QPCA-CLT Net: An Integrated CNN-LSTM-Transformer Framework with Progressive Dimensionality Reduction and Trend Smoothing for Enhanced Multivariate Time Series Forecasting

Forecasting Multivariate Time Series (MTS) data with high accuracy is desired for various decision-making tasks in meteorology, healthcare, and financial analytics. However, traditional statistical models involving ARIMA, VAR, and Holt-Winters cannot capture the true nonlinear dependencies, high dimensionality, and long-range temporal correlations present in real-world MTS data. In this paper, a novel hybrid deep learning framework called QPCA-CLT Net is proposed, which combines Quartet Principal Component Analysis (QPCA), Convolutional Neural Networks, Long Short-Term Memory (LSTM) networks, and Transformers. It follows a multi-stage QPCA strategy where feature dimensions (from n → 4 → 2 → 1) are iteratively reduced while retaining maximal explained variance to preserve the critical temporal information. The preprocessing stage consists of decomposing and stabilizing the trend, seasonal, and level components by using Holt-Winters Exponential Smoothing (HWES) and Exponential Moving Average (EMA) operators to mitigate the problem of non-stationarity. Further, Convolutional Neural Network (CNN) layers extract localized temporal features, stacked LSTMs model the sequential dependencies, and Transformer blocks capture global contextual relations using Multi-Head Self-Attention. QPCA-CLT Net consistently outperformed classical and state-of-the-art deep models with 10–20% less RMSE and MAE and higher R² scores on validation using Air Pollution and ERP datasets. This can be attributed to enhanced scalability, interpretability, and robustness in the proposed model, a promising framework for high-dimensional MTS forecasting applications.