Online forecast reconciliation in wind power prediction

Increasing digitization of the electric power sector allows to further rethink forecasting problems that are crucial input to decision-making. Among other modern challenges, ensuring coherency of forecasts among various agents and at various aggregation levels has recently attracted attention. A number of reconciliation approaches have been proposed, from both game-theoretical and statistical points of view.

However, most of these approaches make unrealistic unbiasedness assumptions and overlook the fact that the underlying stochastic processes may be nonstationary. We propose here an alternative approach to the forecast reconciliation problem in a constrained regression framework. This relies on a multivariate least squares estimator, with equality constraints on the coefficients (denoted MLSE).

A recursive and adaptive version of that estimator is derived (denoted MRLSE), hence allowing to track the optimal reconciliation in a fully data-driven manner. We also prove that our methods by design guarantee the coherency property for any out-of-sample forecasts (reconciliation by design). We show the effectiveness of our forecasting methods using a Danish wind energy dataset with 100 wind farms.