DAFA-BiLSTM: Deep Autoregression Feature Augmented Bidirectional LSTM network for time series prediction

Area of Science:

• Machine Learning
• Time Series Analysis
• Deep Learning

Background:

• Time series forecasting is crucial for real-world applications, but conventional models struggle with complex temporal dependencies and spatial correlations.
• Long short-term memory (LSTM) networks offer improved sequential data handling, yet shallow architectures limit their ability to extract transient characteristics from long intervals.
• Existing methods often fail to fully exploit latent spatial dependence between variables, leading to suboptimal forecasting performance.

Purpose of the Study:

• To propose a novel deep bidirectional LSTM architecture, the deep autoregression feature augmented bidirectional LSTM (DAFA-BiLSTM) network, for enhanced time series prediction.
• To improve the extraction of transient characteristics and spatial dependencies in sequential datasets.
• To demonstrate the superiority and robustness of the proposed DAFA-BiLSTM model in diverse real-world time series forecasting tasks.

Main Methods:

• A vector autoregression (VA) transformation module is employed to represent time-delayed linear and nonlinear properties of input signals in an unsupervised manner.
• Learned nonlinear vectors from VA are progressively fed into multiple BiLSTM layers.
• Outputs from preceding BiLSTM layers are augmented with time-delayed VA vectors to create enhanced input signals for subsequent layers.

Main Results:

• The proposed DAFA-BiLSTM model demonstrates superior performance and robustness compared to conventional time series forecasting methods.
• Extensive experiments on real-world datasets validate the model's effectiveness in capturing complex temporal dynamics.
• Statistical analysis confirms the DAFA-BiLSTM's adaptive performance, even in noisy environmental conditions.

Conclusions:

• The novel DAFA-BiLSTM architecture effectively addresses the limitations of conventional time series models by better exploiting spatial and temporal dependencies.
• The proposed model offers significant improvements in forecasting accuracy and robustness, making it suitable for complex real-world scenarios.
• DAFA-BiLSTM provides a powerful new tool for time series prediction, particularly for datasets with long intervals and noisy characteristics.