waste-predictor 4.0.3

Production-grade machine learning system for industrial waste prediction

Waste Prediction Module V4

A production-grade machine learning system for predicting industrial waste compositions based on production volume and environmental parameters.

📊 Model Performance

Metric	Score
R² Score	0.98
MAE	1,607
RMSE	3,092
CV R² (5-Fold)	0.974 ± 0.005

Per-Target R² Scores

Waste Type	R²
Total_Waste_kg	0.96
Solid_Waste_Limestone_kg	0.98
Solid_Waste_Gypsum_kg	0.99
Solid_Waste_Industrial_Salt_kg	0.98
Liquid_Waste_Bittern_Liters	0.99
Potential_Epsom_Salt_kg	0.98
Potential_Potash_kg	0.99
Potential_Magnesium_Oil_Liters	0.98

🚀 Quick Start

Installation

Install from source (local development)

pip install -e .

Install from wheel file

pip install waste_predictor-4.0.0-py3-none-any.whl

Install from GitHub (if hosted)

pip install git+https://github.com/yourusername/waste-predictor.git

Install from PyPI (if published)

pip install waste-predictor

📈 Usage

Making Predictions

The package provides a simple get_waste_prediction function:

from waste_predictor import get_waste_prediction

# Prepare input data
input_data = {
    'production_volume': 50000,
    'rain_sum': 200,
    'temperature_mean': 28,
    'humidity_mean': 85,
    'wind_speed_mean': 15,
    'month': 6
}

# Get prediction
result = get_waste_prediction(input_data)

print(result)
# {
#   'Total_Waste_kg': 101804.91, 
#   'Solid_Waste_Limestone_kg': 7322.23,
#   'Solid_Waste_Gypsum_kg': 32448.03,
#   'Solid_Waste_Industrial_Salt_kg': 62034.64,
#   'Liquid_Waste_Bittern_Liters': 40547.23,
#   'Potential_Epsom_Salt_kg': 2128.86,
#   'Potential_Potash_kg': 379.34,
#   'Potential_Magnesium_Oil_Liters': 4055.32
# }

Training from MongoDB

Users can train their own models using data from MongoDB:

from waste_predictor import train_from_mongodb

# Train with local MongoDB
results = train_from_mongodb(
    mongo_uri='mongodb://localhost:27017',
    database='waste_db',
    collection='training',
    output_model_path='my_custom_model.pkl'
)

print(f"Model R²: {results['metrics']['r2']:.4f}")
print(f"Model saved to: {results['model_path']}")

Training with MongoDB Atlas (Cloud)

from waste_predictor import train_from_mongodb

results = train_from_mongodb(
    mongo_uri='mongodb+srv://cluster.mongodb.net',
    database='waste_production_db',
    username='your_username',
    password='your_password',
    collection='training',
    output_model_path='waste_predictor_custom.pkl',
    verbose=True
)

Training with Full Connection String

from waste_predictor import train_from_mongodb

connection_string = "mongodb+srv://user:[e-mail]/dbname?retryWrites=true"

results = train_from_mongodb(
    mongo_uri=connection_string,
    database='waste_db',
    collection='training'
)

Training from DataFrame

You can also train from a pandas DataFrame:

import pandas as pd
from waste_predictor import train_from_dataframe

# Load your data
df = pd.read_csv('training_data.csv')

# Train model
results = train_from_dataframe(
    df=df,
    output_model_path='custom_model.pkl',
    verbose=True
)

print(f"Training R²: {results['metrics']['r2']:.4f}")

📋 Required Data Format

MongoDB Document Format

Each document in your MongoDB training collection should have:

{
  "Year": 2000,
  "Month": 1,
  "production_volume": 43163.99,
  "rain_sum": 270.87,
  "temperature_mean": 26.53,
  "humidity_mean": 100,
  "wind_speed_mean": 19.68,
  "Total_Waste_kg": 96664.3838,
  "Solid_Waste_Limestone_kg": 5080.5243,
  "Solid_Waste_Gypsum_kg": 23189.168,
  "Solid_Waste_Industrial_Salt_kg": 68394.6915,
  "Liquid_Waste_Bittern_Liters": 31725.5458,
  "Potential_Epsom_Salt_kg": 1605.914,
  "Potential_Potash_kg": 206.5851,
  "Potential_Magnesium_Oil_Liters": 3163.2316
}

