Green-NAS: A Global-Scale Multi-Objective Neural Architecture Search for Robust and Efficient Edge-Native Weather Forecasting

AI summary

Overview

Research area: Machine learning for weather forecasting, specifically multi-objective neural architecture search (NAS), transfer learning, and edge/Green AI deployment.

Technical level: Intermediate. The paper is readable for someone with basic familiarity with neural networks and time-series forecasting, since the methods are described at a high level and the results are summarized in comparison tables.

Scope: The paper proposes Green-NAS, a multi-objective evolutionary NAS framework that searches for lightweight, accurate weather forecasting models suitable for deployment on low-power edge devices and describes how transfer learning lets those models be adapted to new cities with little historical data.

What This Paper Is About

Conventional numerical weather prediction and large deep learning forecasting models are computationally expensive, require high-end GPUs, and are often hand-crafted for particular climates, which limits their use in low-resource regions and on distributed edge or IoT hardware. Green-NAS attacks this by automatically searching for small forecasting architectures that trade off accuracy, parameter count, and interpretability, using hourly weather data from 24 cities as a case study. The goal is to show that competitive short-term local forecasts can be produced by models small enough to run on edge devices, and that knowledge learned from data-rich cities can be transferred to data-poor ones.

Key Contributions

1. Global-scale NAS: A multi-objective architecture search performed on a dataset covering 24 cities across Köppen climate zones (Tropical, Arid, Temperate, Continental, described in the abstract as six climate zones), producing a Pareto front of architectures that improve on common baseline architectures.
2. Efficiency breakthrough: NSGA-II identified a Pareto front of 20 architectures ranging from an ultra-compact CNN with 1,064 parameters and RMSE 0.1019 to a high-accuracy GRU with 153K parameters and RMSE 0.0988, each improving on the authors' manually designed baselines.
3. Rigorous validation: The discovered models were evaluated on held-out target cities to test generalization to new climate zones.
4. Edge-native robustness: A deployment assessment covering uncertainty quantification (targeting 95% coverage) and Green AI metrics, including inference latency reported as sub-0.5 ms in the contribution list and 0.30–0.52 ms in the deployment measurements.

Main Findings

• Best model accuracy and size: Green-NAS-A, a 2-layer GRU (GRU_128-GRU_128), reached RMSE 0.0988 on target cities with 153,096 parameters. The paper states this is within 1.4% of the manually tuned baseline and uses 239 times fewer parameters than GraphCast (36.7M parameters).
• Balanced model: Green-NAS-B, a single-layer CNN_128, reached RMSE 0.0996 with only 4,232 parameters, reported as the fastest model at 0.30 ms latency (0.29 ms in the comparison table) and roughly 15–17 KB in size.
• Ultra-compact model: Green-NAS-C, a CNN_32, reached RMSE 0.1019 with 1,064 parameters and a 4 KB model size, described as 36 times smaller than Green-NAS-A and using roughly 144 times fewer parameters than the manual baselines.
• Baseline comparisons: Persistence scored RMSE 0.1842 and climatology 0.1523; the Manual GRU scored 0.1004 with 153K parameters, 0.29 ms latency and 598 KB; the Manual Transformer scored 0.0974 with 135K parameters, 1.89 ms latency and 527 KB. GraphCast's RMSE is marked as not directly comparable because it solves global gridded forecasting.
• Architecture preferences: Pure recurrent designs outperformed hybrid CNN-LSTM designs, and attention mechanisms were found to be universally inferior for the 24-hour lookback window, adding parameters without accuracy gains.
• Transfer learning gains: Pre-training on 18 source cities and fine-tuning on target cities beat training from scratch at every data fraction tested: +5.7% at 1% of data (3,155 samples, 0.1083 to 0.1020, p < 10⁻⁸), +3.5% at 10% (31,550 samples, 0.1016 to 0.0980, p < 10⁻⁸), +4.7% at 50% (157,752 samples, 0.0967 to 0.0921, p < 10⁻¹⁴), and +5.2% at 100% (315,504 samples, 0.0948 to 0.0898, p < 10⁻¹²).
• Uncertainty quantification: Using split conformal prediction with a calibration set of 20% of the target city test data (63,000 samples), Green-NAS-A achieved 93.9% coverage with interval width 0.067, while Green-NAS-B and Green-NAS-C each achieved 94.9% coverage with interval widths of 0.085 and 0.092 respectively.
• Deployment metrics: Latencies were measured as 0.52 ms for Green-NAS-A, 0.30 ms for Green-NAS-B and 0.33 ms for Green-NAS-C, with model sizes of approximately 600 KB, 15 KB and 4 KB; the paper notes all models perform sub-millisecond inference and that Green-NAS-C fits on microcontrollers.
• Feature importance: Permutation importance showed Temperature and Surface Pressure as dominant predictors, with Wind Direction also used by Green-NAS-A, which the authors present as evidence that the model captures subtle multivariate dependencies.
• Multi-step degradation: Preliminary recursive multi-step tests showed RMSE rising from 0.03 at 1 hour to 0.16 at 12 hours due to error accumulation.

