (≤10 points), cubic spline interpolation is used; when the missing time period is long (>10 points), similar operating condition intervals are extracted (based on the Euclidean distance of power, speed, and head), and weighted average interpolation is performed according to similarity.

[0029] For example, if 12 points of cooling water flow rate Fc are missing at a certain moment, the corresponding sequences within similar operating conditions are selected and weighted with similarity weights of 0.72, 0.19, and 0.09 to obtain the interpolated value of 52.8 L / min. A combined statistical threshold and isolated forest detection method is used: if the cooling water temperature Tc deviates from the 3σ range, it is judged as abnormal; if the isolated forest score of the lubricating oil pressure Po is >0.65, it is judged as an isolated point; if both are judged as abnormal, backsliding interpolation (using the mean of the neighborhood window) is performed.

[0030] For example, at a certain moment, Po = 0.04 MPa is detected, which is outside the normal distribution range (0.1-0.3 MPa). The isolated forest score is 0.81, so it is judged as an outlier and replaced with the neighborhood window mean of 0.18 MPa.

[0031] Based on lubrication theory, heat dissipation mechanism, and operating condition mechanism, the extended characteristics are constructed, including but not limited to: cooling water temperature change rate: ΔTc=Tc(t)-Tc(t-1); lubricating oil temperature gradient: ; Coupling characteristics of power and head: P / H; Cooling water heat load index: P×Tc; Shaft vibration energy characteristics: Va²; Temperature hysteresis characteristics: T_actual(tk), k = 1, 3, 5, etc. The above characteristics are generated in real time according to a 1s sliding window.

[0032] S2: Based on the selection of feature subsets, the operating conditions of multi-source data are identified, the operating status is divided into multiple operating condition subclasses, and an XGBoost sub-model is trained for each operating condition subclass.

[0033] Furthermore, the operating condition subclasses include low load-high cooling condition, rated stable condition, overload-low cooling condition, and insufficient lubrication condition; for each operating condition subclass, the corresponding feature distribution, feature weights, and typical temperature rise patterns are extracted to construct operating condition labels and operating condition feature templates; based on the divided operating condition subclasses, the corresponding XGBoost sub-models are trained, and the training process includes sample balancing, hyperparameter search, and cross-validation.

[0034] For example, in this embodiment, a 35-dimensional feature subset output by S1 is selected, including key variables such as power P, head H, rotational speed n, cooling water temperature Tc, cooling water flow rate Fc, lubricating oil temperature To, vibration Va, and oil film thickness estimation feature h.

[0035] To identify differences in operating conditions, four months of historical operating data were first standardized, and K-means clustering was used as the initial method for classifying operating conditions. Elbow rule analysis showed that a cluster size of four resulted in the optimal intra-group density, consistent with unit operating experience. Based on clustering results and operator experience, the operating conditions are divided into the following four categories: Low load - high cooling condition (C1): Characteristics: P is below 30% of rated power, Fc ≥ 90% of rated cooling flow, Tc is stable but low; Rated stable condition (C2): Characteristics: P is close to the rated value of 0.85~1.05, n is stable, cooling capacity is normal, and temperature rise is stable; Overload - low cooling condition (C3): Characteristics: P ≥ 1.1 of rated power, Fc ≤ 75% of rated, and temperature rise rate increases; Insufficient lubrication condition (C4): Characteristics: Po fluctuation decreases, oil temperature To is high, oil film thickness h deviates from the design value, and Va occasionally increases; Subsequently, for each operating condition subclass, the following are extracted: feature distribution, feature weight, and typical temperature rise pattern; A mapping of "operating condition feature template - operating condition label" is constructed. For example: Typical temperature rise pattern under C3 operating condition: average temperature rise rate ΔT / Δt over a 5-minute sliding window > 0.18 ℃ / min; cooling water temperature difference rises by 2-4℃. This information will be used for operating condition identification and sub-model selection during online operation.

[0036] To achieve higher accuracy in identifying operating conditions, a classifier model (such as an XGBoost classifier) ​​is trained based on the clustering results. The input is the filtered 35-dimensional features, and the output is the operating condition label {C1, C2, C3, C4}.

