Comparison of Multilayer Perceptron with an Optimal Activation Function and Long Short-Term Memory for Rainfall-Runoff Simulations and Ungauged Catchment Runoff Prediction

Artificial intelligence (AI) models have recently become increasingly prevalent in hydrological research. Multilayer perceptron (MLP) and long short-term memory (LSTM) are the most popular AI models for rainfall-runoff modeling. LSTM, a recurrent neural network, is more commonly used than MLP, which is a feedforward method, owing to its ability to retain important information from previous data that can support simulations. However, a suitable activation function can overcome this issue when the MLP model is used. In this study, we aimed to identify the best activation function for the MLP model and evaluate its applicability for the prediction in ungauged catchments. We selected seven catchments widely located in Korea, with sizes ranging from 763–6648 km2. To achieve our objectives, we applied and analyzed five activation functions. The performance of the optimal MLP model was compared with that of the LSTM model to assess the ability of MLP to simulate rainfall-runoff and predict runoff in ungauged catchments. Consequently, we used the exponential linear unit (ELU) activation function to develop the MLP-ELU model, which was the most suitable for simulating runoff. The performance of the MLP-ELU model in simulating rainfall-runoff was similar to that of the LSTM model. The prediction performance of runoff of the MLP-ELU model in ungauged catchments was high (Nash–Sutcliffe efficiency ranged from 0.83–0.93 in the seven catchments). Simulated runoff using parameter values estimated for different catchments yielded results that were comparable to those of the calibrated catchment. Therefore, if suitable activation functions are applied, the MLP model can be used for rainfall-runoff studies and water resource design in ungauged catchments. Additionally, the use of MLP models can broaden the range of AI options for hydrology research.