TLDR: A new method combines Convolutional Neural Networks (CNNs) with ensemble weather prediction models to improve medium-range surface temperature forecasts. By first correcting individual low-resolution ensemble members with CNNs for systematic errors and super-resolution, and then averaging them, the approach significantly reduces both systematic and random errors. This technique outperforms existing high-resolution models and is particularly beneficial for operational centers with limited computational resources, though it may smooth out extreme temperature predictions.
Accurate temperature forecasts for periods longer than five days are crucial for various sectors, including agriculture, public health, and energy. However, current operational methods often rely on low-resolution numerical weather prediction (NWP) models, which are prone to systematic and random errors, leading to less reliable predictions.
A recent study proposes an innovative approach that integrates convolutional neural networks (CNNs) with ensemble NWP models to significantly enhance surface temperature forecasting for these medium-range periods. The core idea is to address both systematic and random errors by applying a two-step post-processing method.
The first step involves using CNNs to correct systematic errors and perform spatial super-resolution on each individual ensemble member of the low-resolution NWP model. This process reconstructs high-resolution temperature fields from the coarser model outputs. CNNs are particularly effective here because they can learn and correct spatial structures and biases, including those induced by complex terrain.
The second step reduces random errors by performing ensemble averaging across these newly CNN-corrected members. This sequential approach—CNN correction before ensemble averaging—was found to consistently achieve higher accuracy compared to the reverse order (averaging first, then CNN correction).
The research, conducted by Takuya Inoue and Takuya Kawabata from the Meteorological Research Institute in Japan, demonstrated remarkable improvements. The proposed method, even when based on low-resolution ensemble forecasts from the Japan Meteorological Agency’s (JMA) Global Ensemble Prediction System (GEPS), notably outperformed high-resolution deterministic NWP models like the JMA’s Meso-Scale Model (MSM) and their operational Kalman filter-based method (GSM-KF).
Specifically, the method reduced the Root Mean Square Error (RMSE) by 1.3 K (a 49% improvement) and the Mean Error (ME) by 0.13 K (a 28% improvement) compared to the original GEPS control run. This improvement is equivalent to approximately 24 hours of extended lead time compared to the second-best performing model. The study highlights that the significant reduction in RMSE is largely attributable to the suppression of random errors through ensemble averaging after the CNN correction.
Case studies further illustrated the method’s effectiveness. Over complex terrain, the CNN-based super-resolution technique enhanced spatial details, accurately resolving temperature distributions along valleys and ridgelines. In a critical snowfall prediction scenario, the method provided more reliable near-freezing temperature forecasts, reducing the overestimation of snowfall events that can disrupt societal activities.
While highly effective, the study acknowledges certain limitations. Ensemble averaging inherently smooths data, which can diminish the capture of extreme temperatures, such as heatwaves. Additionally, CNNs may struggle with meteorological conditions rarely encountered in their training datasets. Future work aims to address these limitations by using longer-term training datasets, exploring probabilistic forecasting frameworks, and investigating alternative deep learning architectures like Vision Transformers for better long-range spatial dependency capture.
Also Read:
- MedSymmFlow: Enhancing Medical Imaging with Integrated AI Capabilities
- Enhancing Traffic Insights: Inferring Lane-Level Data from Road Information
This research presents a practical and scalable solution for improving medium-range temperature forecasts, especially valuable for operational centers with limited computational resources that cannot run high-resolution ensemble forecasts. For more details, you can refer to the full research paper.
