Next-generation demand response forecasting – increasing the accuracy and speed of your forecasts
This author no longer works for DNV GL.
Does this sound familiar?
Your utility has demand response programs that shave peak when deployed, by between 40 and 80 MW total, depending on the day. Some of this wide variation is due to weather differences, and then there’s the “random”—i.e., unexplained—error that makes up the rest. For an upcoming event, your system planning group receives a MW performance forecast, but without a direct way to measure the program’s performance and the wide variability observed in post-event performance calculations compared with the forecasts, this program resource is seen as “fuzzy,” “soft,” or maybe even “cuddly”…ok maybe not cuddly, but certainly the other two.
What can be done to firm up the DR Resource?
Option 1: Beef up the measurement and communication systems
For direct load control programs, two-way communication switches have been available from vendors for a number of years. These two-way switches can send information back to dispatchers and enable targeting of a specific number of committed MW for the program, through active monitoring or automated controls. There are also non-intrusive load monitoring (NILM) systems which enable near real-time visibility into the power consumption by important appliances used within homes or businesses, which will provide the load shed through auto DR, direct load control, or other curtailment programs. Communication back to the system operator can travel through AMI systems or dedicated cellular channels. These technologies can be used to provide dramatic improvements in the accuracy of near real-time forecasts of load impacts. But unfortunately they still do not include a crystal ball-like device for seeing what is going to happen tomorrow.
Option 2: Micro-forecasting analytics
Whereas option 1 was focused on technology solutions, option 2 involves extracting more predictive power from smart meter and other data that is already being collected, and therefore has no additional data collection cost. In a recent research demonstration with our utility partners, we were able to reduce day-ahead forecasting uncertainty for DR programs from a previous best-case accuracy of ±20% down to ±12% on average by using adaptive analytics. The key to the adaptive analytics is to better understand and account for serial correlation, or, in other words, the degree to which data that is in close proximity in a time series is related. So it’s not just developing relationships between data fields over a broad span of time, e.g., an event season, it’s about finding extra insights about what’s about to happen tomorrow, from cues in the data from yesterday and today.
Option 3: Combined approach
Both option 1 and option 2 can be used independently to dramatically improve the predictability of DR resources. But by combining them the resource planner can get a sum that is greater than its parts. In addition, overlaying the analytics on top of enhanced data collection will likely reduce hardware and communications investments needed in option 1.
Is it all worth the investment?
The equipment and analytics still comes with capital and service costs that need to be taken into account. Utilities considering one or the other should first conduct an analysis of the market value for enhancing the predictability of its DR programs on a day-ahead and real-time basis. What are the wholesale market opportunities? Are we paying for fast response generation services that could be replaced or complemented with more predictable and reliable DR? These and other questions should be addressed in an assessment study before diving head first into the options above.
For more information about what DNV GL offers, read more here.