An analytics framework for interpretable subseasonal forecasting under decadal climate variability

Subseasonal forecasts (SSFs) are essential for managing climate risk. However, their precision is influenced by decadal climatic variability and long-term climate change. As climate change continues, this challenge is expected to grow. To address this, we apply explainable artificial intelligence (XAI) to improve transparency and understanding of deep learning models in climate forecasting. XAI helps reveal how models make predictions, supporting interpretability and scientific reliability. This study uses a Convolutional Long Short-Term Memory 2D (ConvLSTM2D) deep learning model to predict European SSFs and compares it to a baseline Random Forest model. ConvLSTM2D is a deep learning architecture that integrates Convolutional Neural Networks (CNNs) for capturing spatial patterns with Long Short-Term Memory (LSTM) networks for modeling temporal dependencies. Both models are trained on high-resolution European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) data. We apply Gradient-weighted Class Activation Mapping (Grad-CAM) and Feature Importance Analysis (FIA) to explain the model decision-making process and identify key meteorological drivers. For 28-day lead 2 m temperature anomalies, the deterministic skill of the ConvLSTM2D model is modest ( R 2 = − 0 . 0424 ) and lower than that of the Random Forest baseline ( R 2 = 0 . 5624 ), highlighting the strength of relatively simple tree-based methods in this data- and noise-limited subseasonal prediction setting. Nevertheless, XAI provides useful insights by identifying meteorologically relevant patterns, such as synoptic systems and moisture gradients. FIA highlights important variables such as sea level pressure, 2 m temperature, and total column water vapor. Temporal feature importance also shows the model’s reliance on historical data, with both a recency effect and a persistent influence from past time steps. These findings show that XAI can provide valuable diagnostic information for deep learning models in climate science, helping to refine model design and offering a path toward more trustworthy and transparent SSF systems for Europe and elsewhere in a changing climate. • Apply deep learning to subseasonal forecasting with interpretability methods. • Identify meteorological drivers influencing subseasonal forecast accuracy. • Provide transparent insights into the decision-making of forecasting models. • Compare analytical forecasting models to highlight strengths and limits. • Support decision-making in climate risk management through interpretable analytics.