[0037] Example training parameters: Sample size: approximately 420,000 records; Data split: 70% training set + 15% validation set + 15% test set; Feature weights: initially assigned based on the feature importance of S1; Classification accuracy: 94.7%. This model is used for real-time operating condition identification. After the operating condition template is established, sub-models for predicting the temperature of the guide tile are constructed for each of the four operating condition sub-classes.

[0038] Since the sample size of C4 (insufficient lubrication condition) is significantly smaller than that of C1 / C2, SMOTE oversampling and random undersampling are used to balance the sample ratio of the four classes to 1:1:1:1. A grid search method is used for automatic optimization: tree depth max_depth ∈ {4,6,8}; learning rate η ∈ {0.03,0.05,0.1}; subsample rate subsample ∈ {0.6,0.8,1.0}; number of trees n_estimators ∈ {150,200,300}. Five-fold cross-validation is used to finally determine the optimal parameters for each sub-model.

[0039] For example, the optimal parameters for the C3 "overload-low cooling condition" sub-model are: max_depth=8, η=0.05, subsample=0.8, n_estimators=250, and the final validation set RMSE is 0.87 ℃.

[0040] S3: Construct an integrator that outputs multi-step predictions and generates confidence intervals and confidence indices based on residual distributions or integrative differences.

[0041] Furthermore, the global XGBoost model and the XGBoost sub-models corresponding to multiple working condition subclasses are used as candidate model sets; the ensemble weights are calculated based on the current identified working condition, historical residuals, and model stability; the ensemble weights are fused to obtain multi-step predicted values ​​of the lower guide tile temperature; prediction confidence intervals are generated based on the residual distribution, and prediction uncertainty is estimated using the statistical results of near-real-time residuals; an ensemble uncertainty index is constructed based on the model output differences, and the confidence intervals and confidence index are used together as the prediction credibility output.

[0042] Furthermore, the formula for calculating multi-step predicted values ​​is as follows: ;in, To predict the temperature of the lower conductor tile in the next k steps, Let be the predicted value of the i-th model. Let M be the ensemble weight of the i-th model, and M be the number of models; the confidence interval is calculated using the following formula: ;in, , is the confidence coefficient The standard deviation of the residuals within the sliding window; the confidence index is calculated using the following formula: ;in, This is the preset maximum acceptable variance.

[0043] For example, in this embodiment, the following M=5 candidate models are selected as the basic model set of the integrator: XGBoost model G1: trained using all working condition data, with cross-working condition generalization ability; working condition subclass XGBoost models C1, C2, C3, and C4.

[0044] These models maintain a consistent structure but differ in training data, hyperparameters, and feature weights. To improve the adaptability of predictions, this embodiment uses three factors—work condition matching degree, residual statistics, and model stability—to jointly determine the ensemble weights: ① Work condition matching degree weight w g (i) Based on the operating condition identification result at the current time t (e.g., C3), assign higher weights to models that are consistent with the operating condition.

[0045] For example, when the unit is in "overload-low cooling condition (C3)", the following weights can be set: C3 sub-model weight ratio: 0.45; non-current condition sub-model weight ratio: 0.15 (average distribution); XGBoost model weight ratio: 0.25; ② historical residual weight w r (i) Perform sliding window calculation on the model residuals of the most recent 30 minutes. The smaller the residual, the higher the model weight.

[0046] Example calculation: Prediction variance of 5 model outputs: 0.42, preset maximum acceptable variance V armax =2.0; Conf t+1 =1-0.42 / 2.0=0.79, which means the confidence level is 0.79, falling into the "medium-high confidence" range.

[0047] S4: Determines abnormal risks according to multi-level alarm rules, and automatically provides the cause and maintenance suggestions when an alarm is triggered.

[0048] Furthermore, if the predicted value is within the confidence interval and the temperature rise trend is stable, it is judged as normal; if the predicted value is close to the upper limit of the confidence interval or the short-term temperature rise rate increases, it is judged as concerning; if the predicted value exceeds the upper limit of the confidence interval or the rate of decrease in confidence increases, it is judged as a warning; if the predicted value exceeds the confidence interval and the prediction deviation exceeds the threshold, it is judged as serious.

