Akademik Veri Yönetim Sistemi
IEEE ACCESS, cilt.13, ss.164130-164155, 2025 (SCI-Expanded, Scopus)
The critical need for accurate electricity demand forecasting in a country like Iraq, where power generation is frequently disrupted, has driven the development of a novel hybrid framework. The methodology is designed to systematically compare multiple preprocessing strategies: it combines Singular Spectrum Analysis (SSA) to reconstruct the signal by emphasizing regularly occurring patterns, removing anomalies, and compensating for potential missing data, followed by various filtering techniques (Exponential Moving Average (EMA), Kalman Filter (KF), and Kalman Smoother (KS)) to further refine the signal, and then applies multiple deep learning algorithms such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BiLSTM), and Bidirectional GRU (BiGRU) to predict hourly electricity demand over a 72-hour horizon, providing a strategic buffer for operational planning and unexpected demand spikes. A comparative analysis of these combinations revealed that the SSA-KS-BiLSTM model achieves state-of-the-art performance, reducing MAE by 54% (from 1531.5 MW to 699.2 MW) and improving R2 by 19% (from 0.811 to 0.967) compared to BiLSTM on raw data. This robust, scalable approach is suitable for other regions aiming to improve power demand forecasting in noisy or uncertain conditions.