Proof of learning model
Main plot w/ correction factor
Main plot
Zoomed in prediction

treehacks-energy Challenge

Submission for TreeHacks 2018 - Stem's Challenge.

Challenge

To forecast one site's energy consumption and simulate the activity of a behind-the-meter energy storage system in order to reduce this client's energy bill through peak shaving.

Getting Started

The analysis and prediction model (as well as the dataset provided) can be found in this repo. To run, simply modify the path of the file in:

analysis.py and seeding_LSTM.py

Running the code will yield the results shown below. Alternatively, you can reproduce or enhance the plots:

Using: data_points.p
pickle.dump([trainPredict, trainY, testY, testPredict], open(dir+'data_points.p', 'wb'))

Our Solution

Using long term short term memory layer (LTSM), we are able to learn sequence patterns better by having memory of what it has seen before.

Learning

As the train decreases over epoch, it's evident that the model is learning.

Train Score: 44.57 RMSE
Test Score: 46.17 RMSE

Since train and test scores are close, the model can correctly predict on data it hasn't seen yet.

Main Plot

Here, is comparison of our known data (in blue), our training output (in orange), and our test data (in green). Main Plot

One issue that's evident in this plot is that our algorithm consistently underestimates the values. But since the understimation is consistent across time, we can design a correction factor to compensate for the offset.

Correction Factor This plot shows what a simple error correction would look like. Notice how the red lines (magnitude of peaks predictions) match the the actual data much more accurately now.

Zoomed Prediction It's evident and remarkable when zoomed in that the shape closely matches the actual data. The only issue is the biasing issue, which is why we have a correction factor. This essentially shows the the model can learn qualitative geometric features of the usage by training on past pulses.