Multi-family wavelet-based feature engineering method for short-term time series forecasting.

Area of Science:

• Climate Science
• Energy Management
• Time Series Analysis

Background:

• Accurate short-term forecasting of natural phenomena is crucial for climate science and energy management.
• Traditional forecasting methods can be enhanced through advanced feature engineering, particularly wavelet transformations.

Purpose of the Study:

• To introduce a novel feature construction method for short-term time series forecasting using the Stationary Wavelet Transform (SWT) and multi-family wavelets.
• To improve the accuracy of forecasting models by augmenting time series data with detailed wavelet coefficients.

Main Methods:

• Application of the Stationary Wavelet Transform (SWT) with multiple wavelet families (Daubechies, Symlets, Coiflets, Haar, Meyer).
• Feature engineering by supplementing original time series data with wavelet coefficients, preserving dimensionality.
• Utilizing Long Short-Term Memory (LSTM) neural networks for forecasting tasks.

Main Results:

• Wavelet-augmented LSTM models demonstrated consistent error reductions across multiple datasets.
• Mean Absolute Error (MAE) decreased by 13.6%, Mean Squared Error (MSE) by 17.7%, Root Mean Squared Error (RMSE) by 9.5%, and Symmetric Mean Absolute Percentage Error (SMAPE) by 13.9%.
• The multi-family SWT features proved effective in enhancing forecasting accuracy for meteorological variables, electricity demand, and wind-power output.

Conclusions:

• The proposed multi-family SWT feature engineering approach offers a dataset-agnostic method for improving short-term forecasting accuracy.
• This technique enhances the informativeness of time series data, leading to more reliable predictions in critical domains.
• The integration of SWT-derived features represents a significant advancement in time series forecasting for natural phenomena.