Concetti Chiave
Wavelet analysis combined with machine learning methods improves time series forecasting performance.
Sintesi
The article explores the use of Daubechies wavelets and non-decimated wavelet transforms for time series forecasting.
Wavelet features show significant benefits over lagged features in non-temporal methods for one-step-ahead forecasting.
Deep learning models benefit from multivariate inputs using wavelet coefficient vectors, especially with NWPT features.
Experiment results demonstrate improved forecasting performance across various datasets with wavelet features.
Recommendations include considering wavelet features for all time series forecasting tasks.
Statistiche
第1の実験では、NDWT特徴量セットを使用すると、XGBoostモデルのSMAPEが11%低下し、MLPモデルのSMAPEが31%低下します。
第2の実験では、NWPT特徴量セットを使用したGRUアーキテクチャが最も優れた予測結果を示しました。
Citazioni
"Using NDWT and NWPT multivariate inputs result in superior forecasts for seven out of nine deep learning models."
"Wavelet features provide most benefit for wind electricity supply and humidity forecasting."