MLFlow Oil Production Time-Series Forecast Stack

This repository contains a microservice stack based on MLFlow for training and executing Oil production time-series forecasting Machine-Learning Models. It is a proposal for a digital twin module.

1. The project tree has a docker-compose.yml for deployment in a docker swarm. The architecture consists of:
   • mlflow_client
     • app

       • service

         • src
           • forecast.py: API service that receives user parameters and executes the MLProject backcast in projects/train_inference

       • main.py: main application file that executes and keeps all FastAPI endpoints

     • projects
       • train_inference
         • train.py: Trains forecasting models
         • backcast.py: Executes model inference
         • utils.py: Keeps utility functions used throughout the code
         • config.yml: Contains all adjustable parameters for model training (hyperparameter ranges, size of windows, num of trials, etc)
         • data_types.yml: Contains a dictionary with a datatype of each hyperparameter
         • registered_models.yml: Contains a list of registered models for each combination of lookback, horizon and sampling frequency
     • requirements
     • tests
   • mlflow_tracking
   • mysql
   • minio