Capabilities of deep learning models on learning physical relationships: Case of rainfall-runoff modeling with LSTM

This study investigates the relationships which deep learning methods can identify between input and output data. As a case study, rainfall-runoff modeling in snow-dominated watershed by means of long short-term memory (LSTM) network is selected. Daily precipitation mean air temperature were used as model to estimate daily flow discharge. After training verification, two experimental simulations conducted with hypothetical inputs instead observed meteorological data clarify response trained inputs. The first numerical experiment showed that even without precipitation, generated discharge, particularly winter low high during snow melting period. effects warmer colder conditions on discharge also replicated precipitation. Additionally, reflected only 17–39% total mass accumulation period annual revealing strong lack water conservation. results this indicated method may not properly learn explicit physical target variables, although they are still capable maintaining goodness-of-fit results.