Hybrid deep-learning model shows favorable hourly PM2.5 prediction performance in modeling study

A hybrid deep-learning framework reported favorable prediction accuracy and generalization performance for hourly fine particulate matter concentrations in a modeling study summarized in a Scientific Reports abstract. The framework combines decomposition methods with three time-series sub-models and uses a metaheuristic strategy to improve Adam optimization of selected model parameters.

The proposed IMSTL-FVMD-HLDT-SRA Adam framework first applies IMSTL decomposition to divide a PM2.5 concentration series into a trend component, multiple seasonal components and a residual component. The residual is then divided further through FVMD decomposition. The trend, combined seasonal and combined residual components are modeled independently by LSTM, DLinear and TimesNet sub-models, respectively. Their predicted components are added to reconstruct the final PM2.5 concentration estimate.

During training, the authors used a Schrödinger-search-based metaheuristic optimization strategy to strengthen the global search capability of the Adam optimizer. According to the abstract, the strategy optimized output-layer parameters in the DLinear and TimesNet sub-models with the aim of reducing prediction errors. The abstract reports that the resulting framework achieved favorable prediction accuracy and generalization performance, although it does not provide the underlying dataset, monitoring location, study period, sample size, comparison methods or numerical performance measures.

Because only the abstract is available in the supplied evidence, the scope and limitations of the modeling study cannot be assessed in detail. The summary does not report uncertainty intervals, error metrics, baseline models, validation procedures or evidence from independent locations or periods. It therefore does not establish whether the approach will maintain the reported performance in operational forecasting, nor does it demonstrate improvements in air-pollution early-warning or governance outcomes. The reported results concern model-predicted PM2.5 concentrations rather than measured effects of any intervention on air quality or health.

Subject of Research: Hourly PM2.5 concentration prediction using a hybrid decomposition and deep-learning framework

Article Title: A novel hourly PM2.5 concentration prediction model using IMSTL-FVMD-HLDT hybrid architecture with Adam optimized by Schrödinger search strategy

Article References: Zhao, X., Qiao, P., Li, H., Yang, S., Shi, Q., Qin, Q., & Chen, J. (2026). A novel hourly PM2.5 concentration prediction model using IMSTL-FVMD-HLDT hybrid architecture with Adam optimized by Schrödinger search strategy. Scientific Reports. https://doi.org/10.1038/s41598-026-72883-3

Image Credits: AI Generated

DOI: 10.1038/s41598-026-72883-3

Keywords: PM2.5, air-quality forecasting, deep learning, time-series decomposition, IMSTL, FVMD, LSTM, DLinear, TimesNet, Adam optimizer, novel, hourly