The main limitation of PMF is the modeling of static factor profiles that stands in contradiction with a dynamic system, where source profiles are supposed to vary over time. To account for a possible temporal variation of the factor profiles a rolling approach can be used. In this strategy a PMF window is moving/rolling it over the entire dataset, allowing the PMF factor profiles to adapt over time. For running the rolling PMF approach efficiently, the possible present sources need to be roughly known.
A very general rolling PMF strategy could be:
- Seasonal (conventional) PMF: Know which sources are present during which season. Start with unconstrained PMF and if primary sources are not well separated, also run constrained PMF. In case factor profiles need to be constrained, optimize the factor profile so it fits best the current dataset (e.g. use the random a-value and bootstrap technique approach). Different number of factors during different seasons are fine.
- Rolling PMF: Set the number of factors based on the seasonal solution. It is possible to have different number of factors. For constrained factors, the optimized factor profiles from the seasonal PMF analysis should be used. For the rolling PMF approach make use of all features in SoFi, use the bootstrap technique and for constraints the random a-value approach. Set the rolling PMF window length and shift as well as the PMF repeats per window.
- Define environmentally reasonable solutions: Use the criteria approach to define environmentally reasonable solutions. Use criteria based on time series or profile features, correlations with tracers, trends, …
- Average: average the environmentally reasonable solutions and have access to statistical information on your solution.