基于ConvLSTM的四川省逐小时气温预报模型

ConvLSTM-Based Hourly Temperature Forecast Model for Sichuan Province

摘要: 准确的逐小时气温预报对于农业生产、能源调度、灾害预警等领域具有重要意义。四川省地形复杂多样，气候特征区域差异显著，传统数值预报方法在该区域的预报精度面临挑战。本研究基于ERA5再分析数据，构建了轻量级卷积长短期记忆网络(LightConvLSTM)模型，对四川省逐小时2米气温进行预报。模型输入包括历史24小时的气温、纬向风、经向风和比湿数据，输出未来1小时的气温预报。研究采用2000~2023年数据进行模型训练，以2024年全年数据进行独立验证。结果表明：模型在测试集上的平均绝对误差(MAE)为0.608℃，均方根误差(RMSE)为0.892℃，相关系数(R)达到0.997。模型在夏季(6~8月)表现最优，逐月MAE为0.16℃~0.27℃；冬季(12~2月)相对较差，逐月MAE为0.47℃~0.51℃。空间分析显示，模型在四川盆地东部和南部地区预报精度较高，而在西部高原地区误差略大。结果表明，LightConvLSTM对四川复杂地形下逐小时气温具有良好预报能力，可为精细化气温预报提供技术参考。

Abstract: Accurate hourly temperature forecasting holds significant importance for agricultural production, energy dispatch, and disaster early warning. Sichuan Province features complex and diverse topography, with substantial regional variations in climatic characteristics, posing challenges to the forecast accuracy of traditional numerical prediction methods in this region. This study develops a lightweight Convolutional Long Short-Term Memory (LightConvLSTM) model based on ERA5 reanalysis data for hourly 2-meter temperature forecasting in Sichuan Province. The model inputs comprise historical 24-hour data of temperature, zonal wind, meridional wind, and specific humidity, with outputs being 1-hour-ahead temperature forecasts. The model was trained using data from 2000 to 2023 and independently validated against data from the entire year of 2024. The results demonstrate that the model achieves a Mean Absolute Error (MAE) of 0.608˚C, a Root Mean Square Error (RMSE) of 0.892˚C, and a correlation coefficient (R) of 0.997 on the test set. The model exhibits optimal performance during summer (June-August), with monthly MAE ranging from 0.16˚C to 0.27˚C, whereas relatively inferior performance is observed in winter (December–February), with monthly MAE ranging from 0.47˚C to 0.51˚C. Spatial analysis reveals that the model attains higher forecast accuracy in the eastern and southern regions of the Sichuan Basin, while slightly larger errors are found in the western plateau region. These results indicate that the LightConvLSTM model possesses robust forecasting capability for hourly temperature under the complex terrain of Sichuan Province, offering a technical reference for refined temperature prediction.