Improving algorithms for predicting electric vehicle energy consumption to accurately estimate power reserve based on real terrain parameters and current meteorological factors

Background. Accurate forecasting of the energy consumption of electric vehicles is a critically important task for improving the efficiency of vehicle operation and reducing drivers’ anxiety about power reserve. Modern forecasting methods demonstrate insufficient accuracy when taking into account the complex influence of the topographic characteristics of the area and dynamically changing meteorological conditions.

Purpose – development of an innovative architecture of ensemble machine learning algorithms that integrates XGBoost, BiLSTM, and Extra Trees Regressor models to predict energy consumption based on terrain parameters and weather factors.

Materials and methods. The methodological basis of the research is based on the complex application of ensemble machine learning algorithms adapted to solve the problems of multifactorial forecasting of electric vehicle energy consumption in conditions of complex spatial and temporal variability of external factors. The choice of methods is due to the need to process heterogeneous high-dimensional data and ensure the robustness of forecasts in the presence of noise and omissions in the source data. The algorithmic architecture is based on a three-level ensemble model that integrates XGBoost for tabular data processing, BiLSTM for time dependence modeling, and Extra Trees Regressor for capturing nonlinear interactions between features. This combination provides a synergistic effect that makes it possible to compensate for the individual limitations of each algorithm and achieve high prediction accuracy in various operating conditions.

Results. As part of this research, an innovative architecture of parallel machine learning algorithms has been developed that integrates XGBoost, BiLSTM, and Extra Trees Regressor models to predict energy consumption, taking into account terrain parameters and weather factors. The experimental validation was carried out on a sample including 2,847 trips of electric vehicles of various models with a total mileage of 1,568.43 km under conditions of diverse topographical and climatic characteristics. The proposed hybrid model achieves an average absolute error of 4.2 kWh/100 km and a termination coefficient of R2 = 0.971, which exceeds the basic algorithms by 23.8%. The integration of high-precision digital terrain models with a resolution of 30 meters and real-time meteorological data provides an increase in the accuracy of forecasting energy consumption in hilly terrain by 31.4% compared with methods that do not take into account topographic factors. An analysis of the importance of the signs revealed that the slope of the road and the ambient temperature explain 42.6% and 18.3% of the variance in energy consumption, respectively. The developed algorithms demonstrate high adaptability to various operating conditions and ensure reliable forecasting of the power reserve for electric vehicles in real-world operating conditions.