Page 835 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

1. Sequential Flow:

a. Input: Start and end GPS coordinates.[15]
b. Route Planning: Mapbox API determines the optimal route.
c. Terrain Analysis: The OpenTopography API provides elevation and curvature data. Transport 4.10: Intelligent
d. Energy Forecasting: Chronos predicts energy consumption and optimal speed/RPM for
upcoming segments.
e. Behavior Adjustment: Mistral generates natural language instructions based on
predictions and driver behavior data.
2. Hierarchical Structure:

a. Base Layer: Data acquisition and preprocessing.
b. Middle Layer: Predictive modeling (Chronos). [16]
c. Top Layer: Decision-making and instruction generation (Mistral).

Terrain-Aware Processing: Translating Topographical Data
Data Source: OpenTopography provides high-resolution elevation data. [17]

Processing Method:
• Rule-Based Functions: We apply domain-specific rules to translate elevation and curvature
data into energy coefficients.
• GIS Data Integration: Geospatial data is integrated to enhance the accuracy of terrain
analysis.

Outcome: This approach allows for accurate estimation of energy requirements based on
terrain, improving the reliability of our energy forecasts.

Real-Time Performance on Edge Devices

Deployment Platforms: NVIDIA Jetson Nano and Xavier.[18]
Optimization Techniques:

• TensorRT-LLM: Utilized to optimize Mistral's inference performance on NVIDIA hardware.
[19]

Performance Metrics:
• Latency: Mistral 7 B-Instruct-v0.1 achieves a latency of approximately 10.2 seconds per
inference on Jetson AGX Orin. [20]
• Power Consumption: Jetson AGX Orin operates within a configurable power range of
15W to 50W, suitable for edge deployments.[20]

Evaluation Metrics

Chronos:

• RMSE: 0.001443 [21]
• MAE: 0.001105 [21]
Mistral:

• Latency: Approximately 10.2 seconds per inference on Jetson AGX Orin.[22]

799