Pumped storage power station unit trend prediction method based on big data and machine learning

Abstract

The application is named as pumped storage power station unit trend prediction method based on big data and machine learning, and belongs to the technical field of pumped storage power station equipment operation and maintenance and fault diagnosis. The technical problem to be solved is that the traditional unit monitoring relies on single parameter fixed value alarm, the false alarm and missed alarm rate is high, and there is a lack of horizontal comparison under the same working condition and performance degradation trend analysis. The technical solution points are that the multi-dimensional operation data and maintenance records of the unit are collected, the training data set with fault label is constructed, and the structured feature vector is converted through feature engineering; the random forest algorithm is used for modeling, the model is optimized through cross validation and grid search, the classification threshold is determined, the real-time data is predicted in the online stage, the maintenance strategy is output by matching the fault knowledge base, the horizontal comparison of the same type unit is carried out synchronously, and the visual report is generated, so as to realize the intelligent prediction and state operation and maintenance of the unit fault.

Description

Technical Field

[0001] This invention belongs to the field of operation and maintenance and fault diagnosis technology of pumped storage power station equipment, specifically involving a trend prediction method for pumped storage power station units based on big data and machine learning. Background Technology

[0002] Pumped storage power stations, serving as "stabilizers" and "power banks" for the power system, are core facilities for peak shaving and valley filling, frequency regulation, phase adjustment, and emergency backup. Pumped storage units are characterized by frequent start-ups and shutdowns, complex operating conditions (supporting bidirectional operation of pumps / turbines), and rapid load regulation response. This unique operating environment causes critical components of the unit, including thrust bearings, guide bearings, turbine runners, and main shaft seals, to be subjected to alternating loads and operating condition shocks for extended periods, making them highly susceptible to fatigue damage and performance degradation. This places extremely high demands on the unit's condition monitoring, fault diagnosis, and operation and maintenance management.

[0003] Currently, conventional equipment monitoring and fault diagnosis solutions for pumped storage power station units mainly rely on independent monitoring systems such as SCADA (Supervisory Control and Data Acquisition) systems, CMS (Vibration Monitoring System), and oil chromatography analysis systems. These systems primarily employ a fixed threshold alarm mode for a single operating parameter, combined with the on-site experience of maintenance personnel to determine faults and make repair decisions. Some existing diagnostic systems only perform longitudinal trend comparisons of historical operating data for a single unit to assess changes in equipment operating status. These existing technical solutions suffer from the following insurmountable technical shortcomings in practical applications:

[0004] (1) The alarm logic relies on human experience and single parameter setpoint alarms, resulting in high false alarm and false alarm rates. Traditional monitoring systems only set fixed alarm thresholds for independent parameters such as temperature and vibration, lacking in-depth analysis of the coupling relationship between multiple operating parameters. They cannot effectively distinguish between normal parameter changes caused by unit operating condition fluctuations and parameter deviations caused by equipment abnormalities. The effectiveness of alarms highly depends on the human experience of operation and maintenance personnel, leading to prominent problems of false alarms and false alarms.

[0005] (2) Lack of horizontal comparison dimension under the same operating conditions, and insufficient ability to calibrate the status reference value. Existing diagnostic systems are mostly limited to the vertical comparison analysis of the historical data of a single unit, completely ignoring the natural horizontal comparability of the operating data of multiple units of the same model in the same power plant under the same operating conditions. They cannot use the healthy operating status of "brother units" in the same field to calibrate the status reference value of the unit to be monitored, and cannot eliminate the individual deviations caused by the manufacturing and installation of different units. This can easily lead to the inaccuracy of the unit status judgment benchmark and make it impossible to accurately identify the early performance degradation of the equipment.

[0006] (3) Fault warning is seriously delayed and lacks early prediction capability of equipment degradation. Existing technology only triggers alarms after the equipment operating parameters reach the preset alarm value or even after the fault has occurred. It does not conduct forward-looking trend analysis based on the time sequence of equipment performance degradation, lacks early prediction capability of faults, and cannot achieve early intervention in the bud stage of faults. It can only passively carry out post-fault maintenance, which is very likely to cause unplanned unit shutdowns and seriously affect the operation safety and availability of the power plant.

