Multi-Domain Features Enhanced Time Series Prediction Framework Based on Bi-Directional Lstm and Graph Attention Network

Most time series in many real-world applications exhibit irregular, non-stationary patterns that are not well handled by traditional methods. Accurate identification of critical events within such irregular time series is therefore a major challenge. We propose a multi-domain feature-enhanced time series classification framework based on graph attention networks and bidirectional LSTM. In our framework, the node feature generation module extracts time-domain statistics, frequencydomain components, and complex-network characteristics; their interrelations are fused via a multi-head attention mechanism. For relational modeling, we construct a spatial adjacency matrix from the intrinsic topology of measurement channels. Node features and the adjacency matrix are fed into a graph attention network to capture spatial dependencies, while the raw sequences are processed by a bidirectional LSTM to model temporal dynamics. Finally, spatial and temporal representations are integrated to perform binary classification on time series segments. Experimental validation on the CHB-MIT benchmark reveals that our proposed approach attains superior performance in irregular time series classification.