Methodology in Plain English

The researchers treated local weather forecasting as a regression problem: given 24 hours of past hourly observations across eight features (temperature at 2 m, relative humidity, precipitation, surface pressure, cloud cover, wind speed at 10 m, wind direction, and shortwave radiation), predict the next time step. They built a search space of building blocks — recurrent layers (LSTM/GRU), 1D convolutional layers, multi-head attention, and dense layers — allowing 1 to 4 layers, 32 to 256 units per layer, and dropout between 0.0 and 0.5.

Rather than hand-designing a model, they used NSGA-II, an evolutionary algorithm, to breed candidate architectures over a population of 20 individuals for 10 generations (114 total evaluations). Each candidate was scored on three conflicting objectives: minimizing validation RMSE, minimizing parameter count, and maximizing interpretability, which was proxied by the inverse of network depth. This produced a Pareto front of 20 architectures representing different accuracy-versus-size trade-offs.

Data came from the Open-Meteo Historical Weather API, covering January 1, 2019 to December 31, 2024 at hourly resolution, with 1,072,693 total samples split into 757,000 from 18 source cities for training and 315,000 from 6 target cities for testing. Data was Min-Max scaled to [0,1] per city to support transfer learning. All experiments used PyTorch on a single NVIDIA RTX 3060 Ti workstation, with the Adam optimizer (learning rate 1E-3), early stopping patience of 10 and a maximum of 50 epochs — the same protocol for NAS-discovered and manual baselines. To produce uncertainty estimates, they applied split conformal prediction to generate prediction intervals targeting 95% coverage.

Why This Matters

Impact on research: The paper argues that automatic architecture search can match manually designed models for time-series forecasting while being far smaller, and that pure recurrent architectures are preferable for short-horizon temporal forecasting. It also extends transfer learning to cross-continental weather forecasting, where the underlying physics is shared but regional manifestations differ — a setting the authors describe as a new application.

Real-world applications:

• Local urban, agricultural, and emergency-response forecasting in low-resource regions, where the paper motivates the work by the increasing frequency of extreme weather events.
• Deployment on IoT and edge climate-monitoring sensors, since Green-NAS-C is a 4 KB model with 1,064 parameters and 0.33 ms latency.
• Rapid model rollout in new cities: transfer learning means a competitive forecast model can be deployed after collecting roughly a month of historical data (about 1% of the total data collection), per the authors' claim.
• Safety-critical use with calibrated prediction intervals, since all three models approached the 95% coverage target with narrow intervals.

Industry relevance: The framework targets the economics of forecasting infrastructure — reducing dependence on high-end GPUs and centralized cloud inference in favor of distributed, private, edge-native prediction, which matters for operators of sensor networks and for climate services in the Global South.

Future Directions

1. Longer forecast horizons: Extending beyond one-step-ahead prediction to direct multi-step sequence-to-sequence models for 24–48 hour horizons, since recursive multi-step forecasting degraded from RMSE 0.03 at 1 hour to 0.16 at 12 hours.
2. Spatial resolution: Moving beyond point-based stations by incorporating satellite image fusion to enable spatial convolutions.
3. Real deployment: Going beyond simulated federated learning to heterogeneous physical hardware fleets mixing microcontrollers and CPUs.
4. Broader objectives and coverage: Adding energy use and carbon emissions as explicit NAS objectives, building models for under-represented climate zones such as Polar and Sub-Arctic areas, integrating NAS with automated hyperparameter optimization, and developing specialized models for extreme events such as heatwaves, storms, and flooding.

Target Audience

This paper is most useful to machine learning researchers working on NAS, efficient time-series models, or Green AI; to meteorology and climate-informatics practitioners interested in lightweight alternatives to large forecasting models; and to engineers building edge or IoT deployments where model size, latency, and on-device inference matter. Readers seeking a fully operational, multi-step regional forecasting system will find the work positioned as a proof of concept, with the authors noting the RTX 3060 Ti workstation, point-station data, and one-step horizon as limitations.

Authors’ abstract

We introduce Green-NAS, a multi-objective NAS (neural architecture search) framework designed for low-resource environments using weather forecasting as a case study. By adhering to 'Green AI' principles, the framework explicitly minimizes computational energy costs and carbon footprints, prioritizing sustainable deployment over raw computational scale. The Green-NAS architecture search method is optimized for both model accuracy and efficiency to find lightweight models with high accuracy and very few model parameters; this is accomplished through an optimization process that simultaneously optimizes multiple objectives. Our best-performing model, Green-NAS-A, achieved an RMSE of 0.0988 (i.e., within 1.4% of our manually tuned baseline) using only 153k model parameters, which is 239 times fewer than other globally applied weather forecasting models, such as GraphCast. In addition, we also describe how the use of transfer learning will improve the weather forecasting accuracy by approximately 5.2%, in comparison to a naive approach of training a new model for each city, when there is limited historical weather data available for that city.