Deep Learning-Based Temperature Forecast Postprocessing Model Integrating Temporal Bias Correction and Spatial Downscaling

To address systematic biases and insufficient spatial resolution in numerical weather prediction models, this paper proposes a two-stage deep learning model (LSTM-SR) that integrates bias correction and spatial downscaling using ECMWF-IFS high-resolution deterministic forecasts and CLDAS reanalysis data. In the first stage, a Long Short-Term Memory (LSTM) network produces hourly corrected temperature fields at 0.05° resolution. In the second stage, an improved U-Net downscaling model incorporating residual connections and a Transformer module refines the spatial resolution to 0.01°, generating high-resolution hourly temperature forecasts for the next 72 hours. To validate the proposed model, we conducted an ablation experiment in which the downscaling module of LSTM-SR was replaced by nearest-neighbor interpolation to create a benchmark model (LSTM-NN), and compared the results with those of ECMWF-IFS. The results show that: (1) For 1–3-day lead times, LSTM-SR improves the temporal resolution of ECMWF-IFS forecasts from 3/6-hourly to hourly, and reduces the mean absolute error (MAE), root mean square error (RMSE) and bias substantially relative to ECMWF-IFS and LSTM-NN, effectively mitigating the growth of forecast errors at longer lead times. (2) Spatially, the forecast errors of ECMWF-IFS and LSTM-NN exhibit a clear terrain dependence, with significantly larger errors in high-altitude areas than in plains, whereas LSTM-SR achieves an MAE ≤ 1.5°C for 88.5% – 99.2% of grid points, markedly improving forecast accuracy in high-altitude regions. (3) LSTM-SR demonstrates strong generalization ability across seasons: its median MAE is reduced by 19.10%–33.81% relative to ECMWF-IFS, and by an additional 6.89%–8.91% compared to LSTM-NN in spring, summer, and autumn. Error propagation analysis reveals that the second-stage downscaling inherits residual errors from the first-stage bias correction but does not cause error explosion. In summary, the LSTM-SR model achieves systematic bias correction and refined spatial reconstruction of numerical temperature forecasts, providing reliable technical support for operational medium-range gridded temperature forecasting.