2.1

Data preprocessing

In this paper, the SCADA system operation data of a wind farm in Inner Mongolia is used, and the data mainly includes fault data and normal operation data, and the data characteristics are mainly running time, fan wind speed, active power and engine speed. Table 1 lists some of the important feature parameters used in this document

Table 1.

Characteristic data used in wind turbine fault prediction

There is a part of the data of 0 in the data, this kind of data is generally the data when the wind turbine is stopped and the manufacturer will set part of the data to 0 data, which often produces interference in modeling training, so it needs to be eliminated, and the principle of exclusion is mainly as follows: speed is 0, power is 0, The wind speed does not meet the data between the cut in and cut out wind speeds.

2.2

Correlation analysis of data

There is a strong correlation between SCADA data and data, such as state variables such as wind speed, power and generator rotor speed. This is a fixed parameter of learning for model prediction, which has little impact on the training of the model, the lower the correlation of the data in the small change often has a great impact on other data, we often need some relatively low data as an important data for model learning optimization, so the correlation of the data needs to be analyzed.

The Pearson correlation coefficient is a linear correlation coefficient that is mainly used for the analysis of relationships between data. For a set of variables, the correlation coefficient r is calculated as:

In this paper, the SCADA data used will be used for correlation experiments, and the resulting data correlation curve chart is shown in Figure 1.

Figure 1.

Histogram of SCADA data correlation

It can be seen from Figure 1 that the correlation between the front power, wind speed, speed and generator winding temperature is very strong, while the correlation between the bearing temperature of the generator drive end, the bearing temperature of the generator non-drive end, and the cabin temperature and the previous 4 sets of variables is relatively weak, and the subsequent training will have a greater impact on the model, so in the subsequent failure prediction time, we tend to use the latter 3 sets of faulty data to verify the experiment to ensure the accuracy of the experiment.

3.1

LSTM neural network

LSTM networkIt is a special recurrent neural network¹⁰. With the advantages of RNN neural network for sequence processing, it has many more gating switches than RNNs, which has a screening effect on information. It can solve a series of problems such as gradient damage and extremely large or minimal overfitting during training due to slow update of the weight relationship. The network has three control gate structures of output gate, input gate and forget gate, the work of the forget gate is to determine whether the memory unit at the previous moment enters the network for calculation, the input gate determines whether the candidate memory unit is used, and the output gate determines whether the hidden state is used. It is effectively controlled by the activation function. The internal structure of the LSTM network as a whole is shown in Figure 2.

Figure 2.

Internal structure diagram of LSTM network

The LSTM is calculated as follows:

Formula: f_t for the Forgotten Door; σ is the sigmoid activation function; x_t Enter for the vector at the current moment; w_xf is the weight between the input vector and the forget gate; h_t−1 is the hidden layer state of the previous moment;w_hf is the weight between the hidden layer and the forgetting gate;b_f for the bias of the Forgotten Gate;i_t is the input door;w_xi is the weight between the input vector and the input gate;w_hi is the weight size between the hidden layer and the input gate;b_i is the bias of the input gate;c_t’is a candidate memory unit;w_xc is the weight size between the input vector and the candidate memory cells; w_hc is the weight size between the hidden layer and the candidate memory cells;b_c is the bias of the candidate memory cells;o_t is the output gate;w_xo is the weight between the input vector and the output gate;w_ho is the weight size between the hidden layer and the output gate;b_o is the bias of the output gate;c_t is the memory unit;c_t−1 is the memory unit of the previous moment;h_t is the hidden state of the current moment.

3.2

seq2seq neural network

the sequence-to-sequence model was mainly used in natural language processing tasks such as machine translation and speech and text recognition¹¹, and later studies applied the model to time series forecasting tasks and achieved good prediction results. SEQ2SEQ can be divided into encoder and decoder as a whole, the basic unit uses the LSTM model¹², and the encoder and decoder expansion diagram is shown in Figure 3.

Figure 3.

seq2seq encoder and decoder unfolded

In the figure, the encoder input sequence length is t, the decoder output sequence length is t’, the encoder obtains the final hidden layer state ht as the input of the decoder, and finally the output of the decoder transforms the dimension through the fully connected layer to obtain the final output output.

The hidden layer state in the encoder at the current moment is calculated as follows:

The hidden layer state at the current moment in the decoder is calculated as follows:

Formula:h_tis the hidden layer state under the encoder at the current moment;x_tis the input vector at the current moment; h_t−1is the hidden layer state of the previous moment in the encoder; LSTM（）is the internal calculation function of the LSTM model, represents the implied layer state at the current moment in the decoder; f_t−1 is the output vector at the previous moment; is the implied layer state at the last moment in the decoder.

