This paper proposes a new model

which consists of a mathematical morphology (MM) decomposer and two long short term memory (LSTM) networks

to perform ultra-short term wind speed forecast. The MM decomposer is developed in order to improve the forecast accuracy

which separates the wind speed into two parts: a stationary long-term baseline and a nonstationary short-term residue. Afterwards

two LSTM networks are implemented to forecast the baseline and residue

respectively. Besides

this paper makes an integrated forecast that takes into account multiple climate factors

such as temperature and air pressure. The baseline

temperature and air pressure are used as the inputs of baseline network for training and prediction

and the baseline

residue

temperature and air pressure are used as the inputs of residue network for training and prediction. The performance of the proposed model has been validated using data collected from the Australian Meteorological Station

which is compared with least squares-support vector machine (LS-SVM)

back-propagation artificial neural network (BPNN)

LSTM

MM-LS-SVM

and MM-BPNN. The results demonstrate that the proposed model is more suitable to solve non-stationary time-series forecast

and achieves higher accuracy than the other models under various conditions.