Air pollution is one of the key challenges to both human health and our environment, and managing it requires collective systematic efforts to prevent and mitigate future effects. Fundamentally, this required a better understanding of sources that generate pollution and forecasting models to predict current and future air pollution levels. In this work, we investigated features inspired PM2.5 prediction based on a dataset collected in Northern Ireland, UK. We analysed the influence of different features available in the dataset and newly generated with approaches such as Variational Mode Decomposition (VMD) and evaluated single-step forecasting model performance. We found that a single Long Short Term Memory (LSTM) layer model with a small number of cells and integrated features are sufficient to achieve a good forecasting performance. The combination of VMD integrated features enabled the forecasting model to achieve R2 score over 85% and achieve a gain of 6% when compared with lag based prediction only.