How to systematically develop an effective AI-based bias correction model?

TL;DR

This work tackles systematic biases in global numerical weather prediction by introducing ReSA-ConvLSTM, a physics-aware bias-correction framework that fuses dynamic climatological normalization, ConvLSTM with temporal causality, and residual self-attention. The model learns a nonlinear, physics-consistent mapping between ECMWF forecasts and ERA5 reanalysis, achieving up to a $20\%$ reduction in $RMSE$ for $1$–$7$ day forecasts of $T2m$ and extending effective correction to $U10$, $V10$, and $SLP$, with a lightweight 10.6M-parameter footprint enabling rapid cross-variable transfer and downstream impact on ocean models. Ablation studies confirm the added value of dynamic normalization, attention, and residual connections, highlighting the importance of variable-aware architectural design. The framework shows promise for operational integration and downstream applications, though extending accuracy to subseasonal-to-seasonal timescales remains a challenge due to chaotic dynamics and the need to capture slow climate modes.

Abstract

This study introduces ReSA-ConvLSTM, an artificial intelligence (AI) framework for systematic bias correction in numerical weather prediction (NWP). We propose three innovations by integrating dynamic climatological normalization, ConvLSTM with temporal causality constraints, and residual self-attention mechanisms. The model establishes a physics-aware nonlinear mapping between ECMWF forecasts and ERA5 reanalysis data. Using 41 years (1981-2021) of global atmospheric data, the framework reduces systematic biases in 2-m air temperature (T2m), 10-m winds (U10/V10), and sea-level pressure (SLP), achieving up to 20% RMSE reduction over 1-7 day forecasts compared to operational ECMWF outputs. The lightweight architecture (10.6M parameters) enables efficient generalization to multiple variables and downstream applications, reducing retraining time by 85% for cross-variable correction while improving ocean model skill through bias-corrected boundary conditions. The ablation experiments demonstrate that our innovations significantly improve the model's correction performance, suggesting that incorporating variable characteristics into the model helps enhance forecasting skills.

Figures (11)

• Figure 1: An overview of the proposed bias correction architecture. (a) Overview of the core architecture illustrating the bias correction mechanism. Three distinct downstream applications are integrated with the framework: (b) Parameter-adaptive fine-tuning for cross-variable optimization by leveraging the pre-trained backbone model, (c) Output-level correction through direct rectification of AI-model forecast results, and (d) Input-level correction through atmospheric variables serving as oceanic boundary conditions.
• Figure 2: Evaluation of Bias Correction Performance. (a) Temporal evolution of 2m temperature (T2m) Root Mean Square Error (RMSE; units: K) and (b) Anomaly Correlation Coefficient (ACC) across 7-day forecasts, comparing ECMWF predictions with four correction AI methods: UNet (orange), UNet-TimeWindow (blue), Vision Transformer (green), and our proposed method (red). (c-f) Spatial validation analysis showing (c) ERA5 reanalysis ground truth, (d) bias-corrected predictions from our model, with corresponding error distributions for (e) raw ECMWF forecasts versus ERA5 and (f) corrected predictions versus ERA5. All spatial maps represent composite means of 7-day forecasts initialized on the first day of each month during validation years (1981, 1991, 2001, 2011, 2021).
• Figure 3: Ablation experiment results. (a) Validation of the data processing method. The black curve shows the ECMWF forecast, the blue curve represents the Z-score normalization, and the red curve shows the spatiotemporal dynamic normalization (our method). The x-axis is forecast days, and the y-axis is the RMSE of T2m, with a 7-day forecast period. (b) Validation of the unidirectional time vector arrow. The three graphs compare the results of our method, UNet, and UNet-TimeWindow on the same dataset. The x-axis represents forecast lead time, and the y-axis shows T2m RMSE. Yellow, blue, and green bars indicate 3-/5-/7- day lead times. (c) Validation of the model architecture optimization. The curves show correction results for ECMWF (black), ConvLSTM (blue), SA-ConvLSTM (green), Residual-ConvLSTM (yellow), and our proposed ReSA-ConvLSTM (red).
• Figure 4: Downstream tasks correction performance. (a) Correction performance of partially fine-tuned models on U10, V10, and SLP variables. We adapt the T2m backbone model through parameter freezing and partial fine-tuning for downstream meteorological variables. (b) Comparison between original AI-based weather forecasting outputs (black) and corrected predictions (red) using our method. (c) Ocean variable predictions (SST and U) generated by feeding corrected atmospheric forcing fields (T2m, U10, and V10) into the ocean model. Black curves in all panels represent baseline predictions from ECMWF or original AI models, while red curves show corrected results after our pre/post-processing.
• Figure B1: The mean and standard deviation (std) representing time points for each season. Taking T2m as an example, the time span is from 1981 to 2021. For each year within this period, we compute the mean and standard deviation for the same day of the year. The representative time points for the four seasons are selected as follows: December 22 for winter, March 21 for spring, June 22 for summer, and September 21 for autumn.