L-GRU: A novel hybrid AI framework for improving time series forecasting

Abstract

Recurrent neural network (RNN) architectures such as LSTM and GRU have demonstrated remarkable effectiveness in modeling sequential data. However, LSTM possesses a complex structure with three distinct gating mechanisms, resulting in high computational cost, whereas GRU, though more streamlined, merges the hidden state and cell state, which may reduce its ability to manage long-term memory. This paper proposes the Long-term Gated Recurrent Unit (L-GRU), a novel hybrid architecture designed to overcome these limitations by integrating the strengths of both models. Experiments were conducted on four diverse datasets, including three financial datasets (Tesla stock, S&P 500 index, Bitcoin) and one meteorological dataset (Jena Climate) with an extended evaluation period of over 580 days to verify long-term stability. The results indicate that L-GRU achieves superior and stable performance across all datasets, outperforming traditional LSTM, GRU, and Transformer models, and particularly demonstrating strong generalization capability on both stochastic and physically structured data.