Wind turbine gearbox temperature prediction based on improved whale optimized long short-term memory network

Aiming at the problem of low prediction accuracy of wind turbine gearbox temperature, the improved whale algorithm is used to optimize the long and short term memory network to predict the oil temperature of gearbox. The input data is based on the grey correlation degree to mine the parameters that are highly correlated with the oil temperature of the gear box, and the prediction model under healthy state is established. Finally, the experimental analysis shows that the method used in this paper has good accuracy in the oil temperature prediction of gearbox.


Introduction
Under the background of double carbon, wind power generation has developed rapidly due to its achieve pollution-free power generation [1] , and has become an important part of new energy power generation [2] .Gear box is an important part of wind turbine, the gearbox directly affects the reliability and operation safety of the system.There is a close relationship between the temperature change of the gearbox and the fault condition.High temperature will lead to the degradation of lubricating oil inside the gearbox, accelerate the wear of components, and cause serious faults [3] .Therefore, the real-time and accurate prediction of the temperature of the gearbox is of great significance to enhance the operational stability of the wind turbine.
Due to the complexity of gearbox, its feature data is difficult to extract [4] .Many scholars have realized the evaluation and early warning of the running state of the gearbox through machine learning algorithms [5] .Among them, the prediction based on deep learning method has received a lot of attention.Reference [6] monitored the state of wind turbine gearbox through LSTM.Reference [7]  utilized convolutional neural network to diagnose wind turbine faults; in Reference [8], CNN and LSTM were combined to predict wind turbines, and the results showed that the combined model had higher accuracy.
In summary, this paper uses the LSTM to predict the temperature, and uses the improved whale algorithm [10] to optimize its arguments to handle the issue of slow convergence speed.Because the gearbox oil temperature will rise when the gearbox fails, this paper selects the oil temperature by the way of the predictor.Finally, through the verification of the measured data of 2.5 megawatt wind turbine in Hubei, the prediction model of this paper has high accuracy.

improved whale algorithm
Whale algorithm is presented by Mirjalili and his colleagues.It is widely used because of its simplicity and few parameters.WOA includes three types of location updating: prey enclosure, hunting behavior, and random search.However, due to its slow convergence speed and poor global search ability, three improvement methods are added to improve its search optimization ability.
(1) adaptive weight The adaptive inertia weight ω is added to the position update of whale, and the constriction and convergence the algorithm running speed are improved by its mathematical characteristics.The weight formula and the improved algorithm location renewal equation are as below: max ( ) 0.2cos 1 2 The variable spiral position update part is aimed at the problem that the original calculation tends to fall into part optimality.The parameters of control spiral are changed from constants to variables to improve the global search ability and optimization accuracy.The enhanced spiral cell update formula is as below: max 5 cos( ( 1)) ( 1) ( ) *( ) cos( 2) (3) optimal neighborhood perturbation The optimal neighborhood perturbation strategy is aimed at the disadvantage of low search ability of the work.The best value is randomly searched near the current optimal position, and judged by the greedy strategy: if this position is better than the original position, it is retained, otherwise it is abandoned.The nearby perturbation math is as below: *( ) 0.5 1 *( ), 2 0.5 () *( ), 2 0.5

X t rand X t rand Xt
X t rand In the formula, X(t) is the generated new place, and rand1 and rand2 are random digit between [0,1].

LSTM algorithm
RNN has been widely used as a neural network for processing sequence data in recent years, but its model output is only related to the recent input.Therefore, the question of gradient explode is prone to occur when dealing with long sequences, resulting in the model cannot effectively learn long-distance dependencies.
As a variant of RNN, LSTM solves the problem of gradient explosion much better than RNN.The LSTM consists of three gates to control the cell state, which are called the forgetting door, the input door and the output door, respectively, and the cell state is deleted or added information through the gate structure.The gate structure is a combination of a Sigmoid layer and a dot multiplication operation, which selectively determines which information is passed.

Improved whale optimization LSTM
In this paper, account of the LSTM model, the oil temperature of gearbox is predicted.Aiming at difficulty of setting LSTM hyperparameters, the improved WOA is used to determine the global optimization, and an improved whale optimized LSTM memory network model is built.Figure 1 is a flow diagram of the model.The basic steps are as below: Step 1: The processed wind turbine characteristic data is zoned.
Step 2: Initialize the improved WOA, and use the improved whale algorithm to optimize the number of hidden layers, training times, regularization parameter L2 and learning rate in the LSTM neural network to obtain the best prediction model.
Step 3: Randomly generate the initial whale position, and set MSE of the prediction result as the fitness of the whale to find the optimal location.
Step 4: The improved WOA algorithm is used to update the whale position, and the self-adaptability is recalculated to compare whether the new position is the optimal position.
Step 5: Repeat step 4 until the global optimum is found, find the optimal parameter as output.
Step 6: Use the optimal LSTM network model for forecast.

Parameter selection based on grey correlation degree
The wind power data used in this paper comes from the actual operation data of Hubei Province for one year.Sampling data includes dozens of parameters of each component of the unit.The data collection interval was 10 min.For gearbox oil temperature, there are many factors affecting its change trend, but the degree of influence is indeed different.Therefore, this paper uses the grey correlation method to analyze the characteristics that affect the oil temperature of the gearbox, and selects 10 features with the highest correlation as the input of model.Table 1 shows each correlation degree coefficient.For the sake of verify prediction accuracy of the model, the article employ MAE, RMSE, MAPE, and R2 as model accuracy evaluation index to compare with the original LSTM network.Table 2 shows the results.From the above results, compared with the results of the improved WOA-LSTM and LSTM, the values of the evaluation indexes decreased by 0.633, 0.98 and 2.595, severally, and the R2 increased by 0.028.This indicates that the prediction accuracy can be improved by adding improved WOA to the LSTM network.

Conclusion
In this paper, an improved whale optimization LSTM forecast model is proposed.The improved whale optimization algorithm is used to automatically optimize LSTM hyperparameters, which solves problem that the original LSTM is difficult to adjust the parameters.By contrasting prediction consequence of improved model with the original LSTM model, which is verified that the improved model can improve the prediction accuracy.

3. 2
Prediction and evaluation of the model Select 4320 data within one month of feature data for analysis, and allocate them as training data and test data according to the ratio of 8: 2. The first diagram in Figure.2 is the comparison diagram between the predicted value and the true value of the LSTM model.The second diagram in Figure.2 is the prediction error value of the LSTM model.It can be seen from the second diagram that most of the LSTM error values are between ±2℃, and the prediction results are more accurate.However, in the face of rapid changes in oil temperature (320-410 sets of data), the LSTM model has a very obvious error due to the dependence problem, resulting in a decrease in prediction accuracy.

Figure 3
Figure3are the comparison chart of improved WOA-LSTM prediction results and the prediction residual chart, respectively.It can be seen that in the face of rapid changes in unit conditions, the improved model overcomes the problem of LSTM 's dependence on time sequence, which cause the overall predictive accuracy higher and error fluctuation smaller.

Table 1 .
Grey correlation coefficient of each parameter.