CO₂-Conditioned Temporal Convolutional Networks for Robust Climate Forecasting under Distribution Shift

Machine learning approaches for climate forecasting face a major challenge because the statistical characteristics of climate data evolve over time as greenhouse gas concentrations increase. Many data-driven forecasting models assume that training and deployment environments follow similar statistical patterns. However, future climate conditions associated with rising atmospheric CO₂ levels introduce distribution shifts that can reduce the reliability of conventional deep learning forecasting systems.

This study introduces a Feature-wise Linear Modulated Temporal Convolutional Network (FiLM-TCN) that incorporates atmospheric CO₂ concentration into the forecasting process. The proposed architecture integrates physical forcing information through FiLM conditioning, which dynamically modulates intermediate neural features during prediction. By treating CO₂ concentration as an external conditioning signal, the model adapts internal feature representations as forcing conditions evolve, improving robustness under non-stationary climate regimes.

Experiments use ERA5 reanalysis data from 1940–2024, including key atmospheric variables such as near-surface temperature, precipitation, and surface solar radiation. To evaluate performance under future climate conditions, the dataset is extended synthetically from 2025 to 2100 following an SSP3-7.0-like CO₂ pathway. The proposed FiLM-TCN model is compared with a baseline Temporal Convolutional Network under both historical in-distribution conditions and future high-CO₂ out-of-distribution scenarios.

Results show clear forecasting improvements. Under historical conditions, the FiLM-conditioned model reduces RMSE from 7.14 to 2.85, while under future forcing conditions RMSE decreases from 5.70 to 2.00. Statistical testing confirms that the improvements are highly significant with large effect sizes.

These findings demonstrate that incorporating physically meaningful forcing variables can improve the robustness and generalization of deep learning models for climate forecasting.