A plot of a larger sample of post-hoc known wind and other constraints over a regular service would help enormously to sell the idea.
One point at 45min faster and 25% less fuel is of course huge, but how does it map out over a year of routing?
One point at 45min faster and 25% less fuel is of course huge, but how does it map out over a year of routing?
One difference from the OpenAP/scikit-decide setup: my router doesn't search a fixed grid with A*. It runs a continuous isochrone search that expands a fan of headings against the live forecast each step. That's closer to how routing is actually done for this problem, and it's what both the live product and the backtest above use.