How milonigada09/supply-chain-forecasting-deep-learning works
Project Type and Purpose
• Type: A data-science/ML pipeline for time-series sales forecasting.
• Purpose: Ingest historical sales and inventory data, engineer features, train and compare multiple ML and DL models (Linear Regressor, XGBoost, CNN, LSTM, CNN+LSTM, GRU, Transformer), select the best model, and produce future-sales forecasts.
Main Components
a. Data Ingestion & Storage
• Historical sales data, inventory data (could be CSV/Parquet files or a database).
b. Data Preprocessing & Feature Engineering
• Cleaning, normalization, windowing, creation of lag features, train-test split.
c. Model-Training Modules
• Seven parallel subcomponents:
• Linear Regressor
• XGBoost
• CNN
• LSTM
• CNN+LSTM
• GRU
• Transformer (see Transformer.py)
• Each takes preprocessed data and outputs a trained model file (.h5 for DL models).
d. Model Evaluation & Comparison
• Compute error metrics (RMSE, MAE) on validation/test sets.
• Compare models to choose best (XGBoost and CNN+LSTM are highlighted).
e. Model Persistence
• Save the chosen model (and optionally all models) to disk (CNN_MODEL.h5, etc.).
f. Forecasting Interface
• Notebook or script that loads saved model, runs inference for next-day forecasts.
g. (Optional) Deployment / Service Layer
Architectural Patterns / Principles
• Pipeline pattern: clear, linear dataflow from ingestion to deployment.
• Modularity: Each model training is its own interchangeable component.
• Parallel training: six DL/ML models can be trained simultaneously.
• Separation of concerns: distinct stages for preprocessing, modeling, and evaluation.
• Linear Regressor
• XGBoost
• CNN
• LSTM
• CNN+LSTM
• GRU
• Transformer