基于lstm的风力发电机健康监测

Predicting Remaining Useful Life of Wind Turbines Using LSTM

Abstract: Wind power has become one of the most promising and rapidly developing renewable energy sources. However, the reliability and stability of wind turbines are critical factors for ensuring the stable operation of wind power systems. Accurately predicting the remaining useful life (RUL) of wind turbines is of great significance for reducing maintenance costs and ensuring safe and efficient operation. In this study, we propose an LSTM-based RUL prediction method for wind turbines. The proposed method is trained on historical sensor data collected from wind turbines and is capable of accurately predicting the RUL of wind turbines. Experimental results on a real-world wind turbine dataset demonstrate the effectiveness and superiority of the proposed LSTM-based RUL prediction method compared to other state-of-the-art RUL prediction methods. The proposed method has great potential for practical applications in wind power systems, and can significantly improve the reliability and stability of wind power systems.

Keywords: LSTM, wind turbine, remaining useful life prediction.

0    Introduction

Renewable energy has become an important part of the global energy mix due to its low carbon emissions and abundant availability. Wind energy, in particular, has experienced rapid growth in recent years. However, the reliability and maintenance of wind turbines remain major challenges in the wind energy industry. Accurate prediction of the remaining useful life (RUL) of wind turbines is essential to optimize maintenance schedules and minimize downtime. In this paper, we propose an LSTM-based approach for RUL prediction of wind turbines. The proposed method can effectively capture the complex temporal dependencies in sensor data and provide accurate RUL estimates. The experimental results demonstrate the effectiveness and superiority of the proposed method compared to existing approaches.

1    Description of the Dataset

For most complex systems like wind turbines , finding a suitable model that allows the injection of health related changes certainly is a challenge in itself. In addition, the question of how the damage propagation should be modeled within a model needed to be addressed. Secondary issues revolved around how this propagation would be manifested in sensor signatures such that users could build meaningful prognostic solutions. In this paper we first define the prognostics problem to set the context. Then the following sections introduce the simulation model chosen, along with a brief review of health parameter modeling. This is followed by a description of the damage propagation modeling, a description of data, and a discussion on performance evaluation.

1.1  Determination of input factors

The dataset description for the turbofan engine degradation simulation dataset is that each time series represents an individual engine. The initial wear and manufacturing variation at the start of each engine's operation are unknown. The engine operates normally at the start of each time series and then develops a fault at some point within the sequence. The fault size increases progressively in the training set until a system failure occurs. Each column represents a different variable. These columns correspond to the following data: : engine ID; : cycle time; : operational settings; : sensor measurements.

1.2  Determination of output factors

To avoid confusion, we define prognostics here as the estimation of remaining useful component life. The remaining useful life (RUL) estimates are in units of time (e.g. hours or cycles). End-of-life can be subjectively determined as a function of operational thresholds that can be measured. These thresholds depend on specifications to determine safe operational limits

2    Data preprocessing

2.1  Fill in missing values

During the training process of deep learning models, sorting the training data by sequence length can reduce the amount of padding to be used. Then, selecting a batch size that can evenly divide the training data and reducing the amount of padding in each batch can effectively improve the utilization efficiency of training data and reduce the performance loss caused by padding, as shown in Figure 1. This method significantly reduces the performance loss caused by padding.

2.2  Delete constant features

Translation: Features that remain unchanged at all time steps may have a negative impact on the training. To mitigate this, we identify rows in the dataset where the minimum and maximum values are the same, and remove them. Removing these outliers during residual value evaluation can help eliminate noise and abnormal fluctuations in engineering data, resulting in a more accurate
reflection of the actual conditions of the engineering project.

Figure 1     Data set after data preprocessing


2.3  Data normalization

This experiment uses min-max normalization to linearly transform the raw data x and constrain the data size to [0,1], i.e.,

                                                                                           

  

where x* is the normalized value, xmax is the maximum value in the sample data, and xmin is the minimum value in the sample data.

2.4  Evaluation metrics

RMSE is a commonly used metric for evaluating the fitting performance of neural networks [5]. It is the square root of the mean of the sum of squares of differences between predicted values and actual values. By computing the differences between predicted and actual values, RMSE can measure the fitting degree of the model. Its formula is given by:

                                                          

In this formula, Forecast represents the predicted values of the neural network, Observed represents the true observed values, and n represents the sample size. A smaller RMSE value indicates a better fit of the neural network to the true data [6,7]. Usually, the value of RMSE is related to the range of the actual observed values.

2.5  Clipping response

Figure 2     Predicted and raw data


To learn more from sequence data of engines when they are close to failure, the response is clipped at a threshold of 150. This causes the network to treat instances with higher RUL values as equivalent. As shown in the figure, it displays the observed values and clipped response.。

3   Experimental Results

Figure 3 presents four sets of wind turbine life prediction and observation data. Among them, three sets of data show good fitting results, indicating the superiority of the Long Short-Term Memory (LSTM) algorithm in predicting non-linear relationships and its applicability in wind turbine structural health assessment. From Figure 4, it can be observed that the Root Mean Square Error (RMSE) values are within a reasonable range, with the maximum probability distribution near 0. However, LSTM has limitations. There is one set of data that shows poor fitting results, particularly when the wind turbine remains in a healthy state for an extended period, LSTM fitting tends to lean towards sub-healthy status. This may be due to the subjective factors in selecting hyperparameters of LSTM, which play a critical role in the prediction of neural networks.


 By comparing these two figures in figures4 we can conclude the significance of training data and the limitation of LSTM to fit data in certain circumstances. Future studies can explore using optimization algorithms to optimize LSTM hyperparameters, such as algebra, learning rate, number of hidden layers, etc., to enhance the LSTM model's fitting degree to observation values.

Figure 3     Prediction Error Comparison Chart


4    Conclusion

This paper proposes a method for predicting engineering value time series based on the LSTM recurrent neural network, which includes training, prediction, and data preprocessing of the LSTM model. Experimental results show that: 1) the LSTM model can fit and predict engineering value well. 2) The loss change and model accuracy during the training process of the LSTM model are sensitive to the learning rate. Too low or too high learning rates may lead to underfitting or overfitting problems, affecting the predictive performance of the model.。

 

Figure 4     Prediction Error Comparison Chart

5    References

[1] Altunok E, Taha MMR, Epp DS, Mayes RL, Baca TJ. Damage pattern recognition for structural health monitoring using fuzzy similarity prescription. Computer-Aided Civil and Infrastructure Engineering. 2006;21:549-60.

[2] Ananda Rao M, Srinivas J, Murthy BSN. Damage detection in vibrating bodies using genetic algorithms. Computers and Structures. 2004;82:963-8.

[3] Aoki S, Fujino Y, Abe M. Intelligent bridge maintenance system using MEMS and network technology.  Proceedings of SPIE - The International Society for Optical Engineering2003. p. 37-42.

[4] Au FTK, Cheng YS, Tham LG, Bai ZZ. Structural damage detection based on a micro-genetic algorithm using incomplete and noisy modal test data. Journal of Sound and Vibration. 2003;259:1081-94.

  • 0
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值