- I use USDA data about snow pack in Utah to train a neural network to predict snow pack seven days into the future.
- I construct an LSTM with 16 nodes and train the model for 20 epochs. The model predictions have an RMSE of 3 inches on test data.
My goal with this project was to built a model to predict snowpack in the mountains of utah. In the linked notebook I import, clean, and format snowpack data collected throughout the state of Utah. I then train an LSTM model to use 30 days of past snow pack data to predict the expected snow pack 7 days into the future, within the Big Cottonwood snowbasin in northern Utah. The model works well, yeilding an RMSE of 3 inches when the predicted and actual data for 2019--2023 are compared. Most of this disparity comes from the 2023 snowyear, which was a record breaking year and far outside of what characterizes normal snopack in Utah.
To achieve this result I used a Tensorflow implementation of an LSTM with 16 nodes, trained over 20 epochs. Further development to improve this model might include incorporating reinforcement learning to improve the model's accuracy. I collected the data used to train, validate, and test the model from this portal on the USDA website.