State Reliability of Wind Turbines Based on XGBoost–LSTM and Their Application in Northeast China

Abstract

The use of renewable energy sources, such as wind power, has received more attention in China, and wind turbine system reliability has become more important. Based on existing research, this study proposes a state reliability prediction model for wind turbine systems based on XGBoost–LSTM. By considering the dynamic variability of the weight fused by the algorithm, under the irregular fluctuation of the same parameter with time in nonlinear systems, it reduces the algorithm defects in the prediction process. The improved algorithm is validated by arithmetic examples, and the results show that the root mean square error value (hereinafter abbreviated as RMSE) and the mean absolute error value (hereinafter abbreviated as MAPE) of the improved XGBoost–LSTM algorithm are decreased compared with those for the LSTM and XGBoost algorithms, among which the RMSE is reduced by 8.26% and 4.15% and the MAPE is reduced by 24.56% and 27.99%, respectively; its goodness-of-fit R2 value is closer to 1. This indicates that the algorithm proposed in this paper reduces the existing defects present in some current algorithms, and the prediction accuracy is effectively improved, which is of great value in improving the reliability of the system.

1. Introduction

China is a major global carbon emitter and is at an important stage of rapid industrialization and urbanization, with a high energy demand. Considering the wastage of resources caused by the shortage of fossil energy and environmental pollution [1], the government actively seeks solutions; encourages researchers to conduct fundamental research on system operation reliability throughout the life cycle of important equipment and components in key areas, such as new energy, energy conservation, emission reduction, and environmental protection; and accelerates the construction and development of renewable energy power generation. Wind power is a significantly advanced technology in renewable energy generation. China has abundant wind energy resources, especially on the southeast coast, in the Liaotung Peninsula, and in Northeast China. Compared with fossil fuels, the use of clean energy, such as wind power, can effectively reduce carbon dioxide emissions and mitigate global climate change trends. China started to use wind energy in the 1970s when the country began to develop and introduce foreign wind turbines, build wind power demonstration sites, and conduct research on wind turbine systems. Since then, the use of wind energy has increased rapidly. According to the National Energy Board’s 2023 data, by the end of 2022, the country’s cumulative installed wind power capacity was about 370 million kilowatts, which represents an increase of 11.2% year over year. According to the Global Energy Internet Development Cooperation (GEDC), the development trend of installed wind turbine capacity from 2025 to 2060 has been forecasted bearing in mind China’s strategic goals of carbon peaking and carbon neutrality, and is shown in Figure 1. The distinct points where the curve folds indicate the year of installation versus the installed capacity for four datasets (2025, 5; 2030, 8; 2050, 22; 2060, 25).

The above graph shows that the development trend of Chinese wind turbines indicates rapid progress, and the reliability of the operating states of wind turbine systems is crucial. However, the state reliability of a wind turbine system is predicted with poor accuracy, resulting in a generally low wind energy utilization efficiency and a large space for improvement in power generation efficiency due to factors such as the uncertainty of wind speed, the inability to store large amounts of electricity, the environment, and many other factors [2]. According to a study [3], nearly 80% of power plants in Asia have lost more than 30% of their wind energy potential since 1979. Especially in recent years, intelligent technological advances have driven the development of wind turbine system design toward complexity, precision, and comprehensive integration. However, the monitoring technology has not developed simultaneously, making wind turbine system failures frequent in various forms, and system status abnormalities are not accurately judged. In one case of a wind farm, a wind turbine was in a normal state after a day of operation, and then it appeared to exhibit alarm warnings such as “Paddle 1 fast paddle too slow”, “vibration band value is high”, and other warnings, and was shut down; then, it was repaired and started again, and the wind turbine underwent alarm-free operation, but it collapsed after 10 h. An investigation was conducted; it was found that the collapse of the wind turbine was caused by cracks in the cylinder wall due to broken bolts, and the monitoring system did not provide any advanced predictions of the prospective system status problems that might arise. In this case, the internal manifestation of the phenomenon is associated with the reliability problem of the current state of the system, but the reliability of the system’s state is dynamically changing, and its changing characteristics need to be described using the state information characteristic parameter.

Therefore, this study proposes that an improved wind turbine system state reliability prediction and assessment model should be constructed to predict the degree of reliability of the state that the system will exhibit in a specific future period, through an in-depth assessment of state parameter information features and a value mining analysis performed to solve the problem of the accuracy of the wind turbine system’s state reliability prediction. This proposed model was validated on a wind farm in Northeastern China to promptly formulate operation and maintenance strategies, reduce the incidence of system state abnormality, and improve the efficiency of wind energy utilization.