[0007] (4) Multi-source data forms information silos, and the value of the data has not been fully explored. In the existing technology, multi-source monitoring data such as SCADA system, CMS system, and oil chromatography analysis system are independent of each other and have not achieved deep integration and correlation analysis. It is impossible to build a complete health status profile of the equipment based on full-dimensional operation data, and the potential value of monitoring data cannot be fully released, making it difficult to support high-precision, multi-dimensional equipment status assessment and fault diagnosis.

[0008] In view of this, the present invention is hereby proposed. Summary of the Invention

[0009] To address the aforementioned technical problems in existing technologies, this invention provides a trend prediction method for pumped storage power station units based on big data and machine learning, which solves the problems of easy fatigue damage to key components of the units, difficulty in early identification of performance degradation, reliance on manual experience for operation and maintenance, and inefficient maintenance modes.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows: A trend prediction method for pumped storage power station units based on big data and machine learning includes: S1. Construct a training dataset with fault labels for the target fault types of pumped storage power station units, collect historical operating data and corresponding maintenance records of the units, and perform binary classification labeling on the corresponding data segments of historical operating data based on the maintenance records. S2. Perform feature engineering on the labeled dataset to extract multidimensional features from the data fragments and transform the original time-series running data into structured feature vectors. S3. Perform stratified sampling and feature standardization on the dataset corresponding to the feature vectors, and save the standardization parameters. S4. The random forest algorithm is used to construct independent binary classifiers for each target fault type. The structured feature vector is used as the input feature and the corresponding fault label is used as the target value to complete the model training. During the training process, the hyperparameters of the model are tuned through cross-validation and grid search. S5. Perform performance evaluation on the models corresponding to each fault type after training, and optimize and determine the classification threshold for each model. S6. Store the model that meets the evaluation criteria, the corresponding standardized parameters, and the model metadata; S7. In the online prediction stage, the model and standardized parameters corresponding to the fault type are loaded. After preprocessing, feature extraction and standardization transformation of the real-time collected unit operation data, the corresponding model is input and the fault prediction result is output. S8. Based on the output fault prediction results, match the preset fault knowledge base and output the corresponding maintenance recommendation strategy.

[0011] Furthermore, in step S1, the target fault types include guide bearing wear, cooler blockage, oil-water mixing, insulation aging, and cavitation. During the two-category labeling process, different labeling window lengths are set according to the fault mechanisms of different fault types. The labeling window length for oil-water mixing and cavitation is 1 day before the fault occurs, the labeling window length for guide bearing wear and cooler blockage is 3 days before the fault occurs, and the labeling window length for insulation aging is 7 days before the fault occurs. Data segments within the annotation window before the fault occurs are marked as abnormal, while data segments of equal length within the normal operating period are marked as normal.

[0012] Furthermore, the collected operational data includes vibration measurement point data, swing measurement point data, temperature measurement point data, pressure measurement point data, electrical quantity data, and oil quality monitoring data, with a data sampling frequency of 1 second to 1 minute.

[0013] Furthermore, the preprocessing of the operational data includes data cleaning, noise reduction, outlier removal, and missing value removal.

[0014] Furthermore, in step S3, the datasets corresponding to each fault type are randomly divided into training set and test set in a ratio of 8:2. The division process adopts stratified sampling so that the ratio of positive and negative samples in the training set and test set is consistent with that in the original dataset. The feature standardization process uses Z-score normalization, and the saved standardization parameters include the mean and standard deviation of the features.

[0015] Furthermore, in step S4, the model training process uses 5-fold cross-validation to perform hyperparameter tuning. The hyperparameters optimized through grid search include the number of decision trees, the maximum depth of the trees, the minimum number of samples required for node splitting, and the minimum number of samples required for leaf nodes. The Gini coefficient was used as the node splitting criterion, and the weighted F1-score was used as the evaluation index for hyperparameter optimization. The hyperparameter combination with the highest average weighted F1-score after cross-validation was selected as the final model configuration.

[0016] Furthermore, in step S5, the calculation metrics for model performance evaluation include accuracy, precision, recall, and F1-score, while plotting the ROC curve and calculating the AUC value. The optimization process for the classification threshold is as follows: iterate through all thresholds in the range of 0.1 to 0.9 with a step size of 0.05, and select the threshold that maximizes the F1-score of the corresponding model as the final classification threshold of the model.

[0017] Furthermore, in step S6, the evaluated and qualified models are stored in pickle format, and the stored model metadata includes model version number, training time, feature list, and model evaluation metrics.

