Bottom-up energy forecasts that change with the times and adapt to the grid
Why do my time-tested and well-performing forecasts need to change?
It’s sometimes said that load forecasting (especially in the short term) is essentially a “solved” problem. This false sense of success typically comes about from repeatedly achieving a relatively low forecasting error and perhaps even successfully leveraging a well-balanced selection of approaches. Some of the more advanced applications incorporate GIS into their top level forecast as well, in order to capture trends that may be isolated to different pockets of geography. Even at this regional level, energy is aggregated, smoothed, and becomes a very well behaved variable. It’s true that the last few decades of research surrounding forecasting electricity demand have provided the load forecaster with an arsenal of weapons at his or her disposal. If research has effectively solved this problem, then why do utilities have such a hard time quantifying the impact that recent technologies will have on the grid? Utilities and system operators alike are wondering what the impact of distributed generation, power storage, electric vehicles and micro grids will be on capacities and peak demands. It’s certainly a complicated question, but one that is not beyond the capabilities of the industries’ seasoned forecasters.
Ok, so what exactly is my forecast missing?
Forecasters often don’t have the luxury of time; predictions of energy, whether it be in the short, medium, or long term, can’t wait for a full blown impact evaluation or federal study to measure changes in load behavior and incorporate into the forecast. Even when this impact information is available, it is not always clear how best to practically incorporate it into existing approaches. Conversely, perhaps it is the situation where there exists enough historical information to explicitly account for the impact that certain technologies and behavior trends have on energy usage. So how is the forecast going to account for technologies that doesn’t necessarily exist yet? It’s a fact that appliance stocks and consumption patterns change over time. How much forecast inaccuracy stems from the inability to account for (or even know about) future technologies? This may change, however, with the advent of low cost methods capable of monitoring, managing and predicting this uncertainty.
The question remains – how do we account for an ever changing landscape of energy consumption and peak demand? How do we manage the uncertainty of a changing grid with ever evolving technologies and load profiles? A solution only exists when we stop thinking about changes at the aggregate and focus on building up the forecast from the bottom up.
We have the technology, but can we (re)build the models?
Contrary to popular belief, bottom up forecasting does not necessitate placing a smart meter in every home in America and tracking occupancy behavior like some well-meaning manifestation of Big Brother. The term ‘bottom up’ is generic enough to allow flexibility in its implementation, and yet does not relent in its defining principal of granularity.
Bottom up forecasting includes information from the following levels:
- Distribution Grid: At this level, load from substations and feeders can be used to identify geospatial clusters or regional pockets where load behaves very differently from one another. A bottom up approach which uses only distributional characteristics will not be able to predict how load changes at the household or lower level, however.
- Premise or Service Point: Sometimes referred to as the Holy Grail that is AMI data collection, household level consumption patterns are granular enough to detect acute changes in behavior and technology adoption, and variable enough to cause problems in statistical modeling. Directly modeling premise level data makes for inaccurate site level models, but whose forecasts can be aggregated and weighted to as to provide much more accurate forecasts at the top level. This is particularly true when drivers of load occur differently across households or if there is subset of utility customers that are treated in a demand response program for example. Whether the goal is short- or long- term, a top down forecast cannot incorporate household level drivers. Premise level modeling does.
- End Use: End use data is an even worse behaving variable than premise level, and is often times recorded at zero consumption: hopefully you don’t constantly operate your clothes washer or leave your lights while at work all day. However, this level of data provides for the identification of difference in load patterns at the highest possible level and is unique to end use metering. These differences include the following:
- Economic Sectors: How do residential appliance loads compare to those of industrial heavy machinery or commercial office buildings?
- Building Types: How does small business stack up against large office parks? How do single family homes compare against multi-family apartment complexes?
- End Use Categorization: What are the saturation levels of central AC units across low income housing? What types of end use behavior can we observe at government buildings that we can’t observe at similar-sized large commercial complexes?
There are distinct end uses categories across building types, and there are distinct building types across economic sectors; quantitative and qualitative information is available at each level.
This is not news to many – more and more utilities are adopting AMI across their service territories, and many are choosing to take that extra step to make the investment in end use metering. Service providers are choosing to develop deep data warehouses and even richer load profiles. In addition to whole house metering, these investments include technologies related to non-intrusive load metering (NILM). Monitoring end use levels of consumption for entire service areas is not practice or feasible, but a representative sample can be drawn and used to produce 8760 (e.g. annual) and other load shapes. The data collected from these load research samples paired with advanced analytics provides information such building stock, end use saturation, end use UEC’s and end use coincident load factor. With these characteristics developed for each end use within each building type or customer class, a highly competitive bottom up forecast can be produced.
An Illustrative Example
Figure 1 shows a bottom up forecast. Specifically this is a long term, bottom up energy forecast using simulated end use data from a US utility. For sake of illustration, the figure shows 25-year forecasts for four major end uses among all single family homes in the utilities’ service territory. We depict three extremely simplistic scenarios:
[A] – Business As Usual: End Uses (e.g. UEC’s, Saturation Levels) and housing stocks follow a fixed rate of growth.
[B] – Introduction of an AC efficiency Standards: Housing stocks follow a fixed rate of growth, but the efficiency level of air conditioners increase by 30% between 2021 and 2022.
[C] – Increased Saturation of Dishwashers: Housing stocks follow a fixed rate of growth, but the saturation level of dishwashers increases by 10% between 2025 and 2026.
Figure 1: Long Term Scenario Energy Forecasting with End Use Data
How can DNV GL Help You Develop Bottom up Forecasting
The amount of available quantitative and qualitative customer data is growing significantly as AMI solutions are increasingly adopted by utilities and primary and secondary data collection mechanisms are available at lower cost. This yields an increasingly granular comprehension of what happens at the individual household appliance level. Computing and storage power allow for Big Data analytics that enable data and event driven predictive modeling of customer behavior. As a result, more precise forecasting of energy consumption at the individual customer level becomes a reality. It opens up a whole new world of accurate bottom up forecasting starting at the individual appliance level via the aggregated household all the way to utility service areas and complete territories.
DNV GL can help utilities with a deep bench of knowledge and experience in all data science related matters. We know what it takes to discover data, develop use cases and algorithms, and identify customer behavior models to drive reliable bottom up forecasting. For more information about DNV GL’s capabilities in data analytics and modeling services, read more here.