Deep Learning Approaches for Seasonal Precipitation and Runoff Prediction: A Case Study of TaqTaq-Dokan Basin

Abstract

The TaqTaq-Dokan River Basin on the Little Zab River in northern Iraq faces seasonal water shortages in summer and recurrent flooding in winter due to high runoff downstream of Dokan Lake. To address this, this study applied deep learning neural network models to forecast annual precipitation and estimate surface runoff up to 2030. Monthly rainfall data from Koya Meteorological Station for 1987–2024 were used to model runoff via the Soil Conservation Service Curve Number method. Five deep learning architectures Multilayer Perceptron, Convolutional Neural Network, Recurrent Neural Network, Long Short-Term Memory, and Gated Recurrent Unit were trained on 70% of the data and tested on 30%. Model performance was evaluated using MSE, RMSE, MAE, AIC, and BIC. Results show RNN outperformed other models with the lowest errors (MSE = 1009.65, RMSE = 31.78, MAE = 23.09), followed by CNN, LSTM, GRU, and MLP. Forecasts indicate a consistent seasonal pattern: peak rainfall and runoff occur in winter, with zero runoff from June to September. The observed average annual runoff for 1987–2021 was 396.56 million m³, while the predicted runoff for 2024–2030 averages 331.37 million m³ annually. The findings confirm that substantial rainfall water over 330 million m³ per year—is discharged unused into the Little Zab River. The study concludes that applying deep learning algorithms is a reliable tool to predict the annual precipitation and surface runoff in TaqTaq-Dokan basin. Consequently, this study recommends that relevant organisations in Iraq take necessary actions and implement strategic plans and policies to conserve water by constructing ponds, dams, and reservoirs in areas with high precipitation.