Abstract

Crucial to wind energy penetration in electrical systems is the precise forecasting of wind speed, which turns into accurate future wind power estimates. Current trends in wind speed forecasting involve using Recurrent Neural Networks to model complex temporal dynamics in the time-series. These networks, however, have problems learning long temporal dependencies in the data. To address this issue, we devise a Multi-scale Model Based on the Long Short-Term Memory for the day-ahead hourly wind speed forecasting task. Our model uses dense layers to build sub-sequences of different timescales which are used as input for multiple Long Short-Term Memory Networks (LSTM), which model each temporal scale and integrate their information accordingly. An experiment with altered wind speed data shows that our proposal is better able to learn long term dependencies than the stacked LSTM. Furthermore, results on four wind speed datasets of varying length from northern Chile reveal that our approach outperforms several models in terms of MAE and RMSE. Training times also exhibit that adding depth to the model does not increase computational times substantially, making it a more efficient approach than the stacked LSTM.