Is Artificial Intelligence the Ultimate Solution for Hydrological Modelling?

Artificial intelligence plays an increasingly significant in many areas of our lives. Its applications in hydrology are becoming more common, and many authors have reported excellent results in modelling rainfall and predicting floods. However, alongside the successes, it is also important to understand the limitations of these models. This study presents various issues with potential significant impacts on applications, using an LSTM model and the CAMELS-US and CAMELS-GB datasets.The first important point is the problem of data quality. Hydrological observations are uncertain, with the largest error in observed discharge occurring with the highest measurements and the largest relative error with the smallest values. The error structure can change considerably due to alterations in riverbed geometry. Furthermore, areal rainfall is estimated based on point observations and is often biased, especially for extreme values (Bárdossy and Anwar 2023). Poor or variable quality of observational data can lead to suboptimal model outcomes. LSTM models act as bias correctors for many catchments by violating physical principles. For instance, water balances in catchments in the CAMELS-GB data are incorrect in more than 30% of the cases because evaporation is unrealistically high, which is compensated for by the LSTM models.The purpose of modelling is not to repeat what is already known but rather to predict behaviour under varying weather conditions or changing catchment characteristics. Thus, it is important to investigate how these models respond under altered conditions. An increase in precipitation results in inappropriate increases in evaporation in more than 60% of cases in the CAMELS-GB test series. Therefore, the use of these models for climate change studies is questionable.A major advantage of using LSTMs for hydrology is their ability to provide regional models for a large number of catchments. This is significantly different from the usual modelling for individual catchments. Several studies use static catchment attributes for regional modelling. However, integrating these static attributes changes the model structure. It is shown that a similar number of random numbers as attributes instead of catchment attributes can yield comparably good results. Therefore, the models may not be reliably applicable to uncalibrated catchments or changes within the catchments.A frequently discussed problem with the application of AI to hydrological prediction of extreme events is its tendency not to extrapolate beyond the range of its training data. However, this is only a limited issue due to regional modelling. By modelling specific discharges, insights from catchments where extreme floods have occurred can be transferred to other catchments. This allows for the simulation of scenarios exceeding the maximum values previously observed in a single catchment.