2. Materials and Methods

2.1. Previous Related Work

Wind energy is an important component of clean energy. In recent years, the world has rapidly increased the utilization of clean energy, such as wind energy. Wind power generation typically exhibits unpredictable characteristics, and its degree of reliability is dynamic, while the stability of a wind system directly affects the stable operation of the entire power system. System stability requirements are especially high for large electricity-consuming countries. Thermal power generation has better stability, but it is more environmentally polluting. Researchers [4,5,6,7,8] have made certain progress by conducting in-depth studies on the reliability of a wind turbine system’s operation status from the perspectives of wind speed, power generation, and weather.

Existing studies [9,10,11,12] have shown that the climate has obvious geographical characteristics at different geographical scales, such as the south and north of China, and different environments have diverse effects on the operational statuses of wind turbines. According to the existing research [13,14,15,16], the randomness and uncertainty of wind volume have a large impact on the volatility of wind power generation, and wind speed transmission is different at different heights such as 20 m, 50 m, 80 m, and 100 m. Researchers have proposed algorithms that consider the effect of wind speed on the operating states of wind turbines to solve the problem of the accuracy of wind speed prediction. Some researchers [17] have argued that the uncertainty of wind speed has a great impact on the reliability of wind turbines and have verified the inapplicability of normal distribution and the need for spatial and temporal distribution studies with arithmetic examples, while previous studies have ignored the effects of temporal and spatial distributions on wind speed. Other researchers [18] have proposed wind speed stochasticity to account for the theoretical results of Weibull and Bimbaum–Saunders’ wind speed distributions and found them to be suitable for real-time data collected at two different locations. Related studies have verified the necessity of a spatial and temporal distributivity analysis, which has laid the foundation for wind turbine reliability studies. In addition, some researchers [19] have proposed considering the difference between wind speed intermittency and wind power intermittency for a better evaluation of wind power generation, which has an auxiliary effect on system reliability. Existing algorithms have achieved reliable results when considering the effect of wind speed on the operating state of a wind turbine.

In addition, in the studies considering weather factor conditions, there is a significant impact of the icing of wind turbine blades on the operation of wind turbines in winter, especially in high-altitude humid climates. Researchers [20] have proposed indicators such as rotor speed, output power, and ambient temperature to monitor the icing of fan blades, by using the XGBoost algorithm to construct a blade behavior model, and have introduced the sequential probability ratio test to analyze the prediction residuals, with the proposed algorithm able to generate an icing warning 5 h in advance, thereby exhibiting a certain early warning effect. Under the influence of the weather, the XGBoost short-term wind power prediction model is considered to be a more suitable one for weather changes [21], and this has been analyzed by researchers by comparing and contrasting with an inverse neural network, classification and regression tree, random forest, support vector regression, and a single XGBoost model. The result showed that the prediction accuracy proposed in this article has the highest accuracy compared to that of several other models. In addition, Bayesian hyperparameter optimization-based algorithms have also been proposed to be applied in forecasting [22], and a comparative analysis of XGBoost, SVM, and other algorithms was carried out. The results showed that the hyperparametric optimization improved the XGBoost algorithm such that it significantly improved its prediction capabilities in challenging environments, such as extreme weather and low wind speed. Some researchers have proposed the hyperparametrically optimized LSTM algorithm [23], which is also effective for making similar predictions. Related studies have shown that the existing algorithms XGBoost, LSTM, etc., can initially solve the problem of system reliability by considering weather factors.

Based on the preliminary solution proposed for the impact of wind speed and weather factors on the reliability of wind turbine systems, researchers [24,25,26,27,28,29] believe that short-term generation power prediction is an important issue in power grid integration analysis, and propose LSTM algorithm, logistic regression analysis, CNN algorithm, etc., combined with a spatial–temporal correlation, to study it, which illustrates that it produces a good effect on short-term wind power prediction. Meanwhile, some studies [30,31] have illustrated that reliability-based power allocation methods for wind farms should focus on the problem of wind turbine failure probability, and that the accurate characterization of uncertainty in wind power models has an impact on the system steady-state operation, and these studies have great practical value. In terms of nonparametric models [32,33], algorithms such as GP, RF, and SVM can be used for power analysis modeling, and there are differences in the power analysis curves under different algorithms in the case of system anomalies or failures. It is necessary to choose the matching algorithm according to the actual situation.

