Optimizing deep learning predictive models: A comprehensive review of RNN and its variant architectures
Andrew Adewunmi Adekunle, I. Fofana, Patrick Picher, Esperanza Mariela Rodriguez-Celis, Oscar Henry Arroyo-Fernàndez, Ryad Zemouri
Abstract
Accurate prediction of an engineering system behaviour is essential for ensuring a stable and secure long-term operation. It enables proactive problem solving, prevents disruption, enhances safety, and facilitates the seamless integration of new technologies such as digital twins. Consequently, several approaches have been employed to enhance system behaviour prediction by utilizing conventional machine learning models. Moreover, the advent of deep neural networks has proven to be more effective in several scenarios as they offer enhanced prediction accuracy and capacity in handling complex and high-dimensional data. Despite their advantages, deep neural networks encounter challenges in determining the optimal configuration for model structures. Therefore, various optimization techniques such as hyperparameter optimization, activation functions, framework search algorithms, algorithm optimizers, and hybrid frameworks have been proposed to mitigate these challenges. Hence, this study emphasizes recurrent neural networks and their variants, as one of the most popularly utilized frameworks for predictive algorithms. Also, several strategies and techniques for improving the performance of these predictive frameworks have been holistically discussed. By analyzing the state-of-the-art optimization approaches, it serves as a valuable resource for researchers, providing a comprehensive understanding of the approaches that can be employed to optimize prediction accuracy for specific applications and tasks. • Prognostics and health management frameworks predict remaining useful life in systems. • LSTM and GRU variants address RNN limitations like vanishing and exploding gradients. • BiLSTM leverages both forward and backward data flows, enhancing time-series prediction. • Transformer models use self-attention to efficiently capture long-term dependencies. • Hyperparameter tuning, activation functions, and optimization techniques are crucial for model accuracy.