Monke

Monkeys galore!

Inspiration

The main inspiration was with Thomas's parents both of whom work as engineers in HVAC. When Thomas saw this challenge he was ecstatic because he could learn from his parents about how different attributes effect IAQ. For example: the room capacity effect IAQ regulations. Seasons change the way that air is filtered and relative humidity differences between summer and winter. After that the whole group was hyped to try out this QuadReal challenge and see whether what we had learnt matched the analysis of the training data.

What it does

Our project takes in the last 10 5 minute readings from an IAQ sensor as well as date-time and predicts the next 5 minute average readings for IAQ.

How we built it

The project was built using pandas for handling data, tensorflow LSTM for creating the model and Seaborn, Matplotlib and Plotly for the exploratory data analysis. To create the model we first used the training data and resampled it into 5 minute timesteps. Then, we created features depending on whether the timestap is business hours or not as well as seasonality. Finally, we grabbed the last 10 x 5 minute readings and fed it into our LSTM which we trained on all of the QuadReal training data.

Challenges we ran into

The biggest challenge we ran into was finding the best way model the data such that we could finish in time. We had come up with various ideas such as a multivariate LSTM for all buildings and value_type_ids, 12 univariate LSTMs for each value_type_ids, and 12 x 29 univariate LSTMs for each building and value_type_ids. The best performance was with the 12x29 univariate LSTMs. However, it would take a really long time for us to finish predictions.

Accomplishments that we're proud of

We are proud of being able to finish this hackathon at all. For many of us it was our first hackathon and our first datascience hackathon too! We were able to learn alot by attending workshops and, although extremely late. Managed to finish our predictions on time. We are especially proud of learning how to use the functional api in tensorflow to create our LSTM. It allows us to create more flexible models and was extremely exciting to figure out.

What we learned

During this hackathon we learnt alot from data preprocessing to model creation. It was interesting watching the data preprocessing methods we learnt such as one hot encoding change the accuracy of our models and applying them as well. We learnt alot about feature generation and its effects on neural networks. Having minimal experience, it was really cool trying out different models such as random forests and LSTMs and seeing how they performed. Furthermore, we were able to experiment with different EDA tactics such as using Seaborn's correlation matrices to visualize correlations between value_type_ids and ADFuller test for seasonality on our dataset.