Speaker: Jacob Thomas, Partner - Distribution Services, GDS Associates Inc
Say your name, company and title, and how you are knowledgeable about load forecasting.
Jacob Thomas: Hey everyone. My name is Jacob Thomas. I'm a principal at GDS Associates. I have been doing load forecasting for over 25 years in my career, including developing load forecast for Public Power Utilities and Cooperative's Municipals and also reviewing load forecasts that have been prepared by others in regulatory proceedings.
What is your favorite, ideal approach to forecasting future electric loads?
Jacob Thomas: My preferred approach especially when we're talking about long term load forecasting is two fold. Number one, I prefer a bottom up approach in which you project out the various sector level energy sales and number of customers, residential, commercial and industrial. And then within that framework, I'm also a fan of the statistically adjusted in use or SAE modeling approach, especially for residential average usage and small commercial usage. The SAE approach has the advantage of incorporating economic and weather data, which is also used in traditional econometric approaches but also incorporates in things like market share of in use, electric appliances, appliance efficiencies, changes to codes and standards, home efficiency, home size, and a lot of other variables that can really drive electricity consumption within the home or within a firm.
What is the most common approach you see at electric utilities for forecasting future electric loads?
Jacob Thomas: This is a good question. And what you see in the industry really kind of depends on which classes, which sectors we're talking about for the residential and the small commercial sectors, we often see econometric or regression based statistical modeling used to forecast out electricity consumption. And I should mention that the SAE model that I discussed in the last question is regression based. I may not have made that clear. But essentially in that approach, you develop inputs into a regression model to reflect the end use efficiency, housing efficiency, those types of characteristics and then you run it through a regression. So you do get the same statistical models. In the industrial sector or a large power sector, you see more of kind of an expert opinion approach where you might canvass key account reps, you may reach out to those individual consumers and you're really trying to get an understanding of whether or not there'll be any expansions in service in the near future or contractions as well. The likelihood of those expansions or contractions coming to fruition. Are there new potential large loads that might connect in the next year too? And if so how big are they? What kind of load factor are they? So it's often just kind of more of an expert opinion that's pulled together. There are some utilities that try to project out large industrial and show kind of unnamed growth if you will as well. And sometimes it just depends on your use of the load forecast. For cooperatives that are borrowing from the RUS, for instance, the RUS prefers you to only include something kind of like a known and measurable industrial load into your large commercial forecast. But investor owned utilities that are doing integrated resource planning and you know, heavily economic developing areas, they may try to put in some unknown growth to make sure that they have capacity in their plans to meet that load.
Why is the forecast always wrong and how can we learn from that for better forecasting?
Jacob Thomas: This is a good question. I'm fond of telling my clients that their load forecast is gonna be wrong or their money back. And in 25 years, I've yet to have anybody get their money back. So I've consistently been wrong. It's one of the things in the industry where it's ok to be wrong. And so that makes it pretty nice. We're actually fortunate in the energy industry, that our load forecast are more accurate than a lot of industries. And that's because electricity consumption is really driven a lot by weather and by basic economics like adding new members. So, household growth is not the hardest thing in economics to project. And so certainly in the short term, it'll be reasonably accurate. and then you just apply weather to that and you get a pretty good sense of energy sales. What has made it complicated in recent years is that there are a lot of factors like conservation and codes and standards and that has tended to put some downward pressure on average consumption and sort of broken the link between economic activity and consumption. What we can learn from our errors is how to get better at understanding when there's divergence. How quickly can we adapt to those divergences? For many, many years? in the nineties, a simple econometric, just household income and weather was the best indicator of residential average usage. And then that disconnect started to happen. And so if you are observing your load forecast and you're watching your load forecast on a regular basis and really digging in and trying to understand why it's wrong and by how much it's off, then you can be adaptable and revise your methods in a way to maybe create more accurate forecasts quicker than those around you. I think it's also important here when we're talking about forecast accuracy to realize that a base case forecast isn't the only forecast and that any good forecaster should be preparing a variety of scenarios or a set of range forecast or using something like a Monte Carlo simulation to develop probability forecast to go around the base case. So from a planning perspective, it's always important to think about the various scenarios and what they might mean to your plants.
Open question - Any forecasting point you’d like to make for future utility leaders?
Jacob Thomas: Sure, I'd like to talk about a couple of ways that forecasting is getting more difficult in the short term and then maybe we'll talk about here in a second something that might make it a little bit easier for us in the long term. So there's several things going on in the industry that might make it harder for a load forecaster. And, and two good examples of that is electric vehicles. and so we've been kind of grappling with this for a few years now in terms of how do we project out vehicle adoption in a local service territory, how do we understand charging behaviors, driving behaviors and how that might impact electricity sales and even more importantly, how it might impact peak demand. And how are we going to encourage electric vehicle homeowners to charge off peak? The the other big interesting topic is related to the data center and Artificial Intelligence needs of the country in the world that's coming. We're seeing continued interest and requests for massive loads of very high load factors that could be cited throughout the country to serve the needs for data centers and I think we need to be careful about defining three different types of data centers. in my mind. There's the traditional just data warehouse for the Amazons and the Facebooks of the world, the metas of the world where we're just trying to store the data and make sure it stays stored. And that type of load is very high load factor. Very stringent about keeping the lights on. They are not interested in interrupt ability. They're gonna have multiple redundant backups to make sure that the power stays on. The second type of data center is more of a Bitcoin mining type of a load. And those operations are looking to get a lot of power quickly. It tends to be more modular. They can kind of drop in and drop out servers, truckloads of servers and they're usually amenable to interruptible type arrangements if it can save dollars and cents on their power costs. And then the third type that we're starting to see is artificial intelligence type data warehouses. And AI is really a consumer of electricity much in a much greater way than traditional data centers. I believe it's something like a simple chat GPT request consumes something like 15 to 20 times more electricity than a basic Google search. And and we're just going to see more and more AI. And so the question is, how do we project that load? What are the odds that gonna connect to our system as a forecaster? How much is it gonna be? Is it gonna be gigawatts megawatts and how many and how fast? And that makes it very challenging. So those are just a couple of things that's making forecasting harder in the short term. And I mentioned the AI. I also think that there's opportunity there through machine learning algorithms and more advanced statistical techniques to bring to bear on the forecasting problem. And so maybe we can use AI to solve its own energy issues. That'd be kind of cool.
