CloudCast -- Total Cloud Cover Nowcasting with Machine Learning

TL;DR

CloudCast tackles the challenge of nowcasting total cloud cover by leveraging a U-Net CNN trained on five years of satellite-derived effective cloudiness data to forecast up to five hours ahead. The method outperforms traditional NWP and optical-flow baselines, achieving a $MAE$ reduction of about 24% and a 46% reduction in multi-category errors, with ablation showing $MAE$ losses and four input frames plus date/time features as optimal. It is integrated into FMI's operational nowcasting system for Northern Europe, providing faster, higher-fidelity cloud forecasts that support agriculture, solar power, and aviation. However, skillful lead time is currently limited to around 3 hours, and forecast blur increases with lead time; future work includes exploring more complex architectures (e.g., transformers) and higher-resolution data to extend lead time and accuracy.

Abstract

Cloud cover plays a critical role in weather prediction and impacts several sectors, including agriculture, solar power generation, and aviation. Despite advancements in numerical weather prediction (NWP) models, forecasting total cloud cover remains challenging due to the small-scale nature of cloud formation processes. In this study, we introduce CloudCast, a convolutional neural network (CNN) based on the U-Net architecture, designed to predict total cloud cover (TCC) up to five hours ahead. Trained on five years of satellite data, CloudCast significantly outperforms traditional NWP models and optical flow methods. Compared to a reference NWP model, CloudCast achieves a 24% lower mean absolute error and reduces multi-category prediction errors by 46%. The model demonstrates strong performance, particularly in capturing the large-scale structure of cloud cover in the first few forecast hours, though later predictions are subject to blurring and underestimation of cloud formation. An ablation study identified the optimal input features and loss functions, with MAE-based models performing the best. CloudCast has been integrated into the Finnish Meteorological Institute's operational nowcasting system, where it improves cloud cover forecasts used by public and private sector clients. While CloudCast is limited by a relatively short skillful lead time of about three hours, future work aims to extend this through more complex network architectures and higher-resolution data. CloudCast code is available at https://github.com/fmidev/cloudcast.

Figures (12)

• Figure 1: Left: Northern half-hemisphere as produced by FMI's NWCSAF installation. Finland highlighted in red. Right: CloudCast domain covering Northern Europe in Lambert conformal conic projection
• Figure 2: Histogram showing the distribution of effective cloudiness in the training dataset for Scandinavia from November 2018 to October 2023.
• Figure 3: Illustration of the CloudCast architecture. Layout follows the usual U-Net shape where input data is contracted in the encoder (left hand side) and subsequently expanded in the decoder (right hand side). The model’s input consists of seven channels: four consecutive effective cloudiness images, lead time, and date-time information. The output is a single-channel image predicting the cloudiness value.
• Figure 4: Five-hour prediction from each of the models and the corresponding ground truth. Models are sorted from best to worst based on overall verification score. Forecast initialisation time is June 17 2024, 1200 UTC.
• Figure 5: Mean Absolute Error Skill Score as a function of lead time for a) seasonal average for a five hour forecast, b) season-wise scores for CloudCast only, and c) intra-hour prediction for the first hour. Higher values indicate better performance.