[0018] Furthermore, in step S7, during the analysis of real-time operating data, a horizontal comparison of the operating data of the same type of generating units within the same power station under the same operating conditions is performed simultaneously to identify performance differences between the units.

[0019] The horizontal comparison section displays the trend status of different units on a visual chart page. Users can freely select the data to be compared according to their needs, and then present the selected data in the page's curve chart for comparison.

[0020] Furthermore, it also includes step S9, which generates a unit health status trend chart and intelligent report, displays the current unit's operating status comparison data with other units in the same power station, and completes the visual interaction.

[0021] The beneficial effects of this invention are as follows: (1) By comparing the same type of generator units in the same power plant, the deviation of the state reference caused by the manufacturing and installation errors of different units can be eliminated, so that old units can also obtain accurate health status references. (2) By combining the trend analysis of equipment operation data, the trend of equipment performance degradation can be detected in advance before the parameters reach the traditional alarm set value, so as to realize early warning of faults; (3) By integrating multi-dimensional operational data and adopting a multi-model judgment mechanism, false alarms caused by instantaneous interference from sensors are effectively filtered out, reducing the false alarm rate of fault alarms; (4) Provide fault prediction and maintenance suggestions based on the actual operating status of the equipment, so that the unit operation and maintenance can be changed from regular maintenance to condition-based maintenance, thereby reducing operation and maintenance costs and improving unit availability. Attached Figure Description

[0022] Figure 1 A flowchart of a trend prediction method for pumped storage power station units provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0024] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0025] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0026] Example 1 See Figure 1 , Figure 1 This is a flowchart of the trend prediction method for pumped storage power station units based on big data and machine learning proposed in this invention. Specific steps may include: Before officially starting the data acquisition and analysis process, complete the system initialization and core parameter configuration, including: configuring the core operating parameters of the sensor network, including the sampling frequency and acquisition accuracy of each measuring point; loading the pre-trained machine learning model parameters, standardized parameters, and fault knowledge base association rules; configuring the unit benchmarking range and operating condition screening conditions for horizontal comparison under the same operating conditions, and completing the preparation work before the system runs through the entire process.

[0027] S1. Construct a training dataset with fault labels: Construct a training dataset with fault labels for the target fault types of pumped storage power station units. Collect historical operating data and corresponding maintenance records of the units. Perform binary classification labeling on the corresponding data segments of historical operating data based on the maintenance records to complete the construction of the training dataset.

[0028] Among them, the target fault types are five core faults that are common in pumped storage units and have high data availability. Specifically, they include guide bearing wear, cooler blockage, oil-water mixing, insulation aging, and cavitation, comprehensively covering the core fault scenarios of the unit's bearing system, cooling system, oil system, electrical insulation system, and overcurrent components.

[0029] The collected historical operating data is fully integrated with multi-source monitoring data from SCADA (Supervisory and Data Acquisition) system, CMS (Vibration Monitoring) system, and oil chromatography analysis system, breaking down the data silos of existing technologies. Specific data types include: vibration measurement point data (upper guide X / Y direction, lower guide X / Y direction, water guide X / Y direction), sway measurement point data, temperature measurement point data (temperature of each guide bearing bearing, lubricating oil temperature, inlet and outlet temperatures of cooling water), pressure measurement point data (volute pressure, tailpipe pressure), flow parameters, electrical quantity data (voltage, current, power, insulation resistance, partial discharge), and oil quality monitoring data (water content, oil level). The sampling frequency of each type of data varies from 1 second to 1 minute to adapt to the monitoring characteristics and fault mechanisms of different parameters.

[0030] Data annotation employs a binary classification method. The specific annotation rules are as follows: For each fault type, the fault occurrence time is determined based on maintenance records. A data segment within a window preceding the fault occurrence is extracted and labeled as 1 (abnormal). During normal operation periods with no fault records, a data segment of the same length as the abnormal segment is extracted and labeled as 0 (normal). The annotation window length is set differently based on the fault mechanism of different fault types: oil-water mixing and cavitation faults develop rapidly, so the annotation window length is 1 day before the fault occurs; guide bearing wear and cooler blockage are progressive faults, so the annotation window length is 3 days before the fault occurs; insulation aging is a long-term, gradual fault, so the annotation window length is 7 days before the fault occurs. This ensures a high degree of match between the annotated data and the fault development pattern, improving the accuracy of model training.

