We forecast On-site WUE
We leverage 4 years of hourly onsite WUE data from 1000+ locations to build scalable forecasting models for any location in the US. Instead of training expensive, city-specific models, we use clustering techniques like Dynamic Time Warping and Catch22 to group cities with similar patterns. By creating representative series for each cluster and applying SARIMA models, we enable efficient, accurate forecasting. This multilevel approach is cost-effective and adaptable for future scalability.
As well as Off-site WUE
We are equipped with hourly data for 28 geographically partitioned EGRID regions, where all cities in the same region share the same offsite WUE. This massively reduces our problem space compared to Onsite WUE, and allows us to experiment with, and compare/contrast different modelling methodologies i.e. LSTMs, SARIMA, and TimesFM (Google's new foundation model for forecasting).
