Willis Danielson

The 2025 to 2026 winter season was among the worst ever recorded for snow accumulation along the Wasatch Front. I wanted to see if I could predict a reasonable SWE given how unxonventional/historic a season it was. The model I used was simple. It used rolling 60-day windows of SWE observations as features along with sine and cosine features to encode where each date sits in the hydrological year. The network itself was a multilayer perceptron with only two hidden layers (128 and 64 neurons). These layers used ReLU activations and a 15% dropout layer sat between them. Training data included SWE observations from the Brigthon SNOTEL site going back to the start of the 1988 hydrologic year. Leap days were removed so that each hydrologic year had the same length and a consistent day-of-season.

I am a strong believer in ensembles. No one model is going to understand the function. Normally, I would stack and average the results of multiple methodologies: GLMs, GBMs, ANNs, etc., but MLPs are so easy. Normally, a trained network would produce the same outputs for every run, but to introduce uncertainty, I used Monte Carlo dropout in inference: the dropout layer remained active across 300 forward passes, producing a distribution of possible futures for each forecast step. I condensed these prediction runs into three main metrics: the median, 10th percentile, and 90th percentile values. Forecasting was recursive and each date’s median prediction was fed back into the model. This way every prediction step can use the next day’s complete window.

The first forecast run was completed on December 22, 2025, and the process was repeated weekly through the 5th of April, 2026. The full sequence was rerun on June 23, 2026 so that each forecast round could be evaluated against a large set of observed outcomes.

Preliminary results were encouraging. Early in the season, the model tracked accumulation surprisingly well and produced reasonable uncertainty bands. It appeared to react reasonably well to storm days and respected quiet periods. Its weakness appeared when the snowpack entered into an extremely early melt phase. The model lagged behind for roughly two weeks before adjusting to the new downward trajectory. Even so, that behavior was interesting. The model required no retraining to adapt, suggesting that it had learned the broad seasonal structure of the hydrologic cycle and could use both recent observations and seasonal timing to reorient once the melt rate became clear.

Beyond the lagging adaptation, the limits of a site-specific, minimally featured MLP became clear to me when I attmepted to use this model for other SNOTEL sites. The model performed well at Brighton, but this did not transfer to other sites along the Wasatch Front. When focusing the model on Snowbird substantial retunings of hyperparameters and architecture were needed for the model to perform at all sensibly. Improving generalizability is what I want to use as a north-star when refining this project. For next season I will incorporate additional meteorological features like temperature, cloud cover, and elevation. I also want to train the network using data from multiple sites, and I want to use a broader forecasting ensemble that includes MLPs, GBMs, ARIMA-style models, and even newer transformer based time series setups. Hopefully, these additions and refinements will improve performance and generalizability.