[0031] S2. Perform feature engineering on the labeled data: Perform feature engineering on the labeled dataset to extract multidimensional features from the data fragments and transform the original time-series running data into structured feature vectors; Specifically, feature engineering is performed on the labeled dataset. First, data cleaning, noise reduction, outlier removal, and missing value removal are performed on the collected raw running data and labeled data segments to improve data quality. Then, multi-dimensional time-domain and frequency-domain features are extracted from each preprocessed data segment to transform the raw time-series running data into structured feature vectors that can be efficiently processed by algorithms such as random forest and time series analysis. Feature vectors of device operating status are constructed to complete the feature engineering process.

[0032] S3. Dataset partitioning and standardization: Perform stratified sampling partitioning and feature standardization on the dataset corresponding to the feature vectors, and save the standardization parameters; Specifically, the structured feature vector dataset corresponding to each fault type obtained in step S2 is randomly divided into a training set and a test set in a ratio of 8:2. Stratified sampling is used during the division to ensure that the ratio of positive and negative samples in the training set and the test set is consistent with that in the original dataset, so as to avoid model training distortion caused by sample distribution bias.

[0033] Simultaneously, the features of the training and test sets are standardized. In this embodiment, Z-score normalization is used to complete the standardization transformation, and the standardization parameters, including the mean and standard deviation of the corresponding features, are saved in time so that the same standardization transformation can be performed on the real-time collected running data to ensure the consistency of the model input data.

[0034] S4. Model Training and Hyperparameter Tuning: The random forest algorithm is used to construct independent binary classifiers for each target fault type. The structured feature vector is used as the input feature and the corresponding fault label is used as the target value to complete the model training. During the training process, the model hyperparameters are tuned through cross-validation and grid search. Specifically, the feature vector constructed in step S3 is used as the input feature X, and the fault type label marked in S1 is used as the target value Y to train the training set. During training, 5-fold cross-validation is used for hyperparameter tuning, and grid search is used to optimize the key hyperparameters of the random forest.

[0035] Specifically, for training the random forest classifier, 5-fold cross-validation was used for hyperparameter tuning during training, and grid search was used to optimize the key hyperparameters of the random forest. The searched hyperparameters included: the number of decision trees (n_estimators), the maximum depth of the trees (max_depth), the minimum number of samples required for node splitting (min_samples_split), and the minimum number of samples required for leaf nodes (min_samples_leaf). The evaluation metric for hyperparameter optimization was the weighted F1-score, which fully considered the positive and negative sample balance problem. The hyperparameter combination with the highest average weighted F1-score after cross-validation was selected as the final model configuration. The Gini coefficient was used as the node splitting criterion to complete the training and hyperparameter optimization of the classifiers corresponding to each fault type.

[0036] Meanwhile, in order to predict the long-term trend of unit performance degradation, the LSTM time series analysis algorithm is used to construct a predictive model of equipment performance degradation, so as to realize the forward-looking trend analysis of equipment operating status and provide time-series support for early fault prediction.

[0037] S5. Model Evaluation and Threshold Determination: Perform performance evaluation on the models corresponding to each fault type after training, and optimize and determine the classification thresholds corresponding to each model. Specifically, on the test set divided in step S3, the trained model for each fault type is independently evaluated. The core evaluation metrics include: accuracy, precision, recall, and F1-score. At the same time, ROC curves are plotted and AUC values ​​are calculated to fully verify the classification performance and generalization ability of the model.

[0038] To balance false alarms and false negatives in fault prediction, the classification threshold is optimized individually for each model on the test set: all thresholds in the range of 0.1 to 0.9 with a step size of 0.05 are traversed, and the threshold that maximizes the F1-score of the corresponding model is selected as the final classification threshold for that model, thus adapting to the early warning and control requirements of different fault types.

[0039] S6. Model Storage: Store the models that have passed the evaluation, their corresponding standardized parameters, and model metadata in a standardized manner to support model invocation in the subsequent online prediction stage; Specifically, the trained and evaluated random forest model file is exported in pickle format and stored in the system-specified path along with the standardized parameters saved in step S3. The model metadata is recorded and stored synchronously, including the model version number, training time, input feature list, core evaluation metrics, corresponding fault types, and classification thresholds. This enables traceable management of the entire model lifecycle, facilitating subsequent model iteration, deployment, and compliance auditing.