Existing studies have concluded that system failures affect system reliability and that there is a correlation and complexity relationship between the multi-failure characteristics of the system state [34,35,36,37]. The accuracy of prediction directly affects the reliability and efficiency of the wind turbine system operating state. For example, some researchers [38] have proposed a multivariate correlation learning network based on the attention mechanism for extracting the features of wind turbine state monitoring data. Furthermore, the complex relationship between the features and the outputs was initially obtained with the vibration signal as a parameter. Other researchers [39] constructed a wind farm fault prediction model by using a Gaussian process meta-model to determine the key factors affecting the performance of the system and proved the accuracy of the constructed model by comparing it with the actual observation results. Therefore, the factors affecting the key performance characteristic indexes of the system can be deciphered in subsequent studies. In addition, the application of the XGBoost algorithm under different weather conditions and deep learning to predict the abnormal state of wind turbines [40,41,42] has been studied, and they can be efficiently utilized for prediction; however, there is still a need for further improvements in prediction accuracy.

From the existing studies, it can be seen that there is variability in the operation of wind turbines in different environments, but regardless of wind power intermittency, spatial–temporal correlation, uncertainty, and stochasticity, the state reliability under data feature characterization can be determined based on the feature extraction of data information. Relevant studies by researchers have also shown that state reliability under feature characterization is gradually receiving attention. Therefore, in this study, based on the existing research, the weaknesses of the algorithms such as XGBoost and LSTM are improved and studied in the hope of improving the prediction accuracy.

2.2. Methods

In this study, the algorithm research is divided into two phases, i.e., the research model and process design phase and the improved prediction model construction phase. In the first stage, a research model that combines qualitative prediction and quantitative prediction is proposed, and the prediction of system state reliability is divided into two steps: state identification and state prediction. In the second stage, the improved XGBoost–LSTM system state reliability prediction evaluation model is constructed according to the process, and the model is proposed in the first stage.

2.2.1. Research Model and Process Design

Under the continuous intelligent development of the system, the state information appears to have a large volume and is multi-modal and dynamic, which makes the system reliability assessment based on state analysis imprecise. This makes researchers question whether the original quantitative method of state analysis is sufficient to support the current demand for multi-modal and large-volume system reliability assessments. The diverse alarm environment of wind turbine systems has an especially important impact on the system state reliability analysis. Therefore, for the problem of the poor prediction of wind turbine system state abnormality, a research model combining qualitative prediction and quantitative prediction is proposed. Qualitative prediction is mainly based on human subjective experience and ability, and common methods include the brainstorming method, Delphi method, analogical prediction method, etc. Quantitative prediction mainly uses historical data to reveal future change patterns, which requires higher data feature extraction but has a low prediction capability for unexpected events. This study adopted a model primarily based on quantitative prediction and secondarily on qualitative prediction for predicting the system state reliability, which overcomes the interference of subjective factors and simultaneously improves the flexibility and accuracy of prediction. The block diagram of the research model is shown in Figure 2.

The prediction of the reliability of the system state can be divided into two steps: state identification and prediction.

The first step identifies historical states and laws and analyses state reliability criteria through historical data cleaning, standardized processing, analysis, and training.

The second step extracts the key features of the operation state under the consideration of information uncertainty and conflict. Then, it constructs the system state reliability prediction model from the perspective of algorithm improvement and fusion, to effectively predict the system state reliability.

The state identification process is mainly performed to extract the effective features in the data, including consistency features, conflicting features, correlation features, and a quantitative analysis of uncertainty factors. Through the data training stage, the laws and features of system state reliability are effectively recognized. According to the content of information entropy, the weights of different indicators are adjusted, and the modified feature data are used as the input of the prediction model.

The state prediction process is mainly utilized to define the prediction model and improve the fusion of the prediction model for the characteristics of the wind turbine system, to fuse the multi-factor feature prediction results at the decision level, to improve the model accuracy, and to evaluate and analyze the validity of the prediction model by using the evaluation function. The specific prediction process is shown in Figure 3.

2.2.2. Construction of Improved Prediction Model Based on XGBoost–LSTM

There are differences in the applicability and scope of different algorithms in the actual system. The LSTM algorithm solves the problems of some algorithms in terms of long-term dependence and gradient anomaly and has better temporal characteristics. It is well adapted to the state reliability assessment of the system with sequence modeling requirements. The XGBoost algorithm integrates the advantages of statistics, combining regression, regularization, Taylor, and parallel computing; solves the defects of traditional algorithms in terms of a priori probability and overfitting; and improves the accuracy of the loss function. Both the XGBoost and LSTM algorithms have reliable prediction capabilities, but a single algorithm has a preference for effective feature extraction and feature analysis; for example, LSTM is lacking parallel computing capability, and XGBoost traverses all data to find split points for a long time. Since both algorithms are better applied in system state prediction, this study analyzed the fusion of XGBoost and LSTM algorithms under dynamic weights.

