ZPF-DLSTM: An Efficient Deep Network with Low Time Latency

Abstract

Accurate prediction of time series is crucial for various applications in finance, energy, transportation, and other fields. This article proposes an improved LSTM model which uses fully-connected layers to extract long-distance features, enhances the basic LSTM network, and ultimately improves the fitting and delayed prediction capabilities of the model. Furthermore, this paper draw upon knowledge from the communication field and introduce zero-phase filters during data processing to smooth our time series data and establish the ZPF-DLSTM network structure. Our approach aims to handle various delay patterns and effectively model real-time accurate time series data. This paper evaluate our method on three benchmark datasets and demonstrate its superiority in terms of accuracy and latency performance over existing commonly used methods. The experiments show that proposed method can serve as a valuable tool for accurate time series prediction in practical applications.