LSTM-Based Battery Level Prediction Using Flight Data from a Glider Unmanned Aerial Vehicle

Gliders are unpowered aerial platforms that sustain flight by exploiting atmospheric phenomena such as thermal updrafts and wind currents. Although they descend under gravity, they can extend flight duration and range by efficiently harnessing ambient airflow. Glider-type Unmanned Aerial Vehicles (UAVs) emulate these principles to achieve prolonged endurance with reduced reliance on propulsion. In this study, flight data obtained from a glider-type UAV were systematically collected, processed and analysed to evaluate its operational performance. The dataset includes airspeed, altitude, acceleration, position, angular velocity, wind conditions and battery status, enabling a holistic assessment of the UAV's behaviour under varying atmospheric and operational conditions. The integrity, reliability and suitability of the data for modelling were also evaluated. Using these data, flight performance was modelled with a Long Short-Term Memory (LSTM) network, and dynamic flight variables were used to estimate the battery level during flight. The LSTM model reached R² = 0.99, RMSE = 0.011 and MAE = 0.008 on the training partition, and R² = 0.24, RMSE = 0.148 and MAE = 0.143 on the chronologically separated test partition. All errors are expressed in min–max normalized units, corresponding to approximately 0.5% and 7.4% of the battery level range, respectively. The model closely reproduces the overall discharge behaviour of the battery, while its accuracy decreases in the strongly oscillatory final segment of the flight. By integrating real flight data with predictive modelling, the study bridges theoretical methods and practical UAV operations, offering insights for energy management, flight planning and performance analysis of glider-type UAVs.