Multi-Step Hourly Power Consumption Forecasting in a Healthcare Building with Recurrent Neural Networks and Empirical Mode Decomposition.

However, this behavior was not observed with the other four models, where the errors showed a slightly increasing trend at the beginning of the predictions and stabilized as the time horizon increased.

also in 801 other papers

When the errors were analyzed considering their hourly evolution, it was concluded that both LSTM and GRU alone followed the same behavior as for the univariate case: the error increased as the time step increased for the first predictions but changed to a decreasing trend from the 12th prediction onwards.

also in 47,180 other papers