Wind power forecasting using ensemble learning for day-ahead energy trading

Wind power forecasting is a field characterised by sudden weather-related events, turbine failures and constraints imposed by the electricity grid. Nowadays, different energy markets add the extra challenge of requiring predictions at the minute level a day forward for bidding-processes. This is to avoid trading energy as a bulk and match demand. In this context, we present a novel approach to predict power generation at high frequencies one day in advance, which handles constraints such as curtailment and turbine degradation. This has been tested over historical data from SCADA systems and historical forecasts from wind speed providers for eight windfarm locations in Ireland over two years. Our work was performed in two phases. First, we undertook a preliminary study to analyse the relationship between all combinations of observed wind, forecasted wind and electrical power. Secondly, a wide variety of Machine Learning algorithms were run over each of the locations in order to assess the degrees of predictability of different algorithms and regions. Most of the algorithms benchmarked improve linear wind to power mappings besides the high degree of noise in this domain. Our analysis and experimental results show how boosting ensembles are a cost-effective solution in terms of runtime among other Machine Learning algorithms predicting wind power a day ahead.