4.1

Data analysis of wind turbine operation

The wind turbine had caused the SCADA system to malfunction due to generator bearing problems, and then shut down the wind turbine for maintenance. In order to verify the effectiveness of the proposed algorithm, the data of 3 months of the corresponding time period in the year before the failure is used as the training set to train the model, and the data of 13 months, including the fault time point, after the period of the failure, is used as the test data of the test set to verify the time of failure.Since there are abnormal data such as downtime data and fault data, data visualization is first used to preprocess the data. The variables used in this prediction experiment are plotted sequentially from A to G using the features in Table 1, as shown in Figure 4.

Figure 4.

Timing diagram of wind turbine generator related characteristics

As can be seen from the figure, there is a lot of data with 0, which is recorded by the SCADA system when the wind turbine is in a shutdown state. In the F sub-diagram, it can be seen that the temperature of the non-drive bearing of the generator is in a stable state as a whole, but it fluctuates significantly in the red elliptical area, and the temperature rises significantly due to the failure of the non-drive bearing of the generator.4.2 Model parameter optimization

4.2

Model parameter optimization

The data of 3 months of the corresponding time period in the year before the above generator bearing failure was extracted, and 19464 sample data were obtained after processing for SEQ2SEQ network training. Set the batch size to 64, optimize the initial learning rate of 0.001, the sequence length seq_len is 12, and the number of training times is 50. By adjusting the number of neurons in the hidden_dim hidden layer for prediction, the number of neurons in the hidden_dim was 8, 16, 32, 64, 128 for comparative prediction experiments, and the MAE and RMSE were used for evaluation. The specific calculation method is as follows:

MAE calculation formula:

RMSE calculation formula:

where n is the number of data samples; y_i is the actual value; is the predicted value.

The specific experimental results are shown in Table 2.

Table 2.

Model prediction effect under different hidden_dim

With the increase of neurons, the model effect will be better and better, but when the number of neurons is not as good as the number of neurons at 128, this can indicate that the number of neurons is not as much as possible, the curve is concave curve, and the number of neurons used here is 32.

4.3

Alarm threshold determination

The above 3 months of training data were used for failure prediction, and then the residual between the predicted and actual values of the active power was calculated, as shown in Figure 5(a). The residual data are normally fitted, and the fitted normal distribution plot is shown in Figure 5(b).

Figure 5.

(a)Predict the residual between active power and actual power, (b)Distribution of residuals in a healthy state

The distribution of residuals is mainly concentrated between ±0.5, and the main part still tends to 0. Using the nature of the normal distribution to set a suitable threshold a, so that the interval [-a,a] contains more than 99.7% of the data, the data between the two green dashed lines can be determined as normal data, and the data outside the dashed line is partial abnormal data, From this we can determine that the alarm threshold is ±0.3584.

4.4

Verification of failure prediction methods

Using the above trained model to predict and verify the test data in the next 13 months, using the number of hidden layer neurons 32 as the subsequent experimental parameters, the distribution curve of the predicted value of the active power output of the test set model and the actual value is shown in Figure 6(a), in order to verify the accuracy of the model prediction, the difference is calculated here, and the residual distribution plot is shown in Figure 6(b).

Figure 6.

(a)Test set predicted power and actual power, (b)The test set predicts the residual power from the actual power

As can be seen from the figure, the predicted value output by the model basically matches the actual value. Only some of the data have obvious deviations. Through the determination of the alarm threshold, the red dotted line is determined as the alarm line, and only part of the data in front of more than 90,000 data points intermittently exceeds the alarm line, as shown in the black elliptical area in the figure, but the residual in the red rectangular area has exceeded the alarm threshold of the red dotted line a lot, at this time it has been possible to judge that the wind turbine is in an abnormal state until the subsequent complete exceeding of the alarm threshold. The point where the threshold is exceeded three times in a row for the first time is at 23476, and it should be discarded because it does not meet the reality, and the point where the threshold is exceeded for the second three consecutive times is at 88890 points, at which point it can be determined that the wind turbine has been abnormal, and then the residual exceeds the threshold over time, until the subsequent residuals continue to exceed the threshold. According to the correspondence between data points and time and combined with the alarm records provided by the SCADA system, the wind turbine failure prediction through the seq2seq neural network can know that the wind turbine is abnormal about 6 days in advance, which can replace or repair the relevant components early to avoid unnecessary losses.