[0049] For example, the judgment is made step by step according to the rules: Step 1: Determine whether it is normal: The monitored value of 54.8℃ falls within the confidence interval (52.1-54.9) → meets the normal conditions, but the short-term temperature rise slope of 0.32℃ / min>0.2 → does not meet the normal conditions, → is judged not to be normal.

[0050] Step 2: Determine the level of concern: Either of the following conditions must be met: the predicted value is ≤0.5℃ from the upper limit → Yes (0.1℃), or the temperature rise slope is ≥0.2℃ / min → Yes (0.32℃ / min), → it is determined to be at least "concerned".

[0051] Step 3: Determine if the warning level has been reached: A warning must meet any of the following conditions: the predicted value exceeds the upper limit of the confidence interval → No, the confidence level decrease rate ≥ 0.05 / min → Yes (0.07 / min), → Determine to upgrade to "warning".

[0052] Step 4: Determine if the severity level is reached: The following conditions must be met simultaneously: Predicted value exceeds the confidence interval → No, prediction deviation ≥ 2.0℃ → No, → Not met.

[0053] Final alert level: Warning.

[0054] In this embodiment, by analyzing the feature contribution, operating condition information, and residual changes, the system automatically generates the following explanations: The significantly increased short-term temperature rise slope (0.32℃ / min) indicates an accelerated temperature rise trend; the high confidence rate decrease (0.07 / min) may correspond to fluctuations in lubrication status or a decrease in cooling efficiency; the increased contribution of "cooling water flow" and "lubricating oil pressure" to the prediction deviation in the feature contribution indicates that the temperature rise is related to cooling and lubrication; the operating condition identification result is "overload-low cooling condition," indicating that the heavy operating conditions lead to a high oil film heat load.

[0055] S5: Deploy sliding window residual and concept drift detection strategies. If distribution drift or residual abnormalities are detected, prioritize incremental fine-tuning, sample reweighting, or triggering offline retraining.

[0056] Furthermore, a residual monitoring mechanism based on a sliding window is deployed to perform dynamic statistical analysis on the online predicted residuals, calculate the residual mean, standard deviation, and skewness index, and determine residual abnormalities when the residuals continuously exceed the set confidence interval; a concept drift detection strategy is introduced to detect the input feature distribution and output distribution in real time based on distribution comparison test or drift detection algorithm; when feature distribution drift and changes in the frequency of operating condition categories are detected, it is determined to be a concept drift event.

[0057] This embodiment also provides a computer device applicable to a machine learning-based adaptive prediction method for the temperature of the lower guide tile, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based adaptive prediction method for the temperature of the lower guide tile as proposed in the above embodiment.

[0058] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0059] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a machine learning-based adaptive prediction method for the temperature of the lower guide tile as proposed in the above embodiment.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based adaptive prediction method for the temperature of a lower guide tile, characterized in that, include: Real-time acquisition and synchronization of multi-source data, fusion of multi-source data, and selection of feature subsets through XGBoost feature importance; Based on the selection of feature subsets, operating conditions are identified from multi-source data. The operating status is divided into multiple operating condition subclasses, and an XGBoost sub-model is trained for each operating condition subclass. An ensembler is constructed to output multi-step predicted values ​​and generate confidence intervals and confidence indices based on residual distribution or ensemble differences. Anomaly risks are judged according to multi-level alarm rules. When an alarm is triggered, the cause and maintenance suggestions are automatically given. A sliding window residual and concept drift detection strategy is deployed. If distribution drift or residual abnormality is detected, incremental fine-tuning, sample reweighting, or offline retraining is prioritized.

2. The machine learning-based adaptive temperature prediction method for lower guide tile as described in claim 1, characterized in that, The process involves real-time acquisition and synchronization of multi-source data, fusion of the multi-source data, and selection of feature subsets using XGBoost feature importance evaluation. This includes: performing time synchronization, time series alignment, and unit normalization on the real-time acquired multi-source data, fusing it into a unified format multi-dimensional time series feature set, and performing missing value imputation, outlier detection, and noise filtering on the fused multi-dimensional time series feature set; constructing extended features based on lubrication theory and hydropower unit operation mechanism, and using the feature importance evaluation mechanism of the XGBoost model to select the extended features, thereby obtaining a feature subset characterizing the factors affecting the temperature change of the lower guide tile.

