Drift-Aware Dynamic Neural Network for Improving Short-Term Load Forecasting

Abstract

Load forecasting methods, including statistical models and conventional machine learning techniques, often face nonstationary and volatile grid load data challenges, leading to limited forecasting performance. This study presents DRAINOT, an advanced framework for grid load forecasting, enhancing the DRift-Aware dynamIc neural Network (DRAIN) by incorporating a Temporal convolutional network (TCN) for parameter optimi-sation and drOpout layers for improving generalisation across diverse domains. By replacing the original Long Short-Term Memory with TCN, DRAINOT significantly enhances learning capabilities and adaptability, effectively capturing temporal shifts and evolving load patterns. DRAINOT achieves superior gener-alisation, forecasting accuracy, and reduced computational time compared to state-of-the-art models such as Transformer and Informer, as demonstrated on public load data across Belgium and four Australian states.