Short-Term Wind Power Forecasting by a Long Short Term Memory Ensemble Approach

Abstract

Wind power plants have attracted increasing attention during the last decades due to their environmental and economic benefits. However, the wind resource is inherently unpredictable, bringing important challenges to the stable and safe operation of the power grid. In this context, various computational and statistical approaches have been reported in the literature to perform short-term forecasting of wind power generation, and more efficient strategies are still demanded. In this paper, a hybrid framework that includes a statistical pre-processing stage with an enhanced deep learning (DL)-based strategy is proposed to address the limitations of reported forecasting methodologies to predict multi-seasonal wind power time series. The integrated approach applies a suitable transformation to obtain a normal distribution of data and removes multiple seasonalities in wind power time series. Subsequently, it supplements a set of stacked Long Short Term Memory (LSTM) Recurrent Neural Network (RNN) models for each month of the year. The proposed approach is validated using real hourly wind power data from the Spanish electricity market for the period 2008-2019. A comparative analysis with a well-established DL-based model shows the superior performance of the proposed forecasting method. The experimental evaluation is conducted for 1-3 hours ahead of wind power predictions. Keywords—Long Short Term Memory; Deep Learning; Wind Power Forecasting; Recurrent Neural Networks; Time Series Decomposition