Energy-Efficient User Association and Sub-Channel Allocation in Ultra-Dense Networks: A Clustering-Based Approach With LSTM-Driven Channel Prediction

Ultra-dense networks (UDNs) are highly effective in improving network throughput and energy efficiency (EE); however, unplanned small-cell deployment introduces severe intra-tier interference, leading to complex combinatorial challenges that impact network EE. This paper presents a clustering-based framework to optimize EE while managing interference among densely spaced small base stations (SBSs), ensuring quality of service (QoS) for user equipment (UEs). First, we propose the nearest neighbor interference-aware SBS clustering algorithm to optimize SBS clustering into disjoint cell clusters. Then, we introduce a crow search algorithm (CSA) to address energy-efficient joint user association and sub-channel allocation. Additionally, a long short-term memory (LSTM) model is integrated to predict channel conditions, optimizing sub-channel allocation by leveraging historical channel state data. Simulation results demonstrate that our method outperforms PSOUARA, DQN, GA, and Greedy algorithms across key metrics, achieving an 87.64% higher network data rate than PSOUARA, 74.5% higher than Greedy, 71.16% higher than DQN, and 70.2% higher than GA. It improves user acceptance by an average of 23.83% over PSOUARA, 9.83% over DQN, 18.72% over Greedy, and 10.65% over GA, while enhancing EE by 79.49% over PSOUARA, 31.02% over DQN, 32.15% over GA, and 89.27% over Greedy.