An intelligent deep learning-based approach for downscaling atmospheric general circulation model outputs

Currently, as carbon emissions rise and global temperatures increase, climate change models have gained significant importance. These models are capable of yielding insights into historical and forthcoming climate variables. Despite their extensive scope, these models continue to exhibit considerable uncertainties. Selecting suitable methodologies can be crucial and beneficial for effective downscaling to ensure precise forecasts. In this study, an LSTM model was employed to downscale the outputs of two global climate models, i.e., MRI-ESM2-0 and GFDL-ESM4. The climate data (e.g., precipitation and temperatures) from two global climate models, as well as observations from Urmia, Zanjan, Zahedan, and Birjand stations, were collected for the 1995–2014 period. Some indices were employed to evaluate the performance of the LSTM network for the downscaling of climate data over employed stations. As per the results, the LSTM model indicated acceptable downscaling outputs. The indices exhibit higher improvement values in the precipitation variable, outperforming the LSTM network in this variable. The LSTM neural network improved the GFDL-ESM4 precipitation correlation coefficient values by up to 350%. Furthermore, the maximum temperature remained unchanged and exhibited minimal improvement as the values were already high in the original state. Similarly, the minimum temperature improved insignificantly. The GFDL-ESM4 model demonstrated a 75% improvement in the mean square error of precipitation values after the LSTM was used. Additionally, the maximum temperature increased by 75%, whereas the minimum temperature decreased by 50%. The MRI-ESM2-0 model’s correlation coefficient values for precipitation were enhanced by up to 110% following the implementation of the LSTM neural network. There was no significant alteration in the maximum temperature, and the correlation changes observed in both models across all four stations remained below 10%. This stability can be attributed to the initial conditions being within the optimal range, while the minimum temperature exhibited only marginal improvements across the various stations. The implementation of the LSTM resulted in a significant improvement in RMSE, 70%, 75%, and 60% for precipitation, maximum and minimum temperatures, respectively. The results of this research underscore the capability of the LSTM model in enhancing the precision of downscaled climate data, especially concerning precipitation. These advancements indicate that the LSTM model may serve as an effective instrument for refining regional climate forecasts, which could be essential for multiple purposes such as managing water resources, planning for agriculture, and preparing for disasters. By delivering more precise local climate predictions, this method can facilitate improved decision-making in areas that are vulnerable to climate fluctuations and transformations.