Probabilistic Prediction of Energy Demand and Driving Range for Electric Vehicles with Federated Learning

Connected vehicles and backend infrastructures comprise a distributed system, in which large amounts of data are generated. Machine learning algorithms can use this data to improve predictions and forecasts of future events. However, the distribution of the data is a challenge.
Today's drivers of battery electric vehicles must deal with limited driving range in a sparse charging infrastructure. An accurate prediction of energy demand and driving range is therefore important and enables reliable routing and charge planning applications.

Machine learning algorithms use a large amount of data and have high computational requirements. A traditional placement of the software within a vehicle's electronic control unit (ECU) could lead to high latencies and thus detrimental to user experience. High latencies can be prevented with intelligent distribution of the algorithm parts over the vehicle fleet and backend. A distributed system should take the uncertainty of data and predictions into account. Predictive uncertainty can be regarded directly with the use of probabilistic machine learning algorithms.

In this dissertation, distributed and probabilistic prediction algorithms as well as system architectures for distributed systems are investigated. ... mehrDifferent system architectures and module placements are analyzed in terms of latency, network usage, energy usage and cost.
Federated learning algorithms such as FedAvg can be applied in this distributed setting, but predictive uncertainty is typically not considered. In this dissertation, an extension of the federated averaging algorithm, FedAvg-Gaussian, is applied to train probabilistic neural networks and linear regression models in a communication-efficient and privacy-preserving manner. The prediction algorithms are validated using real driving data.

The results show that a system distributed between the vehicle and the cloud reduces the end-to-end latency by up to a factor of 10, when compared to a traditional vehicle-based placement. Likewise, the network usage is decreased to the same degree.
The performance advantage of probabilistic prediction models over deterministic prediction models is demonstrated using proper scoring rules and it is shown that federated learning can improve the standard, driver-individual learning.
Using probabilistic predictions, routing and charge planning based on destination attainability can be applied. Using a simulation framework with real road and charging infrastructure, it is shown that accurate probabilistic predictions lead to increased effective driving range and reduced travel time.