Using AI to integrate weather into electrical and natural gas load forecasts

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 5 February 2014: 9:00 AM
Room C204 (The Georgia World Congress Center )
John K. Williams, NCAR, Boulder, CO; and G. Wiener, W. Myers, S. E. Haupt, T. Brummet, S. Dettling, S. Linden, and J. M. Pearson

Accurate forecasts of electrical and gas net load – the amount of electric energy or gas required by a utility's customers – are crucial for utilities' management of energy production and distribution as well as profitable trading operations. This is becoming even more true as renewable energy production increases. Utilities must be prepared to satisfy their customer's energy requirements at all times, but the complexity of doing so is exacerbated when customers are making increased use of distributed, variable renewable sources to produce some of the energy they use. For instance, on a sunny spring afternoon, a customer with a rooftop solar installation may be producing excess power, running their electrical meter backwards and helping the utility provide electricity to their neighbors. When a storm blocks the sun, the solar panels' power production drops, and the utility must be prepared to instantly make up the difference. Thus, weather is a significant factor both in customer demand for energy and, increasingly, in distributed energy production that affects consumers' net energy demand.

Artificial intelligence techniques are well suited to combining disparate data to provide targeted decision support information. This paper describes work done at NCAR to utilize historical and real-time load datasets, METAR observations, and weather forecasts of key meteorological variables from the NCAR-developed Dynamic Integrated foreCast system (DICast®) to produce real-time forecasts of electrical and natural gas load for several geographical domains. The approach includes careful data quality control, training predictive models on historical load and observed weather data, and continual analysis of recent load data and forecast performance to ensure that forecasts keep up with trends due to population growth, distributed renewable energy penetration or exogenous factors. The load forecasts are produced hourly, with lead-times out to 7 days, and are verified based on subsequent load observations.