With the increase of the scale of wind farms and the proportion of wind power in energy system, the power grid has higher and higher requirements for the voltage stability of wind farms. The actual operation status of wind farms is focused, and the key issues of the reactive power and voltage control of wind farms are summarized, including how wind turbines adjust the reactive power, how to allocate the reactive power in large scale wind farms, how to maintain the stability of the internal node voltage of wind farms, how to solve the problem of voltage control lag, the transient voltage control strategy under fault conditions, and so on. In view of these issues, the methods and characteristics of reactive power and voltage control technology in wind farms are systematically summarized, and the realization process of reactive power and voltage control technology is expounded from the aspects of the characteristics of wind turbines and reactive power compensation equipment, reactive power and voltage steady-state control, internal node voltage control, model predictive control, and transient control under fault conditions. The research can provide reliable technical means for safe and stable operation of wind farms.

In order to effectively monitor the abnormal tower vibration and ensure the unit operation safety, a data-knowledge-driven variable condition tower vibration prediction method based on long-short term memory (LSTM) and empirical mode decomposition (EMD)-eXtreme gradient boosting (XGBoost) algorithm step-by-step modeling is proposed. Firstly, the relationship between environmental and operational variables is stripped out based on the analysis of the unit's operating mechanism and the wind turbine SCADA operating parameters that affect tower vibration are identified. Then, the ultra-short term prediction of unit environmental wind speed and operating power is realized based on LSTM, and the unit data knowledge model is established based on the full working condition historical operating data. Finally, Hilbert-Huang transform (HHT) is used to decompose the vibration signal and extract the low frequency vibration of the tower, and build a tower vibration prediction model based on XGBoost algorithm. Through inputting the predictive variables, the prediction results of the tower low frequency vibration are output, and the prediction interval is determined. The results show that, the tower vibration prediction model can effectively predict the tower vibration, determine the tower operation condition, and ensure the smooth operation of the unit.

In view of the high failure rate of the insulated gate bipolar transistor (IGBT) of wind turbine converter and the fact that the failure occurs on a short time scale, a health state assessment method of the IGBT based on dynamic regularization and Park vector centrifugal change rate is proposed. The similarity calculation model is established by using the dynamic regularization algorithm to calculate the minimum regularization distance and waveform similarity of three-phase waveforms to judge the condition of the converter. The centrifugal rate and change rate of Park vector ellipse are used to evaluate the IGBT status and set the evaluation index. Moreover, the practicability is verified by simulation data and operation data, respectively. The results reveal that, these two methods have good practicability, the waveform similarity decreases gradually before the fault occurs, and the change trend of Park vector eccentricity continues to increase, which proves that the two algorithm models can clearly distinguish between normal and abnormal waveforms. Using these two methods can timely feed back of the converter health status, thus to effectively avoid the shutdown or damage of the power electronic system due to the IGBT fault and avoid the property loss caused by equipment fault.

Large number of bolted structures exist in wind turbine equipment, once the bolt hole is defective, it may lead to fracture of the entire matrix and cause major accident, but the current bolt hole defect detection method has the situation of missed detection and misjudgment. Aiming at solving this problem, through investigation, theoretical analysis and physical research, two methods which combine special tooling with probe for bolt hole defect detection are developed, namely the direct beam method and the sector scanning deflection method. Taking the bolt hole of pitch bearing in wind turbine as the research object, the CIVA software is used to simulate the two detection methods, and the rectangular simulated crack can be detected. Experiments were carried out on the defects of rectangular grooves in the actual pitch bearing bolt holes, which showed that the direct beam method could realize the defect detection of bolt hole cracks. The research provides a new method for monitoring the service status of bolt holes, which is beneficial to ensure the safe service of bolted structures.

China's large-scale clean energy base has formed a regional power-heat combined system with multiple randomness, such as wind power, photovoltaic power generation and power/heat load, which highly depends on flexible and adjustable resources. Its low-carbon and economic operation is also challenging. In order to fully integrate the resources from source and load sides and maximize the consumption of new energy power, a day-ahead optimal dispatching method of the regional integrated electric-heating operation system considering demand response of electric, heating loads and prediction error scenario is proposed. Firstly, sliding time window multivariate Gaussian mixture distribution and Monte Carlo are used to generate and correct the random source-charge day-ahead prediction error to further optimize the forward scheduling results. Secondly, the demand response models of electric and heating loads are analogically defined. A regional electric-thermal system model considering the source-charge interaction is established. Then, considering multiple costs, a low-carbon and economic pre-ahead scheduling expectation model of the joint system is established in all typical scenarios. Finally, by taking the winter energy supply in a region in northern China as an example, the multi-scenario optimization operation results of electricity with and without electricity and heat load demand response are gradually compared. The results show that, fully introducing the electric and heating load demand responses and molten-salt thermal energy storage boiler can promote the consumption of large-scale wind and photovoltaic power, and significantly enhance the low-carbon and economic operation capability of the regional integrated electric-heating system.

