Optimal Takeoff Trajectory Prediction of Electric Drones Based on a Fully Automated Optimal Experimental Design Method

Electric vertical takeoff and landing (eVTOL) aircraft is attracting great interest as a viable solution to promote urban aerial mobility with promising flexibility as well as emission reductions. However, the low specific energy of the current battery is still a strong constraint on the range and endurance of eVTOLflights, especially considering the significant power demands during the takeoff process. Engineering design optimization permits promising solutions for the minimum takeoff energy consumption but can be computationally intensive due to iteratively evaluating simulation models. Surrogate-based design optimization is efficient but still relies on optimization iterations which prohibit real-time decision-making. To fill this gap, we introduce an inverse mapping concept to optimal takeoff trajectory design of eVTOL aircraft. The inverse mapping means that trained surrogate models directly predict optimal takeoff control profiles based on given design requirements (i.e., flight conditions and design constraints). A potential challenge of inverse mapping is that each training sample costs a simulation-based design optimization, which can make the training cost excessive. Thus, we implement a fully automated optimal experiment design (OED) based proper orthogonal decomposition (POD) for intelligent training data acquisition. Specifically, the full automated OED integrates two individual OED strategies (i.e., POD basis improvement and POD coefficient improvement) and automatically switches between these two strategies based on a potential metric. We demonstrated this approach on optimal takeoff trajectory design of the Airbus ��3 Vahana and comparedtheperformanceagainstsinglestrategyOEDandrandomsampling. Resultsexhibited that the automatedOED-baseddeepneuralnetworksurrogatesconsistentlyoutperformedsingle strategy OED-based and random sampling-based surrogates, given the same computational budget and neural architecture. Specifically, the fully automated OED-based surrogate models achieved over 99.5% accuracy for predicting optimal trajectories of electrical power and wing angle, as well as total takeoff time, using around 300 training samples, whereas the single strategy OED-based and random sampling-based surrogates cannot reach that accuracy level with 400 training samples.