The explosive growth of rooftop photovoltaics (PV) caught many electric utilities (and most others) by surprise, and now electric vehicles and battery storage threaten to do the same. Declining cost trajectories and favorable policies for distributed energy resources (DERs), such as rooftop solar PV, battery storage, and electric vehicles, are expected to continue to contribute to rapid adoption of these technologies. As a result, utilities are increasingly attempting to account for DERs in their planning processes, in some cases all the way down to the distribution system level.
For instance, in February 2017, California began requiring that all electric investor owned utilities (IOUs) in the state factor DERs into their distribution system planning (via rulemaking 14-08-013), and has formed a Distribution Forecasting Working Group to advance the cause (in which we participated). Additionally, in June 2018, the Nevada Public Utility Commission proposed a ruling that would require a similar degree of accounting for DERs in distribution system planning -- including DER adoption forecasting, evaluation of non-wires alternatives, and assessment of cost-effectiveness. And most recently, the state of Minnesota ratcheted up its requirements for considering DERs in its utility distribution planning via its August 30, 2018 order (Docket Number E-002/CI-18-251).
Other states are expected to follow suit as DERs continue to take hold across the country. In recognition of the challenges and pitfalls of DER forecasting, the National Renewable Energy Laboratory published a study of the value of improved distributed PV forecasting, which endeavored to quantify the costs of uncertainty in PV forecasting, and noted specifically that:
"The utility-cost impacts of misforecasting [distributed] PV adoption can be non-trivial." (NREL)
Considering that the methodologies for forecasting DERs at the granular level required for distribution planning are still evolving, we endeavor in this post series to identify key factors to consider when developing methodologies or selecting tools to forecast DER adoption in any given service territory. We agree with the assessment of the Interstate Renewable Energy Council (IREC), which noted:
"As more states and utilities evolve their plans to reflect the fact that consumer- and community-driven energy resources are a growing part of our energy picture, DER forecasting will continue to improve and be a cornerstone of future grid planning." (IREC)
This first post in a multi-part series discusses overall DER forecasting considerations; subsequent posts will dive deeper into the nuances of forecasting rooftop solar PV, electric vehicles, and battery storage. Though energy efficiency and demand response are also often categorized as DERs, we do not address them in this series owing to their longer history and somewhat better understanding of forecasting and planning for their adoption.
The first consideration in spatially forecasting any DER is the underlying forecasting architecture. We characterize two different approaches here: 1) static, and 2) dynamic.
Static approaches to forecasting DER adoption can include any number of techniques intended to fit a closed-form formula to historical adoption data. It is well understood that new technologies tend to follow an S-shaped growth pattern for cumulative adoption, and several formulae exist that can generate an S-shaped curve.
A few examples of static architectures include the Logistic, Gompertz, and Richards curves. The Logistic curve is the most prominent of the static architectures owing to its ability to be transformed and estimated using the most common regression technique -- ordinary least squares. Each of these static approaches has the advantage of simplicity due to the ease and speed of estimation (since they are effectively a simple curve-fitting exercise). Thus, in certain applications they can be very useful. However, static forecasting architectures come with the limitation that one cannot readily deal with parameters or phenomena that may change over time rather stay constant over the forecast period of interest. Thus, they are largely unable to respond to complex and rapidly changing market drivers -- such as technology cost declines, electric rate structure changes, expiration of tax credits, and changing incentives. That is where dynamic forecasting architectures come into play, as discussed next.
... static forecasting architectures come with the limitation that one cannot readily deal with parameters or phenomena that may change over time rather stay constant over the forecast period of interest.
In general, dynamic forecasting architectures include any simulation technique that allows for key model input parameters to vary over time and whose forecast will respond accordingly. For instance, as technology costs continue to come down, one can reasonably expect that the long-run market size for a product will increase, and that the rate of adoption of that technology will also increase as a result. Likewise, if expiration of a tax credit several years hence is foreseen, a dynamic model would react in a manner consistent with that.
