Non-dominated solutions for time series learning and forecasting

While measuring forecasting accuracy has been intensively studied by researchers, the consideration of sets of non-dominated solutions has not yet been explored in the literature. Even when considering measures that are in harmony, small changes in model parameters can result in a trained system capable of forecasting different time series characteristics. In this regard, this study shows how a set of non-dominated forecasting models can assist decision makers to pick the most suitable model for operations on dynamic environments. In particular, the system presents an interesting trade-off provided with assistance of a new strategic measure, which forces the forecasting model to learn the highest values of the time series. An automatic self-adaptive forecasting framework, calibrated with Multi-Objective VND inspired techniques, and able to perform k-steps-ahead forecasting, is considered. There is a growing demand for learning big-data time series, such as those derived from sensors in energy mini/microgrid systems, in which real-time decisions should be made quickly and can vary according to the available energy resources. Thus, a case of study is considered using data from disaggregated power readings of a typical microgrid due to the fact that load demand forecasting, with different components of a house, is of crucial importance for management and operation of new emerging decentralized systems.