Abstract
The CFSv2 forecast products have been widely used in climate prediction operation all over the world. Although the real-time forecast is able to basically capture large pattern of climate anomaly, there still exists obvious bias, which may have enormous impacts on predicted result and thus cannot be used directly. This paper describes the development of a daily-updating system for Zhejiang province that utilizes a climatology-recovery technique to correct systematic biases in CFSv2 outputs, providing skillful extended-range forecasts.
Introduction
While short-term weather forecasts (1–7 days) have reached high levels of reliability, there remains a significant gap in "extended-range" (10-30 days) and monthly (1-30 days) climate trend predictions. These windows are critical for agricultural planning and disaster mitigation. Currently, most operational climate products are updated only once every ten days, a frequency insufficient for responding to rapid weather anomalies.
Furthermore, global models like NCEP's Climate Forecast System version 2 (CFSv2) exhibit stable systematic biases, particularly in localized regions with complex geography like China's Zhejiang province. Direct application of raw model outputs often leads to inaccurate predictions. To address this, researchers developed an automated, daily-updating platform that applies real-time bias correction to localized climate forecasts.
Methods and Data
The system utilizes a statistical downscaling strategy known as climatology-recovery. This method acknowledges that while CFSv2 may have a biased baseline (climatology), its ability to predict deviations (anomalies) is often more robust. The process involves extracting the forecasted anomaly from the CFSv2 model, discarding the model's internal climatology, and adding the forecasted anomaly to the actual observed 30-year historical climatology (1981–2010).
Data Sources and Ensemble Strategy
The system integrates observational data from 90 county-level meteorological stations in Zhejiang (1982–2020) and real-time CFSv2 ensemble forecasts. To ensure stability, the system employs an ensemble of 9 daily perturbed runs from the 0600, 1200, and 1800 UTC cycles, extending out to 45 days.
Key Results and Discussion
Evaluation of the system over 335 independent days in 2020 demonstrates significant improvements in forecast accuracy across multiple metrics.
Temperature and Precipitation Skill
The climatology-recovery method drastically improves Spatial Correlation Coefficients (SCC) and reduces Root Mean Square Error (RMSE) compared to raw model outputs. For 1-30 day temperature forecasts, the system achieved a mean Ps score of 81.6 and an Anomaly Correlation Coefficient (ACC) of 0.2, with over 75% of days showing positive skill.
Precipitation remains more challenging due to its chaotic nature, yet the system still outperformed traditional subjective forecasts. The 1-30 day precipitation forecasts achieved a mean Ps score of 78.1 and an ACC of 0.2, which is a notable improvement over the historical average of <0.1 for monthly rainfall in this region.
Seasonal Variability and Lead Time
Analysis reveals that forecast bias is not uniform throughout the year. Temperature RMSE is higher during autumn and winter, while precipitation RMSE peaks during the summer monsoon. Additionally, 1-30 day forecasts consistently outperform 10-30 day forecasts; the inclusion of the highly accurate initial 10-day weather window significantly bolsters the overall monthly statistical skill.
Conclusion
The automated platform proves that high-frequency, objective climate prediction is operationally viable at the provincial level. While the climatology-recovery strategy significantly reduces bias, it does not eliminate it entirely. Future iterations of the system will explore Cumulative Distribution Function (CDF) transforms, predictable component extraction, and multi-model ensemble integration to further refine forecast accuracy.
Summary updated