Dynamic models also have the ability to incorporate feedback loops. For instance, as more people become aware of a technology, the more people there are who can communicate the benefits and attributes of that technology to friends, co-workers and colleagues through word-of-mouth -- which increases awareness and further adoption, leading to a virtuous cycle. And, a good dynamic model will be causal in nature -- meaning that the physical, economic, social, and regulatory factors that truly drive adoption are endogenous (internal) to the model itself.
... a good dynamic model will be causal in nature -- meaning that the physical, economic, social, and regulatory factors that truly drive adoption are endogenous (internal) to the model itself.
A graphical representation of one such model is provided in the figure below, which depicts a fully generalized Bass diffusion model implemented in a System Dynamics framework. This type of dynamic model permits incorporation of feedback and explicit representation of the causal factors impacting adoption (e.g., advertising, word-of-mouth, system costs, electric rates, load shapes, generation shapes, and non-economic attributes of the technology).
The two graphs below illustrate several scenarios for DER adoption, each of which differs depending on potential future electric rates (though the scenario could be any input variable). Complex dynamics can arise under these situations, such as a collapse of a market resulting from a step change in a rate structure (or tax credit, incentives, or other mechanisms), followed by a resurgence in out years once costs decline to the point where the technology economics improve. In contrast with static forecasting architectures, a good dynamic forecasting architecture can readily capture these complex phenomena. In today's DER environment, these changes over time are the rule rather than the exeption.
Further, a good dynamic model can not only respond to changing adoption drivers in its forecast, but also in a backcast, which is useful in calibrating a model in jurisdictions where potentially rapid or dramatic shifts may have occured in the market (e.g., significant electric rate or incentive changes), as discussed in the next section.
As with the stock market, past performance of DER adoption is no guarantee of future performance. However, there is a wealth of information embedded in past performance that can, and should, be leveraged when forecasting DER adoption. Though it is possible to forecast DER adoption in the absence of calibration to historical data, such forecasts should be considered to be far more uncertain. Some have argued that early in the S-curve of adoption, there is too little data to make use of historical trends. However, we disagree. Though the historical data when adoption is early on the S-curve is certainly insufficient to tell you what the long-run market share will be, it can tell you a lot about the current rate of growth of adoption, which is useful for near-term forecasting.
When fitting a model to historical data, the near-term forecast is somewhat insensitive to assumptions about the ultimate long-run market share. It is often only after a long period of time has passed that S-curve models fitting the same historical dataset diverge, and the point of divergence may beyond the planning horizon of interest. If, for instance, a particular feeder upgrade has a 3-5 year lead time, you need not overly concern yourself with the value of the S-curve 15 years from now.
When fitting a model to historical data, the near-term forecast is somewhat insensitive to assumptions about the ultimate long-run market share.
When one is further along on the S-curve, as illustrated below, calibration to historical data becomes even more valuable. For instance, if it is clear that one has passed the inflection point of the S-curve, as evident by inspection of the incremental adoption data (e.g., in installations/month), it significantly bounds the uncertainty on the long-run adoption. Uncertainty is still inevitable, of course, but calibration of model parameters to historical data can minimize the uncertainty.
Fortunately, historical data on DER adoption are often readily available. For instance, rooftop solar PV typically requires an interconnection agreement, each of which is typically stored in a database that can be imported into DER forecasting models. Likewise, time-series electric vehicle registration and sales data are typically available for use in model calibrations. These time-series data are valuable and should be factored into any DER forecasting process.
Forecasting adoption of DERs to facilitate electric distribution system planning necessitates spatial disaggregation. However, the granularity of the forecast needs to consider the limitations of predictive capability when disaggregating, since the uncertainty in any DER forecast inevitably increases the more granular the forecast. Overly granular models are an exercise in false precision, which will be the topic of a separate post (as it is our pet peeve in the world of modeling).
...the uncertainty in any DER forecast inevitably increases the more granular the forecast.
