Water-turbine generator set bearing oil level early warning method based on TimesNet model

Abstract

The invention relates to a water-turbine generator set bearing oil level early warning method based on a TimesNet model, and the method comprises the steps: obtaining monitoring data of a water-turbine generator set, converting the monitoring data of the water-turbine generator set into structural data, and carrying out the steady-state screening of the data of the water-turbine generator set; analyzing the business characteristics of the unit based on the unit operation condition data; aiming at the data characteristics of the bearing oil level of the unit, combining the business characteristics of the unit, and using a TimesNet model for fitting; calculating a reference value of a dynamic threshold value band of the bearing oil level of the unit based on a fitting result; the trained TimesNet model is used for predicting the bearing oil level of the water-turbine generator set in the production environment, and according to the oil level prediction value of the TimesNet model, the oil level dynamic threshold value judgment and the oil level change trend judgment are carried out. According to the method, the early warning accuracy of the bearing oil level of the unit is remarkably improved, and the method can adapt to the change of the running state of the unit by setting the dynamic threshold value band of the oil level.

Description

Technical Field

[0001] The present invention belongs to the field of hydropower generation monitoring, and particularly relates to a bearing oil level warning method for a hydro-generator unit based on the TimesNet model. Background Art

[0002] Due to the operation of the hydro-generator unit being affected by various factors such as hydraulics, machinery, and electricity, it may cause safety risks such as deterioration of equipment operating conditions, abnormal vibration of components, fatigue damage of parts, and even structural damage.

[0003] The bearing is a mechanical support component of the hydro-generator unit and plays an important role in the safe and stable operation of the unit. The turbine oil lubricates and dissipates heat for the bearing, and the bearing oil level of the hydro-generator unit needs to be at a reasonable level to ensure the efficient functioning of the turbine oil. How to predict the bearing oil level of the hydro-generator unit and give a timely fault warning is of great importance.

[0004] Most of the existing oil level warning methods are mechanism-based methods, which set thresholds for the oil level data of the bearing oil tank of the hydro-generator unit for warning. The method of setting fixed thresholds has a certain effect on the warning of the bearing oil tank oil level, but there are also the following deficiencies: (1) It can only alarm the data at the current moment; (2) The technical requirements for threshold setting are relatively high. If the threshold is too low, false alarms are likely to occur frequently; if the threshold is too high, missed alarms are likely to occur; (3) The model generalization ability is insufficient. If the current oil level is higher than the historical oil level and does not exceed the threshold, it is impossible to determine whether the oil level is normal. Summary of the Invention

[0005] The purpose of the present invention is to address the above problems, and provide a bearing oil level warning method for a hydro-generator unit based on the TimesNet model, which combines the operating condition data and business characteristics of the unit, and uses the advanced TimesNet model for fitting and prediction to improve the accuracy of the bearing oil level warning of the unit; design a dynamic threshold band, calculate the residual between the predicted oil level and the actual oil level, and use its standard deviation as the reference value of the oil level dynamic threshold band to achieve dynamic monitoring and warning of oil level changes.

[0006] The technical solution of the present invention is a bearing oil level warning method for a hydro-generator unit based on the TimesNet model, including the following steps: Step 1: Obtain the monitoring data of the hydro-generator unit, including the active power, bearing temperature, and bearing oil level of the unit; Step 2: After converting the unit data obtained in Step 1 into structured data, perform steady-state screening on the unit data; Step 3: Analyze the business characteristics of the unit based on the unit operating condition data, and the business characteristics include whether the bearing oil level of the hydro-generator unit continues to increase; Step 4: Divide the unit data into a training set and an evaluation data set; Step 5: Fit the data characteristics of the unit bearing oil level by using the TimesNet model in combination with the business characteristics of the unit; Step 6: Based on the fitting result obtained in Step 5, calculate the reference value of the dynamic threshold band of the unit bearing oil level; Step 6.1: Input the evaluation data set into the TimesNet model, and use the TimesNet model to calculate the predicted oil level of the unit bearing; Step 6.2: Calculate the residual between the predicted oil level value of the TimesNet model and the oil level value of the evaluation data set, and take the standard deviation of the residual as the reference value of the dynamic threshold band of the oil level; Step 7: Use the trained TimesNet model for predicting the bearing oil level of the hydro-generator unit in the production environment, and make a judgment on the dynamic threshold of the oil level and the trend of oil level change according to the oil level prediction value of the TimesNet model.

[0007] Preferably, the specific steps of Step 2 include: Data conversion: Convert the time series data into structured data that is easy for users to analyze; Data filling: Perform data filling processing for unit data defects; Steady-state screening: Screen out the data when the unit operates in a steady state; Data downsampling: Align the timestamps of the unit data by means of data downsampling; Outlier removal: Remove the abnormal data caused by external interference.

