Raphasha27/EskomSense-AI

ML-powered load shedding predictor & backup energy optimizer for South African homes and businesses. # EskomSense AI > Predictive Load Forecasting for the South African Power Grid ## Overview EskomSense AI is a production-grade machine learning pipeline that forecasts South African electricity demand 24 hours ahead using an LSTM (Long Short-Term Memory) neural network. The system ingests historical Eskom load data, trains a sequence model, and exposes predictions via a FastAPI REST endpoint. ## Architecture ` +---------------+ +----------------+ +-------------+ +---------+ +-----------+ | Data Pipeline |-->| Preprocessing |-->| LSTM Model |-->| FastAPI |-->| Dashboard | | (CSV load) | | (MinMaxScaler) | | (PyTorch) | | REST | | | +---------------+ +----------------+ +-------------+ +---------+ +-----------+ ` ## Model Architecture | Layer | Configuration | |-------------------|----------------------------------------| | Input Projection | Linear(1 -> hidden_size) | | LSTM | hidden_size units, num_layers stacked | | Dropout | Applied between LSTM layers & output | | Fully Connected | Linear(hidden_size -> 1) | **Default hyperparameters:** hidden_size=128, num_layers=2, dropout=0.2, lr=1e-3 ## Quick Start `ash # Install dependencies pip install -r requirements.txt # Generate synthetic training data python -m src.data.generator # Train the model python scripts/train.py --data data/sample_data.csv --epochs 50 # Run predictions python scripts/predict.py ` ## API Start the server: `ash uvicorn src.api.main:app --reload ` | Endpoint | Method | Description | |--------------|--------|----------------------------------------------------| | /health | GET | Liveness probe - returns service status | | /model/info| GET | Model metadata (architecture, param count, device) | | /predict | POST | Accepts sequence data, returns predicted MW load | ### Exampl …