aws-samples/sagemaker-rl-energy-storage-system

Examples of RL for the energy arbitrage use case.

sagemaker-rl-energy-storage-system's Issues

Deploy trained agent endpoint using any existing framework

Currently SageMaker RL is unable to deploy DQN (tf based) trained agent as live endpoint. Is there a workaround?

Duplicate file: `battery_env_sm.py`

The battery environment seems to be duplicated:

1. src/share/battery_env_sm.py
2. src/energy_storage_system/envs/_battery.py

Is it safe to remove 1) ?

Cleanup unused file in smnb_utils

Data download process clarification

A bit of a nit question:

The instructions in the first workbook state:

Download dataset from here and placed into data folder as follows:

|-- data
|   `-- PRICE_AND_DEMAND_202105_NSW1.csv

However, in the next section, it says to use full year data.

Use full year data: "FILEPATH":"data/PRICE_AND_DEMAND_2020FULL_NSW1.csv"

I don't see the link to download a year's worth of data.....

Just wanted to clarify: are users are required to download the individual monthly files and concatenate them together before running notebook?

Document how to run Streamlit demo

Somewhere to show streamlit run src/demo/streamlit_demo.py -- INPUT_DIR [--update-seconds N]

Streamlit shows three charts briefly, then always two charts

On start, Streamlit shows three charts very briefly (screenshot-01), then subsequently down to two charts only (screenshot-02).

Screenshot-01: three charts on start

---

Screenshot-02: two charts for the rest of time

Fix bug in notebook

Document how to run Streamlit on SageMaker notebook instance

Document how to quick-run the Streamlit app directly on a SageMaker notebook instance (without needing additional Lambda, API Gategay, and ECS instance).

Do not use Python's `site-packages` as default sample data.

Default sample data is DATA = Path(__file__).parent.parent / "data/sample-data.csv" which means to download the sample data to Python's site-packages.

We should use a safer default; proposed to use "./data".

Example on using hyper-parameter tuning

Provide example on using hyper-parameter tuning

Example how to run Streamlit demo

Running Streamlit demo as-is produces error about a few missing files data/....

Need to document:

• pre-Streamlit steps: agents to trained, data / reports to produce, etc.
• where to store data/ for Streamlit app

Documentations can be in the form of readme md, or sample notebooks.

Refactor SageMaker RL common files

Refactor SageMaker RL common files by keeping the bare minimum for Ray, and track in separate repo.

Refactor src/share to src/energy_storage_system

src/share/ contains multiple top-level Python package. To relocate these under a new module called energy_storage_system

pip package to cover gym environment setup

Package to cover:

• gym environment registration
• data download

Migrate streamlit demo code here

Migrate streamlit demo code here and use this as master code repo

Support full year data for reproducibility

setup.py for the battery gym

Create setup.py for users to install battery gym to their Python environment.

Why SageMaker train entrypoint sets use PyTorch = True?

Notebook-01 uses Tensorflow container, but the entrypoint script sets PyTorch to True?

sagemaker-rl-energy-storage-system/src/source_dir/train_battery_sm.py

Line 18 in 431af55

"use_pytorch": True,

Train using custom container with SageMaker RL