Smart Trash Bin Prediction

ISEF Category: Environmental Engineering

Subcategory: Recycling and Waste Management  ·  Difficulty: Intermediate  ·  Setup: School Lab  ·  Time: 1 to 2 Months

The Hook

A full trash bin does not warn you before it overflows. That makes pickup schedules messy and wasteful. If you can predict fill level early, you can help crews skip half-empty bins and catch the full ones on time. That is a real engineering problem with a clean data science twist.

What Is It?

This project asks a simple question, can you predict how full a trash bin will be before it gets there? You use an ultrasonic sensor, which sends out sound waves and measures how long they take to bounce back. That time tells you the distance to the trash surface. As the bin fills, the distance shrinks, like watching a bathtub drain in reverse.

The second part is the model. An LSTM, or long short-term memory model, is a type of neural network that looks for patterns over time. Instead of only reacting to one sensor reading, it can learn how fill levels change across days, weekdays, event days, or lunch periods. That makes it a good fit for prediction, not just measurement.

Why This Is a Good Topic

This is a strong science fair topic because you can measure a real-world process, collect your own data, and test whether prediction beats simple guessing. It connects to waste collection, route planning, and cleaner campuses. You can learn sensor calibration, time-series data, and model evaluation without needing a university lab.

Research Questions

• How does using an LSTM compare with a simple moving average for predicting trash-bin fill level?
• What is the effect of bin location on the accuracy of fill-level predictions?
• Does adding day-of-week information improve prediction of school trash-bin use?
• To what extent can one sensor per bin estimate fill level for bins with different shapes?
• Which feature set, past fill level alone or fill level plus school schedule data, gives the lowest prediction error?
• How does sensor placement inside the bin change the stability of ultrasonic distance readings?

Basic Materials

• Ultrasonic distance sensor module compatible with a microcontroller.
• Microcontroller board such as Arduino or Raspberry Pi Pico.
• Breadboard and jumper wires.
• Trash bin or labeled test container.
• Meter stick or tape measure for calibration.
• Notebook or spreadsheet for recording fill measurements.
• Assorted safe test materials such as paper, clean packing material, or plastic cups for simulated waste.
• Laptop for data logging and model training.
• Cardboard or mounting bracket for fixed sensor placement.
• Digital camera or phone for documenting bin setup.

Advanced Materials

• Ultrasonic sensor array for multiple sensing positions.
• Microcontroller with data logging support and timestamp capture.
• Load cell and amplifier for weight-based comparison data.
• Depth gauge or laser distance sensor for ground-truth validation.
• GPS or route data if studying pickup optimization.
• Python environment with ML libraries.
• Local server or single-board computer for edge inference tests.
• Protective enclosure for field deployment.
• Thermal or humidity sensor if you want to test environmental effects on readings.
• Access to multiple bins for cross-site comparison.

Software & Tools

• Python: Cleans the sensor data, builds features, and trains your prediction model.
• Pandas: Organizes timestamped fill-level records into tables you can analyze.
• Matplotlib: Plots fill trends, errors, and model comparisons over time.
• Scikit-learn: Tests baseline models and calculates error metrics.
• TensorFlow: Trains an LSTM model for time-series prediction.

Experiment Steps

1. Define the exact prediction target, such as current fill level, next-day fill level, or time until full.
2. Choose one bin design and one sensor placement so your first dataset stays consistent.
3. Plan a calibration method that turns raw distance readings into a fill-level percentage or volume estimate.
4. Build a baseline model first, then compare the LSTM against a simpler time-series method.
5. Decide what extra inputs you will test, such as day of week, event days, or bin location.
6. Plan evaluation metrics before you collect data, so you can compare models fairly.

Common Pitfalls

• Mounting the sensor loosely, which changes the reading every time the bin shifts.
• Mixing bin types with different shapes, which makes one fill model fail across containers.
• Using only one cleanup pattern, which gives the model too little variation to learn real trends.
• Letting reflective trash surfaces confuse the ultrasonic signal, which creates sudden false spikes.
• Training and testing on the same time period, which makes the model look better than it really is.

What Makes This Competitive

A strong version of this project goes beyond basic prediction. You can compare multiple baselines, test whether schedule data improves accuracy, and check how well the model transfers to another bin or another campus area. You can also frame the work as an operations problem, not just a coding task, by estimating how much pickup time or overflow risk your model could reduce.

Project Variations

• Test whether a single-bin model works better than a multi-bin model across different school locations.
• Compare ultrasonic sensing with load-cell based fill estimation for the same trash bin.
• Study whether event days, lunch periods, or weather data improve bin fill prediction.

Learn More

• MIT OpenCourseWare: Search for free lectures on machine learning, neural networks, and time-series modeling.
• TensorFlow Tutorials: Find beginner-friendly guides for sequence models and LSTMs in the official TensorFlow documentation.
• US EPA Sustainable Materials Management: Search for school waste and diversion resources to connect your project to real waste problems.
• USGS Science Data and Tools: Use open data and methods pages to practice data handling and environmental measurement thinking.
• PubMed: Search for review articles on smart waste management, ultrasonic sensing, and fill-level prediction.