Load Forecasting Challenges in the Wake of a Pandemic

Abhishek Mandhana

We in the retail power industry rely heavily on accurate hourly load forecasts to minimize variance across several downstream business operations: Pricing, Hedging, Scheduling, Risk, Compliance, and Management Reporting.

Some Load Serving Entities (LSEs) outsource load forecasting to experienced vendors while others build it for themselves. Many approaches exist with varying levels of accuracy, including but not limited to:

    • Multiple linear regression
    • Time series analysis
    • Machine learning
    • Splines
    • Top-down/Bottom-up
    • Arithmetic (just kidding)

The level of error against actual load is largely dependent on variables like meter type (interval data recorder or non-interval data recorder), customer industry and load factor, amount of historical usage available, data reliability, etc. Changes in these variables need to be consistently fed back to the forecasting algorithm. Some customers may get peakier, others may retain their hourly load shape but have lower baseload, or meters may be added or dropped unexpectedly. Higher forecast error leads to more variability in the business operations above and ultimately the company’s bottom line. For example, forecasting load too high will result in potentially higher capacity and transmission costs to serve for customers on fixed price contracts. That is, a supplier today forecasted a fixed unitized ($/MWh) forward cost to serve at an expected volume for a Commercial & Industrial (C&I) customer and the actual volume realizes dramatically lower when the contract begins. Because capacity and transmission costs are a fixed dollar amount, not volumetric, their cost to serve has increased on a $/MWh basis. Consequently, the supplier collects less than anticipated and is now underwater on the contract. Ouch.

C&I load forecasting is complicated and a data integrity minefield. But it is especially challenging after a pandemic. Normalizing a time series of customer usage back to a pre-pandemic level is relatively straightforward: review the pre-pandemic usage if available and adjust for weather. But this is insufficient. Will power consumption and behavior return to what it was before the pandemic? Or will it gradually ramp back? Maybe the customer’s usage is no longer what it used to be altogether and the recent lower usage is the new normal.

This is the point where the algorithms and tools reach their ceiling. The problem transitions to how power consumption will change based on customer insight. This can only be accomplished if there is a direct line of communication with the customer. How is that tackled with a book of many customers? Hammer the phone lines with your customer service team and pray is one option. Or reach out to a representative subset of customers and extrapolate to the remaining customer segment. At David Energy, it certainly helps to have what we call “flexible load” – if we know it is going to be a scorcher tomorrow and our weather sensitive load could increase considerably, we can offset the worst of it by remotely adjusting HVAC’s, thermostats, and lighting for customers on our platform. There is no perfect answer, but better information and control leads to less variability.

TL/DR: In the retail power business, load forecasting is paramount. And it gets significantly harder after a pandemic. Any level of rocket surgery and artificially learning T1000 neural machine modeling is only as good as the data being fed in. The best retail energy providers are the ones who recognize this, have better insight into their customer portfolio, and are proactive rather than reactive.

Back To Blog