A Heterogeneous Ensemble Framework Using Hybrid Deep Learning and XGBoost for Hourly Electrical Load Forecasting

DOI

https://www.doi.org/10.59256/indjcst.20260503027

Published In

Volume 5, Issue 3 (2026)

Published Online

September-December 2026

Pages

213-223

Article Type

Original Article

Check for updates

Accurate short-term electrical load forecasting is essential for reliable power-system operation, reserve planning, economic dispatch, and the effective integration of variable renewable generation. The nonlinear and time-varying nature of electricity demand makes it difficult for a single forecasting architecture to represent all relevant temporal patterns. This paper presents a heterogeneous ensemble model that combines two structurally different attention-enhanced deep-learning base learners, namely CNN-LSTM-MHA and CNN-GRU-MHA, with an XGBoost meta-learner. The input pipeline uses historical demand, meteorological variables, calendar indicators, engineered 24-hour and 168-hour demand lags, and a 24-hour rolling mean. Interquartile-range outlier capping and Min-Max normalization are fitted using training data only, followed by chronological construction of 24-hour input windows for one-hour-ahead forecasting. Both neural base learners are optimized exclusively with mean squared error loss and trained using Adam with manually explored hyperparameters on Google Colab. The ensemble uses out-of-fold base-model predictions, residual errors, recent demand values, and summary statistics as meta-features. On the held-out test partition, the proposed model reports RMSE = 28.1609 MW, MAE = 20.0483 MW, MAPE = 1.6612%, and R² = 0.9767, outperforming the eight reported baseline architectures under the same evaluation setting. The results indicate that heterogeneous representation learning and nonlinear meta-level fusion can substantially improve point forecasting accuracy relative to any single base architecture evaluated. The study also highlights the importance of training-only preprocessing, purged out-of-fold stacking, and careful experimental design for reproducible time-series experiments.