%0 Journal Article %T 一种优于动力次季节温度预测的机器学习模型
A Machine Learning Model Superior to Dynamic Subseasonal Temperature Forecasting %A 薛翔海 %J Hans Journal of Data Mining %P 176-183 %@ 2163-1468 %D 2025 %I Hans Publishing %R 10.12677/hjdm.2025.152015 %X 可靠的次季节温度预测对极端温度事件的防灾减灾至关重要。然而，现有的次季节温度预测动力模型常受到初值问题和边值问题的影响，导致其预报能力相对薄弱。尽管近年来机器学习模型在次季节预测中逐渐展示出超越动力模型的潜力，但中国次季节温度预测仍主要依赖于动力学模型。鉴于此，本研究基于Lasso (Multi-task Lasso)机器学习算法，构建了覆盖中国所有格点的次季节温度预测模型，并采用余弦相似度指标评估Lasso和CFSv2 (The Climate Forecast System version 2)动力模型在2018~2022年测试期内的预测性能表现。结果表明：Lasso在整体预测技能上显著优于CFSv2，其在未来3~4周和5~6周的平均余弦相似度较CFSv2分别提升了0.33和0.34；并且，在常规温度情景下，Lasso能够更精准地捕捉温度变化的规律，80%以上月份的平均CS高于CFSv2；其仅在极端低温情景下存在一定局限性，预测技能略逊于CFSv2。
Reliable subseasonal temperature forecasting plays an important part in extreme temperature events prevention and mitigation. However, current dynamical models for subseasonal temperature forecasting are often influenced by initial value and boundary value problems, resulting in relatively weak forecasting performance. Although machine learning models have shown potential in surpassing dynamical models for subseasonal forecasting in recent years, subseasonal temperature forecasting in China still mainly relies on dynamical models. Under this background, the study constructs a subseasonal temperature forecasting model covering 957 grid points across China based on the Lasso (Multi-task Lasso) machine learning algorithm and uses the cosine similarity metric to evaluate the performance between the Lasso and CFSv2 (The Climate Forecast System version 2) dynamic model during the test period from 2018 to 2022. The results show that the Lasso significantly outperforms CFSv2 in overall forecasting performance. The average cosine similarity of the Lasso is 0.33 and 0.34 higher than the CFSv2 at the forecast horizon of weeks 3~4 and 5~6, respectively. Moreover, in normal temperature scenarios, the Lasso can more accurately capture temperature variation patterns with the average cosine similarity for over 80% of the months higher than that of the CFSv2. However, the Lasso has some limitations in forecasting extreme low temperature scenarios, where its forecasting skill is slightly inferior to that of the CFSv2. %K 动力模型， %K 机器学习模型， %K 次季节温度预测， %K 中国
Dynamic Model %K Machine Learning Model %K Subseasonal Temperature Forecasting %K China %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=111709