Enhancing Hydrological and Climate Predictions Through Artificial Intelligence

Abstract

The integration of artificial intelligence (AI) into hydrological and climate forecasting represents a significant advancement in enhancing prediction accuracy and reliability. This study explores how coupled land-atmosphere interactions influence soil moisture memory and water availability trends. In the Southeastern United States, the increased water use efficiency effect under elevated CO2 concentration is counteracted by the plant growth effect. To address the computational challenges posed by climate model experiments, specifically storage and runtime issues, this study developed two AI models:

- Hybrid Physics- AI model to improve hydrological forecasting.

-Deep learning model to improve ENSO prediction.

Building on insights into the interplay between biogeophysical variables, this research addresses the challenges presented by the National Oceanic and Atmospheric Administration’s (NOAA) high-resolution streamflow forecasting system using the National Water Model (NWM). We propose a hybrid Physics-AI model that integrates biophysical attributes such as topography, land use, and soil types with NWM forecasts, utilizing a deep learning approach to predict forecast errors. This hybrid model captures the complex interdependencies between biophysical attributes and hydrological processes, making it useful for predicting errors in areas lacking observational data.

Furthermore, integrating machine learning (ML) and deep learning (DL) models into ENSO index time series forecasting offers a promising approach to enhancing climate predictions. Traditional methods often struggle with accuracy and lead time, especially for long-term forecasts. However, recent advancements in DL models, such as Convolutional Neural Networks (CNNs), Bi-LSTM, hybrid CNN-1D and LSTM models, Gated Recurrent Units (GRUs), and Long Short-Term Memory (LSTM) networks, have demonstrated significant improvements in predictive skill. These advancements are influenced by lookback periods, where analyzing dependency on different historical data spans optimizes model accuracy. High-quality datasets further enhance model performance, leading to more reliable forecasts.

By comparing various DL models, we identified unique advantages, such as Bi-LSTM and LSTM networks' ability to capture temporal dependencies and GRUs' strength in feature extraction. Advanced training techniques like transfer learning bolster these models' efficacy. In conclusion, this research significantly enhances hydrological and climate predictions by integrating AI with traditional models and employing advanced ML/DL techniques for ENSO forecasting. The hybrid Physics-AI model improves forecast reliability from 21\% to 82\% compared to the NWM. For the 1-12 month lead period, skill transfer learning improved ENSO prediction skill, achieving an anomaly correlation metric (ACM) consistently above 0.8, comparable to Wang et al. For the 12-24 month lead period, the hybrid CNN-1D-LSTM model maintains an ACM above 0.5, though Wang et al.’s model shows superior long-term prediction skill with a higher and more stable ACM. The hybrid model, trained on the CESM2-LE-EOF dataset, excels in short-term forecasts, while Wang et al.’s model is more robust for long-term predictions. These complementary strengths will aid in better water resource management and climate risk mitigation.