Real-Time Predictive Optimization for Energy Management in a Hybrid Electric Vehicle

Alexander David Styler, Illah Reza NourbakhshAAAI 2015 · DOI

TL;DR

This paper introduces a real-time predictive optimization strategy for energy management in hybrid electric vehicles (HEVs). By leveraging historical driving data to predict upcoming power loads, the system achieves near-optimal energy efficiency and improved battery longevity, outperforming traditional reactive control methods.

What Problem Does This Paper Address?

The increasing adoption of hybrid and electric vehicles presents a significant opportunity to reduce energy consumption and emissions. A key challenge in optimizing energy management for these vehicles lies in effectively predicting and responding to future power demands. Conventional approaches often rely on reactive policies or pre-programmed strategies that don't adapt to real-time driving conditions, driver behavior, or evolving routes. This leads to suboptimal energy usage, increased battery wear, and reduced overall efficiency. The authors aim to bridge this gap by developing an adaptive, real-time system that learns from past driving experiences to make intelligent energy management decisions.

Methods

The core of the proposed solution is a four-step control loop: Measure, Predict, Optimize, and Execute.

  • Measure: The system continuously collects vehicle state data, including GPS, speed, acceleration, and power load.
  • Predict: Using a k-Nearest Neighbors (kNN) approach, the current vehicle state is matched to similar historical states in a dataset. The power loads that followed these historical states are then used as predictions for the upcoming load. This method leverages the assumption that current driving states are informative of future demands.
  • Optimize: The system employs a dynamic programming approach to calculate a cost-to-go function for each prediction. This function estimates the future cost (e.g., battery degradation) associated with different control actions. At each timestep, the algorithm selects the control action that minimizes the immediate cost plus the expected future cost, considering the distribution of predictions. Crucially, these cost-to-go functions are precomputed offline, allowing for rapid real-time execution.
  • Execute: The chosen optimal control action is implemented.

The system is designed to be adaptive, continuously updating its dataset with new driving data, thereby improving prediction accuracy and control performance over the vehicle's lifetime.

Datasets

The authors utilized a dataset of real-world driving trips collected from eight different petrol vehicles over a period of ten months. This data was then used to simulate longitudinal vehicle models and calculate expected power loads for electric vehicle (EV) batteries. The data was collected at a 1Hz frequency and included GPS coordinates, time, elevation, speed, acceleration, and power load.

Key Insights and Ideas

  • Leveraging Historical Data for Prediction: The fundamental insight is that past driving patterns, when matched to the current state, can provide surprisingly accurate predictions of future power demands, which is crucial for proactive energy management.
  • Offline Precomputation for Real-Time Performance: By precomputing cost-to-go functions offline, the system can achieve near-optimal control in real-time with minimal on-board computation, overcoming the computational burden of dynamic programming at runtime.
  • Adaptive Learning for Continuous Improvement: The system's ability to continuously learn and expand its dataset over the vehicle's lifetime allows it to adapt to changing routes, driver behaviors, and traffic trends, leading to progressively better performance.
  • The Importance of Prediction Uncertainty: The use of a set of weighted predictions, rather than a single average prediction, allows the system to account for uncertainty. When predictions disagree, the system adopts more conservative controls, mitigating risks associated with prediction errors.

Key Results

The proposed predictive optimization strategy achieved performance within 10% of the optimal upper bound calculated using a priori knowledge of upcoming loads. Specifically, the system achieved over 91% of the improvement possible with a prescient optimal controller when using 30 neighbors for prediction. The study also highlighted that larger supercapacitor sizes offer more resilience to prediction errors, while increasing the number of neighbors generally improves performance up to a point of diminishing returns. The dynamic learning system, which learns chronologically, performed comparably to a batch learning system that used all available data, demonstrating the effectiveness of continuous adaptation.

Limitations and Future Work

The authors acknowledge that future uncertainty and prediction error will inherently bound the performance of their approach. They also note that the performance of the dynamic system was generally slightly lower than the batch system, which had the advantage of complete data coverage.

Future work includes implementing the algorithm on a high-fidelity plug-in HEV simulator. A key area of focus will be on managing charge sustainment in plug-in HEVs, where the algorithm can optimally deplete the battery and distribute its use for charging events. The authors also suggest that this spatiotemporal load prediction and optimization approach could be applied to other domains, such as HVAC control in office buildings, where human behavior influences energy demand.

Summary updated

BibTeX
@inproceedings{styler2015real,
  author = {Alexander David Styler and Illah Reza Nourbakhsh},
  title = {{Real-Time Predictive Optimization for Energy Management in a Hybrid Electric Vehicle}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9232},
}
Figure