Create and Evaluate Forecasting Models
The many interrelated factors which influence demand for electricity cannot be directly modeled
by closed-form mathematical formulas. However, historic weather, demand, and price data
contain implicit relationships between independent input values – for example, weather
conditions – and electricity use. Neural networks are particularly well-suited for finding
meaningful patterns in such data, and Modeler is the framework for creating, evaluating,
and deploying neural networks which learn the data patterns that underlie electricity
demand and price forecasts.
Modeler can use virtually any quantity of historic data to train neural networks, and it includes built-in pre-processing to handle invalid or missing data. It supports building monthly, seasonal, and annual models, any of which can be used in operations or for long-term strategic planning. Modeler accepts historic daily input, such as the day-ofweek, and historic hourly input from multiple sources and multiple types of data from each source, such as temperature, wind speed, and relative humidity. Lead times for hourly inputs are specified through a simple “point-and-click” grid interface.
After forecasting models in a Schema are trained, Modeler provides powerful analysis and visualization tools for evaluating model performance and the quality of forecasts. Models are easily ranked by standard metrics, and sensitivity analysis indicates which inputs are most influential for any particular model. The Schema which demonstrates the best performance (smallest error) across any given training dataset is deployed for use by Forecaster and Scheduler to generate ongoing daily demand and price forecasts.
