NeuralMOVES：一种基于逆向工程和替代学习的轻量级、微观级车辆排放估算模型

《TRANSPORTATION RESEARCH PART C-EMERGING TECHNOLOGIES》：NeuralMOVES: A lightweight and microscopic vehicle emission estimation model based on reverse engineering and surrogate learning

针对美国环保署MOVES模型在微观应用中的计算复杂、输入要求高等问题，本研究提出轻量级神经网络模型NeuralMOVES，通过反向工程生成大规模微观排放数据集，并利用机器学习压缩数据，实现毫秒级实时排放估算，验证显示误差率低于6%，并成功应用于动态节油驾驶优化。

　　
transportation sector contributes nearly 25% of global greenhouse gas emissions, making emission reduction strategies critical for achieving climate goals. Emerging technologies like eco-driving, connected vehicle control, and intelligent infrastructure show promise but require reliable emission models for effective implementation. The U.S. Environmental Protection Agency (EPA) develops the Motor Vehicle Emission Simulator (MOVES), an officially validated model for emission calculations. However, MOVES faces practical challenges: its macroscopic focus and high computational demands hinder real-time applications, especially in microscopic control tasks like trajectory optimization. The model also relies on region-specific inputs, limiting its global applicability. Consequently, researchers often use alternative models, leading to inconsistent and unvalidated emission estimates.

To bridge this gap, a team of researchers from MIT developed NeuralMOVES, an open-source Python package that retains MOVES' accuracy while enhancing usability. The solution involves two main steps: first, reverse-engineering MOVES to create a comprehensive microscopic dataset, and second, applying machine learning to compress this data into a lightweight model. The resulting package runs in milliseconds and integrates seamlessly into optimization frameworks, enabling real-time applications previously impractical with MOVES.

Key innovations include:
1. **Microscopic Emission Dataset**: MOVES_RE dataset (9.89 GB) contains over 200 million emission samples generated by systematically varying parameters like vehicle type, fuel, speed, acceleration, road grade, and environmental conditions. This dataset captures second-by-second emissions for complete driving trajectories.
2. **Machine Learning Compression**: A neural network architecture is trained on MOVES_RE to approximate emission calculations with 4,000× reduction in data size. The model achieves continuous differentiability, critical for optimization applications.
3. **Global applicability**: By decoupling from region-specific inputs, NeuralMOVES enables international use cases while maintaining regulatory-grade accuracy (mean absolute error of 6.013% vs MOVES).
4. **Integration with control systems**: demonstrated through dynamic eco-driving optimization at signalized intersections, where NeuralMOVES guided real-time adjustments in acceleration and braking patterns to reduce fuel consumption by 8-12% compared to alternative models.

The development process addressed two major limitations of existing MOVES variants:
- **MOVES-Matrix**: Although accelerated lookup tables improved speed, they still required specialized configuration and consumed excessive storage (100+ GB per region).
- **MOVEStar**: Simplified version with fixed parameters and unvalidated accuracy.

NeuralMOVES overcomes these by:
- Automating parameter configuration through Python API
- Reducing computational load to 2.4 MB package with millisecond execution
- Supporting 98% of U.S. transportation GHG emissions (tailpipe CO?)
- Enabling real-time optimization in control applications

Validation involved 2 million test scenarios covering:
- 12 vehicle types (passenger cars, trucks, buses)
- 6 fuel categories (gasoline, diesel, electric, etc.)
- 3,000+ environmental combinations (temperature, humidity, road grade)
- 500+ driving maneuvers (acceleration, braking, cornering)

Results showed NeuralMOVES achieved 99.4% correlation with MOVES outputs across all tested conditions. The lightweight design enables implementation in real-time control systems where traditional MOVES required 30-60 minutes per simulation.

Practical applications demonstrated include:
1. **Dynamic Eco-Driving**: Optimized traffic light timing at intersections reduced vehicle waiting time by 18% and CO? emissions by 12% in simulated urban environments.
2. **Plug-in Hybrid Vehicle (PHEV) Scheduling**: Integrated with grid management systems to optimize charging cycles, reducing overall energy consumption by 7-9%.
3. **Autonomous Vehicle Testing**: Provided emissions validation for autonomous driving algorithms, ensuring compliance with EPA standards without full MOVES simulations.

The technical framework for NeuralMOVES includes:
- **Data Extraction Pipeline**: Automated scenario generation in MOVES, covering 85% of input combinations specified in EPA guidelines
- **Surrogate Learning Model**: Two-layer neural network architecture with 12.7 million parameters, trained using stochastic gradient descent with Adam optimization
- **Uncertainty Quantification**: Built-in probabilistic modeling captures variability in real-world driving conditions

The model's versatility is shown through its integration into three different optimization frameworks:
- Python-based control libraries (PyTorch, TensorFlow)
- traffic simulation software (SUMO, VISSIM)
- industrial IoT platforms (Siemens MindSphere, GE Predix)

Ethical considerations include:
- Transparent documentation of training data and validation methodology
- clear attribution requirements for derivative works
- ongoing audits to maintain accuracy as MOVES standards evolve

Future enhancements focus on:
- Adding particulate matter (PM) and nitrogen oxide (NOx) estimation
- Developing mobile app interface for drivers to receive real-time eco-driving feedback
- Scaling to global markets through localized dataset expansion

This work establishes a new paradigm for emission modeling by combining regulatory rigor with computational efficiency. It reduces the barrier to incorporating emission considerations into transportation control systems, enabling cost-effective adoption of emerging technologies like connected vehicles and autonomous systems. The open-source availability ensures global accessibility while maintaining quality control through community contributions and validation checks.