Identifying the anomalies of wind turbine (WT) and maintaining in time will improve the reliability of wind turbine and the efficiency of energy use, however it is difficult toidentify the wind turbine’s abnormal operation by the traditional threshold settings because the anomalies can be induced by multiple factors.Therefore, this paper presents an anomaly identification model for wind turbine state parameters,and the model can identify abnormal state which the fluctuation range of the condition parametersis within the SCADA alarm threshold. The main work is as follows: 1) in order to increase the accuracy of the prediction model, a novel BPNN model integrated genetic algorithm (GA) was employed to optimize the training method (called GABP method), data samples, and input parameter selection, respectively; 2) on this basis, the distribution characteristics of state parameter prediction errors were depicted by a T-location scale (TLS) distribution with the shift factor and elastic coefficient; 3)error abnormal index (EAI) is defined to quantify the abnormal level of the prediction error, which is used as an indicator of the wind turbine anomaly. The proposed method has been applied on areal 1.5 MW wind turbine, and the analysis shows that the proposed method is effective in wind turbine anomaly identification.

Introduction

Wind power has attracted global attention in recent years as a clean and renewable energy generation (Li et al., 2017). Wind turbines (WTs) are an emerging renewable energy technology that have the potential to provide low carbon intensity power in the future (Cruz and Martín, 2016; Demir and Taşkın, 2013; Li et al., 2016; Ortegon et al., 2013). Rapid developments of wind energy in recent years have drawn attention to issues of operation and maintenance (O&M) of wind farms (Li and Chen, 2013). Andthe detection of wind turbine faults is considered to be a cost-effective approach to improve the reliability of WTs and reduce the O&M costs of the wind farms (Li et al., 2012). Resently, the detection of wind turbine faults becomes a hot problem. In the study (Kusiak and Verma, 2012a; Kusiak and Li, 2010), three wind turbine condition parameters, including a main bearing temperature, a lubrication oil temperature of the gearbox, and the winding temperature of the generator, were modeled in a back propagation neural network (BPNN) for the fault detection of WTs based on SCADA data (Zaher et al., 2010). How to utilize the BPNNs to model the wind turbine parameters with SCADA data was also investigated in publications (Garcia et al., 2006a; Lapira et al., 2012; Schlechtingen and Santos, 2011a; Stickland, 2012; Xiang et al., 2009). A comparative analysis of two BPNN-based models and a regression-based model was presented (Schlechtingen and Santos, 2011a) for modeling parameters of gearbox bearing temperature and generator stator temperature. Besides, certain thresholds of prediction errors are usually set to identify the anomalies in the WTs (Kusiak and Verma, 2012b; Schlechtingen et al., 2013; Schlechtingen and Santos, 2011b). Intelligent anomaly identification systems, such as the multi-agent system (Zaher and Mcarthur, 2007) and SIMAP (Garcia et al., 2006b) were developed using the prediction models of the wind turbine condition parameters such as gearbox bearing temperature, gearbox oil temperature, and generator winding temperature. However, most of the previous studies identify the wind turbine’s abnormal operation by the traditional threshold settings, and it is difficult to identify abnormal state which the fluctuation range of the condition parameters is within the SCADA alarm threshold(Chandola et al., 2009; Sun et al., 2016; Sun et al., 2016). To solve the problem, the paper developed an anomaly identification model for the wind turbine state parameters. The main work is as follows. First, in order to increase the accuracy of the prediction model, a novel BPNN model integrated genetic algorithm (GA) was employed to optimize the training method (called GABP method), data samples, and input parameter selection, respectively. On this basis, the distribution characteristics of state parameter prediction errors were depicted by a T-location scale (TLS) distribution with the shift factor and elastic coefficient. After that, the estimated residual anomaly intensity of the turbine state parameter was quantized and indicated. Finally, the proposed method has been applied to a real 1.5 MW wind turbinefor verification.