Towards Robust Multivariate Time Series Analysis: A Novel Attention-based Deep Learning Framework

Multivariate Time Series (MTS) forecasting predicts future values of interdependent variables using historical data.MTS forecasting considers relationships between multiple variables that are often correlated, this demands models that capture these relationships effectively which is complex and computationally intensive.To address these challenges, we propose a novel framework that enhances MTS forecasting by leveraging advanced pattern recognition and feature selection techniques.The framework integrates an Exponential Identity Maxout Bidirectional Long Short-Term Memory (EIM_BiLSTM) model with a Covariance-based Context Vector Hierarchical Attention Network (CCV_HAN) to effectively capture intricate temporal and structural patterns.CCV_HAN is employed to model the complex patterns, represented via an adjacency matrix that evaluates structural similarities across the data.We introduce the Spontaneous Mutation Correlation-based Osprey Optimization Algorithm (SMC_OOA) for precise and effective feature selection.The framework leverages the Levenshtein Distance Jaccard KMeans (LDJ_KMeans) clustering technique to meaningfully group similar data points.Feature extraction is efficiently handled by the Autonomous Feature Retrieving Autoregressive Network (AFRAN).The EIM_BiLSTM model is then used to predict future values based on these selected features.Extensive experimentation on four standard datasets, viz., Stock Market India, Beijing Multi-Site Air Quality, Hungarian Chickenpox Cases and Appliances Energy Prediction demonstrates that the proposed framework achieves the best prediction accuracy around 97% compared to State-of-Art techniques including Transformers and Graph Neural Networks and significantly outperforms existing methods.This framework offers a robust solution for MTS prediction, with potential applications across diverse fields such as finance, energy, and healthcare.