Scholarlink

<p>Accurate climate forecasting in regions with complex topography and sparse observational data, such as Iran, remains a significant challenge. This study introduces an innovative hybrid framework integrating Wavelet Transform (WT) with Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) to improve the downscaling of Global Circulation Model (GCM) outputs. Applied to two GCMs—AWI-CM-1-1-MR and BCC-CSM2-MR—across 58 stations in eight Iranian regions, the method uses WT to decompose GCM data into frequency components. These components are then processed by a CNN-LSTM model to capture intricate climate patterns. Four model configurations were evaluated using key metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Percentage Error (%Error), Coefficient of Determination (R²), and Combined Accuracy (CA). The results show substantial improvements over raw GCM outputs. For AWI-CM-1-1-MR, the LSTM model achieved high accuracy in precipitation forecasting (R² = 0.7721, CA = 0.3249), the Wavelet-db1-CNN-LSTM model improved minimum temperature (Tasmin) prediction (CA = 0.4286), and the LSTM model enhanced maximum temperature (Tasmax) accuracy (R² = 0.8004). For BCC-CSM2-MR, the Wavelet-db1-CNN-LSTM model excelled in precipitation forecasting (R² = 0.7869, CA = 0.3194), while the Wavelet-Haar-CNN-LSTM model performed best for maximum temperature (Tasmax; R² = 0.7854, CA = 0.3336) and minimum temperature (Tasmin; CA = 0.4188). Incorporating WT significantly enhanced performance, particularly for non-stationary and noisy data common in data-scarce regions. The Haar and db1 wavelet functions demonstrated distinct advantages for temperature and precipitation forecasting, respectively, highlighting the method’s adaptability. This framework advances climate downscaling by providing a robust, scalable solution for regions with limited observational data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Downscaling of two selected GCM data using a hybrid deep learning method of Wavelet-CNN-LSTM in Iran

  • 期刊论文

  • 发表日期：

    2025年8月11日

查看全文
  • Saeed Hosseinpour,
  • Ahmad Sharafati,
  • Hirad Abghari

摘要

Accurate climate forecasting in regions with complex topography and sparse observational data, such as Iran, remains a significant challenge. This study introduces an innovative hybrid framework integrating Wavelet Transform (WT) with Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) to improve the downscaling of Global Circulation Model (GCM) outputs. Applied to two GCMs—AWI-CM-1-1-MR and BCC-CSM2-MR—across 58 stations in eight Iranian regions, the method uses WT to decompose GCM data into frequency components. These components are then processed by a CNN-LSTM model to capture intricate climate patterns. Four model configurations were evaluated using key metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Percentage Error (%Error), Coefficient of Determination (R²), and Combined Accuracy (CA). The results show substantial improvements over raw GCM outputs. For AWI-CM-1-1-MR, the LSTM model achieved high accuracy in precipitation forecasting (R² = 0.7721, CA = 0.3249), the Wavelet-db1-CNN-LSTM model improved minimum temperature (Tasmin) prediction (CA = 0.4286), and the LSTM model enhanced maximum temperature (Tasmax) accuracy (R² = 0.8004). For BCC-CSM2-MR, the Wavelet-db1-CNN-LSTM model excelled in precipitation forecasting (R² = 0.7869, CA = 0.3194), while the Wavelet-Haar-CNN-LSTM model performed best for maximum temperature (Tasmax; R² = 0.7854, CA = 0.3336) and minimum temperature (Tasmin; CA = 0.4188). Incorporating WT significantly enhanced performance, particularly for non-stationary and noisy data common in data-scarce regions. The Haar and db1 wavelet functions demonstrated distinct advantages for temperature and precipitation forecasting, respectively, highlighting the method’s adaptability. This framework advances climate downscaling by providing a robust, scalable solution for regions with limited observational data.