at-tan/Forecasting_Air_Pollution ? reverse-engineered prompt

Build me a notebook based project that predicts the next hour PM 2.5 air pollution level from the Beijing air quality dataset.

I want it to start with some basic exploration and cleaning of the data, then train several different models for time series forecasting, combine them into a stacked ensemble, and compare the results against a simple persistence baseline. Please keep the workflow focused on the actual data, handle missing values and obvious outliers reasonably, and split the data in time order so the latest chunk is used as a holdout test set.

It would be great if the notebooks show the full process clearly, from exploring the data to training, cross validation, and final evaluation with MAE and RMSE. If you need to check anything about the best way to set it up, look up current docs online.