🍽️ Forecasting School Meal Production Costs: A Comparative Study of Machine Learning and Deep Learning Time-Series Models

This project develops a time-series forecasting system for Fairfax County Public Schools (FCPS) to estimate daily meal production costs and analyze waste-related patterns. It compares multiple machine learning and deep learning models-including LSTM, GRU, XGBoost, Linear Regression, and Feed-Forward Neural Networks and visualizes the results through an interactive Streamlit dashboard.

The goal is simple: 👉 Reduce food waste, improve planning, and optimize meal production costs across the district.

🏷️ Badges

---

📋 Table of Contents

• Overview
• Dataset Workflow
• Key Features
• System Architecture
• Model Pipeline
• Getting Started
  • Prerequisites
  • Installation
  • Environment Setup
  • Running the Models
• Dashboard (Streamlit App)
• API Endpoints
• Troubleshooting
• Research & Performance
• Technology Stack
• Contributing
• License
• Acknowledgments

---

🔄 Dataset Workflow

Our pipeline transforms raw FCPS Production Records + POS data → clean, structured forecasting dataset.

---

1️⃣ HTML → CSV Parser

✔ Reads dozens of messy FCPS breakfast & lunch HTML files
✔ Auto-detects school sections
✔ Extracts production, leftover, planned, served, discarded values
✔ Cleans currencies, percentages, and item names
✔ Standardizes headers

Outputs generated:

• src/Data/Output/breakfast_combined.csv
• src/Data/Output/lunch_combined.csv
• src/Data/Output/meals_combined.csv

---

**2️⃣ Data Cleaning & Preprocessing **

✔ Cleans $ & % → float
✔ Converts & sorts dates
✔ Handles missing values
✔ Outlier removal using 99th percentile
✔ Encodes meal types
✔ Produces final ML-ready dataset for:

• Univariate Forecasting
• Multivariate Forecasting
• Streamlit Dashboard

---

⭐ Final Dataset Columns

Column	Description
school_name	FCPS school
meal_type	breakfast/lunch
date	daily record
served_total	meals served
planned_total	planned meals
discarded_total	wasted meals
left_over_total	leftover meals
production_cost_total	$$ spent per item-day

---

🎯 Key Features

🍽️ 1. Meal Demand Forecasting

• LSTM & GRU deep learning models
• Univariate forecasting (district-level daily time-series)
• Multivariate forecasting (served/planned/discarded/leftovers → cost)

---

♻️ 2. Waste Optimization

• Predict discarded + leftover quantities
• Waste ratio analytics
• Identify high-waste menu items

---

💲 3. Cost Forecasting

• Predict production cost for next 10 days
• Scenario modeling using “What-If” adjustments

---

📊 4. Interactive Streamlit Dashboard

• School-wise filtering
• Cost trends
• Waste ratio analysis
• What-if ML predictions
• Benchmark model comparison

---

🧠 5. Machine Learning Benchmarking

• Linear Regression
• XGBoost
• Feed-Forward Neural Network
• GRU & LSTM

---

🏗️ System Architecture

┌──────────────────────────┐
│    Raw FCPS HTML Files   │
└───────────────┬──────────┘
                │
     (HTML Parser + Normalizer)
                │
                ▼
┌──────────────────────────┐
│    meals_combined.csv    │
└───────────────┬──────────┘
                │
       (Data Preprocessing)
                │
   ┌────────────┼───────────────┬──────────────┐
   ▼            ▼               ▼
Univariate   Multivariate     Benchmark  
   LSTM          GRU           Models
   │             │               │
   └───────┬─────┴───────┬──────┘
           ▼             ▼
      Forecasts   Performance Charts
           │             │
           └───────┬────┘
                   ▼
        Streamlit Dashboard


---

🤖 Model Pipeline

📌 Univariate Forecasting (LSTM / GRU)

Uses district-wide daily production costs:

➡️ [Cost(t−7) … Cost(t−1)] → Predict Cost(t)

Models:

• LSTM
• GRU
• Feedforward baseline
• XGBoost
• Linear Regression

---

📌 Multivariate Forecasting

Features:

• served_total
• planned_total
• discarded_total
• left_over_total

Target:

• production_cost_total

Models:

• Linear Regression
• XGBoost
• FeedForwardNN
• GRU (sequence-based, school-wise)
• LSTM

---

🚀 Getting Started

✔️ Prerequisites

Module Installation

You can install the following modules through pip

pip install -r src/requirements.txt

Install:

• Python 3.10+
• pip
• PyTorch
• XGBoost
• pandas
• numpy
• beautifulsoup4
• lxml
• pdfplumber
• PyPDF2
• tqdm
• statsmodels
• scikit-learn
• matplotlib
• streamlit

🌱 Environment Setup

Important source files:

src/component/preprocess.py     
src/component/EDA.py            
src/component/univariate/        
src/component/multivariate/     
src/tests/combine_csv.py     
src/maincode/main.py         

