Due to current structural change in energy systems (energy transition), the construction of wind turbines in order to provide a sustainable production of energy rose to prominence in the last couple of years. While investment in renewable energy has been supported by governments in one way or the other in the past, ceased subsidy may strongly influence profitability in the future. The costs of construction as well as for operations and maintenance (O&M) can be significant. Unscheduled maintenance, often caused by breakdown of the turbine, has been identified as a major part of overall (O&M) costs, owing to increased logistical expenditure and lost revenue. As a consequence condition monitoring systems (CMS) more and more have manifested themselves to measure behavior of the turbine or individual components in order to provide diagnostic information. This in turn facilitates the scheduling of repair work as well as allows the prediction and consecutive avoidance of component failures that may lead to a breakdown. CMS can have access to various data sources in form of time series, that include vibration, voltage levels or performance related information. Common CMS approaches and related research likewise are often focused on anomaly detection in time series by applying physical models and various transformations on the data (e.g. Wavelet transformation). The usage of machine learning algorithms for prediction lacks a more thorough investigation. The subject of this thesis is to contribute to the closure of this gap by applying two different machine learning algorithms to performance related data of a wind turbine in order to detect anomalies and predict events in defined future time intervals. Despite the technical nature of this specific case, we aim to present the applied approach as valuable procedure for any time series prediction as often occurring in the domain of predictive analytics. Besides physical time series data also sequential event data were provided. Because of it's proclaimed applicability for modeling time series, a restricted Boltzmann machine (RBM) with Gaussian visible units has been used. As it is a generative model the trained RBM can estimate the probability of a given set of input variables, that are the sensor values at a specific point in time. In case the probability falls below a threshold the occurrence of an event is predicted. The second applied algorithm was a support vector machine (SVM), a binary linear classifier which is trained through supervised learning. To examine the performances of the methods for the respective time interval a receiver operating characteristic (ROC) was used. Thereby not only false event predictions but also the number of missed events could be investigated. The results substantiate the approach to be valuable for data analysis, although further improvements are possible. Furthermore, the derived predictions can be enhanced by decision prescriptions in the sense of prescriptive analytics. A link to entrepreneurial circumstances could enlarge practical benefits.