Energy Consumption Prediction for EV Charging Sessions

1. Project Objective:

To develop a predictive model for estimating energy consumption during electric vehicle (EV) charging sessions using historical data from a workplace charging program. The model will aid in optimizing energy management, understanding usage patterns, and improving the efficiency of charging infrastructure.

1.1 Project structure

    ├── 01-intro
    │   └── Notes.md
    ├── 02-experiement-tracking
    │   └── Notes.md
    ├── 03-orchestration
    │   └── Notes.md
    ├── README.md
    ├── images
    ├── project
    └── README.md

The folder contains the updated notes for the first three chapters, and the project for MLops course. README.MD is also provided for understanding the project structure and intruduction of the project.

2.Dataset Description:

The dataset comprises 3,395 high-resolution EV charging sessions, with data collected from 85 EV drivers across 105 stations located at 25 different workplace sites. These sites include various facilities such as research and innovation centers, manufacturing plants, testing facilities, and office headquarters. The dataset is in CSV format, with timestamps recorded to the nearest second, allowing for precise analysis.

2.1. Key Features:

• sessionId: Unique identifier for each charging session.
• kwhTotal: The amount of energy consumed during the session (in kWh).
• dollars: The amount of money paid for energy charging.
• created: Timestamp for when the charging session started.
• end: Timestamp for when the charging session ended.- dollars: The amount of money paid for energy charging.
• startTime: The min of the created time.
• endTime: The min of the end time.
• chargeTimeHrs: The total charging mins.
• weekday: Indicates if the session occurred on a weekday.
• platform: The platform used for accessing the Charging.
• distance: The distance from a user's home to the charging location, expressed in miles except where user did not report address..
• userId: Unique identifier for each user.
• stationId: Unique identifier for each station.
• locationId: Unique identifier for each location.
• managerVehicle: A ambiguous identifier which is described as "Firm manager vehicle indicator".
• facilityType: The type of facility a station is installed at (manufacturing = 1, office = 2, research and development = 3, other = 4). As indicated in the dataset description, all facility types are at workplaces..
• reportedZip: A ambiguous identifier for the Zip of the reported location.

3. How to Use:

3.1 Environment preparetion

Using Pip

1. Clone the repository: https://github.com/Muhongfan/MLops-zoomcamp-2024.git

2. Navigate to Project directory: cd project

3. Install the required packages using Pipefile:pipenv shell

3.2 Model training, tracking and registry

1. Run jupyter notebook energy_forecast.ipynb to reproduce the model (if need)
   • Set up the aws profile: cd .aws to config aws credentials including aws_access_key_id, aws_secret_access_key, region, etc.
   • Turn on MLflow server on EC2 by seting up the virtual env:

     sudo apt-get install python3-pip
     sudo apt install python3-venv
     python3 -m venv myenv
     source myenv/bin/activate
     pip install mlflow boto3 psycopg2-binary
     pip show mlflow boto3 psycopg2-binary
     

   • Start the MLflow server: mlflow server -h 0.0.0.0 -p PORT_LAUNCH_MLFLOW --backend-store-uri postgresql://DB_USER:DB_USER_PWD@DB_ENDPOINT:PORT/DB_NAME --default-artifact-root s3://S3_BUCKET_NAME
   • Run jupyter notebook energy_forecast.ipynb to retrain the model
   • Check the registryed model on terminal with MLflow server asTRACKING_SERVER_HOST:5000, where TRACKING_SERVER_HOST is the DNS of your EC2 instance.

The datasets and model will be stored and registryed in S3 and MLflow after model training. Check at AWS S3 and MLflow.

Note: As the modle and artifacts are deployed on AWS, In model_registry, it requires your AWS profile info.

3.4 Deployment with Flask and docker

Docker was used to build a container-based infrastructure, which packaged the entire environment into containers, making it portable and consistent across different machines.

1. Setup your AWS profile

2. Set up the pipenv for the deployment

3. Build the docker with docker build -t energy-consumption-prediction-service-mlflow:v1 . for webservice under service/web-service-mlflow-with-Docker

4. Run the docker

```
docker run -e RUN_ID=your_run_id -e EXP_ID=your_exp_id -e AWS_PROFILE=your_aws_profile -v ~/.aws:/root/.aws:ro -p 9696:9696 your-image-name
```
with which requires the configure of your run_id, experiment_id and AWS profile

5. Test the docker with python test.py

All details are in the web service for energy consumption prediction project

Note:

• The run_id, experiment_id can be obtained from step 4.2 on Mage-ai, Model_registry process.
• AWS profile is expected to be your own AWS profile.
• Run the test file via python test.py

• docker logs -f to show the realtime logs.
• docker exec -it container_id /bin/bash run within the container.
• docker ps list the containers.
• docker build -t container_name:vresion . build docker image
• docker run -v ~/.aws:/root/.aws:ro -p host_port:docker_port container_name:vresion run the docker

3.5 (or) Deployment with lambda and docker

All details are in the serverless service for energy consumption prediction project

1. Setup your AWS profile

2. Set up the pipenv for the deployment

3. Build the docker with docker build -t stream-model-duration:v1 . for serverless service under service/streaming

4. Run the docker

   docker run -v ~/.aws:/root/.aws:ro -p 1234:9696 stream-model-duration:v1
   

   with which requires the configure of your run_id, experiment_id and AWS profile in Dockerfile

5. Test the docker with python test_docker.py

4. Orchestration

Set up Orchestration with Mage AI.

• Go to mlops and run ./scripts/start.sh to set up the orchestration with mageai.
• Go to http://localhost:6789/ to open the mageai on your browser
• Go to project_energy_consumption->Pipelines to run each pipelines. (data preparetion -> xgboost_training -> model_registry)

4.1 The Orchestration Structures

1. EDA and Data preprocessing The primary goal of the first step is to gain a basic understanding of the dataset related to energy consumption during electric vehicle (EV) charging sessions. Additionally, the analysis will explore and identify relationships between key variables such as energy consumption, associated costs, distance traveled, charge time, and user behavior.

All details are in EDA for energy consumption dataset

2. Modelling and tracking The next step involves building the modeling process based on the exploration and setting up an experiment tracking pipeline using MLflow. The entire service will be deployed on AWS, utilizing EC2 for the tracking server, RDS for metadata storage, and an S3 bucket for storing datasets and models. Deployment will be managed using Docker to ensure consistency and scalability across the environment. This setup will streamline model experimentation and tracking, enhancing the efficiency of the development workflow.

All details are in this jupyter notebook

3. orchestration

• Data preparatio
• Model training
• Model registry

All details are in mlops folder

4. Monitoring results (some)