Simulating runoff under changing climatic conditions: A comparison of the long short-term memory network with two conceptual hydrologic models

Hydrologic models are commonly used to assess climate change impact on water resources. Several studies have reported that hydrologic models often experience severe performance degradation under climatic conditions different from calibration periods. With the advancement of artificial intelligence technology, the long short-term memory (LSTM) network has recently shown great potentials in rainfall-runoff modeling. However, little is known about the robustness of the LSTM network when used in changing climatic conditions. In this study, we compare the robustness of the LSTM network and two conceptual hydrologic models in runoff prediction in changing climatic conditions in 278 Model Parameter Estimation Experiment (MOPEX) basins. For calibration periods, the two hydrologic models have better performance in wet periods than in dry periods, while the LSTM network shows little performance difference under different climatic conditions. For validation periods, the three models suffer the largest performance loss when calibrated in a wet period and validated in a dry period. The performance losses of the LSTM network are primarily affected by the climate contrast between calibration and validation periods, while the performance losses of the two hydrologic models are mainly dependent on the climatic condition of validation periods. We also find that the length of the calibration period is an important factor affecting the relative performance of the models. Increasing the length of the calibration period has little effect on the validation performance of the two hydrologic models but enhances the LSTM network's performance. If sufficient calibration data is available, the LSTM network is a preferred tool for runoff simulation. On the other hand, the hydrologic models could have more advantages over the LSTM network in case of limited calibration data available.