[Paper Review] Deep multi-stations weather forecasting: explainable recurrent convolutional neural networks

Ismail Alaoui Abdellaoui, Siamak Mehrkanoon|arXiv (Cornell University)|Sep 23, 2020

Hydrological Forecasting Using AI31 references11 citations

TL;DR

This paper proposes Deep Attention Unistream Multistream (DAUM) networks, a novel recurrent convolutional architecture with self-attention for multi-station weather forecasting across 18 European cities over 15 years. It demonstrates that multistream input representation and self-attention significantly improve prediction accuracy for temperature and wind speed, while occlusion and score maximization analyses reveal key features and cities—such as Brussels for temperature and coastal cities for wind speed—enhancing model interpretability.

ABSTRACT

Deep learning applied to weather forecasting has started gaining popularity because of the progress achieved by data-driven models. The present paper compares two different deep learning architectures to perform weather prediction on daily data gathered from 18 cities across Europe and spanned over a period of 15 years. We propose the Deep Attention Unistream Multistream (DAUM) networks that investigate different types of input representations (i.e. tensorial unistream vs. multistream ) as well as the incorporation of the attention mechanism. In particular, we show that adding a self-attention block within the models increases the overall forecasting performance. Furthermore, visualization techniques such as occlusion analysis and score maximization are used to give an additional insight on the most important features and cities for predicting a particular target feature of target cities.

Motivation & Objective

• To improve deep learning-based weather forecasting by comparing unistream and multistream input representations across multiple European cities.
• To evaluate the impact of integrating a self-attention mechanism on forecasting performance for temperature and wind speed.
• To enhance model interpretability by identifying the most influential weather features and cities using occlusion analysis and score maximization.
• To provide actionable insights into spatio-temporal dependencies in weather prediction using explainable AI techniques.

Proposed method

• Proposes two deep learning architectures: Att-Unistream and Att-Multistream, combining Conv-LSTM with self-attention for spatio-temporal feature learning.
• Uses tensorial input representations—unistream (concatenated features per city) and multistream (separate streams per feature)—to compare modeling efficiency.
• Integrates a self-attention mechanism within the network to dynamically weigh the importance of different spatial and temporal features.
• Applies occlusion analysis by masking input regions (spatial and temporal) to identify critical features and cities influencing predictions.
• Employs score maximization to visualize which input feature-city pairs most influence the model’s output for specific lags.
• Trains models for 6-day-ahead forecasting of temperature and wind speed using daily data from 18 European cities over 15 years.

Experimental results

Research questions

• RQ1 Does the multistream input representation outperform the unistream representation in multi-station weather forecasting?
• RQ2 How does the integration of a self-attention mechanism affect forecasting accuracy in recurrent convolutional networks?
• RQ3 Which weather features and cities are most influential for predicting temperature and wind speed in specific target cities?
• RQ4 How do temporal lags influence prediction performance, and do different architectures favor different lag ranges?
• RQ5 Can occlusion and score maximization techniques effectively explain the model’s predictions in a geospatial context?

Key findings

• The multistream input representation consistently outperformed the unistream representation in forecasting accuracy for both temperature and wind speed.
• The integration of a self-attention mechanism improved model performance across all evaluation metrics, indicating better capture of long-range dependencies.
• For temperature prediction, dew point and pressure were identified as the most influential features, with Brussels being a key city due to its central geographic position.
• For wind speed prediction, maximum wind speed and wind speed features were most critical, with coastal cities like Barcelona and Amsterdam playing a significant role.
• The Att-Unistream model favored recent temporal lags (1st–2nd), while the Att-Multistream model showed higher importance for older lags, indicating different temporal modeling behaviors.
• Score maximization revealed that visibility was consistently the most important feature across all lags for wind speed prediction, with pressure also contributing significantly.