LSTM‐KAN‐Based Energy Consumption Prediction and Optimization Method for Quadcopter Drone Under Unknown Load

With the rapid expansion of drones in last‐mile delivery, high energy consumption has become a critical challenge, affecting both operational efficiency and sustainability. Among the various factors influencing unmanned aerial vehicle (UAV) energy consumption, payload weight plays a significant role in determining power usage. However, most existing studies assume known payload conditions, while in real‐world applications, payload is often unknown, limiting the accuracy of energy consumption prediction and optimization. To address this issue, this study proposes a long short‐term memory (LSTM)‐Kolmogorov–Arnold network (KAN) energy consumption modeling approach, integrating deep learning with prior knowledge of quadrotor dynamics to enhance UAV energy prediction under unknown load conditions. The proposed framework consists of a load prediction model based on an improved LSTM network and an energy consumption model employing the KAN for power estimation, effectively capturing the impact of payload variations on energy consumption. Additionally, a particle swarm optimization (PSO) algorithm is applied to dynamically adjust flight parameters, further improving energy efficiency. Experimental validation on a customized quadrotor platform demonstrates that the proposed model achieves R 2 values of 0.9852 and 0.9235 for load and power prediction, respectively, indicating high predictive accuracy. When combined with PSO, the approach enhances energy efficiency by 23.14%–28.21% under various payload conditions. This study provides an effective solution for UAV energy consumption prediction and optimization, contributing to improved operational efficiency and the sustainability of drone‐based logistics.