Fuel Dispatching for Haul Trucks

In an open-pit mine, dispatchers decide which haul truck to send for fuel mostly by intuition and radio calls. Pull the wrong truck at the wrong moment and you starve a shovel, quietly dragging down the productivity of the whole fleet. My MSc thesis turned that judgment call into an optimization problem.

It has two coupled parts.

1. Predict fuel consumption (ML). I benchmarked six regression models to predict liters consumed from effective flat-haul distance and tonnage. An artificial neural network won with $R^2 \approx 0.90$ ; multilinear regression was close and far cheaper, so it ships as a fast pre-trained fallback.

2. Schedule refueling (optimization). A binary integer program decides which truck fuels and when, maximizing the total truck–shovel match factor so refueling disrupts production as little as possible:

\max \sum_{t}\sum_{\tau} m_{t}\,y_{t,\tau}

where $y_{t,\tau} = 1$ if truck $t$ is sent to fuel in period $\tau$ . Constraints keep it physical: at most one truck fueling at a time, each truck fuels at most once, a 15-minute non-overlapping fueling window, and nobody runs dry.

Solved with Gurobi, it runs in under a second for a fleet of 57 trucks, improving the average match factor by about a point versus current practice. The whole pipeline is wrapped in a Streamlit app so a dispatcher can pick a shift, run the optimizer, and compare its plan against what the mine actually did.