[0008] Further, the judgment method for the steady-state screening in Step 2 is: If the fluctuation of the unit active power is not greater than 20 MW within 30 minutes, it is judged that the unit is in a steady state.

[0009] Preferably, the method of using the 3σ criterion to judge the outliers in the unit data in Step 2 is: If the difference between a single item of unit data and the mean value of this item of unit data is greater than 3 times the standard deviation, it is judged as an outlier. Further, in Step 4, the unit data is divided into a training set and an evaluation data set according to a ratio of 7:3, that is, 70% of the unit data is used as the training set, and 30% of the unit data is used as the evaluation data set.

[0010] Further, in Step 5, the TimesNet model analyzes the time series changes of the unit from a multi-period perspective, specifically including: 1) Fold the one-dimensional time series based on multiple periods to obtain multiple two-dimensional tensors, and the columns and rows of each two-dimensional tensor respectively reflect the time series changes within and between periods; 2) Use the Inception module of the TimesNet model to extract two-dimensional time-series change features from the two-dimensional tensor; 3) Transform the two-dimensional time-series change features into one-dimensional space for convenient information aggregation.

[0011] Furthermore, the Inception module includes the following parallel convolution calculation units: (1) A convolutional layer with a 1×1 convolutional kernel; (2) A convolutional layer with a 1×1 convolutional kernel and a convolutional layer with a 3×3 convolutional kernel; (3) A convolutional layer with a 1×1 convolutional kernel and a convolutional layer with a 5×5 convolutional kernel; (4) A max-pooling layer with a 3×3 convolutional kernel and a convolutional layer with a 1×1 convolutional kernel.

[0012] Preferably, step 7 specifically includes: Step 7.1: Use the TimesNet model to perform real-time prediction on the oil level of the unit bearing, and judge whether the predicted oil level value is within the range of the oil level dynamic threshold band. If the predicted oil level value exceeds the range of the oil level dynamic threshold band, execute step 7.2; if the predicted oil level value is within the range of the oil level dynamic threshold band, repeat step 7.1; Step 7.2: Observe and judge the change trend of the oil level of the unit bearing. If the oil level of the unit bearing shows an upward trend, issue a warning signal and execute step 7.1.

[0013] Compared with the prior art, the beneficial effects of the present invention include: 1) By combining the unit operation condition data and business characteristics, the present invention uses the TimesNet model for fitting and prediction. The TimesNet model is an advanced prediction model that can capture the time-series characteristics and non-linear relationships in the data, significantly improving the accuracy of the oil level warning of the unit bearing; calculate the residual between the predicted oil level and the actual oil level, and use its standard deviation as the reference value of the oil level dynamic threshold band. By setting the oil level dynamic threshold band, the method of the present invention can adapt to the changes in the unit operation state and enhance the reliability and stability of the oil level warning method.

[0014] 2) By converting the monitoring data into structured data and performing steady-state screening, the present invention improves the accuracy and reliability of the data.

[0015] 3) Based on the oil level prediction value of the TimesNet model and combined with the dynamic threshold band, the present invention can timely issue an early warning signal for the oil level value. When the overall trend of the oil level shows an upward trend, it can timely issue a trend early warning signal. Accurate oil level early warning can help operation and maintenance personnel timely discover and handle potential bearing faults, optimize maintenance strategies, and reduce unplanned downtime. Through real-time monitoring and early warning, abnormal conditions during the operation of the unit can be timely discovered and handled, improving the operation efficiency and safety of the hydro-generator unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the drawings and embodiments.

[0017] Figure 1 It is a schematic flowchart of the method for warning the bearing oil level of a hydro-generator unit based on the TimesNet model according to an embodiment of the present invention.

[0018] Figure 2 It is a schematic diagram of the data processing process of the TimesNet model according to an embodiment of the present invention.

[0019] Figure 3 It is a schematic diagram of the Inception module according to an embodiment of the present invention.

[0020] Figure 4 It is a diagram of the steady-state screening parameter configuration interface according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] As Figure 1 shown, the method for warning the bearing oil level of a hydro-generator unit based on the TimesNet model includes the following steps: Step 1: Obtain unit monitoring data. At regular intervals, communicate with the big data platform of the hydropower station to obtain the monitoring data of the hydro-generator unit, including data such as active power, bearing tile temperature, and bearing oil level.

[0022] Step 2: Perform data processing in sequence through data table conversion, data value filling, steady-state screening, data downsampling, and outlier screening.