First, the input data are cleaned and normalized, the extracted effective features for correlation are analyzed, and they are used as the input of XGBoost and LSTM algorithms, respectively. Second, the input for fitness is analyzed and trained, and the system state is judged from the perspective of data fusion. Then, the prediction models of the algorithms predict the input parameters according to the training models. The system states corresponding to different parameters are defined, and the predicted values of state parameters are clustered and analyzed, to evaluate the reliability prediction of system states. From the perspective of algorithm fusion, XGBoost and LSTM are fused to construct a new prediction model to predict the system state and output the prediction results, considering the dynamic adaptive adjustment of weights. The block diagram of the prediction model is shown in Figure 4.

Therefore, the construction of the XGBoost–LSTM fusion model can be divided into two main stages: (1) single-prediction model construction and (2) fusion prediction model construction.

(1)

Single-prediction model construction

The LSTM has a cell layer with functions such as memory and forgetting information processing. The LSTM divides the process into three stages: input, forgetting, and output, corresponding to three control gates. The process of analyzing the forgetting gate, input gate, memory cell state, and output gate involves the weight matrices of W_f, W_i, W_c, and W_o, and the input and output weight parameters of the system. The specific calculation formula is shown in Equation (1).

$$\begin{gathered} \mathrm{Forgetting} \mathrm{Gate}: f_t=\sigma \left(W_f\ast \left[h_{t-1},x_t\right]+b_f\right) \\ \mathrm{Input} \mathrm{gate}: i_t=\sigma \left(W_i\ast \left[h_{t-1},x_t\right]+b_i\right) \\ \mathrm{Cell} \mathrm{state}: {\tilde{C}}_t=\tanh \left(W_C\ast \left[h_{t-1},x_t\right]+b_C\right) \\ \mathrm{Output} \mathrm{gate}: o_t=\sigma \left(W_o\ast \left[h_{t-1},x_t\right]+b_o\right) \end{gathered}$$

(1)

where h_t−1 is the cell output value at the previous moment, x_t is the current input value, b is the bias parameter, and σ is the sigmoid function.

Finally, the system output is obtained as shown in Equation (2):

$$h_t=o_t\ast \tanh \left(C_t\right)$$

(2)

In the calculation process of LSTM, the weights are continuously corrected according to the error propagation, and the adjustment of the weights is proportional to the inverse of the error gradient. Using $\Delta w$ to denote the degree of weight change, the weight change can be shown in Equation (3). In the cyclic process, the input weight of one cyclic unit is equal to the sum of the output weight and the weight adjustment value of the previous cyclic unit. Similarly, in the LSTM algorithm, the step-by-step correction of the threshold value $\theta$ follows the same principle.

$$\Delta w=-\eta \frac{\partial E}{\partial w}$$

(3)

Therefore, the output layer function can be considered as a function influenced by weights w, thresholds $\theta$, and error transmission e, and it can be expressed in the form of Equation (4):

$$y=f(w,\theta ,e)$$

(4)

Eventually, the system outputs the state prediction values of the target parameters and further analyzes them according to the thresholds in different states, predicting the operating state that will be achieved by the system at a specific time in the future.

The XGBoost algorithm uses the idea of regression, which uses the new function obtained to fit the residuals of the last function analysis, to complete the training and fitting analysis process of the data. When the data training is completed, the different feature cases of the data features in different periods will be distributed to each leaf node corresponding to it, which will be summed up to obtain the final output prediction value.

Assuming that the objective function has been learned for a total of m iterations, the initial objective function model is as shown in Equation (5):

$$\begin{array}{l}ob{j}^m={\displaystyle \sum _{i=1}^{n}l\left(y_{i,}{{\widehat{y}}_{i,}}^{\left(m\right)}\right)}+{\displaystyle \sum _{j=1}^m\Omega \left(f_m\right)}\\ ={\displaystyle \sum _{i=1}^{n}l\left(y_i,{{\widehat{y}}_i}^{\left(m-1\right)}+f_m\left(x_i\right)\right)}+{\displaystyle \sum _{j=1}^{m-1}\Omega \left(f_m\right)}+\Omega \left(f_m\right)\end{array}$$

(5)

where ${\displaystyle \sum _{j=1}^{m-1}\Omega \left(f_m\right)}=C$, C is a constant, y_i is the prediction result of sample i after m iterations, y_i^(m−1) is the prediction result of the previous (m − 1) trees, $\Omega$ is regularized to prevent overfitting, and l is the loss function.

