Comparative Streamflow Forecasting Using LSTM, GRU, and 1D-CNN Models with PSO-Based Hyperparameter Optimization and MANOVA Analysis

Daily flow estimation is very important for situations such as sudden floods, low flows, and drought estimation. Therefore, daily flow estimation should be performed for the planning and protection of water resources. The existence of numerous methods that can be used for daily flow estimation makes it necessary to find the best method. In this study, three deep learning methods (Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and one-dimensional Convolutional Neural Network (1D-CNN)) were compared using 31 years of observation data for daily flow estimation. These configurations were found using a complete factorial design. Performance was evaluated using standard hydrological metrics, and the influence of hyperparameters and model types was statistically examined through multivariate analysis of variance (MANOVA). The results show that PSO-based hyperparameter tuning significantly improves prediction accuracy across all models, with the number of hidden units proving to be the most influential parameter. Notably, models with 64 hidden units consistently outperformed those with 32, while further increases to 128 units yielded no additional benefit. All models achieved comparable performance when optimised, emphasising the critical role of rigorous hyperparameter selection over architectural preference. Thus, the success of the methods could be statistically evaluated. The study provides compelling evidence for integrating deep learning and metaheuristic optimization in streamflow prediction, along with valuable insights for future hydrological modelling efforts.