3. The machine learning-based adaptive temperature prediction method for lower guide tile as described in claim 2, characterized in that, The missing value imputation adopts a hybrid interpolation strategy based on temporal interpolation and weighted similar working conditions, and the outlier detection adopts a joint identification method of statistical threshold judgment and isolated forest model.

4. The machine learning-based adaptive prediction method for lower conductor tile temperature as described in claim 1, characterized in that: The method involves identifying operating conditions from multi-source data based on a selected feature subset, dividing the operating state into multiple operating condition subclasses, and training an XGBoost sub-model for each subclass. These subclasses include low-load-high-cooling condition, rated stable condition, overload-low-cooling condition, and insufficient lubrication condition. For each subclass, corresponding feature distributions, feature weights, and typical temperature rise patterns are extracted to construct operating condition labels and feature templates. Based on the divided operating condition subclasses, corresponding XGBoost sub-models are trained. The training process includes sample balancing, hyperparameter search, and cross-validation.

5. The machine learning-based adaptive temperature prediction method for lower guide tile as described in claim 1, characterized in that, The construction of the integrator, which outputs multi-step predicted values ​​and generates confidence intervals and confidence indices based on residual distribution or integrator differences, includes: using the XGBoost model and XGBoost sub-models corresponding to multiple operating condition subclasses as a candidate model set; calculating integrator weights based on the current identified operating condition, historical residuals, and model stability; fusing based on the integrator weights to obtain multi-step predicted values ​​of the lower guide tile temperature; generating prediction confidence intervals based on residual distribution and estimating prediction uncertainty using statistical results of near-real-time residuals; constructing an integrator uncertainty index based on model output differences, with the confidence intervals and confidence indices jointly serving as the prediction credibility output.

6. The machine learning-based adaptive prediction method for lower conductor tile temperature as described in claim 5, characterized in that: The formula for calculating the multi-step prediction value is as follows: ;in, To predict the temperature of the lower conductor tile in the next k steps, Let be the predicted value of the i-th model. Let M be the ensemble weight of the i-th model, and M be the number of models; the confidence interval is calculated using the following formula: ;in, , is the confidence coefficient The standard deviation of the residuals within the sliding window; the confidence index is calculated using the following formula: ;in, This is the preset maximum acceptable variance.

7. The machine learning-based adaptive prediction method for lower conductor tile temperature as described in claim 1, characterized in that: The method for determining abnormal risks according to multi-level alarm rules includes: if the predicted value is within the confidence interval and the temperature rise trend is stable, it is determined to be normal; if the predicted value is close to the upper limit of the confidence interval or the short-term temperature rise rate increases, it is determined to be of concern; if the predicted value exceeds the upper limit of the confidence interval or the rate of decrease in confidence increases, it is determined to be a warning; if the predicted value exceeds the confidence interval and the prediction deviation exceeds the threshold, it is determined to be serious.

8. The machine learning-based adaptive prediction method for lower conductor tile temperature as described in claim 1, characterized in that: The proposed sliding window residual and concept drift detection strategy prioritizes incremental fine-tuning, sample reweighting, or triggering offline retraining if distribution drift or residual anomalies are detected. This includes: deploying a sliding window-based residual monitoring mechanism to perform dynamic statistical analysis on online predicted residuals, calculating the residual mean, standard deviation, and skewness index; determining residual anomalies when residuals continuously exceed a set confidence interval; introducing a concept drift detection strategy to perform real-time detection of input feature distribution and output distribution based on distribution comparison tests or drift detection algorithms; and determining concept drift events when feature distribution drift and changes in operating condition category frequency are detected.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based adaptive prediction method for the temperature of the lower guide tile as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based adaptive prediction method for the temperature of the lower guide tile as described in any one of claims 1 to 8.

Images