Wind turbine condition monitoring method based on multi-source heterogeneous SCADA data

Abstract

The present invention provides a wind turbine status monitoring method based on multi-source heterogeneous SCADA data, comprising the following steps: calculating the probability distribution of monitoring quantities of all wind turbines in a wind farm, screening out a wind turbine that can represent the entire wind farm; cleaning the monitoring data; performing feature dimensionality reduction, screening out several sensitive characteristic parameters with a high correlation with the monitoring quantity; normalizing the screened sensitive characteristic parameters, then establishing a sample set, and using the sample set to train a long short-term memory neural network model; calculating the root mean square error between the actual value of the monitoring quantity at the current moment and the predicted value, and constructing a health monitoring status indicator for the wind turbine; and performing sliding averaging by designing a sliding window; and when monitoring the screened wind turbine, issuing an alarm when its health monitoring status indicator exceeds a warning threshold. The present invention achieves wind turbine status monitoring and fault warning.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing, and in particular to a wind turbine generator set status monitoring method based on multi-source heterogeneous SCADA data. Background Art

[0002] In recent years, countries around the world have increasingly focused on issues such as energy security, the environment, and climate change. Accelerating the development of the wind power industry has become a universal consensus and a concerted action for the international community to promote energy transition and address global climate change. Wind turbines are often exposed to extreme weather conditions, including scorching sun, dust storms, rain and snow, thunderstorms, high temperatures, and severe cold. This leads to frequent failures and declining performance, resulting in significant economic losses. To meet national strategic needs, the development of intelligent condition monitoring technology for wind turbines is urgent.

[0003] However, existing wind turbine vibration data requires high sampling frequencies, long acquisition times, and numerous measurement points, resulting in large amounts of vibration monitoring data. This creates challenges for remote wireless data transmission and storage, and makes it difficult to generalize monitoring models for a single turbine to other turbines. Leveraging the small size and ease of storage and transmission of multi-source heterogeneous SCADA data, it is necessary to develop a wind turbine condition monitoring method based on this data. Summary of the Invention

[0004] To address the above issues, the present invention provides a wind turbine status monitoring method based on multi-source heterogeneous SCADA data, constructs a long short-term memory neural network prediction model, fully exploits the time series information of characteristic parameters and the coupling information between characteristic parameters, and realizes wind turbine status monitoring and fault warning. To achieve the above technical objectives, the technical solutions adopted in the embodiments of the present invention are:

[0005] An embodiment of the present invention provides a method for monitoring the status of a wind turbine generator system based on multi-source heterogeneous SCADA data, comprising the following steps:

[0006] Step S10, obtaining monitoring data of the wind turbine as a characteristic parameter, selecting a characteristic parameter that can characterize the operating state of the wind turbine as a monitoring quantity; calculating the probability distribution of the monitoring quantity for all wind turbines in the wind farm, and selecting a wind turbine that can characterize the entire wind farm based on the similarity of the probability distribution of the monitoring quantity;

[0007] Step S20, cleaning the monitoring data to remove invalid data from the monitoring data of the wind turbine generator set;

[0008] Step S30, performing feature dimensionality reduction on all characteristic parameters of the wind turbine generator system based on the Spearman correlation coefficient, and screening out several sensitive characteristic parameters with a high correlation with the monitored quantity;

[0009] Step S40, normalizing the screened sensitive feature parameters, then establishing a sample set, and using the sample set to train a long short-term memory neural network model; the sample set includes a training set and a test set;

[0010] Step S50: For a wind turbine that can represent the entire wind farm, the data in the test set is input into the trained long short-term memory neural network model to obtain the predicted value of the monitoring quantity at the current moment; the root mean square error between the actual value of the monitoring quantity at the current moment and the predicted value is calculated to construct the health monitoring status index HI(t d );and perform sliding average by designing sliding window;

[0011] Step S60: setting an early warning threshold. When monitoring the selected wind turbines, when the health monitoring status indicator HI (t d ) When the warning threshold is exceeded, an alarm is issued.

