LIVE SITE

🔗 https://load-prediction-9zp5.onrender.com

Load Type Predictor

A machine learning model that predicts power system Load Type (Light / Medium / Maximum) from energy consumption data. Deployed as a FastAPI web service on Render.

Problem Statement

Predict the Load_Type of a steel industry power system using historical data. Classification into 3 classes:

• Light_Load
• Medium_Load
• Maximum_Load

Dataset

• 35,041 rows × 9 columns (Jan–Dec 2018)
• Features: Usage_kWh, Lagging/Leading Reactive Power, CO2, Power Factors, NSM
• Validation: Last month (Dec 2018) as test set (time-based split)

Model Pipeline

┌────────────┐   ┌────────────────┐   ┌───────────────────┐   ┌────────────────┐   ┌────────────────┐    ┌─────────┐   ┌─────────────────┐   ┌──────────────────┐
│  Raw Data  │─▶│ Preprocessing  │──▶│ Feature Engineer  │─▶│ Outlier Handle │─▶│ Train/Test Split│─▶ │  SMOTE  │──▶│ 7 Models Trained│──▶│🏆Stacking Ensem.│
│  (35,041)  │   │• Median impute │   │• 31 features      │   │• 1st-99th %ile │   │• Time-based     │   │ Balance │   │ LR | RF | XGB   │   │ Accuracy: 93.95% │
│            │   │• Remove dupes  │   │• Time + Cyclical  │   │  capping       │   │• Last month     │   │ classes │   │ LGBM | ET       │   │ F1-Score: 93.99% │
│            │   │• CO2 zero fix  │   │• Power interact.  │   │                │   │  = test set     │   │         │   │ Voting|Stacking │   │                  │
└────────────┘   └────────────────┘   └───────────────────┘   └────────────────┘   └────────────────┘    └─────────┘   └─────────────────┘   └──────────────────┘

Results

Overall Metrics Metric Score Accuracy 93.95% Precision 94.28% Recall 93.95% F1-Score 93.99%

Per-Class Performance Class Precision Recall F1 Light_Load 0.98 0.92 0.95 Maximum_Load 0.90 0.96 0.93 Medium_Load 0.89 0.97 0.93

Project Structure

├── Assignment2.ipynb          # Full EDA + model training notebook
├── train_and_save.py          # Script to train & export model artifacts
├── app/
│   ├── main.py                # FastAPI app with web UI
│   ├── model.py               # Model loading & prediction
│   └── feature_engineering.py # Replicates notebook's feature pipeline
├── models/                    # Saved model, scaler, encoder
├── data/load_data.csv         # Dataset
├── Dockerfile                 # Docker build for Render
├── render.yaml                # Render deployment config
└── requirements.txt           # Dependencies

API Usage

POST /predict

{
  "Date_Time": "01-01-2018 00:15",
  "Usage_kWh": 8.75,
  "Lagging_Current_Reactive_Power_kVarh": 2.95,
  "Leading_Current_Reactive_Power_kVarh": 0.0,
  "CO2_tCO2": 0.0,
  "Lagging_Current_Power_Factor": 73.21,
  "Leading_Current_Power_Factor": 100.0,
  "NSM": 900.0
}

Response:

{
  "predicted_load_type": "Light_Load",
  "confidence": 0.9953,
  "probabilities": {
    "Light_Load": 0.9953,
    "Maximum_Load": 0.0012,
    "Medium_Load": 0.0035
  }
}

Run Locally

pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 8000

Open http://localhost:8000 — web UI with sample data for testing.

Tech Stack

Python • FastAPI • scikit-learn • XGBoost • LightGBM • Docker • Render

---

Future Scope

                                      Where This Project Can Go Next
       ┌──────────────────────────────────────────────────────────────────────────────────┐
       │                                                                                  │
       │   SMARTER MODELS             MLOps & AUTOMATION           EXPLAINABILITY         │
       │   ─────────────────────      ────────────────────────     ───────────────────    │
       │   LSTM/Transformers for      Auto-retrain on drift        SHAP/LIME to show      │
       │   temporal patterns          MLflow model versioning      WHY a load type was    │
       │   (93.95% → higher)          CI/CD via GitHub Actions     predicted (XAI)        │
       │                                                                                  │
       │   REAL-TIME & IoT            DEMAND FORECASTING           MULTI-PLANT            │
       │   ────────────────────       ───────────────────────      ──────────────────     │
       │   Kafka/MQTT streaming       Predict Usage_kWh ahead      Transfer/Federated     │
       │   Edge deploy (ONNX)         Peak shaving & load          Learning across        │
       │   SCADA integration          shifting to cut costs        multiple facilities    │
       │                                                                                  │
       │   ANOMALY DETECTION          BATCH & CSV UPLOAD            PROD HARDENING        │
       │   ─────────────────────      ─────────────────────────     ───────────────────── │
       │   Flag equipment faults      Bulk predictions via          JWT auth, rate limits │
       │   & energy theft with        file upload + interactive     Kubernetes scaling    │
       │   Isolation Forests          historical dashboards         Grafana monitoring    │
       │                                                                                  │
       └──────────────────────────────────────────────────────────────────────────────────┘