The regression tree in the XGBoost algorithm is analyzed, and the model of a single tree can be represented by Equation (6):

$$f_m\left(x\right)=w_q\left(x\right),w\in R^T,q:R^d\to \left\{1,2,\cdots ,T\right\}$$

(6)

where w represents the value of a leaf node, q represents the corresponding leaf node, and T is the number of nodes.

In the regression tree construction, the threshold value for splitting the tree needs to be set. When the gain is greater than the set threshold value, the tree starts to split to generate a new tree, and the size of the threshold value is determined according to the cut-off point of the maximum gain. The whole splitting process is based on the theory of regression ideas, and a new tree is constructed by iterative analysis based on the residuals of the previous prediction.

The complexity of the tree is quantitatively analyzed, and the regularization idea is introduced, to control and optimize the complexity of the model and effectively improve the efficiency and accuracy, as shown in Equation (7):

$$\Omega \left(f_m\right)=\gamma T+\frac{1}{2}\lambda {\displaystyle {\sum }_{j=1}^T{w}_j^2}$$

(7)

The second-order Taylor’s formula expansion of the objective function yields Equation (8):

$$ob{j}^m\approx {\displaystyle \sum _{i=1}^n\left[l\left(y_{i,}{{\widehat{y}}_i}^{\left(m-1\right)}\right)+g_i{f}_m\left(x_i\right)+\frac{1}{2}{h}_i{f}_m{\left(x_i\right)}^2\right]}+\Omega \left(f_m\right)$$

(8)

where g is the first-order derivative of the loss function, h is the second-order derivative of the loss function, and γ and λ are hyperparameters.

When dealing with regression problems and classification problems, the commonly used loss functions mainly include mean square error and logarithmic error.

The analysis of the target object is transformed into an optimal analysis of the model. For instance, the loss function is minimized for the process. If the loss function uses the mean square error loss function, then, after a series of derivations, the objective function can be finally derived, as shown in Equation (9):

$$ob{j}^m\approx {\displaystyle \sum _{i=1}^n\left[g_i{w}_{q\left(x_i\right)}+\frac{1}{2}{h}_i{w_{q\left(x_i\right)}}^2\right]}+\gamma T+\frac{1}{2}\lambda {\displaystyle \sum _{j=1}^T{w}_j^2}$$

(9)

From the perspective of tree nodes, the data on the same leaf correspond to the same output and, therefore, simplify the above equation, as in Equation (10):

$$ob{j}^m={\displaystyle \sum _{j=1}^T\left[G_j{w}_j+\frac{1}{2}\left(H_j+\lambda \right)w_j^2\right]}+\gamma T$$

(10)

At this point, the optimal leaf node value w_j and objective function value obj* are calculated as shown in Equation (11):

$$\begin{gathered} w_j^{\ast }=-\frac{G_j}{H_j+\lambda } \\ ob{j}^{\ast }=-\frac{1}{2}{\displaystyle \sum _{j=1}^T\frac{G_j^2}{H_j+\lambda }}+\gamma T \end{gathered}$$

(11)

Therefore, the output function can be considered as a function influenced by the number of nodes, node values, and model accuracy, which can be expressed as in Equation (12):

$$y=f(w_j,T,e)$$

(12)

When analyzing algorithm fusion, it is necessary to comprehensively analyze the multiple influencing factors of the algorithms and the research objects. On the one hand, there are differences in the applicability and scope of algorithms, and on the other hand, there are differences in the characteristics, environment, and needs of different research objects. Therefore, considering the characteristics of the wind turbine system, the weight dynamic adjustment mechanism is introduced to analyze the algorithm fusion process.

The fusion model under different weights has different accuracies, and when the system is in different states at different times, the fixed weights slightly reduce the accuracy of the fused model. Therefore, it uses a dynamic weight assignment model to fuse the algorithms. The dynamic analysis of the weights is based on the fluctuating relative errors of different prediction points, and the weight corresponding to each point in the prediction sequence changes with the fluctuation in the errors. The closer the relative error is to zero, the better the fit between the predicted value and the true value. Because of the existence of negative numbers in the relative error, the error is processed to take the absolute value. The dynamic of the error can be characterized as shown in Equation (13):

$$e_{R_i}=\left|\left({\widehat{Y}}_i-Y_i\right)/Y_i\right|$$

(13)

(2)

Fusion prediction model construction

The objective function is analyzed such that w₁ denotes the weights of XGBoost and w₂ denotes the weights of LSTM, and the weights are dynamically changing; then, the new fusion prediction model function can be represented by Equation (14):

$$f_t\left(x\right)=w_{1\ast }{f}_{1t}\left(x\right)+w_{2\ast }{f}_{2t}\left(x\right)$$

