Using artificial neural networks, we are able to accurately predict fuel consumption and emissions of commercial and government fleets. Our computer modeling demonstrates the accuracy of predictive analytics to help fleets manage fuel consumption, decrease their fuel usage, and reduce emissions. The tools we are developing can take incredibly complex real-world problems and turn them into extremely accurate predictions about your fleet.

—Tyler Yadon, director of analytics for Lightning Systems

Yadon oversees Lightning Systems LightningAnalytics intelligent fleet management product. LightningAnalytics is a predictive analytics system that helps fleets monitor vehicle maintenance and track routes, thereby decreasing fuel usage and emissions. The system predicts impending maintenance repairs that are influenced by fuel usage, drive cycles and routes, and driver behavior. This is achieved through high-frequency recording of many real-time parameters from the vehicle.

Lightning Systems recently announced its LightningElectric zero-emissions package for the heavy-duty Ford Transit. The company also offers a hydraulic hybrid energy recovery system called LightningHybrid. The hydraulic hybrid system is retrofitted onto trucks, buses and other large transit and delivery vehicles, providing conventional vehicles with upgraded energy-management and powertrain-control systems. LightningAnalytics is embedded in both of these products.

Artificial neural networks are not only less computationally costly than existing simulation standards, but they are easier and faster to re-train and apply to new vehicles and drive cycles offering the potential for high accuracy estimates with reduced infrastructure requirements.

—Brian Johnston, director of emissions regulation and strategy for Lightning Systems

Johnston is one of the co-authors of a paper on the technology to be published by researchers at Colorado State University and presented at SAE World Congress in April.

The paper describes the training and testing of a time series artificial neural network (NN) using real world, on-road vehicle velocity and battery state of charge data to predict instantaneous fuel economy (FE) and emissions.

First, city and highway real world drive cycles were developed. These drive cycles were then driven in a HEV and velocity, battery state of charge, FE and emissions data were recorded. From the velocity data, FE was determined using a custom developed vehicle simulation model created in Modelica and using the Autonomie vehicle modeling software. Next, NNs were trained first using velocity and battery state of charge as inputs and FE as a target and then using velocity as an input and emissions as a target.

—Asher et al.

The reported results show that the neural network model was computationally faster and predicted FE within a mean absolute error of 0-5%. For emissions prediction, the NN model had a mean absolute error of 0-8% across CO₂, CO, and NO_x aggregate predicted concentrations.

Our research demonstrates significant benefit for designing improved vehicle-control strategies, such as eco-driving and optimal energy management. It also has the potential to reduce the need for physical vehicle testing, because this type of computer modeling accurately captures emissions results from slight drive-cycle variations and improves the understanding of real-world emissions and fuel impacts, enabling high-fidelity learning control in physical vehicles.

Our analysis predicted fuel consumption with a margin for error as low as 0.1 percent, and predicted CO₂, CO and NO_x emissions with 97 to 100 percent accuracy.

—Will Briggs, lead test engineer for Lightning Systems, and co-author

Lightning Systems used data from the operation of hydraulic hybrid fleet vehicles in the United Kingdom and the United States to create the artificial neural network predictions on fuel consumption and emissions. The company also used test-track and dynamometer testing data.

Customer drive cycles are not always reproducible for fuel consumption and emissions research. Our results indicate that artificial neural network models can be used for a variety of research applications due to their economic and computational benefits, such as improving vehicle-control strategies to reduce fuel consumption and emissions in modern vehicles.

—Tyler Yadon

Resources

• Zachary D. Asher, Abril A. Galang, Will Briggs, Brian Johnston, Thomas H. Bradley, Shantanu Jathar (2018) “Economic and Efficient Hybrid Electric Vehicle Fuel Economy and Emissions Modeling Using an Artificial Neural Network” SAE 2018-01-0315