An electric utility’s 3-part guide to peak shaving with neural networks.

Research in coordination with the Open Modeling Framework.

Forecasting technology has given utilities an opportunity to flatten their load curves, raising a whole new family of questions. Below are solutions to important questions that can save utilities a good deal of money by reducing capital and operating expenses from peaking power plants. All testing can be found here.

This research can also be viewed on my website:

⚡️ Load forecasting and peak shaving with neural networks

Part I: What’s tomorrow’s load?

Using neural nets to predict tomorrow’s electric consumption

Main takeaways:

• To get any kind of useful energy consumption forecast, simple machine learning isn’t appropriate. Deep learning, however, can get us the accuracy we need.
• Given historical load and temperature data, a straightforward neural network can give a 24-hour forecast with about 97 percent accuracy.

Predict daily electric consumption with neural networks.

Main takeaways:

• A day-long approach to load forecasting is more accurate than an hour-by-hour approach.
• Despite having only a 1 MAPE difference in error between the two approaches, tests showed the method doubling our savings when peak shaving due to less eratic output.

Part II: But is it the monthly peak?

How short-term forecasting with neural nets can inform long-term decisions.

Main takeaways:

• Making peak shaving dispatches every day can be costly.
• Multi-day forecasts can help us dramatically reduce the number of dispatches per month without the risk of missing the monthly peak
• Utilities would need to set their own priorities, but for the most part, they can dispatch only once every week or so while only missing one peak every few years.

Part III: Okay, we’re dispatching. How much should we trust the forecast?

Peak shaving with unreliable forecasts.

Main takeaways:

• Because there’s inevitably errors in our forecast, the “optimal” dispatch solution for the forecast won’t necessarily be the best dispatch in practice.
• The heat equation can be used to spread out our dispatches (e.g. if our predicted forecast would suggest dispatching 500kW at 12pm, the equation might return 150kW at 11am, 200kW at 12pm, and 150kW at 1pm.)
• This simple approach can save a lot of money. Savings in one region was increased by more than 60%.
• The equation requires two constants as inputs, but they shouldn’t be hard for a utility to optimize.

Questions? Corrections? Contact me and see more projects on my website.