Technical Summary: Green-NAS for Edge-Native Weather Forecasting

Problem Statement

The paper addresses the critical need for accurate, short-term weather forecasting in low-resource environments, particularly within the Global South. While Numerical Weather Prediction (NWP) and recent deep learning "foundation" models (e.g., GraphCast) offer high performance, they face two primary barriers to widespread adoption in developing regions:

1. Computational Cost: Large models with hundreds of millions of parameters require high-end GPUs, making them unsuitable for distributed edge networks or IoT-based climate monitoring.
2. Architectural Rigidity: Models hand-crafted for specific climates (e.g., European) often fail to capture the rapid convective processes of tropical belts.
3. Environmental Impact: The energy consumption and carbon footprint of training and deploying massive models contradict the principles of sustainable "Green AI."

Additionally, many developing nations suffer from limited historical weather data, hindering the training of robust local models from scratch.

Methodology

The authors propose Green-NAS, a multi-objective Neural Architecture Search (NAS) framework designed to automatically generate lightweight, accurate, and energy-efficient forecasting models.

1. Search Space and Optimization

The framework utilizes NSGA-II (Non-dominated Sorting Genetic Algorithm II) to search for optimal architectures balancing three conflicting objectives:

• Minimize Validation RMSE: Maximizing predictive accuracy.
• Minimize Parameter Count: Reducing model size for edge deployment.
• Maximize Interpretability: Proxied by minimizing network depth (favoring simpler models).

The search space includes combinations of:

• Recurrent Layers: LSTM and GRU for capturing temporal dependencies.
• Convolutional Layers: 1D CNNs for extracting local short-term features.
• Attention Mechanisms: Multi-head attention for global relationships (though results showed these were often inferior for this specific task).
• Dense Layers: MLPs for non-linear transformation.

The search was conducted over 10 generations with a population size of 20, evaluating 114 total architectures.

2. Dataset and Experimental Setup

• Data: 1,072,693 hourly samples from 24 cities across six Köppen climate zones (Tropical, Arid, Temperate, Continental) spanning 2019–2024.
• Split: 18 cities served as the source (training) set, and 6 distinct cities served as the target (test) set to evaluate generalization.
• Features: Temperature, humidity, precipitation, pressure, cloud cover, wind speed/direction, and short-wave radiation.
• Baseline: Comparisons were made against persistence, climatology, manually tuned DL models (LSTM/GRU/Transformer), and the state-of-the-art GraphCast.

3. Robustness and Transfer Learning

• Conformal Prediction: Split Conformal Prediction was integrated to generate rigorous 95% confidence intervals, ensuring reliability for safety-critical applications.
• Transfer Learning: A methodology was developed to pre-train models on the rich source dataset and fine-tune them on sparse target datasets, addressing data scarcity in specific regions.

Key Contributions and Results

1. Discovery of Pareto-Optimal Architectures

The NSGA-II search identified a Pareto front of 20 architectures. Three representative models were highlighted:

• Green-NAS-A (High Accuracy): A 2-layer GRU architecture with 153,096 parameters. It achieved an RMSE of 0.0988, within 1.4% of the best manually tuned Transformer baseline (0.0974) and significantly outperforming standard baselines.
• Green-NAS-B (Balanced): A single-layer CNN with 4,232 parameters and an RMSE of 0.0996, offering a balance between accuracy and efficiency.
• Green-NAS-C (High Efficiency): An ultra-compact CNN with only 1,064 parameters and an RMSE of 0.1019. This model is 36 times smaller than Green-NAS-A and suitable for microcontrollers.

Comparison with State-of-the-Art:
While GraphCast (36.7M parameters) solves global gridded forecasting, Green-NAS models solve local point forecasting with drastically fewer parameters (Green-NAS-A is 239x smaller; Green-NAS-C is 35,500x smaller).

2. Efficiency and Edge Deployment

• Inference Latency: All models achieved sub-millisecond inference on a standard CPU (Green-NAS-A: 0.52ms; Green-NAS-C: 0.33ms).
• Model Size: Green-NAS-C fits within ~4 KB, making it deployable on ultra-low-power IoT sensors.
• Uncertainty Quantification: All models achieved near-target coverage (93.9%–94.9% for 95% target) with narrow confidence intervals, indicating well-calibrated uncertainty estimates.

3. Transfer Learning Efficacy

The study demonstrated that pre-training on source cities significantly improves performance on target cities, even when target data is abundant.

• Performance Gain: Transfer learning improved accuracy by 5.2% compared to training from scratch on 100% of the target data.
• Data Scarcity: With only 1% of the target data available, transfer learning still yielded a 5.7% improvement, enabling accurate forecasting in new cities with minimal historical data.

4. Architectural Insights

• Temporal Dominance: Pure recurrent models (GRUs) outperformed hybrid CNN-LSTM designs and attention-based mechanisms for 24-hour look-back windows.
• Feature Importance: Permutation importance analysis confirmed that Temperature and Surface Pressure are dominant predictors, aligning with meteorological principles. Notably, the NAS-discovered GRU successfully utilized Wind Direction, a complex cyclic feature often ignored by simpler models.

Significance and Claims

The paper claims that Green-NAS successfully bridges the gap between high-performance weather forecasting and the resource constraints of the Global South. Its significance lies in three fundamental insights:

1. Automated Search Viability: Automated NAS methods can produce models competitive with, or superior to, manually designed architectures while optimizing for efficiency.
2. Efficiency without Sacrifice: Sub-millisecond inference and ultra-compact model sizes (down to 1,064 parameters) are achievable without significantly compromising accuracy.
3. Global Knowledge Transfer: Weather physics are universal enough that knowledge learned from data-rich regions can be effectively transferred to data-sparse regions, democratizing access to climate services.

The authors position Green-NAS as a step toward "equitable, edge-native climate AI," enabling real-time, private, and sustainable weather forecasting directly on IoT devices without reliance on centralized cloud infrastructure.