Advanced spatio-temporal wind power forecasting with distributed wind farms as sensors

International audienceThe large density of wind farm installations in an electricity grid imposes several challenges to the economic and secure power system operation. The capacity of the power system to accommodate large shares of wind generation is enhanced by good quality forecasts (i.e. few minutes up to few days ahead) of the wind production. On the other hand, most wind farm facilities spread over a territory nowadays are equipped with measurement systems (anemometers, SCADA, ...). When this information is collected centrally, it allows monitoring the weather and energy situation with a greater temporal granularity than what can be handled by meteorological models. Considering distributed wind farms as measuring stations presents an opportunity for improved monitoring and forecasting of wind generation. Previous works have shown what type of improvement can be obtained in wind power forecasting with a spatio-temporal model integrating regional aggregated data into a single model. For some regions in Denmark an accuracy improvement up to 15% could be obtained for one hour ahead forecasts. In this paper, we propose to show how to integrate spatio-temporal wind power production data into a statistical model at a finer spatial resolution. The aim is to consider transformation stations or wind farms over an area as simultaneous data entries. A statistical forecasting approach is proposed for this purpose. The problem of dimensionality of the statistical model is addressed by proposing a sparse model with an appropriate dynamical term relating the local wind speed and direction to planar wave propagation term. Beyond the measured power data, the interest of integrating meteorological observations into this dynamic spatio-temporal model is shown. The methodology is applied to the test case of Denmark with a resolution going down to the transformation stations (aggregation points of several wind farms). By this way more than 300 spatial points are taken into account. Historical data sets of more than 5 years at a 15 minutes resolution were used. The obtained accuracy improvement for the entire area considered of Denmark exceeds 15 % and for some regions, it exceeds 25 % for up to 4 hours ahead. These results are very promising and demonstrate that without additional cost , but by using the existing infrastructure , one can obtain considerable improvement in the forecasts accuracy