In the extreme, an individual household will either adopt a technology or not over the timeframe of the forecast. Even if you happened to know that, on average, 50% of households will adopt a particular technology, for any given household you are either right or wrong -- it's binary. If you estimate a given household would install a 6 kW PV system, and they never installed one, you will be off by 100% (or by an infinite percentage if we were off in the opposite direction). Such is the nature of statistics -- you really only benefit from it when we look at data or forecasts aggregated at some appropriate level.
Recognizing the inherent limitation in the predictive capability of any forecasting architecture is critical. In the world of "big data," there is a tendency to add more detail to a model than is actually supported by the data and ability to estimate model parameters. After all, if you know the demographics of every household, shouldn't you be able to better predict adoption if we model adoption at the household level (e.g., using an agent-based model or so-called "customer-level adoption" approaches)? Well, not necessarily. In fact, in many cases having an overly granular adoption model will actually result in worse, not better, forecasts.
... in many cases having an overly granular adoption model will actually result in worse, not better, forecasts.
The reason is that we only garner information about the probability of adoption by looking at aggregate data, such as historical adoption totaled across some geographic region (e.g, substation) or survey data (which only becomes meaningful if we analyze survey statistics of large populations of people). When calibrating a spatial model, you inevitably need a model coefficient that is specific to an appropriate level of spatial aggregation (i.e., that which is supported by the quantity and quality of data) to get a good fit for the spatial region of interest. Customer-level demographics might explain 30-40% of the variance, at best, so you need a spatial coefficient at an appropriate level of aggregation to explain the remaining 60-70%.
However, as soon as you introduce such a coefficient, other granular coefficients you may have added to your model become largely redundant with the spatial coefficient (see page 28 of this SDG&E report for additional explanation). In the meantime, you may have wasted a lot of effort and money purchasing, collecting, cleaning, importing, and analyzing granular data that are ultimately of little value. Thus, good DER forecasting methodologies will balance the desired granularity of the forecast with the quantity, quality, and noise in historical data that result in inherently uncertain model coefficients.
... good DER forecasting methodologies will balance the desired granularity of the forecast with the quantity, quality, and noise in historical data that result in inherently uncertain model coefficients.
Another reason introducing too much granularity into a DER forecast model may actually be detrimental is due to the well-known problem that can occur due to overfitting. Overfitting simply refers to the introduction of additional model parameters that give you a good fit to historical data, but that are not truly supported by the quantity and quality of the historical data. If you fall into this trap, you'll find yourself with a model that can backcast well, but that does a lousy job of forecasting. Paraphrasing a far wiser individual than I:
"[A model] should be as simple as possible, but no simpler." (Albert Einstein)
Though the jury is still out on the most appropriate level of disaggregation of DER forecasts, many utilities (e.g., IOUs in California) are taking advantage of the quantity and quality of historical data, customer counts, vehicle registrations, and other information at the ZIP Code level (see page 11 of this consensus report of California's Distribution Forecasting Working Group, in which we participated). In some cases, sufficient data also exist for substation-level forecasting. These calibrated forecasts can then be allocated further down (e.g., to the feeder level) using simple mechanisms (e.g., proportional to load, customer counts, or past adoption) or more complex mechanisms (e.g., propensity estimation).
Finally, one needs to be ever-aware of increasing sensitivity to the management of data with customer-identifiable-information. While in theory one might be able to take advantage of customer-specific data, there is often a reticence to dive that deeply due to the additional liabilities of and potential restrictions on usage of those data (e.g., by utility commissions or other government agencies).
As DERs such as solar PV, electric vehicles, and battery storage continue the downward trajectory on costs and upward trajectory on adoption, developing methods to forecast them with reasonable accuracy will become increasingly important to electric utility distribution planning. Doing so can reduce overall distribution system costs while maintaining system reliability. We offer several key take-aways that developers or consumers of DER forecast ought to consider:
This article is the first of a series of posts on this topic, which we expect will gain attention in more jurisdictions as the market for DERs evolves. Subsequent posts will focus on key consideration and nuances of modeling adoption of rooftop solar PV, battery storage, and electric vehicles.