Utilizing Machine Learning and Deep Learning for Precise Intensity-Duration- Frequency (IDF) Curve Predictions

Intensity-Duration-Frequency (IDF) curves are crucial for the design and management of engineering infrastructure, including storm sewers, retention ponds, dams, and flood mitigation systems. This study adopts a comparative approach to estimate IDF curves using a combination of traditional statistical methods, machine learning techniques, and advanced deep learning models. Rainfall data from Koya City, Iraq (2005–2022), was used, with the 2005–2015 period for training and 2016–2022 for validation. The models evaluated include the Gumbel Distribution, Linear Regression, Support Vector Regression (SVR), and Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM), assessed based on three metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). Among these, the RNN-LSTM model demonstrated the lowest RMSE (1.44 mm/hr), lowest MAE (0.81 mm/hr), and highest R² (0.99), outperforming the Gumbel Distribution (RMSE: 9.13 mm/hr), Linear Regression (RMSE: 10.76 mm/hr), and SVR (RMSE: 6.19 mm/hr). This establishes RNN-LSTM as the most reliable approach for IDF curve prediction. Leveraging the RNN-LSTM model, rainfall trends for 2023–2043 were forecasted, revealing an expected increase in short-duration, high-intensity rainfall events, heightening flood risks, and emphasizing the need for adaptive stormwater management strategies. The findings underscore the significant potential of deep learning models like RNN-LSTM in enhancing IDF curve predictions and guiding the development of resilient hydraulic infrastructure, particularly in regions like Koya City, where complex topography exacerbates flood challenges during intense rainfall events.