Experiments

Our analysis was designed to be rigorous and consultant of real-world utilization. For each the 30-minute and 60-minute time horizons, we evaluated predictions on 100 randomly chosen stations, sampling their occupancy standing 48 instances each day (each half-hour) for a full week.

The mannequin was benchmarked in opposition to a remarkably sturdy baseline: the “Preserve Present State” strategy. This baseline merely assumes that the variety of accessible ports a sure variety of minutes (H) sooner or later might be precisely the identical as the present quantity.

Whereas easy, this baseline could be very laborious to beat, particularly over quick horizons. For instance, our information confirmed that on the US East Coast, by no means greater than 10% of ports change their availability state inside a 30-minute block. Since more often than not the state would not change, the only prediction — no change — is right more often than not, making the duty of including predictive worth extraordinarily troublesome.

We targeted on two key metrics to measure the mannequin’s accuracy for predicting the precise variety of free ports: imply squared error (MSE) and imply absolute error (MAE). A ratio of MSE/MAE ≥ 1 free port measures the accuracy of essentially the most vital binary job for the consumer: “Will I discover at the least one free port (Sure/No)?”