[Paper Review] A Deep Learning Approach for Macroscopic Energy Consumption Prediction with Microscopic Quality for Electric Vehicles

Ayman Moawad, Krishna Murthy Gurumurthy+5 more|arXiv (Cornell University)|Nov 24, 2021

Transportation Planning and Optimization4 citations

TL;DR

This paper proposes a deep learning model that predicts macroscopic energy consumption for electric vehicles using only high-level route data—without requiring microscopic speed profiles—by leveraging a large-scale, high-fidelity simulation dataset. Despite masking detailed vehicle dynamics, the model achieves high accuracy (R² > 0.98) and low inference latency (≤50 ms per prediction), enabling real-time integration into large-scale transportation simulations like POLARIS.

ABSTRACT

This paper presents a machine learning approach to model the electric consumption of electric vehicles at macroscopic level, i.e., in the absence of a speed profile, while preserving microscopic level accuracy. For this work, we leveraged a high-performance, agent-based transportation tool to model trips that occur in the Greater Chicago region under various scenario changes, along with physics-based modeling and simulation tools to provide high-fidelity energy consumption values. The generated results constitute a very large dataset of vehicle-route energy outcomes that capture variability in vehicle and routing setting, and in which high-fidelity time series of vehicle speed dynamics is masked. We show that although all internal dynamics that affect energy consumption are masked, it is possible to learn aggregate-level energy consumption values quite accurately with a deep learning approach. When large-scale data is available, and with carefully tailored feature engineering, a well-designed model can overcome and retrieve latent information. This model has been deployed and integrated within POLARIS Transportation System Simulation Tool to support real-time behavioral transportation models for individual charging decision-making, and rerouting of electric vehicles.

Motivation & Objective

• To enable accurate macroscopic energy consumption prediction for electric vehicles using only route-level inputs, without requiring detailed speed profiles.
• To overcome the challenge of latent information loss by training deep learning models on large-scale, high-fidelity simulation data where microscopic dynamics are masked.
• To support real-time behavioral modeling in large-scale transportation systems by integrating a lightweight, scalable energy prediction model into POLARIS.
• To enable improved forecasting of charging demand, routing decisions, and grid impact under diverse driving behaviors and scenario conditions.
• To demonstrate that deep learning can retrieve meaningful energy consumption patterns from aggregate-level data even when internal dynamics are obscured.

Proposed method

• The study uses the POLARIS agent-based simulation framework to model trips across the Greater Chicago region under diverse scenarios, including varying vehicle types, driving behaviors, and charging patterns.
• High-fidelity energy consumption values are generated using the Autonomie physics-based simulation tool, producing a large-scale dataset of vehicle-route energy outcomes.
• Microscopic speed dynamics are masked from the input data, leaving only macroscopic route-level features such as link length, average speed, and road type.
• A deep learning model combining convolutional and recurrent neural network architectures is trained on the masked dataset to predict link-level energy consumption.
• Feature engineering is carefully tailored to encode relevant physical and behavioral factors (e.g., vehicle class, battery capacity, traffic conditions) into the input representation.
• The model is deployed within POLARIS to support real-time energy and SOC prediction with minimal performance overhead (≤50 ms per prediction).

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict link-level energy consumption for electric vehicles using only macroscopic route inputs, without access to microscopic speed profiles?
• RQ2 To what extent can a model trained on masked internal dynamics recover latent physical relationships that govern energy consumption?
• RQ3 How does the model’s performance scale across diverse vehicle types, traffic conditions, and scenario variations in a large urban region?
• RQ4 What is the computational efficiency of the model when deployed in real-time transportation simulations at scale?
• RQ5 Can such a model be effectively integrated into large-scale agent-based systems like POLARIS to support dynamic charging and routing decisions?

Key findings

• The deep learning model achieves an R² score of over 0.98 in predicting link-level energy consumption, demonstrating high accuracy despite the absence of microscopic speed data.
• The model maintains low inference latency, with each prediction taking ≤50 ms, enabling real-time deployment in large-scale simulations.
• The model successfully retrieves latent physical relationships from aggregate data, outperforming traditional methods that rely on detailed speed profiles in similar settings.
• The integration of the model into POLARIS maintains high simulation performance, with minimal impact on overall runtime and scalability.
• The approach enables accurate forecasting of regional charging demand and SOC depletion, supporting applications in eco-routing, fleet management, and grid planning.
• The results suggest that large, diverse, and well-structured simulation datasets can compensate for missing microscopic data in deep learning-based energy prediction.