| 摘 要: 针对气象温度预测的非线性与区域复杂性难题,提出融合站点聚类与混合深度学习的多步预测框架。基于K-Means算法对站点地理位置及温度统计特征聚类,划分微气候区域;为各区域构建自回归积分移动平均模型(ARIMA)捕捉线性趋势,并利用BiLSTM-Attention预测残差。创新性地将聚类信息与位置编码注入双向长短期记忆(BiLSTM)初始状态,增强微气候感知能力。BSk气候下的24h预测实验表明:相较于ARIMA、BiLSTM及CNN-LSTM等模型,所提模型的均方误差(MSE)降低了20.1%~58.7%,平均绝对误差(MAE)降低了13.3%~59.8%,R2提升至0.914,显著提升了温度预测精度与鲁棒性,为区域气象预测提供新范式。 |
| 关键词: 温度预测 聚类分析 ARIMA BiLSTM 注意力机制 多步预测 |
|
中图分类号: TP391
文献标识码: A
|
|
| Multi-step Temperature Prediction Based on ARIMA-BiLSTM-Attention Modeland Hierarchical Clustering |
|
CHEN Mengyu, HE Liwen
|
(School of Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing 210003, China)
1223076622@njupt.edu.cn; helw@njupt.edu.cn
|
| Abstract: To address the challenges of nonlinearity and regional complexity in meteorological temperature prediction, a mult-i step prediction framework integrating station clustering with hybrid deep learning is proposed. Stations are clustered via the K-Means algorithm using geographical coordinates and statistical temperature features to delineate micro-climatic zones. An AutoRegressive Integrated Moving Average (ARIMA) model is fitted to each zone to capture linear temporal trends, and the residual component is modeled by a BiLSTM-Attention network. Clustering identifiers and positional encoding are incorporated into the Bidirectional Long Shor-t Term Memory (BiLSTM) initial state to enhance microclimate perception. The 24-hour forecasting experiments conducted under the BSk climate regime demonstrate that, relative to ARIMA, standalone BiLSTM, and CNN-LSTM baselines, the proposed framework reduces Mean Square Error (MSE) by 20.1% ~58.7% and Mean Absolute Error (MAE) by 13.3% ~59.8% , while elevating the coefficient of determination (R2 ) to 0.914. These results substantiate a significant improvement in predictive accuracy and robustness, offering a new paradigm for regional meteorological forecasting. |
| Keywords: temperature prediction cluster analysis ARIMA BiLSTM attention mechanism mult-i step prediction |
