Uncertainty-Driven Data Augmentation for Improved Load Time Series Forecasting

Accurate load forecasting is essential for efficient energy management, yet many deep learning models overlook the impact of measurement uncertainty in input data. This study introduces a novel data augmentation (DA) technique that leverages measurement uncertainty to improve forecasting reliability in a Long Short-Term Memory (LSTM) network. Using the publicly available Ausgrid dataset, which records historical energy demand in substations across Australia, the proposed method generates augmented data by sampling realistic variations based on uncertainty levels. The approach enriches the dataset while maintaining fidelity to real-world measurement variability. Experimental results demonstrate significant improvements in predictive accuracy across multiple error metrics, even when using limited training data. The DA-enhanced LSTM outperforms baseline models with longer training epochs and delivers more precise forecasts at reduced computational costs. This study emphasizes the potential of incorporating uncertainty into DA to advance deep learning-based load forecasting, offering practical insights for optimizing energy networks.