(14)

where t is the fusion moment and i is the number of the sequence points at the fusion moment.

$$w_{1=}{e}_{2R_i}\left(t\right)/\left( e_{1R_i}\left(t\right)+e_{2R_i}\left(t\right)\right), w_{2=}{e}_{1R_i}\left(t\right)/\left( e_{1R_i}\left(t\right)+e_{2R_i}\left(t\right)\right).$$

It can be seen that, with a change in time t, point i moves according to the time series, and the weight function changes with the absolute fluctuation in the relative error value of the prediction sequence; the larger the error fluctuation of the algorithm, the smaller the weight. By dynamically matching the weights of different algorithms for the same parameter at different times, the defects of the algorithms in the prediction process are reduced, and the improvement in the accuracy of the fusion model is achieved through the complementary advantages between algorithms. Finally, in the evaluation of the effectiveness of the fusion prediction model using RMSE, MAPE, R², etc., the closer the goodness-of-fit R² is to one, the better the model is, and the smaller the RMSE and MAPE are, which are preferred outcomes.

3. Results

3.1. Wind Turbine System State Reliability Prediction

A wind farm in Tieling, Liaoning, Northeast China, was considered as an example to study the system state reliability, aiming to improve the reliability of the wind turbine system and the efficiency of wind energy utilization and to reduce the failure rate and maintenance cost. The data used in the algorithm were derived from the internal data of the New Energy Research Institute of China Electric Power Research Institute.

Considering the changing characteristics of the temperate monsoon climate in Northeastern China, as well as the state reliability diversification of the wind farm system in the past year, this study selected several characteristic values of parameters with better calculation characteristics, such as mean, variance, root mean square, and the stiffness of several well-characterized parameters, such as gearbox high-speed bearing temperature, gearbox low-speed bearing temperature, gearbox oil pressure, gearbox inlet oil temperature, and gearbox oil temperature, to preliminarily decipher the data characteristic of the variation trend. Considering the sample size effect, it used 100,000 data points from the last year for the analysis, and the interval between data was about 5 s. The original partial base dataset is shown in

Table 1. Partial base dataset.

Time	Generator Speed	Gearbox Low-Speed Bearing Temperature	Gearbox Oil Pressure	Gearbox Inlet Oil Temperature	Grid Current	Wind Turbine Rotation Speed
1	495	105	1	26	64	46
2	13,134	395	67	50	2030	132
3	12,107	405	61	73	1480	125
4	12,145	403	60	79	1586	132
5	11,246	396	56	74	952	114
6	11,954	381	57	74	1453	125
7	11,669	376	55	71	1352	119
8	6489	359	33	124	275	104
9	8651	355	43	71	624	109
10	11,933	346	56	108	1334	132
…	…	…	…	…	…	…
…	…	…	…	…	…	…
…	…	…	…	…	…	…
n	11,634	491	44	361	1046	125

.

The basic characteristics of the dataset such as mean, variance, root mean square, and cliffiness were analyzed as shown in

Table 2. Basic characterization parameters of the base dataset.

	$\overline{\mathit{X}}$	${\mathit{\sigma }}_{\mathit{x}}^2$	Xrms	Ku
Generator speed	10,521.7620	41,960,495.0503	6477.6922	−1.0436
Gearbox low-speed bearing temperature	504.6761	9946.0789	99.7300	0.9723
Gearbox oil pressure	39.6961	159.9260	12.6462	3.2665
Gearbox inlet oil temperature	288.1725	12,941.4651	113.7606	−0.7909
Grid current	2891.4333	12,926,206.1303	3595.3033	0.7805
Wind turbine rotation speed	104.5015	3774.7921	61.4393	−0.8912

.

There are fluctuations in the data, and failure information is generally displayed in the data characteristics. Therefore, considering the information uncertainty and conflict, the mapping of the component state information to the system state reliability is classified into four types—normal state, hidden risk, explicit risk, and alarm risk—based on the analysis of the correspondence between the historical state anomalies and time series, which are not repeated here.

For the trend of the wind speed–power standard relationship curve, based on the analysis of historical data, this section uses the gearbox high-speed bearing temperature data with better comprehensive performance as the base data and selects the data with a representative interval for the cleaning process. Considering the data fluctuation error to be within ±5% and the wind speed–power standard curve data as the qualified data, excluding the unqualified data, the data were fitted for training and prediction analysis.

