Wavelet-Enhanced Hybrid LSTM-XGBoost Model for Predicting Time Series Containing Unpredictable Events

Abstract

Accurate electricity consumption forecasting is essential for effective power management, especially in the presence of unpredictable events that disrupt typical consumption patterns. Using the COVID-19 pandemic as a case study for such unpredictable events, this study proposes an improved hybrid LSTM-XGBoost model with adapted wavelets to capture complex, irregular fluctuations in energy demand. The model first applies wavelet decomposition to the original data, extracting multiple frequency components that highlight short-term variations and long-term trends. By incorporating these wavelet coefficients as features, the model is sensitized to anomalous events, resulting in more accurate forecasts over a more extended period without the need for frequent retraining. The hybrid approach takes advantage of the LSTM's ability to model temporal sequences and uses XGBoost to adjust for residual errors. Experimental results show that the model can effectively forecast energy demand with minimal error, especially on regular weekdays, and achieves robust performance in the face of unforeseen anomalies. This methodology shows a promising aspect for improving the reliability of energy forecasting models with potential applications in smart grid management and sustainable energy planning.