Phase-based power prediction for quadrotor UAVs with RF-TLATT

To address the problem of low power prediction accuracy and limited adaptability of quadrotor unmanned aerial vehicles (UAVs) under complex operational conditions, this paper proposes a high-precision prediction framework. The framework integrates structural awareness with temporal modeling and is termed Random Forest-Transformer with LSTM and Attention (RF-TLATT), for Time-series. The proposed method innovatively introduces the civil aviation Climb, Cruise, Descent (CCD) phase division strategy into UAV power modeling, segmenting the flight process into four distinct phases: Ascent phase, Cruise phase, Descent phase, and Other phase. A random forest classifier is employed to automatically identify flight phases. A hybrid architecture that combines LSTM and Transformer is then used to build dedicated sub-models for each phase, utilizing multi-head attention to extract deep temporal features and capture phase-specific power patterns. Furthermore, by incorporating both flight dynamics and environmental parameters, the model demonstrates enhanced generalization and robustness in dynamic environments. Experimental results show that RF-TLATT consistently outperforms a range of recent benchmark models, achieving performance improvements ranging from 16.1 % to 58.3 % in RMSE, 18.5 %–74.1 % in MAE, 8.2 %–76.1 % in MAPE, and 12.8 %–28.2 % in R2, depending on the baseline model used for comparison. Further analysis reveals that the proposed framework effectively captures the nonlinear relationships between multi-source features and UAV power consumption under complex and highly dynamic flight conditions, enabling accurate and robust power prediction. This study provides a reliable foundation for UAV energy management, trajectory optimization, and intelligent mission scheduling.