Forecasting a bin so trucks stop driving to empty ones
Forecasting how fast a waste container fills, so collection trucks stop driving to bins that aren't full.
KOCO's existing forecast was an adaptive linear regression: average the fill rate between two collections and extrapolate. It works, but it only ever sees data from collection days, a handful of points a week. The company wanted to know whether eighteen months of sparse weighings could support something better.
Research engineer. Earlier work existed when I arrived; I redid most of it, designed the evaluation, and carried the results in front of the client.
Filter the export down to a single chip, collection point and waste type. The models are per-container, not global.
Each collection becomes a linear segment: divide the weight gained by the days elapsed, and interpolate between measurements to get a kg/day series.
Take the log of the target so one enormous weighing can't dominate, hold back the last stretch as a test set, and fit each model family.
Compare every model against the incumbent on all interpolated points, and again on real collection days only.
A dockerised train-and-infer pipeline driven by a config file, plus a separate analysis toolkit.
One fit/predict surface across statistical and neural models, so the comparison is like for like
Reads the Excel export the client actually sends, unchanged
Backed the RNN, LSTM and Transformer experiments that ultimately lost
Every knob lives in config rather than code, so retraining doesn't need a developer
Training and inference as independently switchable services
A second deliverable: seasonality, measurement intervals and trends per waste type, filterable
Collection days only, the number a dispatcher actually acts on. Mean absolute error and mean absolute percentage error, on the container the final configuration ships pointed at.
| KOCO's linear model | Prophet | FFT | RNN | LSTM | Transformer | |
|---|---|---|---|---|---|---|
| Mean absolute error | 75.0 kg | 48.04 kg (best) | 60.69 kg | 92.37 kg | 61.39 kg | 150.31 kg |
| Mean absolute % error | 40.42% | 20.8% (best) | 30.16% | 36.7% | 27.68% | 57.93% |
| Root mean squared error | 94.93 kg | 65.43 kg (best) | 75.31 kg | 112.57 kg | 76.93 kg | 164.45 kg |
instead of · The finer hourly resolution originally agreed with the client
At hourly resolution about 98% of the series was invented by interpolation, and nothing in the field confirmed that a bin fills linearly within a day. Moving to daily didn't make the model smarter. It made the honest fraction of the data go from 2% to 26%. Choosing the coarser option is what made the numbers trustworthy.
instead of · A single train/test score, which is what any standard loop gives you
Scoring on all points measures agreement with an interpolation we invented ourselves. A model can look good by being good at reproducing a straight line. Scoring on collection days measures the thing the dispatcher acts on. Running both is what exposed the difference.
instead of · Pushing the RNN, LSTM and Transformer harder with more capacity and a real hyperparameter search
They cost the same to train and run and delivered less. With a few hundred daily points per container there is nothing for a large network to find, and the structure that does exist, an annual cycle and a weekly cadence, is exactly what an additive decomposition and a Fourier transform are built for.
instead of · Handing over the notebooks, which is what the first year produced
A model that only its author can retrain is a consultancy dependency, not a deliverable. Config-driven training and inference as separate services means they can retrain monthly without us.
KOCO Solutions sells software for waste collection, and their existing forecast was an adaptive linear regression: average the fill rate between two collections and extrapolate. It works, but it only ever sees data from collection days, which is a handful of points a week.
The first problem wasn't modelling, it was shape. Weights recorded at collection points had to become a continuous series before any time-series model could touch them, so we converted the measurements into a chain of linear regressions and interpolated between them to get a fill rate in kg/day. We started at kg/hour, and abandoned it: only 2.15% of those points were real measurements against 25.6% at daily resolution, and there was no field evidence that filling is linear within a day. Choosing the coarser resolution was the decision that made the numbers trustworthy.
Then we put five model families against the incumbent: Prophet, FFT, RNN, LSTM and Transformers. Every model faced the same two-way protocol: one score on every interpolated point, one on real collection days only. The first measures whether the curve is right, the second whether it is useful on the day a dispatcher makes a decision.
The deep models lost. RNNs, LSTMs and Transformers landed at or slightly below KOCO's linear baseline while costing the same to train and run. Prophet and FFT won, and Prophet won clearly: 48.04 kg mean absolute error against the incumbent's 75.0 kg, and 19.80% MAPE against 40.42%. The remaining recommendations came out of exploratory analysis, not taste: outliers were doing real damage, so a log transform and splitting containers by size went in.
It shipped as something they can run without us: a dockerised train-and-infer pipeline configured by TOML, plus a modular analysis toolkit for seasonality, measurement intervals and trends per waste type. We also answered the operational questions a rollout decision hangs on. Roughly four months of history at one to three measurements a week is enough to train, and for a new container at an existing site, a model from a same-type neighbour lands within ±5%.
Both partners rated the project 1 of 4, very successful, in the final report to Innosuisse.