Optimizing LSTM, Bi-LSTM, and GRU Models with SSA for Daily Electricity Forecasting

Precise forecasting of electricity demand is vital for enhancing energy efficiency and ensuring grid stability. This research extends previous work by integrating Singular Spectrum Analysis (SSA) with enhanced deep learning architectures-specifically, dual-layered Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Longitudinal and Bidirectional LSTM networks have been enhanced through Dropout regularization and Multi-Head Attention mechanisms. The dataset utilized, sourced from the 2016 Electrical Engineering Cup, spans from January 2, 2012, to January 10, 2015, incorporating preprocessing like the exclusion of outliers and filling in missing values. The models show significant improvements in predictive accuracy: the SSA-enhanced BiLSTM model notably achieved an RMSE of 0.1050 and an R2 of 0.7085, a considerable improvement over the original BiLSTM model (RMSE 0.1395, R2 0.4898). Similarly, SSA-enhanced GRU and LSTM models demonstrated substantial performance gains over their original counterparts. modifications enhance the models' capability to effectively discern complex time series patterns, resulting in forecasts that are both more precise and dependable. The advancements underscore the potential of sophisticated model architectures in optimizing electricity demand management and advancing sustainable energy practices.