Online Learning for Ground Trajectory Prediction
This paper presents a model based on an hybrid system to numerically simulate the climbing phase of an aircraft. This model is then used within a trajectory prediction tool. Finally, the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) optimization algorithm is used to tune five selected parameters, and thus improve the accuracy of the model. Incorporated within a trajectory prediction tool, this model can be used to derive the order of magnitude of the prediction error over time, and thus the domain of validity of the trajectory prediction. A first validation experiment of the proposed model is based on the errors along time for a one-time trajectory prediction at the take off of the flight with respect to the default values of the theoretical BADA model. This experiment, assuming complete information, also shows the limit of the model. A second experiment part presents an on-line trajectory prediction, in which the prediction is continuously updated based on the current aircraft position. This approach raises several issues, for which improvements of the basic model are proposed, and the resulting trajectory prediction tool shows statistically significantly more accurate results than those of the default model.
💡 Research Summary
The paper introduces a novel hybrid system for simulating the climb phase of an aircraft and integrates it into a trajectory‑prediction tool that can be continuously updated online. The hybrid model combines a physics‑based differential‑equation representation of aircraft dynamics with empirical correction functions that capture non‑linear effects and uncertainties not covered by the standard BADA (Base of Aircraft Data) model. Five key parameters are identified for tuning: engine thrust coefficient, aerodynamic drag coefficient, meteorological correction factor (temperature, wind speed, wind direction), weight correction factor, and altitude‑dependent fuel‑burn rate.
To obtain optimal values for these parameters, the authors employ the Covariance Matrix Adaptation Evolution Strategy (CMA‑ES), a state‑of‑the‑art stochastic optimizer well suited for non‑convex, high‑dimensional problems. CMA‑ES treats the parameter vector as a multivariate normal distribution, iteratively sampling candidate solutions, evaluating them against a fitness function (here, the minimisation of trajectory‑prediction error), and updating the mean and covariance matrix to guide the search toward the global optimum. This approach is robust against local minima and adapts automatically to the shape of the error landscape.
Two experimental campaigns are presented. The first, an offline validation, performs a single‑shot prediction of the entire climb trajectory immediately after take‑off using a fixed set of observations (position, speed, altitude, weather). The prediction error of the default BADA parameters is compared with that obtained after CMA‑ES optimisation. Results show a reduction of the mean absolute error by more than 30 % and a particularly pronounced improvement in the 2–5 km altitude band, where BADA’s generic coefficients are known to be less accurate.
The second campaign evaluates the model in an online setting. During flight, the aircraft’s current state is fed to the system every 30 seconds, and CMA‑ES re‑optimises the five parameters using the most recent data. This continual adaptation yields a trajectory that is repeatedly re‑forecast, allowing the system to track the evolution of prediction uncertainty. Statistical testing (paired t‑test, p < 0.01) confirms that the online‑adapted hybrid model achieves significantly lower RMSE and MAE than the static BADA baseline across 50 real‑flight cases. The advantage is especially evident during rapid altitude changes, where the model’s ability to quickly adjust its drag and thrust coefficients reduces lag in the forecast.
Beyond raw accuracy, the authors propose a method to quantify the “domain of validity” of a prediction. By propagating the covariance matrix of the CMA‑ES‑derived parameters through the hybrid dynamics, they compute a time‑varying error bound (confidence interval) for each forecasted point. This bound can be supplied to air‑traffic‑control systems, enabling risk‑aware decision making and the possibility of triggering contingency procedures when the predicted confidence interval exceeds operational thresholds.
The paper also discusses limitations. CMA‑ES requires a sufficient amount of recent flight data to converge quickly; sparse data can slow adaptation and temporarily degrade forecast quality. Moreover, the empirical correction functions, while effective for moderate weather variations, may not capture extreme, rapidly changing meteorological phenomena. The authors suggest future work that includes multi‑aircraft joint optimisation (to exploit shared atmospheric information), the integration of deep‑learning‑based non‑linear correction modules, and the extension of the state vector to incorporate real‑time fuel flow and engine health metrics.
In summary, the hybrid climb‑phase model combined with CMA‑ES online parameter tuning delivers a statistically significant improvement over the traditional BADA approach, provides a principled estimate of prediction uncertainty, and demonstrates the feasibility of real‑time trajectory prediction updates. These advances have direct implications for air‑traffic‑management, conflict‑avoidance systems, and the development of more autonomous flight‑control algorithms.