Artificial neural networks, particularly recurrent architectures like Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are widely used for time series prediction. However, achieving high accuracy, especially for complex industrial process data, remains challenging, as standard models may not explicitly capture local temporal trends effectively. This paper proposes a novel data inference structure based on Propositional Linear Temporal Logic (PLTL) designed to capture qualitative data trends within a sliding window. This PLTL inference structure is integrated into LSTM and GRU networks, creating two new models: L-LSTM and L-GRU. The PLTL module provides an interpretable, logic-based representation of recent data dynamics, which modulates the standard recurrent computations, enabling the networks to learn temporal dependencies more effectively, as demonstrated by empirical results. The proposed methods are evaluated on the TAIEX financial benchmark dataset and real-world data from a textile manufacturing process. Experimental results, evaluated using Root Mean Square Error (RMSE), indicate that the L-LSTM and L-GRU models demonstrate statistically significant improved prediction accuracy compared to baseline LSTM, GRU, Deep Belief Networks (DBN), and a Fuzzy Time Series-LSTM (FTS-LSTM) hybrid model on the benchmark dataset, and show strong performance on the industrial data.
Huo et al. (Thu,) studied this question.