Introducing distributed learning approaches in wind power forecasting

Abstract

Renewable energy forecasting is now of core interest to both academics, who continuously propose new forecast methodologies, and forecast users for optimal operations and participation in electricity markets. In view of the increasing amount of data being collected at power generation sites, thanks to substantial deployment of generating capacities and increased temporal resolution, it may now be possible to build large models accounting for all space-time dependencies. This will eventually allow to significantly improve the quality of short-term renewable power forecasts. However, in practice, it is often the case that operators of these generation sites do not want to share their data due to competitive interests. Consequently, approaches to privacy-preserving distributed learning are proposed and investigated here. These permit to take advantage of all potential data collected by others, without having to ever share any data, by decomposing the original large learning problem into a number of small learning problems that can be solved in a decentralized manner. As an example, emphasis is placed on Lasso-type estimation of autoregressive models with offsite observations. Different applications on medium to large datasets in Australia (22 wind farms) and France (85 wind farms) are used to illustrate the interest and performance of our proposal.