Trajectory-Based State-of-Charge Prediction Using LSTM Recurrent Neural Networks

Planning and execution of a UAV operation requires properly monitoring and forecasting available power over the complete flight plan. This is especially true in the case of delivery operations where payloads increase power requirements as a battery discharges. In this paper, we document the development and application of a long short-term memory (LSTM) recurrent neural network to predict UAV battery state-of-charge levels as a function the UAV’s flight trajectory. The model is developed and evaluated using over 134,000 achieved real-world UAV flights. Beyond trajectory information, meta-data regarding the aircraft model type, prior discharges, and battery health is integrated into the LSTM model. Through validation exercises we demonstrate that the resulting LSTM model is able to significantly outperform linear models in predicting battery state-of-charge.