In case you weren’t aware of the importance of forecasting for a utility company, it’s really BIG. Not only are you predicting the amount of power required to keep the lights on for fickle residential customers and for complex industrial businesses, you employ hedging strategies to protect against any unforeseen impacts to your forecasted demand.
At the Analytics2013 conference, Andy Mortimer from RWE nPower shared that a one percent error in their forecast can cost as much as £10M. (That’s nearly $16M US.)
If that’s not enough to get your attention, this financial impact may only increase with the volatility of the energy markets. This comes at a time when Citibank analysts project that the utility business is in for a major overhaul. In the UK alone, there has been a gradual reduction in electricity demand since 2007. Whether this is driven from aging population or improvements in energy efficiency, the business model of getting paid by the kilowatt is in jeopardy. (For more on that, see recent article about the Art of the Possible).
RWE nPower uses a combination of technologies from SAS, including hierarchical forecasts and neural networks, to reduce the risk to the business and potentially capitalize on the upside of energy markets.
After listening to Mortimer's presentation, I can see why load forecasts are such an interesting and ever-changing opportunity for utilities. Take these few examples related to weather into consideration:
- Buildings take time to heat and cool, so there is a lag in demand after a change in temperature.
- The time of the sunrise and sunset, which varies daily, can impact forecast accuracy.
- Regulations for incorporating renewables in the generation portfolio mean that you may need to improve wind and solar forecasts
Just when I began to wonder how Andy sleeps at night, with the ever present need to cut each percentage error off the forecast, he warned us of the risk of “overfitting the model.” That happens when the model becomes too complex. An overfitted model is very good at describing the data used for training the model, but is a poor predictor of future behavior.
I just polished my crystal ball and I guarantee we’ll be hearing more about the importance of forecasting for utilities in the years to come. It’s not a new problem for the business, but the increased volume of data from the smart grid enables more granular forecasting techniques.
For a white paper on the topic, read "When One Size No Longer Fits All - Electric Load Forecasting with a Geographic Hierarchy."