Multi-scale integration framework integrating EEMD with hybrid networks for indoor PM2.5 forecasting

This study introduces a deep learning-based soft-sensor framework integrating ensemble empirical mode decomposition, average Hurst exponent analysis, and multi-scale prediction for PM2.5 early warning. The proposed methodology proceeds in three stages: first, ensemble empirical mode decomposition is employed to decompose raw PM2.5 time series into multiple intrinsic mode functions representing distinct frequency components; second, average Hurst exponent is to compute each timescale to determine the most suitable optimal prediction model for subsequent forecasting tasks; finally, PM2.5 concentration predictions are generated by aggregating forecasts across all decomposed timescales. Experimental results demonstrate that the hybrid model outperforms reference models in extracting discriminative features and modelling complex temporal dependencies within PM2.5 datasets. Specifically, its coefficient of determination exceeds those of the comparative models―convolutional neural network, gated recurrent unit and its versions implemented with empirical mode decomposition and ensemble empirical mode decomposition methods―by 33.8%, 20.29%, 10.67%, and 1.12%, respectively, highlighting the effectiveness of multi-scale feature fusion in enhancing prediction accuracy. With the more effective and accurate soft-sensor model proposed, the research offers a novel perspective on bridging process safety engineering with environmental sustainability, demonstrate how data-driven intelligence can address health-protective problems better and provides a feasible suggestion for resource-efficient infrastructure operations.