Wind power forecasting in distribution networks using non-parametric models and regression trees

Abstract Renewable resources provide viable and advantageous solutions up to a certain integration share. At higher penetration levels, they violate the conventional generation constraints, leading to decentralized uncertainty with respect to bi-directional power flows. This generates an increasing need for smart tools able to predict their output with high accuracy, based on easily accessible input data for forecasting. Based on actual data with respect to load demand and wind power generation, this work presents a realization of decision trees that target on a continuous response, also known as regression trees. Utilizing the speed and direction of wind, the ambient temperature, relative humidity, renewable capacity and renewable energy source curtailment as predictors in distribution networks of different regions, the proposed configuration is able to predict the generated power with high accuracy. According to the obtained results under distinguished scenarios, the inclusion of temperature and humidity to the predictive list greatly improves the accuracy in terms of mean square error, root mean square error and mean absolute range normalized error, whereas the renewable availability offer no relevant changes. However, in the forthcoming de-carbonized power systems, the impact of curtailed energy will play an important role in expert forecasting systems where the input/output association must be modelled with high resolution.