HTML → CSV Preprocessing

python src/tests/combine_csv.py

This script:

• Reads FCPS breakfast & lunch HTML production records
• Extracts → served, planned, discarded, leftover, cost
• Cleans currency & % values
• Standardizes headers
• Generates:

src/Data/Output/breakfast_combined.csv
src/Data/Output/lunch_combined.csv
src/Data/Output/meals_combined.csv

Run the Data Pipeline

Before opening the dashboard, you must generate the data:

python src/maincode/main.py

Download Instructions Navigate to:

LSTM Models: https://drive.google.com/drive/folders/1daVVcrvS2u6rFBcKG_RjkQJh6Q19z4P7?usp=drive_link

GRU Models: https://drive.google.com/drive/folders/1KtWlLppDeVkl30KRoWk2DoWPzYHRoRkN?usp=drive_link

Comparing Models: https://drive.google.com/drive/folders/1GFUuX3pHYzUkceX-JBETsv6FtPSd_x0M?usp=drive_link

✅ 📊 Dashboard (Streamlit App)

Our interactive FCPS Meal Analytics Dashboard provides real-time insights into school meal operations. Run the full interactive dashboard

streamlit run demo/app.py

Folder Structure

├── demo
│   ├── fig
│   │   └── Video.mp4
│   │
│   ├── images
│   │   ├── multivariate_plots
│   │   │   ├── GRU.png
│   │   │   ├── LSTM.png
│   │   │   ├── fnn_model.png
│   │   │   ├── linear_regression.png
│   │   │   └── xgboost_model.png
│   │   │
│   │   └── univariate_plots
│   │       ├── GRU.png
│   │       ├── LSTM.png
│   │       ├── LSTM_train_test_forecast_example.png
│   │       ├── fnn_model.png
│   │       ├── linear_regression.png
│   │       └── xgboost_model.png
│   │
│   ├── .gitkeep
│   └── app.py
│
├── presentation
│   ├── Capstone_Presentation_Group9.pptx
|   ├── Capstone_Presentation_Group9.pdf 
│   └── .gitkeep
│
├── reports
│   ├── Latex_report
│   │   ├── fig
│   │   ├── File_Setup.tex
│   │   ├── Report_PDF.pdf
│   │   ├── references.bib
│   │   └── word_report.text
│   │
│   ├── Markdown_Report
│   │   └── .gitkeep
│   │
│   ├── Progress_Report
│   │   ├── Markdown_CheatSheet
│   │   │   ├── Markdown1.pdf
│   │   │   ├── Markdown2.pdf
│   │   │   ├── Markdown3.pdf
│   │   │   └── Markdown4.pdf
│   │   │
│   │   ├── Progress_Report.md
│   │   └── img_2.png
│   │
│   └── Word_Report
│       └── Final Report.docx
│
├── research_paper
│   ├── Latex
│   │   ├── fig
│   │   │   └── images
│   │   ├── mybib.bib
│   │   ├── research_paper.pdf
│   │   └── research_paper.tex
│   │
│   ├── Word
│   │   └── Conference-template-A4.doc
│   │
│   └── .DS_Store
│
└── src
    ├── Data
    │   ├── Html
    │   │   ├── May 2025 Breakfast production records
    │   │   └── May 2025 Lunch production records
    │   │
    │   └── Output
    │       ├── breakfast_combined.csv
    │       ├── lunch_combined.csv
    │       └── meals_combined.csv
    │
    ├── component
    │   ├── EDA.py
    │   ├── preprocess.py
    │   │
    │   ├── multivariate
    │   │   ├── model.py
    │   │   ├── plot.py
    │   │   ├── training.py
    │   │   └── utils.py
    │   │
    │   └── univariate
    │       ├── comparing_model.py
    │       ├── forecasting.py
    │       ├── model.py
    │       ├── plot.py
    │       ├── training.py
    │       └── utils.py
    │
    ├── maincode
    │   └── main.py
    │
    ├── results
    │   └── all_school_meal_forecasts.csv
    │
    ├── tests
    │   ├── combine_csv.py
    │   └── multivariate_main.py
    │
    ├── .gitkeep
    └── requirements.txt

📡 API Endpoints

Although this project does not use external REST APIs, the internal Streamlit dashboard relies on several Python-based API-like functions that power forecasting and analysis.