For wind power units, both torque control and pitch control are designed based on generator speed, which leads to the coupling between them near the rated speed, further resulting in rotational speed disturbances. Consequently, conventional torque control will frequently change the control logic according to generator speed, thereby causing the torque and power dips. Besides, two kinds of conventional torque control strategies above the nominal wind, constant-torque and constant-power control, have advantages and disadvantages in torque, power and drive-train loads. In order to eliminate the torque and power dips, full load curve was extended to generator speeds below the rated value. Several state variables, relevant to pitch angle, were selected to reduce the frequency of switching the torque control regions near the rated speed. A new full load curve was designed based on weight programming to comprehensively consider the power and torque performances. The transition curve connecting the optimum Lambda curve and full load curve, was optimized dynamically. Moreover, the proposed control strategy was tested under typical turbulent wind conditions. The simulation results show that, the torque and power dips at above nominal wind are eliminated and the power capture increases at rated wind by using the improved torque control strategy. Furthermore, both torque and power achieves good performance by applying the rational weight.

Existing single-channel networks have poor noise immunity during fault diagnosis of rotating machinery due to the many noises associated with the operation of rotating machinery. To address this problem, a two-channel input LetNet-5 convolutional neural network model incorporating a parallel mechanism was proposed. Case Western Reserve University bearing dataset was used for the model plausibility check process, based on which Gaussian white noise with a signal-to-noise ratio of -10 dB was added to simulate the real noise situation. The short-time Fourier transform was used to process the motor fan-side and drive-side vibration data, and the resulting time-frequency images were passed to a two-channel input LetNet-5 convolutional neural network for training and learning. The results show that, the dual-channel input LetNet-5 convolutional neural network model is able to capture the fault features in a strong noise environment well, it has higher efficiency and accuracy than the multi-scale feature fusion residual model, the multimodal coupled input neural network model, the conventional K-nearest neighbour and decision tree model and the single-channel input LetNet-5 convolutional neural network model.

As a critical component for capturing wind energy, wind turbine blades may subject different degrees of damage due to blade manufacturing and operating load, which directly affects the reliability of wind turbine operation. For preventing quality and safety accidents, a fast and easy non-implantable detection method is needed to identify the damages. According to the physical correlation between blade damage and blade operation noise, a blade damage detection method based on acoustic signal and convolutional neural network (CNN) is proposed. The method converts the time-series acoustic signal into a two-dimensional spectral picture and combines the healthy spectral picture to generate a residual spectral picture. Then, the residual spectrogram is used to train the convolutional neural network and detect the damage. The analysis results show that the algorithm eliminates the influence of the inherent blade sweeping sound generated by the impeller rotation on the damage identification and improves the identification accuracy. The algorithm analysis was carried out with the actual measured data of a local wind turbine, and the results showed that the classification accuracy of the algorithm reached 96.9%, which verified the effectiveness and accuracy of the detection method based on convolutional neural network.

The integration of large-scale wind farms will lead to transient stability problems, such as frequency offset, weak feedback and high harmonics, and affect the safe operation of the system. Therefore, research of transient protection technology for large-scale wind farm integration is an urgent requirement. A comprehensive review about grid connection technology and the transient protection technology for large-scale wind farms is carried out. The topology structure of each integration technology is carefully analyzed. Besides, the fault characteristics of three power systems, including permanent magnet synchronous generator, doubly fed induction generator, and traditional synchronous power system are compared and the current research status of transient based protection of onshore and offshore wind farms is illustrated. Finally, perspectives are outlined for future development of relay protection technology in large-scale wind farms, which provides a significant reference for the research of future wind farm integration.

Aiming at the problem of weak internal regularity caused by the characteristics of nonlinear and strong fluctuation of load data, a TCN-WOA-BiLSTM-Attention power load short-term prediction model based on Bootstrap error correction was constructed. Temporal convolutional network (TCN) was used to extract temporal features and the contribution of important information to the features was highlighted through the Attention mechanism. The whale optimization algorithm (WOA) was employed to find the optimal bidirectional long short term memory network (BiLSTM) hyperparameters, thus to reduce the negative impact of manual search hyperparameters and then forecast. Based on Bootstrap analysis on error distribution of the prediction interval, the necessity of correcting the prediction result was judged by whether the PICP was lower than the corresponding confidence, and the reasonable correction range was selected. The results show that, the error correction based on the Bootstrap method can avoid the problem of insufficient correction and excessive correction. Compared with the method of correcting the whole error sequence, it is more scientific and improves the prediction accuracy of the model to the greatest extent.