Accurate fuel consumption prediction is critical for balancing flight safety, routine operational efficiency, and emergency decision-making. This study proposes Adap-Informer, an adaptive time-series framework for aircraft fuel prediction, integrating two core components: Bayesian optimization to identify optimal multi-input–output sequence length combinations and pretrain phase-specific models, and real-time true-value-based dynamic model selection for full-flight progressive prediction. Its key innovations are adaptive multi-sequence-length switching and real-time data-driven model compatibility evaluation. As true data accumulate, the prediction error converges, with MAE decreasing from 0.12 to 0.058, translating to a practical fuel error of no more than 1500 kg—this is far lower than airlines’ typical 2000–5000 kg of redundant fuel. This cuts unnecessary loading in normal flights, reducing fuel burn, costs, and emissions. Unlike conventional fixed-sequence models with poor real-time adaptability, Adap-Informer fuses real-time data with multiscale pretrained models: it optimizes routine fuel use while providing high-confidence support for emergency decisions such as diversion and urgent refueling, balancing efficiency and safety.
Yanxiong et al. (Thu,) studied this question.