🔧 Internal Model Endpoints

Function	Description	Location
forecast_future_dates()	Predicts next k days using trained LSTM/GRU models	src/forecasting.py
load_and_aggregate_district()	Loads CSV + cleans + aggregates district production cost	src/utils.py
safe_time_split()	Chronological train-test split for time-series	src/utils.py
TimeSeriesDataset	Creates sliding windows for univariate LSTM/GRU	src/utils.py
ForecastingModel	LSTM/GRU model class	src/model.py
FeedForwardRegressor	Baseline neural network model	src/model.py
forecast_all_models_in_folder()	Runs forecasts for every school (batch mode)	app.py
school_loss_analysis()	Detects schools with high loss or wastage	app.py

🖥️ Dashboard-Level Actions (Triggered in Streamlit)

Action	Trigger Button	What Happens
Run Forecast	Run Forecast	Loads all LSTM/GRU models and predicts next k days
Compare Models	Compare LSTM vs GRU	Runs both folders → compares total cost curves
AI Recommendations	AI Recommendations	Suggests waste reduction strategies
Wastage Heatmap	Auto-loaded	Creates weekday-based discarded food heatmap
School-Level View	Dropdown Filters	Filters graphs/tables by school + meal type

🔧 Troubleshooting

Quick solutions to the most common issues:

Issue	Cause	Simple Fix
Empty CSV after parsing	Wrong HTML folder path	Check breakfast/lunch folder paths before running preprocess_html.py
Date errors / NaNs	FCPS dates use mixed formats	Use dayfirst=True in pd.to_datetime() (already used in code)
LSTM/GRU model not loading	Wrong .pth path	Ensure model file is inside: univariate/LSTM_models/ or univariate/GRU_models/
Streamlit blank page	Cached old data	Run: streamlit cache clear
XGBoost import error	Not installed	pip install xgboost
Very high forecast values	Outliers in cost	99th percentile cleaning already included—recheck preprocessing
Training too slow	Model too big	Reduce HIDDEN_DIM from 256 → 128
Forecast shows empty for a school	School name mismatch	Filename uses _ (e.g., Aldrin_Elementary), CSV uses spaces → ensure both match
Heatmap blank	Non-numeric waste columns	Convert with pd.to_numeric(errors='coerce').fillna(0)
“Forecast failed” error	Not enough rows for that school	Check if subset CSV has enough data; retrain if needed
Port already in use (Streamlit)	Another app running	Run: lsof -i :8501 → kill -9 <PID>

---

📊 Research & Performance

1️⃣ Univariate Forecasting Results (Cost-Only Models)

These models predict production_cost_total using only past cost values (sliding window of 7 days).

Model	RMSE	R²	Notes
LSTM	⭐ Best	High	Learns long-term temporal patterns extremely well
GRU	Very Good	High	Faster than LSTM, stable performance
XGBoost	Medium	Medium	Strong non-linear baseline, but not sequence-aware
Feed-Forward NN (FNN)	Medium	Medium	Good baseline but ignores temporal structure
Linear Regression	Poor	Low	Cannot model sequential dependencies

---

2️⃣ Multivariate Forecasting Results (School-Level Features)

These models use:

• served_total
• planned_total
• discarded_total
• left_over_total

to predict:

• production_cost_total

Model	Performance	Notes
GRU (Sequence Model)	⭐ Best (if metrics show this)	Captures school-wise temporal patterns across multiple features
LSTM (Sequence Model)	⭐ Best / Very Strong	Multivariate LSTM trained on same features; stable long-range learning
XGBoost	Strong	Excellent for structured/tabular data
Feed-Forward NN	Good	Learns non-linear interactions but not sequence structure
Linear Regression	Baseline	Limited for multi-feature temporal data

---

🗝️ Key Findings (Short)

• Both multivariate LSTM and GRU clearly outperform classical models (XGBoost, FNN, Linear Regression).
• Including served, planned, discarded, and leftover meals improves cost prediction compared to cost-only models.
• Sequence models (LSTM/GRU) handle school-level temporal behavior much better than non-sequence models.
• Outlier removal and proper preprocessing stabilize forecasts and reduce noise.

---

🧰 Technology Stack

Category	Technologies
🤖 Machine Learning	PyTorch · XGBoost · Scikit-Learn
🧠 Deep Learning	LSTM · GRU · FeedForwardNN
🖥️ Dashboard	Streamlit · Plotly Express
🧹 Data Processing	Pandas · NumPy · BeautifulSoup · lxml
📊 Visualization	Matplotlib · Seaborn
🧪 Evaluation	MSE · RMSE · R² · MAE
📁 Utilities	Pickle · Glob · Pathlib · OS
🔧 Version Control	Git · GitHub
🚀 Deployment	Local Machine · Streamlit Cloud
💻 Language	Python

✅ 🤝 Contributing

We welcome contributions from developers, students, and researchers. Steps:

# Create a feature branch
git checkout -b feature/my-feature

# Make changes and commit
git commit -m "Added new improvement"

# Push to repo
git push origin feature/my-feature

Then open a Pull Request on GitHub.

✅ 📄 License

This project is licensed under the MIT License. You are free to use, modify, and distribute the software with proper attribution.

✅ 🙏 Acknowledgments

Special thanks to the contributors who made this project possible:

• Dr. Amir Jafari – Project Guidance (GWU)

• Fairfax County Public Schools (FCPS) – For providing production record structures

• Open-source community – PyTorch, Streamlit, XGBoost

• Team Members – Areena, Chaya, Varshith