FiLM-SimVP: Scalable Uncertainty Quantification in Spatiotemporal Forecasting

Uncertainty quantification in spatiotemporal forecasting is crucial for decision-making. Quantile regression has been proposed as a computationally efficient and assumption-free approach for uncertainty quantification by generating deriving prediction intervals. Yet, existing deep learning based quantile regression methods struggle to scale efficiently with large datasets. This paper introduces FiLM-SimVP, a novel architecture that combines Feature-wise Linear Modulation (FiLM) and recent advances in efficient 2D convolution-based attention mechanisms to improve spatiotemporal forecasting and interval prediction. The proposed approach enables adaptive modulation of spatiotemporal features by directly conditioning intermediate features on confidence levels, allowing the model to learn desired arbitrary predictive intervals. Through extensive experimentation on three diverse datasets, namely TaxiBJ, Traffic4cast, and WeatherBench, we demonstrate that FiLM-SimVP consistently outperforms existing state-of-the-art (SOTA) methods in point estimation accuracy and interval quality metrics. The model achieves an improvement of 1.1% in Mean Absolute Error and 6% in Mean Squared Error compared to baseline approaches with only slightly increased parameters and computations. Additionally, FiLM-SimVP shows superior scalability when modeling multiple intervals, including the entire conditional distribution, effectively capturing the relationship between different confidence levels without requiring separate models.