[0023] The data processing in the embodiment specifically includes: Data table conversion: Convert time series data into structured data that is easy for users to analyze. The specific method is to convert from a long table to a wide table. For example, the header of the long table is id, time, v, and the header of the converted wide table is time, active power, bearing tile temperature, oil level, etc.; Data value filling: Since the timestamps of data at different measurement points in each system may not be aligned, there may be missing data after table conversion; or there may be missing data in on-site collection, so data value filling processing is performed on the data; Steady-state screening: Screen the data when the unit is operating in a steady state. The specific screening method is that when the active power fluctuates less than 20 MW within 30 minutes, the data segment is considered to be in a steady state; Data downsampling to align the timestamps of the data; Outlier screening: The data collected on-site may have abnormal data due to external factor interference. The specific screening method is that according to the 3-sigma principle, in the dataset, if the difference between the data and the mean value is less than 3 times the standard deviation, the data is retained.

[0024] In the embodiment, the parameters of the steady-state filtering operator are set as Figure 4 shown.

[0025] Step 3: Based on the obtained operating condition data, use the look-up table method to analyze its business characteristics. The business characteristics include that under normal circumstances, under stable operating conditions, although the bearing oil level value of the hydro-generator unit will fluctuate, there will be no continuous increase for more than 1 hour.

[0026] Step 4: Divide the data processed in Step 3. Specifically, the first 70% of the data is designated as the training set, and the last 30% of the data is designated as the evaluation dataset.

[0027] Step 5: For the data characteristics of the oil sump oil level, combined with its business characteristics, select the prediction method of TimesNet for long time series to fit the training dataset.

[0028] As Figure 2 shown, the TimesNet prediction model analyzes the time series changes from a multi-period perspective, expands the one-dimensional time series data to two-dimensional space for analysis. Folding the one-dimensional time series based on multiple periods can obtain multiple two-dimensional tensors (2D tensors). The columns and rows of each two-dimensional tensor respectively reflect the time series changes within and between periods, that is, the two-dimensional time series changes (Temporal 2D-variations) are obtained. After converting the data to two dimensions, use the Inception module to extract the two-dimensional time series change representation. The structure of the Inception module is as Figure 3 shown. The Inception module consists of convolution operation and max pooling operation. After extracting the representation data, convert it back to one-dimensional space for information aggregation.

[0029] Step 6: For the fitting result, calculate the reference value of the dynamic threshold band, specifically: Use the algorithm to make predictions on the evaluation dataset, obtain the predicted oil level of the evaluation dataset, and then obtain the residual data of the oil level. The residual of the oil level is calculated by subtracting the predicted oil level from the original oil level. The standard deviation of the residual data is the reference value of the dynamic threshold band.

[0030] Step 7: Apply the trained TimesNet model to predict the bearing oil level of the hydro-generator unit in the production environment. Based on the oil level prediction value of the TimesNet model, perform dynamic threshold judgment of the oil level and judgment of the oil level change trend.

[0031] Deploy the TimesNet model to the production environment through the power industry Internet platform and set it to schedule and run once every half hour.

[0032] Based on the real-time prediction result of the TimesNet model, judge whether to issue a warning signal: (1) Numerical warning: For the predicted numerical result, combined with the dynamic threshold band, perform dynamic threshold warning judgment. If the predicted data exceeds the threshold band, conduct trend observation; the calculation method of the dynamic threshold band is three times the positive and negative dynamic threshold reference values of the real-time data set.

[0033] (2) Trend warning: Calculate the trend of the oil level. If the overall trend shows an upward trend, perform alarm processing on the oil level.

[0034] The system of the above-mentioned hydro-generator unit bearing oil level warning method includes: Data acquisition module: Used to obtain the monitoring data of the hydro-generator unit, including the active power, bearing temperature, and bearing oil level of the unit; Data preprocessing module: Used to preprocess the monitoring data obtained by the data acquisition module. After converting it into structured data, perform steady-state screening on the unit data; Oil level prediction model module: Construct a TimesNet model, train it, and use the trained TimesNet model to calculate the predicted oil level of the unit bearing; Dynamic threshold calculation module: Calculate the residual between the predicted oil level value of the TimesNet model and the oil level value of the evaluation data set, and use the standard deviation of the residual as the reference value of the oil level dynamic threshold band; Warning module: Based on the oil level prediction value of the TimesNet model, perform dynamic threshold judgment of the oil level and judgment of the oil level change trend. If the predicted data exceeds the dynamic threshold band, issue a numerical warning signal; calculate the trend of the oil level. If the overall trend shows an upward trend, issue a trend warning signal.

[0035] In the embodiment, the hydro-generator unit bearing oil level warning method based on the TimesNet model is adopted to achieve over-threshold warning and trend warning of the unit bearing oil level. The method of the present invention performs excellently in long-time series prediction tasks, and the long-time series is exactly the data essence of the hydro-generator unit bearing oil level; the present invention adopts dynamic threshold judgment and trend judgment for comprehensive judgment, and finally forms a warning, which is more reliable and timely than the oil level warning of the fixed threshold method.

