There are two study cases commercial buildings and housholds. We attempted to use six RNN-based models to forecast energy consumption, then compare them with ARIMA. To prove the feasibility of a new service, which can forecast future demand, we used the best performer in RNN-based models for building an API, only for the study case commercials buildings.

• python 3
• flask
• flask_restful
• json
• nbimporter
• pickle
• joblib
• tensorflow 2.3
• keras

• Get Cuda, Cudnn, python3 and virtualenv: To complete the installation, it is required to have Graphic card fron Nvdia, because the project was building on tensorflow GPU, which is compatible to Nvdia hardware.
• Create env:

python3 -m venv tensorflow_env

• Activate:

source tensorflow_env/bin/activate

• Install and upgrade pip

pip install --upgrade pip

• Install tensorflow 2.3 or up

pip install --upgrade tensorflow==version
pip install --upgrade tensorflow-gpu

• Install Jupyter notebook

pip install jupyter

• Data_acquisition: contains DBs and cleaner notebooks for both buildings. The weather cleaner notebook for households is not neccessary.
  • Raw DB
    • commercial
    • household
  • Final DB
    • commercial
    • household
• Model: for each study cases, there are notebooks:
  • analysis: basic analysis of the study case
  • arima: forecasting using ARIMA
  • dnn: forecasting using DNN
  • rnn: forecasting using RNN-based models
• App: contains app.py, the restful API
• api: contains the stored data of the pre-trained model

After finish setting up the environment, one can run:

• Activate environment

source tensorflow_env/bin/activate

• Go to folder App

python3 app.py

• The API is available for observation at: localhost, port 4848. Using command to ask model to predict:

/forecast<int:days>

• Retrain model: go to folder Model and process the notebook forecasting_model