A new hybrid model and its application for forecasting of daily PM2.5 concentrations

The PM2.5 concentration serves as a pivotal indicator in air quality assessment, inherently displaying nonlinearity, irregularity, and uncertainty. So, this study introduces a hybrid forecasting framework grounded on the Synchrosqueezing Wavelet Transform (SST), Sparrow Search Algorithm (SSA), and Bidirectional Long-Short Term Memory (BiLSTM) model. In the initial stages, the SST technique is applied to denoise raw data, setting it up for the forecasting model input. Thereafter, the core parameters of the BiLSTM model are refined using a global search mechanism via SSA. This addresses issues of slow convergence and local optima that plague conventional parameter training algorithms. Our SST-SSA-BiLSTM model targets the prediction of daily PM2.5 concentrations in cities like Chengdu, Kunming, Shenyang, and Nanchang, and specifies a prediction interval for each forecast. Evaluation based on several benchmarks, including the Diebold-Mariano test, shows our proposed hybrid model surpasses competing models in both prediction accuracy and error minimization.