Required Fields

Input Features:

• production_volume - Production volume (numeric)
• rain_sum - Total rainfall in mm (numeric)
• temperature_mean - Average temperature in °C (numeric)
• humidity_mean - Average humidity percentage (numeric)
• wind_speed_mean - Average wind speed (numeric)
• Month - Month number 1-12 (integer)

Output Targets (for training only):

• Total_Waste_kg
• Solid_Waste_Limestone_kg
• Solid_Waste_Gypsum_kg
• Solid_Waste_Industrial_Salt_kg
• Liquid_Waste_Bittern_Liters
• Potential_Epsom_Salt_kg
• Potential_Potash_kg
• Potential_Magnesium_Oil_Liters

📁 Project Structure

local-module/
├── data/
│   └── training/
│       └── training.csv          # Training dataset (312 samples)
├── train_v4.py                   # V4 training (PRODUCTION - R²=0.98)
├── predict_v4.py                 # V4 inference module
├── waste_predictor_v4.pkl        # Trained model checkpoint
├── waste_predictor_v4_metadata.json
├── train_v3.py                   # V3 training (Neural network)
├── model_v2.py                   # Enhanced neural network model
├── feature_engineering.py        # Feature engineering pipeline
├── requirements.txt              # Dependencies
└── README.md

🔧 Input Features

Feature	Description	Range
production_volume	Production volume	0 - 200,000
rain_sum	Total rainfall (mm)	0 - 1,000
temperature_mean	Average temperature (°C)	0 - 50
humidity_mean	Average humidity (%)	0 - 100
wind_speed_mean	Average wind speed (km/h)	0 - 50
month	Month of year	1 - 12

📤 Output Predictions

Output	Description
Total_Waste_kg	Total waste produced (kg)
Solid_Waste_Limestone_kg	Limestone solid waste (kg)
Solid_Waste_Gypsum_kg	Gypsum solid waste (kg)
Solid_Waste_Industrial_Salt_kg	Industrial salt waste (kg)
Liquid_Waste_Bittern_Liters	Bittern liquid waste (L)
Potential_Epsom_Salt_kg	Potential Epsom salt byproduct (kg)
Potential_Potash_kg	Potential Potash byproduct (kg)
Potential_Magnesium_Oil_Liters	Potential Magnesium oil (L)

🧠 Model Architecture (V4)

Weighted Ensemble of 3 Model Types:

1. XGBoost Gradient Boosting (weight ~33%)

   • 500 estimators, max_depth=6
   • Per-target models with log-transformed outputs

2. Stacked Ensemble (weight ~34%)

   • Level 0: XGBoost + LightGBM + Random Forest + GBR
   • Level 1: Ridge regression meta-learner
   • 5-fold stacking with passthrough

3. Deep Neural Network (weight ~33%)

   • Architecture: 256 → 512 → 256 → 128 with skip connections
   • GELU activation, BatchNorm, Dropout
   • Cosine annealing LR schedule

Feature Engineering (30+ features):

• Log/sqrt/squared production transforms
• Cyclical month encoding (sin/cos)
• Weather condition indices (wet, dry, evaporation)
• Production × weather interactions
• Domain-driven ratio features

📈 Performance Comparison

Version	R² Score	MAE	Key Technique
V1 (Original)	0.47	7,873	Simple MLP
V3	0.77	5,575	Log transform + Feature eng
V4 (Production)	0.98	1,607	XGBoost + Stacked + DNN Ensemble

🔄 Retraining

To retrain the model with new data:

1. Add new data to data/training/training.csv
2. Run training:

   python train_v4.py
   

3. Model will be saved to waste_predictor_v4.pkl

📝 License

MIT License