Technical Summary: Stable Attention Response for Reliable Precipitation Nowcasting

Problem Statement

Precipitation nowcasting aims to predict near-future weather evolution from recent observations, a task critical for disaster management and daily planning. Despite advancements in deep learning, reliable nowcasting remains difficult due to the highly localized, rapidly evolving, and heterogeneous nature of atmospheric dynamics.

Recent methods have increasingly adopted attention-based architectures (both unimodal and multimodal) to capture long-range interactions and adaptively focus on informative regions. However, this paper identifies a previously underexplored failure mode: the cross-sample instability of attention-response energy. The authors observe that in scenarios with significant regime changes (e.g., varying precipitation intensity, spatial coverage, or cross-source consistency), the responses of attention heads fluctuate sharply across different samples. This instability is empirically correlated with inaccurate forecasts and theoretically shown to propagate through self-attention, enlarging the lower bound on prediction error.

Methodology: HARECast

To address this, the authors propose HARECast (Head-wise Attention Response Energy-regulated framework), a diffusion-based forecasting model that explicitly stabilizes attention responses.

1. Theoretical Motivation

The paper provides a theoretical analysis demonstrating that cross-sample variability in input features can propagate through self-attention mappings. Under the assumption that self-attention does not collapse input variability and the prediction head is non-degenerate, the variance of the prediction is lower-bounded by the variance of the attention response. Consequently, excessive variance in attention responses across samples increases the Mean Squared Error (MSE) lower bound, degrading forecast reliability.

2. Core Mechanism: Group-wise HARE Stabilization

The central innovation is the definition and regularization of Head-wise Attention Response Energy (HARE).

• HARE Definition: For a given sample and attention head, HARE is defined as the Frobenius norm squared of the attention-induced response ($O_{i,m} = A_{i,m}V_{i,m}$). This quantifies the effective strength of the response delivered to downstream prediction.
• Head Partitioning: Based on batch-averaged HARE statistics, attention heads are partitioned into three groups:
  • Dominant ($H^\uparrow$): Heads with the strongest responses.
  • Inactive ($H^\downarrow$): Heads with weak responses (below a threshold $\alpha$).
  • Contextual ($H^\circ$): Intermediate heads.
• Stabilization Objective: A group-wise regularization loss ($L_{hare}$) is introduced. For each sample, the group-level energy is regularized toward the batch-level mean energy of its respective group. This suppresses harmful cross-sample fluctuations while preserving meaningful differences in response strength between groups.

3. Framework Architecture

HARECast is designed to be generic, applicable to both unimodal (radar-only) and multimodal (radar + satellite) settings. The pipeline consists of:

1. Encoder: Processes input sequences (radar, and optionally satellite) using self-attention or cross-attention blocks.
2. HARE Stabilization: The proposed regularization is applied to the encoder's attention outputs to stabilize cross-sample energy.
3. Reconstruction Branches: Lightweight decoders reconstruct the input modalities from the latent representation. This auxiliary task ensures the latent space preserves sufficient input information.
4. Diffusion Predictor: A conditional diffusion model generates future radar frames based on the stabilized latent representation.

The total training objective combines reconstruction loss, HARE stabilization loss, and diffusion generative loss.

Key Contributions

1. Identification of a Failure Mode: The paper establishes that cross-sample instability of attention-response energy is a significant source of unreliability in precipitation nowcasting, distinct from issues related to representation capacity or generative realism.
2. Theoretical and Empirical Evidence: The authors provide theoretical proofs showing how input variability propagates through self-attention to increase prediction error bounds. Empirically, they demonstrate a correlation between high attention-energy variance and poor forecasting performance.
3. HARECast Framework: They propose a novel, generic framework that stabilizes attention responses via group-wise regularization.
4. State-of-the-Art Performance: The method achieves superior results on standard benchmarks (SEVIR and MeteoNet) in both unimodal and multimodal settings, particularly for high-intensity precipitation events.

Experimental Results

The authors evaluated HARECast on the SEVIR and MeteoNet datasets.

• Unimodal Performance (SEVIR & MeteoNet): HARECast outperformed recent state-of-the-art methods (e.g., AlphaPre, DuoCast, DiffCast). On SEVIR, it improved the Mean Critical Success Index (CSI-M) by approximately 2.0–2.4% and the Heidke Skill Score (HSS) by 0.1–1.2% compared to the best baselines. Gains were particularly pronounced for high-intensity precipitation thresholds.
• Multimodal Performance (SEVIR): In the multimodal setting, HARECast achieved a 7% improvement in CSI-M and a 2% gain in HSS over the previous best multimodal method (PiMMNet).
• Qualitative Analysis: Visual comparisons showed that HARECast better preserves coherent large-scale evolution and fine-scale precipitation structures over long forecasting horizons compared to baselines, which often suffer from blurring or structural deviations.
• Ablation Studies:
  • Removing the HARE stabilization module caused significant drops in performance, especially for high-threshold metrics.
  • Group-wise partitioning of heads was shown to be superior to stabilizing all heads jointly.
  • The method is sensitive to batch size; smaller batches lead to noisier target estimates and degraded performance, confirming the reliance on stable batch statistics.

Significance and Claims

The paper claims that while modern nowcasting models focus heavily on improving representational capacity and generative diversity, explicitly controlling the stability of internal attention responses is a critical, yet overlooked, direction for building reliable systems.

The authors argue that their work demonstrates that "expressive attention alone is insufficient for reliable nowcasting." By stabilizing the energy of attention-induced responses across samples, HARECast reduces the propagation of variance that leads to prediction errors. The framework is presented as a generalizable principle that can be integrated into existing unimodal or multimodal architectures to enhance forecasting reliability, particularly in challenging regimes involving heterogeneous inputs and rapid atmospheric changes. The authors do not claim to solve all aspects of atmospheric modeling but position their method as a specific, effective intervention to address the stability bottleneck in attention-based forecasting.