Claims

1. A method for early warning of bearing oil level of a hydro-turbine generator set based on the TimesNet model, characterized in that: The following steps are involved: Step 1: Obtain monitoring data of the hydro-turbine generator set, including the active power, bearing shell temperature and bearing oil level of the set; Step 2: After converting the unit data obtained in step 1 into structured data, the unit data is screened in a steady state; Step 3: Based on the unit operating condition data, analyze the business characteristics of the unit, wherein the business characteristics include whether the oil level of the bearing of the hydro-turbine generator unit continues to increase; Step 4: Divide the unit data into training set and evaluation data set; Step 5: Based on the data characteristics of the unit bearing oil level and the business characteristics of the unit, the TimesNet model is used for fitting; Step 6: Based on the fitting result obtained in step 5, a reference value of the dynamic threshold band of the bearing oil level of the computer group is calculated; Step 6.1: Input the evaluation data set into the TimesNet model, and use the TimesNet model to calculate the predicted oil level of the unit bearing; Step 6.2: Calculate the residual between the predicted oil level value of the TimesNet model and the oil level value of the evaluation data set, and use the standard deviation of the residual as the reference value of the oil level dynamic threshold band; Step 7: Use the trained TimesNet model to predict the bearing oil level of the hydro-turbine generator set in the production environment. According to the oil level prediction value of the TimesNet model, the dynamic threshold value and oil level change trend of the oil level are judged.

2. The method for early warning of the bearing oil level of a hydro-turbine generator set based on the TimesNet model according to claim 1 is characterized in that: The step 2 specifically includes: Data conversion: convert time series data into structured data that is easy for users to analyze; Data supplement: supplement the data defects of the unit; Steady-state screening: Screen out data of units operating in steady state; Data frequency reduction: The timestamp alignment of unit data is achieved by data frequency reduction; Eliminate outliers: Eliminate abnormal data caused by external interference.

3. The method for early warning of the bearing oil level of a hydro-turbine generator set based on the TimesNet model according to claim 2 is characterized in that: The judgment method for steady-state screening in step 2 is: if the active power of the unit fluctuates by no more than 20MW within 30 minutes, the unit is judged to be in steady state.

4. The method for early warning of the bearing oil level of a hydro-turbine generator set based on the TimesNet model according to claim 3 is characterized in that: In step 2, the Laida criterion is used to determine the outliers in the unit data. If the difference between the single unit data and the mean of this unit data is greater than 3 times the standard deviation, it is determined to be an outlier.

5. The method for early warning of the bearing oil level of a hydro-turbine generator set based on the TimesNet model according to claim 4 is characterized in that: In step 4, the unit data is divided into a training set and an evaluation data set in a ratio of 7:3, that is, 70% of the unit data is used as a training set and 30% of the unit data is used as an evaluation data set.

6. The method for early warning of the bearing oil level of a hydro-generator set based on the TimesNet model according to claim 2, 3, 4 or 5, characterized in that: In step 5, the TimesNet model analyzes the timing changes of the unit from a multi-cycle perspective, specifically including: 1) Fold the one-dimensional time series based on multiple periods to obtain multiple two-dimensional tensors, where the columns and rows of each two-dimensional tensor reflect the time series changes within and between periods respectively; 2) Use the Inception module of the TimesNet model to extract two-dimensional time series change features from the two-dimensional tensor; 3) Convert the two-dimensional time series change characteristics into one-dimensional space to facilitate information aggregation.

7. The method for early warning of bearing oil level of a hydro-turbine generator set based on TimesNet model according to claim 6 is characterized in that: The Inception module includes the following parallel convolution calculation units: (1) Convolutional layer with a convolution kernel of 1×1; (2) Convolutional layers with a convolution kernel of 1×1 and a convolutional layer with a convolution kernel of 3×3; (3) Convolutional layers with a convolution kernel of 1×1 and a convolutional layer with a convolution kernel of 5×5; (4) The maximum pooling layer with a convolution kernel of 3×3 and the convolution layer with a convolution kernel of 1×1.

8. The method for early warning of the bearing oil level of a hydro-turbine generator set based on the TimesNet model according to claim 2, 3, 4, 5 or 7, characterized in that: The step 7 specifically includes: Step 7.1: Use the TimesNet model to predict the oil level of the unit bearing in real time, and determine whether the predicted oil level value is within the range of the oil level dynamic threshold band. If the predicted oil level value exceeds the range of the oil level dynamic threshold band, execute step 7.2; if the predicted oil level value is within the range of the oil level dynamic threshold band, repeat step 7.1; Step 7.2: Observe and determine the changing trend of the unit bearing oil level. If the unit bearing oil level shows an upward trend, issue a warning signal and execute step 7.1.

Images