With the growing scale and heterogeneity of high-performance computing (HPC) systems, energy consumption has become a critical challenge, exerting a substantial impact on operational costs and sustainability. Existing energy prediction methods, however, often exhibit limited accuracy and generalization when confronted with heterogeneous workloads and sparse historical data, particularly for users lacking sufficient records. To overcome these limitations, we first propose a lightweight dual-LSTM-Transformer network (L-DLTNet) that captures short-term dynamics through dual-layer LSTM and models long-range dependencies via a Transformer, thereby effectively adapting to varying data volumes. To further enhance prediction robustness under data sparsity, we introduce a multi-user historical association prediction (MHAP) mechanism that leverages inter-user workload correlations. Building on these components, the prediction model is incorporated into an energy-aware scheduling strategy based on EASY Backfilling, which utilizes workload submissions and power logs to facilitate dynamic scheduling under power constraints. Finally, experiments on real traces from the Marconi 100 supercomputer show that, compared to the existing History method, our approach reduces MAPE and MAE by 22.6% and 20.1%, respectively. Moreover, the integrated strategy decreases power under-utilization by 13.4% and average turnaround time by 17.6%, significantly enhancing energy efficiency without compromising performance. These findings offer valuable insights and practical guidance for advancing energy management in future HPC systems.