[0040] S7. Load the model and perform fault prediction on real-time data: In the online prediction stage, load the model and standardized parameters corresponding to the fault type, and after preprocessing, feature extraction and standardization transformation of the real-time collected unit operation data, input the corresponding model and output the fault prediction results; at the same time, carry out horizontal comparative analysis and trend analysis of the unit operation status.

[0041] Specifically, during the actual online operation phase, the system collects real-time, multi-dimensional operating status data of pumped storage power station unit equipment through the sensor network deployed in the data acquisition layer, including but not limited to parameters such as vibration, temperature, pressure, and flow rate, and transmits the collected real-time data to the system data center via wired or wireless means. Perform the same preprocessing and feature extraction operations as in step S2 on the raw data collected in real time to obtain the feature vector corresponding to the current unit's real-time operation data. After performing a standardization transformation on the feature vector using the standardization parameters saved in step S3, input it into the random forest model corresponding to the fault type stored in step S6. The model outputs the prediction results of the corresponding fault, including fault type, fault occurrence probability, and risk level.

[0042] Simultaneously, two core analyses are executed during real-time data analysis: Horizontal comparative analysis: For multiple units of the same model in the same power plant, the operating data under the same operating conditions are selected and horizontal comparative analysis is carried out. The health operation status of the "sibling units" in the same field is used to calibrate the current unit's status benchmark value, identify the performance differences between units, eliminate individual deviations caused by manufacturing and installation of different units, and solve the problem of no accurate reference benchmark for old units. Trend Analysis: Based on real-time and historical operating data of the unit, trend analysis is carried out through LSTM time series prediction model to fit the time sequence pattern of equipment performance degradation, identify early signs of equipment performance degradation, and achieve proactive early warning before parameters reach traditional alarm settings.

[0043] S8. Fault Database Matching and Repair Recommendation: Based on the output fault prediction results, match the preset fault knowledge base and output the corresponding repair recommendation strategy to provide decision support for operation and maintenance personnel.

[0044] Specifically, based on the output fault type, fault occurrence probability and risk level, the system retrieves the pre-stored association rules in the fault knowledge base. The association rules include the fault occurrence mechanism, fault cause investigation path, standardized maintenance strategy and spare parts suggestions for the corresponding fault. After sorting the retrieved association rules by confidence level, the system recommends graded auxiliary decision-making suggestions and maintenance strategies to the operation and maintenance personnel. For high-risk emergency failures, an immediate shutdown and maintenance strategy is recommended; for medium- to low-risk gradual failures, a planned shutdown and preventative maintenance strategy is recommended to guide maintenance personnel to conduct precise condition-based maintenance, replacing the traditional scheduled maintenance model.

[0045] S9. Visual Interaction and Intelligent Reports: Generates unit health status trend charts and intelligent reports, displays comparative data on the current unit's operating status with other units in the same power station, and completes visual interaction; Specifically, based on real-time analysis results, a visualized "Unit Health Status Trend Chart" is generated, which intuitively displays the current performance degradation trend, fault risk level, and horizontal comparison between the current unit and other units of the same model in the same plant. Customizable interaction on the visualized chart page is also supported, allowing users to freely select the operating data, unit range, and time period for comparison according to their needs, and present the selected data in the comparative graph on the page. Standardized intelligent operation and maintenance reports are generated regularly to record changes in unit health status, fault warning records, and the implementation status of maintenance recommendations, providing data support for the full lifecycle operation and maintenance management of power plant units.

[0046] In this embodiment, after completing a single data analysis and fault prediction, the system loops back to the real-time data acquisition step, continuously performing real-time acquisition, processing, analysis, and fault prediction of unit operation data, thereby achieving full-time, uninterrupted intelligent monitoring and operation and maintenance management of the unit's operating status.

