A data driven approach for real-world vehicle energy consumption prediction

Accurately predicting real-world vehicle energy consumption is essential for optimizing vehicle designs, enhancing energy efficiency, and developing effective energy management strategies. This paper presents a data-driven approach that utilizes machine learning techniques and a comprehensive dataset of vehicle parameters and environmental factors to create precise energy consumption prediction models. The methodology involves recording real-world vehicle data using data loggers to extract information from the CAN bus systems for ICE and hybrid electric, as well as hydrogen and battery fuel cell vehicles. Data cleaning and cycle-based analysis are employed to process the dataset for accurate energy consumption prediction. This includes cycle detection and analysis using methods from statistics and signal processing, and then pattern recognition based on these metrics. K-means clustering and t-SNE were used to influence the design of the prediction model and inform about vehicle and driver behavior, which resulted in a multi-layer perceptron regressor based on the above metrics. This novel data-driven model was able to achieve an average R² over 0.95 and unlocks a new perspective on powertrain analysis for a variety of vehicle types.