[0012] Furthermore, in step S20, the conditions for monitoring data cleaning are as follows:

[0013] ① Average wind speed ≤ cut-in wind speed; or,

[0014] ② Average wind speed ≥ cut-out wind speed; or,

[0015] ③Average spindle speed ≤10rpm.

[0016] Furthermore, the monitored quantity is the average gearbox oil temperature;

[0017] The selected sensitive characteristic parameters include: average generator speed, average cabin temperature, average gearbox low-speed end bearing temperature, average gearbox oil filter inlet pressure, average gearbox oil distributor outlet pressure, average gearbox oil temperature at the previous moment, and average gearbox oil temperature at the previous two moments.

[0018] Furthermore, the hyperparameters of the long short-term memory neural network model include:

[0019] Number of input features, number of neurons per layer, time steps, number of output features, learning rate, number of training times.

[0020] Furthermore, the health monitoring status indicator HI(t d ) is as follows:

[0021]

[0022] Among them, t d is the current moment, r real Indicates the actual value of the average gearbox oil temperature; r predict represents the predicted value of the average gearbox oil temperature; L represents the sliding window length.

[0023] Furthermore, the setting of the warning threshold specifically includes:

[0024] Calculate the health monitoring status index HI(t d ), and the Gaussian kernel density function value when the confidence probability is 99.95% is used as the warning threshold.

[0025] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0026] 1) The automation of wind turbine status monitoring is realized.

[0027] 2) The adaptability is relatively good. The long short-term memory neural network model can be applied to almost all wind turbines in a wind farm.

[0028] 3) The health monitoring status indicators are accurate and can realize real-time monitoring of the health status of wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Flowchart of the monitoring method in an embodiment of the present invention.

[0030] Figure 2 Schematic diagram of changes in health monitoring status indicators of a normal wind turbine generator system in an embodiment of the present invention.

[0031] Figure 3 Schematic diagram of changes in health monitoring status indicators of a faulty wind turbine in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] like Figure 1 As shown, a method for monitoring the status of a wind turbine generator system based on multi-source heterogeneous SCADA data is proposed in an embodiment of the present invention, comprising the following steps:

[0034] Step S10, obtaining monitoring data of the wind turbine as a characteristic parameter, selecting a characteristic parameter that can characterize the operating state of the wind turbine as a monitoring quantity; calculating the probability distribution of the monitoring quantity for all wind turbines in the wind farm, and selecting a wind turbine that can characterize the entire wind farm based on the similarity of the probability distribution of the monitoring quantity;

[0035] In this embodiment, the monitored quantity is the average gearbox oil temperature;

[0036] Step S20, cleaning the monitoring data to remove invalid data from the monitoring data of the wind turbine generator set;

[0037] In this embodiment, the conditions for monitoring data cleaning are as follows:

[0038] ① Average wind speed ≤ 3m / s (cut-in wind speed); or,

[0039] ② Average wind speed ≥ 25m / s (cut-out wind speed); or,

[0040] ③Average spindle speed ≤10rpm;

[0041] Step S30, performing feature dimensionality reduction on all characteristic parameters of the wind turbine generator system based on the Spearman correlation coefficient, and screening out several sensitive characteristic parameters with a high correlation with the monitored quantity;

[0042] In this embodiment, the selected sensitive characteristic parameters include: average generator speed, average cabin temperature, average gearbox low-speed end bearing temperature, average gearbox oil filter inlet pressure, average gearbox oil distributor outlet pressure, average gearbox oil temperature at the previous moment, and average gearbox oil temperature at the previous two moments; a total of seven characteristic parameters;

[0043] It should be noted that the monitoring data of the wind turbine is sampled at a certain time interval, so the monitoring data corresponds to the current moment, the previous moment, the previous two moments...;

