Modern wind turbines gather a wealth of data with Supervisory Control And Data Acquisition (SCADA) systems. We study the short-term mutual dependencies of a variety of observables by evaluating Pearson correlation matrices on a moving time window. Using clustering on these matrices, we identify multiple stable operational states, which characterize the non-stationarity of mutual dependencies at a single turbine. They represent different turbine operational settings. Moreover, we combine the clustering analysis with a construction of a stochastic process to study the switching dynamics of those states in more detail. Calculating the distances between correlation matrices we obtain a time series that describes the behavior of the complex system in a collective way. Assuming this time series to be governed by a Langevin equation, we estimate the deterministic (drift) and stochastic (diffusion) components of the dynamics to understand the underlying non-stationarity. After adapting our method to specific features of our data, we are able to study the dynamics of operational states and their transitions as well as to resolve hysteresis effects.

翻译：现代风力涡轮机通过监控与数据采集（SCADA）系统积累了丰富的数据。我们通过滑动时间窗口上评估皮尔逊相关矩阵，研究多种观测变量之间的短期相互依赖关系。对这些矩阵进行聚类分析后，我们识别出多个稳定运行状态，这些状态表征了单台涡轮机相互依赖关系的非平稳性，代表了不同的涡轮机运行设置。此外，我们将聚类分析与随机过程构建相结合，更详细地研究这些状态的切换动力学。通过计算相关矩阵之间的距离，我们获得了一个以集体方式描述复杂系统行为的时间序列。假设该时间序列服从朗之万方程，我们估计动力学中的确定性（漂移）和随机性（扩散）分量，以理解潜在的机理。在将我们的方法适应于数据的特定特征后，我们得以研究运行状态的动力学及其转换，并解析迟滞效应。