Highly accurate forecasting of particulate matter concentration (PMC) is essential and effective for establishing a reliable air pollution early warning system and has both theoretical and practical significance. To meet this demand, a novel multi-scale hybrid learning framework based on robust local mean decomposition (RLMD) and moving window (MW) ensemble strategy is developed for PM2.5 and PM10 forecasting. In this architecture, the RLMD is adopted to adaptively decompose the PMC time series (PMCTS) into several production functions and one residue with different frequencies. These subseries are simpler than the original PMCTS, but they still work alongside mode aliasing. Thus, following the well-established “linear and nonlinear” modeling philosophy, a novel hybrid learning framework, composed of the autoregressive integrated moving average (ARIMA) and combined kernel function relevance vector machine (RVMcom), is proposed to capture both the linear and nonlinear patterns in the subseries. To obtain better final outputs, based on the definition of the ensemble improvement degree, the MW ensemble method is used to merge the forecasting results of all subseries. A comprehensive experiment is conducted using PM2.5 and PM10 datasets from four municipalities in China to investigate the forecasting performance of our proposed framework, and the results demonstrate that our proposed RLMD-ARIMA–RVMcom-MW (R-A&Rcom-M) model is superior to other considered methods in terms of forecasting accuracy and generalization ability. This means that the developed forecasting architecture has a great application value in the field of PMCTS prediction.