[0044] Step S40, normalizing the screened sensitive feature parameters, then establishing a sample set, and using the sample set to train a long short-term memory neural network model;

[0045] The numerical ranges of the seven sensitive characteristic parameters selected above are as follows:

[0046] Maximum Minimum unit Average generator speed 2000 1000 rpm Average cabin temperature 30 -30 ℃ Average gearbox low speed end bearing temperature 100 0 ℃ Average gearbox oil filter inlet pressure 10 0 bar Average gearbox oil distributor outlet pressure 10 0 bar Average gearbox oil temperature 1 100 0 ℃ Average gearbox oil temperature 2 100 0 ℃

[0047] The normalization formula of sensitive feature parameters is as follows:

[0048]

[0049] Among them, X is the sensitive characteristic parameter value, x min 、x max are the minimum and maximum values ​​of the sensitive characteristic parameters respectively;

[0050] 80% of the sample set is used as the training set and 20% is used as the test set;

[0051] The hyperparameters of the long short-term memory neural network model include:

[0052]

[0053]

[0054] The structure of the long short-term memory neural network model: input layer + 3 hidden layers + output layer;

[0055] The number of input features is 7, namely, average generator speed, average cabin temperature, average gearbox low-speed end bearing temperature, average gearbox oil filter inlet pressure, average gearbox oil distributor outlet pressure, average gearbox oil temperature at the previous moment, and average gearbox oil temperature at the previous two moments.

[0056] The number of neurons in each layer is 128;

[0057] The time step is 1, which indicates the memory duration of the LSTM neural network model. The longer the time step, the higher the risk of gradient vanishing or exploding.

[0058] The output feature quantity is the average gearbox oil temperature at the current moment;

[0059] Step S50: For a wind turbine that can represent the entire wind farm, the data in the test set is input into the trained long short-term memory neural network model to obtain the predicted value of the monitoring quantity at the current moment; the root mean square error between the actual value of the monitoring quantity at the current moment and the predicted value is calculated to construct the health monitoring status index HI(t d ); and by designing a sliding window for sliding averaging; this can reduce indicator glitches, improve trending, and achieve real-time monitoring of the health status of wind turbines;

[0060] In this embodiment, the monitored quantity is the average gearbox oil temperature;

[0061] The health monitoring status index HI(t d ) is as follows:

[0062]

[0063] Among them, t d is the current moment, r real Indicates the actual value of the average gearbox oil temperature; r predict represents the predicted value of the average gearbox oil temperature; L represents the sliding window length;

[0064] Step S60: setting an early warning threshold. When monitoring the selected wind turbines, when the health monitoring status indicator HI (t d ) When the warning threshold is exceeded, an alarm is issued;

[0065] The setting of the warning threshold specifically includes:

[0066] Calculate the health monitoring status index HI(t d), the Gaussian kernel density function value when the confidence probability is 99.95% is used as the warning threshold. In this embodiment, the warning threshold is 0.025;

[0067] In the following specific embodiment, the method proposed in this application is verified using SCADA data of a wind turbine in a wind farm; the rated power of the wind turbine in the wind farm is 2MW, the SCADA data sampling interval is 5 minutes, and the time span of the selected data is from June 1, 2020 to July 27, 2021; each SCADA data includes 100 characteristic parameters;

[0068] Calculate the probability distribution of the average gearbox oil temperature of all wind turbines and calculate the similarity of the probability distributions to select a wind turbine that can represent the entire wind farm. The probability distribution of the average gearbox oil temperature of this wind turbine is highly similar to that of other wind turbines. In this test analysis, wind turbine No. 004 was selected.