3.1.1. Hyperparameter Optimization

Hyperparameter optimization is a key process in machine learning to improve the performance of a model by adjusting the model’s non-learning parameters to find the best model configuration. Hyperparameter optimization can improve the accuracy and generalization ability of the model, avoid overfitting, speed up the convergence of the model, and reduce the consumption of computational resources.

Optuna is an automatic hyperparameter tuning framework that can be used with other frameworks such as Pytorch, TensorFlow, Shlearn, etc. Optuna [43,44] can automatically tune the hyperparameters using samplers such as grid search, random search, and Bayesian search. In this study, after analyzing the hyperparameters of the models, Optuna was finally used to optimize four parameters, such as dropout_rate and neurons of LSTM and eight parameters, such as learning_rate and max_depth of XGBoost model, respectively, and cross-validation was performed for the hyperparameter tuning process. The optimized parameters of LSTM and XGBoost are shown in

Table 3. Parameter optimization for the LSTM model.

Parameter	Value	Parameter	Value
dropout_rate	0.295	neurons	150
num_layers	2	learning_rate	0.0002

and

Table 4. Parameter optimization for the XGBoost model.

Parameter	Value	Parameter	Value
learning_rate	0.008	Lambda	0.002
Alpha	0.017	max_depth	15
Gamma	1	min_child_weight	7
Subsample	0.6	colsample_bytree	0.6

.

3.1.2. System State Reliability Prediction Based on the LSTM Algorithm

For the analysis of the cleaned data using the LSTM algorithm, the sliding window method was adopted to transform the cleaned data into the format of LSTM input data. Meanwhile, the transformed data were normalized to improve the model training efficiency. Considering the size of the sample data, the time step was set to five, and 95% of the data were used as the training set, with the remaining 5% as the test set. The fitting of training and prediction analysis was performed using the determination state mentioned in the previous section as the prediction criterion. Finally, the model was validated.

The data analysis is divided into three stages: training, prediction, and evaluation. First, the dataset is trained. When the obtained training results are better, the training set and the test set are analyzed for prediction, as shown in Figure 5. The graphical fluctuation of the prediction sequence is clearly observed.

Then, on this basis, the wind turbine state reliability is predicted for the next five hours, and the prediction curve is shown in Figure 6. The state prediction assessment is performed using the state definition, considering information uncertainty and conflict as the criteria, i.e., the mapping relationship between system state reliability and the four different types of states. The curve shows that, in the next five hours, the turbine runs smoothly within the normal state thresholds, with no obvious occurrence of state problems. The output prediction dataset is x = [611.4763, 612.5042, 613.5381, 614.5782, 614.5782, 614.5782, 582.3287, 571.6381, 570.9048, 562.3884, 564.1832, 565.4058, 594.1227, 554.6396, 558.9407, 548.3760, 540.8195, 541.8890].

The loss functions of the training and test sets of the LSTM prediction model were calculated, and the loss function curve of the training set effectively decreased with an increase in the number of training times, and finally converged smoothly. The instantaneous function curve of the test set was oscillating, but the overall trend was decreasing and finally converged smoothly with 100 iterations. Figure 7 shows the running curve.

3.1.3. System State Reliability Prediction Based on the XGBoost Algorithm

The data sequence obtained from the previous cleaning step was utilized in the XGBoost model, and 95% of the data sequence was used as the training set, with the remaining 5% used as the test set. The determination state mentioned in the previous section was used as the prediction criterion to fit training and perform the analysis. The prediction sequence schematic of the model is shown in Figure 8.

The XGBoost prediction model was constructed, and a short-term prediction of the turbine state reliability was performed in the next five hours; the prediction curve is shown in Figure 9. Through the output analysis, it can be found that, over the next five hours, the prediction results of the XGBoost prediction model were roughly the same as those of the LSTM prediction model, the turbine’s operation state was smoother, and there was no obvious state problem.

The calculation of the loss function for the XGBoost prediction model shows that the number of iterations is increased compared to that for the LSTM prediction model, but the loss curve is smoother.

The calculation of the loss function for the training and test sets of the XGBoost prediction model shows that the loss function curves for both the training and test sets of XGBoost are smoothly decreasing and eventually smoothly converge. The running curve is shown in Figure 10.

3.1.4. Relative Error Analysis of the LSTM and XGBoost Algorithms

The relative errors of the two algorithms, LSTM and XGBoost, were calculated and compared for evaluation, as shown in Figure 11. Overall, the two relative error curves are relatively close to each other, which indicates that the overall effect is good for both the models. There are fluctuations in some points, and the fluctuation of individual points reaches 30%, which indicates that there are points with low accuracy in the prediction process of individual models.

