Ensemble forecast with machine learning algorithms

The most obvious option is the ensemble mean where every model is given the same weight. In many cases, this approach brings little improvements (if any, depending on the target) that may not be worth the price of running an ensemble. More successful approaches come from data assimilation, especially with the ensemble Kalman filter and its variants. Another class of methods produces a weighted average of the individual forecasts: the weights are computed based on past observations and past model forecasts. They are updated before any new forecast period, hence the approach is referred to as sequential aggregation. It should be noted that, in the present study, the weights do not depend on the position so that they may be applied away from the observation locations (which is actually validated in practice).