Pu Wang: Automatic Singular Spectrum Analysis and Forecasting
The singular spectrum analysis (SSA) method of time series analysis applies nonparametric techniques to decompose time series into principal components. SSA is particularly valuable for long time series, in which patterns (such as trends and cycles) are difficult to visualize and analyze. An important step in SSA is determining the spectral groupings; this step can be automated by analyzing the w-correlations (weighted correlations) of the spectral components. To illustrate, monthly data on temperatures in the United States for about the last 100 years are analyzed to discover significant patterns.

Rui Li: Short-Term Wind Energy Forecasting with Temporally Dependent Neural Network Models
As the penetration of renewable energy into the electrical grid is increasing worldwide, accurate forecasting of renewable energy generation is essential not only for grid operation and reliability, but also for energy trading and long-term planning. In this paper, we focus on short-term wind energy forecasting. The inherent variability and unpredictability of wind energy imposes great challenges upon many models. Conventional time series models, such as ARIMAX, often fail to capture nonlinear patterns in energy output, and a feedforward artificial neural network doesn’t take temporal dependency into account. In this paper, we apply state-of-art autoregressive artificial neural network (AR-ANN) models and recurrent neural network (RNN) models to wind energy forecasting. By capturing both the sequential pattern of energy output and the complex relationship between weather predictors and power generation, we can achieve better forecasting accuracy. These temporally dependent neural network structures can also be easily extended to model other nonlinear time series and temporal data.

Anna Mándli: Time Series Classification for Scrap Rate Prediction in Transfer Molding In this paper, we present and evaluate methods for predicting critical increase in manufacturing scrap rate of automotive electronic products. Along with information on processes such as maintenance cycles, we analyze the sensor time series of the so-called transfer molding process, in which the electronic product is packaged into plastic for protection. Production data are organized in a two level hierarchy of the individual parts and of the sequence of parts. Since the main goal is to predict and warn about the future state of the process, we designed a training and prediction framework over certain production cycles. By using sensor and other information, we adapt known time series classifi- cation methods to predict increase in scrap rate in the near future. By using three months of manufacturing time series, we evaluate both feature based and dynamic time warping based methods that are capable of fusing a large number of production time series. As a main conclusion, we may warn the operators of increase in failures with an AUC above 0.7 by combining multiple approaches in our final classifier ensemble.