LEVERAGING HYBRID MODEL FOR RESOURCE ALLOCATION AND PREDICTION IN DYNAMIC EDGE ENVIRONMENT
Shreya Agarwal, Sonam Seth
Abstract: Edge computing has emerged from the broader paradigm of distributed processing to bring computation closer to end devices, but the dynamic and resource-constrained nature of edge workloads makes resource allocation, utilization forecasting, and energy-aware scheduling considerably more difficult than in centralized cloud settings. Conventional statistical and single-branch recurrent forecasting methods may not jointly capture short-term fluctuations, long-range temporal dependencies, and context-sensitive temporal importance. To address this limitation, the present study proposes an attention-enhanced residual hybrid deep-learning framework, referred to as Proposed-Mega, which integrates a one-dimensional convolutional front end, bidirectional LSTM and GRU processing, multi-head self-attention, layer normalization, dropout regularization, and a persistence-based residual connection for resource/energy forecasting. Extensive preprocessing, temporal feature engineering, and stationarity testing using the Augmented Dickey-Fuller test were performed. On the held-out test set, Proposed-Mega achieves the lowest MAE (0.1137), MSE (0.0198), and RMSE (0.1406), and the highest R² (0.6855) among the five evaluated architectures. The Hybrid model achieves the lowest MAPE (9.9566%), while Proposed-Mega records 10.0245%. When the five metrics are considered jointly using average rank, Proposed-Mega obtains the best overall ranking (1.20), followed by Hybrid (1.80). Five-fold cross-validation of Proposed-Mega gives RMSE values from 0.1679 to 0.2073 kW, with a mean of 0.1804 ± 0.0138 kW. These findings indicate that the attention-enhanced hybrid architecture provides strong predictive performance on the held-out dataset and offers a useful predictive basis for proactive resource-management decisions, while full closed-loop resource allocation remains outside the scope of the present evaluation.
Keywords: Hybrid BiLSTM-GRU networks; Attention-enhanced deep learning; Edge computing; Resource allocation and prediction; Time-series forecasting
DOI: 10.24874/PES08.03.033
Recieved: 05.05.2026 Revised: 17.06.2026 Accepted: 23.07.2026
UDC:
Reads: 7 
