A Comprehensive Survey of SSA-LSTM Models: Signal Decomposition Techniques, Algorithmic Advances, Practical Applications, and Future Directions

Abstract

Achieving optimal long short-term memory (LSTM) performance is highly dependent on the correct selection of hyperparameters. Incorrectly choosing the number of layers, neurons, learning rate, or batch size can significantly reduce the model’s accuracy or even cause instability during training. In this regard, metaheuristic algorithms play a crucial role in optimizing hyperparameters. One of these algorithms is the sparrow search algorithm (SSA), which is inspired by the social behaviour of sparrows in their search for food and response to threats. Studies have shown that combining SSA with LSTM not only increases the prediction accuracy in time series problems and noisy data, but also significantly reduces the time required for human trial and error in parameter tuning. Another critical strategy for improving LSTM is the use of signal decomposition methods such as empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), complete EEMD with adaptive noise (CEEMDAN), and variational mode decomposition (VMD). These methods decompose non-stationary and noisy data into simpler components and then model each component separately using an LSTM. This paper presents a comprehensive analysis of 237 selected papers published between 2020 and 2025, collected from reputable international sources, including IEEE, Elsevier, Springer, and MDPI. The share of scientific publishers shows that IEEE and MDPI publish more than 60% of the papers. The primary purpose of this paper is to investigate the role of SSA and its improved versions in optimizing LSTM networks, analyzing signal decomposition, and reviewing the wide range of applications of SSA-LSTM in various fields, including industrial, medical, financial, and environmental sectors.