Weather forecasting based on hybrid decomposition methods and adaptive deep learning strategy

Many global climate-affecting factors are combined to influence the weather and the most challenging ones are the wind speed and the air temperature. The objective of our study is to build a reliable model, capable of handling both data different behaviors for an efficient 12 hours-ahead prediction of each parameter separately. A hybrid strategy called OVMD-DWT-Attention-based-Adaptive-mLSTM was proposed for this purpose, based on a double decomposition method that combines the optimized variational mode decomposition (OVMD) algorithm with the discrete wavelet transform (DWT) technique, that proved its efficiency in extracting the appropriate features, and pre-processing both datasets, in order to reach the desired data denoised effect. The denoised high and low-frequency sub-sequences of each parameter resulted are then forecasted separately using the Adaptive-Multiplicative-LSTM (Adaptive-mLSTM) model to eliminate the inter-correlations between the sub-signals. Finally, each parameter predicted results are reconstructed and fitted independently to the Attention-based-Adaptive mLSTM proposed model for the final prediction. A benchmark of models was implemented for comparison purpose and evaluated over both wind speed and air temperature datasets characteristics, collected from three different stations, where the proposed strategy showed the best and the more consistent results.

Automated summary

The objective of this study is to build a reliable model for the efficient prediction of wind speed and air temperature changes in the next 12 hours. A hybrid strategy was proposed, which combines the optimized variational mode decomposition (OVMD) algorithm with the discrete wavelet transform (DWT) technique for data preprocessing. The Adaptive-Multiplicative-LSTM (Adaptive-mLSTM) model is used to independently predict the denoised sub-sequences, and the Attention-based-Adaptive mLSTM model is applied for the final prediction.