Where the cost sits
Waste operations are dominated by vehicle and crew costs. Anything that reduces distance travelled, empty collections or return visits goes almost directly to the bottom line.
That makes collection efficiency the first place to look, and it is usually addressable with data the business already has from vehicle tracking and weighbridge records.
Predicting fill levels
Collecting a container that is a quarter full wastes a stop; missing one that overflowed creates a complaint and sometimes a fine. Predicting fill at the point of collection allows dynamic scheduling.
- Historical fill at collection, where weights are recorded per lift
- Site type and size - a restaurant fills differently from an office
- Day of week and seasonality, including holiday periods
- Weather, for open containers and garden waste
- Known events or site activity where available
Sensors give the strongest signal but cost money per container. A useful sequencing is to predict from history first, prove the savings, then fit sensors to the containers where the prediction is least certain.
Routing is an optimisation problem
Route planning is optimisation rather than machine learning, and the distinction matters when scoping. The prediction supplies inputs - which containers need collecting, how long each stop takes - and an optimiser builds the route.
Predicting service time per stop is where machine learning contributes most, and it is usually poorly estimated. Stops vary by access, container type, time of day and whether the crew has been there before, and a planner using an average produces routes that run late by mid-morning.
Contamination detection
Identifying the wrong material in a recycling stream is a genuine image classification problem, applied either at the vehicle or on a sorting line.
| Where | What it enables | What it needs |
|---|---|---|
| On the vehicle at lift | Feedback to the household or business | Cameras, connectivity, in-cab handling |
| At the sorting facility | Quality control, supplier feedback | Line-mounted cameras, lighting |
| On tipping | Load rejection decisions | Fast inference, operator workflow |
The hard part is labelled training images from your own operation. Generic datasets transfer poorly because lighting, angle and the local waste mix all differ. Budget for a labelling exercise on your own footage.
Tonnage forecasting
Forecasting arisings supports capacity planning, contract pricing and regulatory reporting. It behaves like most demand forecasting: strong seasonality, weather sensitivity for green waste, and step changes when contracts are won or lost.
Contract changes need handling explicitly rather than being left to the model. A new commercial contract is a known event with a known start date - it belongs in the forecast as an input, not as a surprise the model discovers three months late.
In waste, the cheapest collection is the one you correctly decided not to make.