Comparing the Performance of Deep Learning, Polynomial Neural Network and HEC-HMS Models in Predicting Daily Runoff

Estimating runoff caused by excess rainfall at the watershed scale is necessary for precise design of water structures, comprehensive watershed, and flood management. In this research, we evaluated deep learning methods in comparison to polynomial neural networks and HEC-HMS models in three watersheds, i.e., Bar-Erieh, Kasilian, and Latian. For this purpose, daily runoff was simulated using a long short-term memory (LSTM) deep learning model and compared to multi-layer perceptron (MLP) as the most common artificial intelligence model, group method of data handling (GMDH) and HEC-HMS as a physically based model for robust neural network modeling. The results showed that the R2 ranged from 0.872 to 0.986, RMSE from 0.086 m3/s to 2.22 m3/s, and NRMSE from 18.9 to 66.0%. The results indicate that the performance of the MLP model is mediocre with an average NRMSE of 51.2%, the performance of the GMDH model is good with an average NRMSE of 44.6%, and the LSTM model is very good with an average NRMSE of 26.8%. Considering the high computational cost of LSTM compared to the GMDH model, it can be recommended that the GMDH model should be used if the user does not expect very high accuracy and precision, and if higher accuracy is required, the user may need to accept a high computational cost and the LSTM model. Also, if the process and scenario-based modeling are the focus, the HEC-HMS model is preferred.