| 摘要: |
| 全球变暖预期下不同区域的未来降水变化，是政府和公众都关心的重要问题，也是全球变化研究的前沿科学问题。预测模型的建立是预测/预估研究的重点和难点，现今不稳定的气候背景对预测模型的精准度提出了更高的要求。为了解决传统方法对长期时间序列预测效果欠佳的问题，本文以泰国南部洞穴石笋δ18O重建的过去270多年（公元1773—2004年）的降水记录为数据集，提出了SSA-XGBoost预测模型。对原始数据去趋势预处理后，采用奇异谱分析法（SSA）提取前部分数据（1773—1964年）的振荡成分以确定数据的最佳谐波个数，并进行准确的周期信号分量分解；之后用去趋势数据减去周期信号得到随机信号，再利用XGBoost模型对随机项进行预测；最后将预测的序列、趋势曲线和周期信号延拓结果相叠加得到最终的预测数据（1965—2004年）。与其他四种模型（XGBoost、ARIMA、SSA-ARIMA、LightGBM）的预测结果相比，SSA-XGBoost的预测结果与真实值最相近，且MAE和RMSE均最小，R2也更接近1，说明该模型具有更高的精度和稳定性。该研究对于泰国南部等热带地区未来的降水变化趋势预测具有较好的指导意义，也可为其他长时间序列的预估研究提供借鉴。 |
| 关键词: 降水变化趋势 预测 机器学习 XGBoost模型 奇异谱分解 |
| DOI：10.7515/JEE202011 |
| CSTR：32259.14.JEE202011 |
| 分类号: |
| 基金项目:中国科学院战略性先导科技专项（B类）（XDB40000000）；国家自然科学基金项目（41991252）；国家重点研发计划（2017YFA0603401） |
| 英文基金项目:Strategic Priority Research Program of Chinese Academy of Sciences (XDB40000000); National Natural Science Foundation of China (41991252); National Key Research and Development Program of China (2017YFA0603401) |
|
| Model based on SSA-XGBoost method for predicting precipitation change trends |
|
XU Lei, WANG Tianli, LIU Songguo, LI Dong, LI Wei, TAN Liangcheng
|
|
1. School of Information Engineering, Chang’an University, Xi’an 710064, China
2. State Key Laboratory of Loess and Quaternary Geology, Institute of Earth Environment, Chinese Academy of Sciences, Xi’an 710061, China
3. University of Chinese Academy of Sciences, Beijing 100049, China
4. Zhejiang Lab, Hangzhou 310000, China
5. Library of Chang’an University, Xi’an 710064, China
6. School of Earth Science and Resources, Chang’an University, Xi’an 710064, China
7. Institute of Global Environmental Change, Xi’an Jiaotong University, Xi’an 710049, China
|
| Abstract: |
| Background, aim, and scope Predicting future climate changes is one of the most important scientific questions in the research field of global change. Precise and reliable prediction results can enable us to plan ahead and minimize the impacts of drought and flood disasters on social and economic developments. However, the establishment of prediction models is both necessary and difficult to achieve. Furthermore, owing to the current unstable climate, the requirement for the accuracy of such models has increased. Hence, this study addresses the problems that conventional methods cannot effectively predict in long-term time series, which in turn can be significant for the prediction of periodic series. Materials and methods In this study, a model based on the singular spectrum analysis (SSA)-XGBoost method was developed by incorporating the precipitation records over the past 270 years (AD 1773—2004) that were reconstructed on the basis of the stalagmiteδ 18O content in southern Thailand. After detrending the original data, the oscillation components of the past data (AD 1773—1964) were extracted through SSA for determining the optimal harmonic number of the data and for reconstructing accurate periodic signal components. Subsequently, subtracting the periodic signal from the detrended original data produced a random signal, which was predicted by employing the XGBoost framework for obtaining a forecast sequence. The final forecast results (AD 1965—2004) were obtained by superimposing the obtained sequence and the periodic signal extension results. Results The obtained results indicate that the proposed model based on the SSA-XGBoost method performs very well while predicting the test date. The proposed model’s mean absolute error (MAE) and root mean squared error (RMSE) values are found to be 0.1334 and 0.1678, respectively. Furthermore, the value of R-square (R2) is found to be 0.536. Discussion In general, the proposed model has revealed satisfactory performance for the predictions of precipitation. Apart from the machine learning frameworks, XGBoost and LightGBM, conventional models such as ARIMA and SSA-ARIMA were employed for predicting the data to demonstrate the advantages of the proposed model. It is observed that each model provides unique features while predicting the final results. Conclusions Compared to the prediction results of the other four models (XGBoost, ARIMA, SSA-ARIMA, and LightGBM), the prediction result of the proposed model is closest to the real value. Both the MAE and RMSE of the SSA-XGBoost method are lowest in value, and the value of R2 is closest to 1, which indicates that the proposed model based on the SSA-XGBoost method has relatively higher precision and stability. Recommendations and perspectives This study serves as an important directive for future studies on predicting precipitation changes in southern Thailand and other tropical regions, and it also provides a reference for future research on predicting long-term time series. |
| Key words: precipitation change trend prediction machine learning XGBoost Model Singular Spectrum Analysis (SSA) |