[0069] According to the probability distribution diagram of the health monitoring status indicators of wind turbine No. 004, the warning threshold is determined to be 0.025 when the confidence level is 99.95%. Figure 2 It can be seen that the health monitoring status indicators of the wind turbine have not exceeded the warning threshold of 0.025, and are generally stable without a significant decline trend, indicating that wind turbine No. 004 is normal, which is consistent with the actual results;

[0070] For a faulty wind turbine (wind turbine No. 002), if Figure 3 As shown in the figure, it can be seen that the health monitoring status indicators show an obvious degradation trend from a certain day, and the health monitoring status indicators exceed the warning threshold of 0.025, indicating that wind turbine No. 002 is faulty, which is consistent with the actual results.

[0071] For other wind turbines in the wind farm, it is only necessary to fine-tune the hyperparameters of the trained LSTM neural network prediction model using the real-time monitoring data of other wind turbines, and the LSTM neural network prediction model will be applicable to other wind turbines. The status monitoring of other wind turbines is the same as above.

[0072] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A wind turbine condition monitoring method based on multi-source heterogeneous SCADA data, characterized in that: The following steps are involved: Step S10, obtaining monitoring data of the wind turbine as a characteristic parameter, selecting a characteristic parameter that can characterize the operating state of the wind turbine as a monitoring quantity; calculating the probability distribution of the monitoring quantity for all wind turbines in the wind farm, and selecting a wind turbine that can characterize the entire wind farm based on the similarity of the probability distribution of the monitoring quantity; Step S20, cleaning the monitoring data to remove invalid data from the monitoring data of the wind turbine generator set; Step S30, performing feature dimensionality reduction on all characteristic parameters of the wind turbine generator system based on the Spearman correlation coefficient, and screening out several sensitive characteristic parameters with a high correlation with the monitored quantity; Step S40, normalizing the screened sensitive feature parameters, then establishing a sample set, and using the sample set to train a long short-term memory neural network model; the sample set includes a training set and a test set; Step S50: For a wind turbine that can represent the entire wind farm, the data in the test set is input into the trained long short-term memory neural network model to obtain the predicted value of the monitoring quantity at the current moment; the root mean square error between the actual value of the monitoring quantity at the current moment and the predicted value is calculated to construct the health monitoring status index HI(t d );and perform sliding average by designing sliding window; Step S60: setting an early warning threshold. When monitoring the selected wind turbines, when the health monitoring status indicator HI (t d ) When the warning threshold is exceeded, an alarm is issued.

2. The wind turbine status monitoring method based on multi-source heterogeneous SCADA data according to claim 1, characterized in that: In step S20, the conditions for monitoring data cleaning are as follows: ① Average wind speed ≤ cut-in wind speed; or, ② Average wind speed ≥ cut-out wind speed; or, ③Average spindle speed ≤10rpm.

3. The wind turbine status monitoring method based on multi-source heterogeneous SCADA data according to claim 1, characterized in that: The monitored quantity is the average gearbox oil temperature; The selected sensitive characteristic parameters include: average generator speed, average cabin temperature, average gearbox low-speed end bearing temperature, average gearbox oil filter inlet pressure, average gearbox oil distributor outlet pressure, average gearbox oil temperature at the previous moment, and average gearbox oil temperature at the previous two moments.

4. The wind turbine status monitoring method based on multi-source heterogeneous SCADA data according to claim 3, characterized in that: The hyperparameters of the long short-term memory neural network model include: Number of input features, number of neurons per layer, time steps, number of output features, learning rate, number of training times.

5. The wind turbine status monitoring method based on multi-source heterogeneous SCADA data according to claim 3, characterized in that: The health monitoring status index HI(t d ) is as follows: Among them, t d is the current moment, r real Indicates the actual value of the average gearbox oil temperature; r predict represents the predicted value of the average gearbox oil temperature; L represents the sliding window length.

6. The method for monitoring the status of a wind turbine generator system based on multi-source heterogeneous SCADA data according to any one of claims 1 to 5, characterized in that: The setting of the warning threshold specifically includes: Calculate the health monitoring status index HI(t d ), and the Gaussian kernel density function value when the confidence probability is 99.95% is used as the warning threshold.

Images