Abstract
This paper introduces a data-driven Model Predictive Control (MPC) framework designed to manage uncertainty in human-driven systems. By employing a modified k-nearest neighbor (k-NN) regressor, the system generates a weighted distribution of potential future power demands rather than a single deterministic prediction. This approach allows the controller to optimize energy management in electric vehicles equipped with heterogeneous energy storage systems. Experimental results using real-world driving data demonstrate that this probabilistic approach significantly outperforms standard mean-based predictors, capturing over half of the performance gap relative to a theoretical prescient oracle.
Introduction
Subsystem controllers in modern technology, such as those managing electric vehicle (EV) power systems, must often operate based on unpredictable human inputs. While human behavior follows certain patterns, it remains a "black box" to low-level controllers. For instance, an energy management system cannot definitively know if a driver is about to accelerate aggressively or decelerate for a turn. Standard Model Predictive Control (MPC) typically relies on a single "best guess" prediction. However, if this deterministic prediction is inaccurate, the resulting control strategy can lead to significant inefficiencies or hardware stress. This research addresses the challenge of building a controller that acknowledges uncertainty and "hedges its bets" against multiple likely future scenarios.
System Architecture and Application
The proposed framework is applied to a heterogeneous energy pack consisting of a high-energy battery and a high-power supercapacitor. The primary goal is to use the supercapacitor to buffer high-power bursts—which are detrimental to battery health and efficiency—while the battery provides the sustained energy needed for the trip.
As shown in Figure 1, the system utilizes an Intelligent Charge Controller and a Regenerative Controller to manage the flow of energy between the storage components and the motor.
Methods: Stochastic Data-Driven MPC
The core innovation of this work is a stochastic, data-driven MPC approach that models "upcoming duty" (power demand) as a probability distribution.
Modified k-Nearest Neighbor (k-NN) Predictor
Instead of complex parametric models, the authors utilize a non-parametric k-NN approach. The predictor identifies the $k$ most similar historical states based on features like GPS coordinates, velocity, and acceleration.
Weighted Distribution and Optimization
Rather than averaging these neighbors into a single mean prediction—which can lead to the "mean trap" where the average of two distinct possibilities results in an impossible middle ground—the model treats them as a set of weighted samples. The controller evaluates potential control sequences against all $k$ samples, selecting the sequence that minimizes the expected cost across the entire distribution.
Figure 2 illustrates how the predictor generates a variety of potential future energy trajectories, allowing the MPC to account for a range of likely outcomes.
Experimental Results
The framework was evaluated using a real-world dataset of seven drivers over one year. Power profiles were derived from GPS and timestamp data using a longitudinal vehicle model. Performance was measured by the reduction in "current-squared" on the battery, a proxy for thermal stress.
The results demonstrate the superiority of the probabilistic approach:
- Mean KNN (Standard): 47.8% reduction.
- Distributed KNN (Proposed): 51.9% reduction.
- Prescient Oracle: 55.9% reduction.
The proposed Distributed KNN captured over half of the remaining performance gap between standard data-driven methods and the theoretical upper bound of a perfect future-knowledge oracle.
Conclusion
By shifting from deterministic to probabilistic predictions, this MPC framework provides a more robust solution for managing human-driven systems. While challenges remain regarding computational scaling for large datasets and performance in data-sparse regions, the use of semantic-free learning and local gradient descent for feature weighting offers a promising path forward for intelligent vehicle energy management.
Summary updated