A physics-informed causal learning framework for power consumption of logistics drones

Accurate modeling and causal analysis of logistics drone power consumption are essential for energy-efficient operations and mission planning. This study develops an integrated framework that combines high-accuracy power prediction with causal effect estimation using logistics drone data in Shenzhen, China. We propose a Transformer with Physics-Informed Manifold Mixup and Mode-Stratified Self-Distillation (PM2D-Transformer). This framework improves power prediction by generating physically consistent mixed samples within each flight mode, thereby reducing the influence of sparse and heterogeneous training data. Embedded within a Double/Debiased Machine Learning module, the model accurately estimates the heterogeneous causal effects of key variables on power consumption. Results demonstrate superior predictive performance relative to baseline models, while revealing complex causal patterns. The framework provides actionable insights for energy-aware flight control, operational scheduling and mission optimization.