Abstract
As climate change exacerbates variability and non-stationarity in rainfall patterns, it is crucial to assess the predictive capabilities of forecasting models. Previous researches on rainfall-runoff modeling have focused on the impact of training dataset size on Artificial Neural Networks (ANNs) results, with limited consideration of hydrometeorological diversity. This study first evaluates the influence of the training dataset length (1 to 15 years) on performance of a Long Short-Term Memory (LSTM) and a traditional conceptual model, Superflex, across 10 validation years. Next, training years are categorized based on hydrometeorological diversity (wetter, standard, drier). This clustering allows for experiments where models are trained on data from similar or different clusters, enhancing understanding of how data diversity, and therefore climate change, can affect model performance. Results indicate that the LSTM model is highly sensitive to training length, showing poor performance with short datasets (below three years), reaches similar performance to Superflex around six training years on average, and overperforms with 15 years of training. Conversely, Superflex maintains rather constant performance levels regardless of the dataset length. LSTM model benefits from diverse 1 training data, achieving higher accuracy and reliability when trained on years with diverse hydrological typology. Despite their potential to outperform traditional models (with six or more training years on average), LSTM models are highly dependent on the quality and diversity of training data. In climate change scenarios, caution is needed when applying LSTM models to unfamiliar conditions, as their predictive accuracy may decline more rapidly than that of more traditional hydrological models.