Verdigris uses neural network models to provide day-ahead forecasts of the whole building load in 15 minute intervals. The model is regularly retrained and queried. The forecast includes an expected mean and a confidence distribution around the mean. The confidence levels are calculated for every quantile between 5 to 95% (Figure 1). The model incorporates inputs such as historical energy use, weather data, and time of day. We benchmark our models on synthetic tests and with open source data.
Figure 1: Forecasting showing confidence interval at 90% and 80%