Time Series Analysis of Building Energy Data

by Aida da Silva and Jason Rafe Miller. This was a student collaboration for WVU CS 677 with Professor Dehzangi.

Getting started

• Create a directory called BuildingEnergy.
• Populate BuildingEnergy with all the Python notebook files (*.ipynb).
• Create a subdirectory called data and move to that directory.
• Download archive.zip from Kaggle into the data subdirectory.
• Rename archive.zip to BuildingData.zip (but do not unzip it).
• Move up to the BuildingEnergy directory.
• Start Jupyter notebook.
• Run the weather notebook as your first test.

Notebooks from Presentation 1

• Note each notebook (ipynb) has a corresponding Python script generated by nbconvert and saved in the scripts subdirctory. All notebooks used our BuildingSet1 data subset: 16 buildings with fairly complete steam usage data from site Eagle.
• Report1.Weather.ipynb finds that air temp is highly correlated to energy usage.
• Report1.Identity_101.ipynb uses a naive model.
• Report1.LinReg_101.ipynb uses linear regression.
• Report1.RNN_107.ipynb (formerly named LSTM_107) uses a SimpleRNN neural net.
• Report1.LSTM_108.ipynb uses an LSTM neural net.
• Report1.CNN_107.ipynb uses a CNN neural net.

Notebooks from Presentation 2

• ConvLSTM.ipynb

Data sources

• Publication Scientific Data
• Kaggle project page
• Download archive.zip from Kaggle

LSTnet

• Paper on ACM or arXiv or arXiv
• LSTNet authors posted their Python 2 and PyTorch code
• A 3rd party Python 3 and Keras implementation here and another that claims to be faster.