<p>The prediction and control of Particulate Matter 2.5 concentrations near tailings dams have become an essential concern in the fields of environmental protection and public health. This study examines Particulate Matter 2.5 concentrations near the Chengmendong tailings dam located in Ma’anshan City. It presents a multi-site Particulate Matter 2.5 prediction model combining the hybrid Whale Optimization Algorithm and Particle Swarm Optimization Algorithm, along with Empirical Mode Decomposition, and reconstruction. The Particulate Matter 2.5 concentrations are influenced not only by antecedent values but also by subsequent measurements. The Bidirectional Long Short-Term Memory is proficient in capturing both forward and backward temporal patterns, thereby facilitating the extraction of periodic features in Particulate Matter 2.5 concentrations data. Beyond meteorological influences, the differential weighting of impacts from correlated monitoring stations in proximity to the target station is also considered. The Attention mechanism was introduced to allocate weights effectively. In addition, an error correction model that integrates statistical methods with Empirical Mode Decomposition and reconstruction is used to decompose and reconstruct the error sequence of initial prediction. The predicted errors are subsequently incorporated into prediction results. The best parameter setting for the Bidirectional Long Short-Term Memory-Attention model were determined using the hybrid optimization algorithm, leading to an optimized prediction model. Experimental results indicate that the proposed model achieves Root Mean Square Error, Mean Absolute Error, and Mean Absolute Percentage Error values of 12.33, 7.70, and 0.12, respectively. Compared to benchmark models, this newly proposed model provides more accurate predictions for high-concentration Particulate Matter 2.5 prediction.</p>