Accurate discharge prediction in hydrological forecasting relies on robust modeling. This study investigates the Long Short-Term Memory (LSTM) model's performance, focusing on training dataset size and hydrometeorological patterns. Convolutional Neural Networks (CNNs) and Artificial Neural Networks (ANNs) are also considered for spatial and temporal dependencies. Data is drawn from the CAMELS-GB dataset (1975-2015, Saxons Lode, UK). Results show that LSTM performance varies, with years surrounding high water events (like 2004) performing poorly in training, and struggles in validation. Training with one year yields 23.03% NSE values above 0.7, but using three consecutive years improves this to 84.42%. Typological differences also affect model performance. This study reveals LSTM sensitivity to training periods, aiding the optimization of training duration for better discharge prediction accuracy. Future research will delve into year selection within typological clusters.