3.2. State Reliability Prediction Evaluation of Wind Turbine System with Improved Algorithm

According to the prediction data and error fluctuation of LSTM and XGBoost algorithms, the improved algorithm described in Section 2.2 was used to construct the model. The dynamic weights were calculated and analyzed for prediction. The predicted sequence values obtained from the XGBoost–LSTM fusion algorithm are shown in Figure 12.

The predicted values of LSTM, XGBoost, and XGBoost–LSTM were fitted and analyzed. In the smoother part, it is evident that the XGBoost–LSTM predicted value represented by the red curve is in the middle of the LSTM and XGBoost curves; in some fluctuation anomaly locations, the fusion algorithm predicts the results better than the two separate algorithms, and the prediction curves are shown in Figure 13.

According to the prediction sequence curve, the system state reliability prediction data of a future period are the output. The state prediction evaluation was carried out using the state definition that considers information uncertainty and conflict as a criterion, and it was observed that the system was in a normal state throughout the next five hours, with a high degree of system reliability.

The MAPE, RMSE, and R² of the LSTM, XGBoost, and XGBoost–LSTM fusion models were calculated and analyzed under dynamic weight assignment, and the results are shown in

Table 5. Comparison of RMSE, MAPE, and R² for the training and test sets of the algorithm.

Algorithms	Training Set			Test Set
	RMSE	MAPE	R2	RMSE	MAPE	R2
LSTM	16.677	1.391	0.8664	16.970	1.422	0.8087
XGBoost	15.961	1.457	0.8776	16.033	1.415	0.8293
XGBoost–LSTM	15.299	1.049	0.8875	15.836	1.297	0.8337

.

Table 3 shows that the RMSE and MAPE of the XGBoost–LSTM fusion model are reduced compared to those for the previous two algorithms. Among them, the RMSE is reduced by 8.26% and 4.15%, the MAPE is reduced by 24.56% and 27.99%, and the goodness-of-fit R² is closer to 1.

4. Discussion

The conclusions drawn from the above study are as follows:

(1)

In the trend of the wind speed–power standard relationship curve, the gearbox high-speed bearing temperature data with better comprehensive performance can be used as the basic data support for system state classification and reliability prediction, which can effectively predict the reliability of wind turbine state for the next 5 h.

(2)

Considering the dynamic variability of weights for algorithm fusion under irregular fluctuations in the same parameter over time in nonlinear systems, the improved XGBoost–LSTM fusion model reduces the defects in gradient anomalies, the lack of a priori probability, and overfitting that exist in the individual models. It is verified by the index parameters of RMSE, MAPE, and R².

(3)

The improved XGBoost–LSTM fusion model effectively improves the short-term prediction accuracy compared to that of the LSTM and XGBoost algorithms. The RMSE is reduced by 8.26% and 4.15%, respectively; the MAPE is reduced by 24.56% and 27.99%, respectively; and the goodness-of-fit R² is closer to 1, which solves the problem of the low accuracy in predicting the state anomaly of the wind turbine system.

(4)

In the comparison of the results analyzed using the same dataset in LSTM, XGBoost, and the improved XGBoost–LSTM algorithm, it can be found that the model accuracy of the improved XGBoost–LSTM algorithm has been effectively improved, and the accuracy of the probabilistic assessment of the state reliability of the system has also been improved, which is of great value for improving the reliability of the system.

5. Conclusions

In this study, by considering the dynamic analysis of the weight changes for the same parameter by time series under different algorithms, the multi-factor weight analysis mode was adopted to construct the XGBoost–LSTM prediction and assessment model based on the dynamic weight method, which decreased the defects of traditional algorithms in the process of system state reliability prediction and assessment. This study carried out the short-term prediction of wind turbine system state reliability, verified the validity of the model, and solved the problem of the poor prediction of system state abnormality.

The theoretical significance of this study is an improvement in the accuracy of the system state reliability assessment algorithm and the provision of a theoretical and algorithmic basis for the reliability assessment of wind turbine systems in Northeast China.

The practical significance of the study is to reduce the incidence of system failures, minimize failure losses and operation and maintenance costs, improve system reliability and operation efficiency, enable managers to reasonably formulate operation and maintenance plans and strategies based on the predicted conditions, and provide support for management decision makers to formulate operation and maintenance strategies.

However, the current research still has limitations; for example, environmental changes are usually reflected in the state information, but the changes in the external environment are sudden and unpredictable, which do not belong within the scope of the definition of conventional prediction models, and the current research is more inclined to the general reliability law. Considering its influence on the precision of the assessment results, there is still room for improvement in the research conducted in this study, especially pertaining to the environmental factors that impact the sudden emergency response.