As distributed PV (DPV) continues to proliferate at the customer end of power grids across many parts of the world, DPV’s impact on short-term load forecasting (STLF) has become a significant emerging challenge for ISO/RTOs. For several decades, STLF tools have provided industry with a key ingredient for the reliable operation of the bulk system, which involves efficient unit commitment/dispatch, effective short-term outage coordination, and operational situational awareness. Legacy STLF tools and practices are now being challenged by rapidly growing amounts of DPV embedded in the distribution system that are masking a growing share of the observed load, and thus its characteristic profile and relationship to weather. Methods of accurately forecasting DPV, and perhaps more importantly, the net load (i.e., native load minus DPV) are now needed.
Addressing these challenges begins with gaining DPV visibility that is often lacking at the ISO/RTO level. Since DPV activity frequently occurs in the retail market space, ISOs are not traditionally aware of DPV interconnection trends or its historical or real-time energy production. Therefore, gaining visibility at the ISO/RTO level includes building systems and/or processes for acquiring both ongoing knowledge of DPV installations and the growing DPV fleet’s ongoing generating performance. This is particularly important as historical DPV data is required to better understand the changing net load profile and to develop and evaluate net load forecasting methods and their implementation. For these reasons, there is a natural cohesion between the methods used to gain DPV visibility and those used for DPV forecasting.
In general, DPV visibility can be developed via simulation methods that use weather data in tandem with knowledge of the DPV fleet to provide realistic estimates of DPV production, or by utilizing a subset of representative DPV sites with metered data to infer (via “upscaling”) the performance of the entire fleet. These approaches can be applied to derive realistic estimates of DPV’s historical performance as well as its output in real-time provided that sufficient telemetry (of weather and power, respectively) is available.
As DPV penetrations increase across a region, the number of individual installations grows into the tens or hundreds of thousands. This raises the issue of whether the preferred forecasting approach should be “bottom-up,” in which the goal is detailed modeling of individual systems and subsequent aggregation of the resulting system forecasts, or a “top-down” one that leverages the statistical characteristics of a large population of DPV systems to forecast their aggregate performance. Specific design parameters of DPV systems (e.g., panel/inverter type and ratings, tilt angle, azimuth angle, shading, etc.) may not be readily available, making pure bottom-up approaches infeasible. Hybrid approaches that incorporate features of each are also possible. Whichever combination of methods is chosen, validation of the results against some amount of actual DPV data is important to ensure the full array of weather effects is well-captured.
STLF tools typically consist of statistical models such as linear regression or artificial neural networks that use a variety of calendar and weather inputs to generate quantitative forecasts. While a variety of STLF methodologies exist, in general they do not currently include the requisite features to capture DPV’s impacts. As a result, STLF performance tends to degrade as DPV penetrations increase. Modification of STLF tools and processes to better forecast load net of DPV can be done using any of the following three general methods:
- Directly incorporating DPV-related explanatory variable(s) in an integrated load/DPV model, which is then used to forecast net load directly
- Reconstituting historical DPV production into historical loads used to train a STLF model, forecasting native load and DPV separately, and subtracting the DPV forecast from native load forecast to derive net load forecast
- Bias correction of the quantitative STLF model known to be caused by DPV
Combining all of these approaches is likely warranted, as they each can be found to have strengths and drawbacks.
Given the relative nascence of DPV and net load forecasting, benchmarking of the various DPV fleet estimation approaches and forecasting methods would be a valuable contribution to industry. And, as the fields of energy forecasting and weather forecasting continue to advance and converge, there are many exciting challenges for the upcoming generation of forecasters!
Jon Black
Manager, Load Forecasting
ISO New England