[0047] The above specific 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 examples, 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 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 method for predicting trends of pumped storage power station units based on big data and machine learning, characterized in that, include: S1. Construct a training dataset with fault labels for the target fault types of pumped storage power station units, collect historical operating data and corresponding maintenance records of the units, and perform binary classification labeling on the corresponding data segments of historical operating data based on the maintenance records. S2. Perform feature engineering on the labeled dataset to extract multidimensional features from the data fragments and transform the original time-series running data into structured feature vectors. S3. Perform stratified sampling and feature standardization on the dataset corresponding to the feature vectors, and save the standardization parameters. S4. The random forest algorithm is used to construct independent binary classifiers for each target fault type. The structured feature vector is used as the input feature and the corresponding fault label is used as the target value to complete the model training. During the training process, the hyperparameters of the model are tuned through cross-validation and grid search. S5. Perform performance evaluation on the models corresponding to each fault type after training, and optimize and determine the classification threshold for each model. S6. Store the model that meets the evaluation criteria, the corresponding standardized parameters, and the model metadata; S7. In the online prediction stage, the model and standardized parameters corresponding to the fault type are loaded. After preprocessing, feature extraction and standardization transformation of the real-time collected unit operation data, the corresponding model is input and the fault prediction result is output. S8. Based on the output fault prediction results, match the preset fault knowledge base and output the corresponding maintenance recommendation strategy.

2. The big data and machine learning based pumped storage power plant unit trend prediction method of claim 1, wherein, In step S1, the target fault types include guide bearing wear, cooler blockage, oil-water mixing, insulation aging, and cavitation. During the two-category labeling process, different labeling window lengths are set according to the fault mechanisms of different fault types. The labeling window length for oil-water mixing and cavitation is 1 day before the fault occurs, the labeling window length for guide bearing wear and cooler blockage is 3 days before the fault occurs, and the labeling window length for insulation aging is 7 days before the fault occurs. Data segments within the annotation window before the fault occurs are marked as abnormal, while data segments of equal length within the normal operating period are marked as normal. 3.The big data and machine learning based pumped storage power plant unit trend prediction method according to claim 1, wherein, The collected operational data includes vibration measurement point data, swing measurement point data, temperature measurement point data, pressure measurement point data, electrical quantity data, and oil quality monitoring data, with a data sampling frequency of 1 second to 1 minute.

4. The big data and machine learning based pumped storage power plant unit trend prediction method of claim 1, wherein, Preprocessing of operational data includes data cleaning, noise reduction, outlier removal, and missing value removal.

5. The method for trend prediction of pumped storage power station units based on big data and machine learning according to claim 1, characterized in that, In step S3, the datasets corresponding to each fault type are randomly divided into training set and test set in a ratio of 8:

2. The division process adopts stratified sampling so that the ratio of positive and negative samples in the training set and test set is consistent with the original dataset. The feature standardization process uses Z-score normalization, and the saved standardization parameters include the mean and standard deviation of the features.

6. The big data and machine learning based pumped storage power plant unit trend prediction method of claim 1, wherein, In step S4, the model training process uses 5-fold cross-validation to perform hyperparameter tuning. The hyperparameters optimized through grid search include the number of decision trees, the maximum depth of the trees, the minimum number of samples required for node splitting, and the minimum number of samples required for leaf nodes. The Gini coefficient was used as the node splitting criterion, and the weighted F1-score was used as the evaluation index for hyperparameter optimization. The hyperparameter combination with the highest average weighted F1-score after cross-validation was selected as the final model configuration.

7. The big data and machine learning based pumped storage power plant unit trend prediction method of claim 1, wherein, In step S5, the calculation metrics for model performance evaluation include accuracy, precision, recall, and F1-score. At the same time, the ROC curve is plotted and the AUC value is calculated. The optimization process for the classification threshold is as follows: iterate through all thresholds in the range of 0.1 to 0.9 with a step size of 0.05, and select the threshold that maximizes the F1-score of the corresponding model as the final classification threshold of the model.

8. The big data and machine learning based pumped storage power plant unit trend prediction method of claim 1, wherein, In step S6, the evaluated and qualified models are stored in pickle format. The stored model metadata includes model version number, training time, feature list, and model evaluation metrics. 9.The big data and machine learning based pumped storage power plant unit trend prediction method according to claim 1, wherein, In step S7, during the analysis of real-time operating data, a horizontal comparison of the operating data of the same type of unit in the same power station under the same operating conditions is performed simultaneously to identify the performance differences between the units. The horizontal comparison section displays the trend status of different units on a visual chart page. Users can freely select the data to be compared according to their needs, and then present the selected data in the page's curve chart for comparison.

10. The big data and machine learning based pumped storage power plant unit trend prediction method of claim 1, wherein, It also includes step S9, which generates a unit health status trend chart and intelligent report, displays the current unit's operating status comparison data with other units in the same power station, and completes the visual interaction.

Images