Big Belly Smart Bin Forecasting

Forecasting system for UC Berkeley's Big Belly smart bin fleet. Predicts time-to-collection-threshold per bin, classifies collection frequency, and produces a rolling weekly operational forecast with calibrated uncertainty bounds.

---

Problem

UC Berkeley's Cal Zero Waste division manages ~248 Big Belly compactor bins across 90 campus locations. Collection is triggered reactively — either when a bin reaches 60% capacity or after 7 days. Analysis of 59,111 collection events (Jan 2023 – Feb 2026) shows 35.1% of all collections are wasted trips, with average bin fullness at collection of 47.1%. The goal is to predict when each bin will next need service so that dispatch can shift from reactive alert-response to proactive scheduling.

---

Dataset

• Source: Big Belly CLEAN Management Console (export-only; raw files excluded from repo)
• Size: 59,111 collection events, 248 bins, 90 locations, 3 waste streams
• Date range: January 2023 – February 2026
• Features per event: bin serial, location, stream type (Landfill / Compostables / Recyclables), collection reason, fill level at collection (20% increments), timestamp
• Key constraint: fill level is only observable at collection time, in coarse 20% increments — no continuous telemetry

Supplementary features built from external sources:

• UC Berkeley academic calendar (semester dates, spring recess, move-in week)
• Cal Bears football home game schedule (2023–2025)
• Daily weather via Open-Meteo API (temperature, precipitation, wind)

---

Models

Fill-Curve Model (analytical)

Per-bin nonlinear curve fitting on (elapsed hours, fill level at collection) pairs. Sigmoid selected for 82% of bins. Analytically solves for T* = time until bin reaches 60%. Used for near-term urgency assessment (0–48h window).

• 5 curve types evaluated: sigmoid, saturating exponential, power law, logarithmic, linear
• Mean R² = 0.23 (reflects 20% quantization noise, not model inadequacy)
• Outputs: hours_to_full, frequency_class (multi-visit / once-a-week / rarely-full)

LightGBM Quantile Regression

Cross-bin tabular model. Target: hours_to_next_collection. Produces P10/P50/P90 estimates per bin.

Model	MAE (hours)	RMSE (hours)
Naive baseline (DOW median)	48.3	65.8
LightGBM Q50	40.3	54.3

Calibration: Raw P10–P90 coverage was 59% (target: 80%). Applied split conformal calibration using 2024 H2 as a held-out calibration set. Correction factor: +/- 17 hours. Post-calibration coverage: 80.8%.

XGBoost Binary Classifier

Predicts: is this bin currently at or above 60% fill? Class imbalance handled via scale_pos_weight. Threshold tuned for recall.

Threshold	AUC	Precision	Recall
0.50 (default)	0.857	0.77	0.74
0.33 (recall-optimized)	0.857	0.70	0.85

Prophet (aggregate demand)

Daily fleet-level collection count forecast by stream type. Weekly + yearly seasonality, 4 exogenous regressors.

Stream	MAE (collections/day)	RMSE
Landfill	7.7	10.2
Compostables	6.8	8.6
Recyclables	6.9	9.1

Forecast Reliability by Horizon (LightGBM)

Horizon	AUC	Precision	Recall
24h	0.737	0.702	0.122
48h	0.771	0.815	0.371
72h	0.809	0.870	0.579
1 week	0.871	0.904	0.948

Short-horizon recall is structurally limited by the data collection design. Fill-curve model used for 0–48h window; LightGBM used for 72h+ planning.

---

Wasted Trip Analysis

The 35.1% wasted trip rate (20,742 events) is decomposed into four categories:

Category	Count	% of Wasted	Eliminable
Pre-event staging	4,761	23.0%	Partially
Frequency mismatch	3,736	18.0%	Yes
Unexplained	12,245	59.0%	Yes
Age-rule forced	0	0.0%	No

~77% of wasted trips are potentially eliminable. The unexplained rate has grown from ~8% of all collections in 2023 Q3 to ~29% in early 2026.

---

Repo Structure

bigbelly-forecasting/
├── notebooks/
│   ├── 01_eda.ipynb                      # Exploratory data analysis, wasted trip decomposition
│   ├── 02_fillcurve_forecast.ipynb       # Per-bin nonlinear curve fitting, T*, frequency classification
│   ├── 03_modeling.ipynb                 # All ML models: naive baseline, Prophet, XGBoost, LightGBM quantile; per-bin error analysis; enriched quantile output
│   ├── 04_calibration.ipynb              # Split conformal calibration; computes q_hat = 16.98 hours
│   └── 05_operational_forecast.ipynb     # Rolling daily forecast, urgency bands, calibrated output table
├── data/
│   ├── README.md                         # Data dictionary; raw files excluded
│   └── bin_fill_curve_coefficients.csv   # Per-bin curve coefficients (pre-computed)
├── report/
│   └── main.tex                          # LaTeX source for final report
├── .gitignore
└── README.md

---

How to Run

Requirements: Python 3.10+, Jupyter

pip install lightgbm xgboost prophet openmeteo-requests requests-cache retry-requests pandas numpy scikit-learn matplotlib

Data setup: Place the three CLEAN export CSVs in the data/ directory. Files are named:

Daily Collection Activity - CLEAN 2024 (1).csv
Daily Collection Activity - CLEAN 2025 (1).csv
Daily Collection Activity - CLEAN 2026 (1).csv

Each file has a 10-row metadata header — the notebooks handle this automatically via skiprows=10.

Execution order:

1. 01_eda.ipynb — standalone EDA; no dependencies on other notebooks
2. 02_fillcurve_forecast.ipynb — generates bin_fill_curve_coefficients.csv
3. 03_modeling.ipynb — trains all models; generates bin_quantile_forecasts_enriched.csv and per_bin_error_q50.csv
4. 04_calibration.ipynb — computes q_hat correction factor; run after 03
5. 05_operational_forecast.ipynb — loads all outputs; applies q_hat; produces operational_forecast_full.csv

Weather data is fetched automatically from the Open-Meteo API on first run and cached locally.

---

Key Results Summary

• 35.1% of collections are wasted trips; ~77% are potentially eliminable
• LightGBM Q50 reduces prediction error 16% over naive baseline (MAE 40.3h vs 48.3h)
• XGBoost binary classifier: AUC 0.857, recall 0.85 at threshold 0.33
• Weekly planning recall: 94.8% — reliable basis for 7-day scheduling
• P10/P90 uncertainty bounds calibrated to 80.8% empirical coverage via split conformal calibration
• Frequency classification: 25% of bins need multiple collections/week; 59% need ~1/week

Urgency Band Definitions

Label	Condition
OVERDUE	Current fill ≥ 60% — collect immediately
MONITOR	No curve fit, or curve asymptote < 60% — bin rarely reaches threshold
HIGH	Calibrated P10 ≤ 1 day — collect within 24 hours
ELEVATED	Curve estimate ≤ 3 days
MODERATE	Curve estimate ≤ 5 days
LOW	Curve estimate > 5 days — no collection needed this week

Conditions evaluated in priority order top to bottom.

Acknowledgments

This project was completed as part of IND ENG 180 (Industrial Engineering Capstone) at UC Berkeley. We thank Samantha Bunke (Zero Waste Specialist, Cal Zero Waste) for providing data access and operational guidance, and Professor Candace Yano for advising the project.