Behind-the-Meter PV Fleet Forecasting and Influence on Utility Load Forecasts

The penetration of customer-sited PV systems continues to grow as a result of reduced costs to the consumer, innovative business models such as third-party ownership and state-mandated renewable portfolio standards. This expansion of installed PV systems is increasing the amount of on-site energy generation leading to a change in the shape of the load profile during critical parts of the day. The increased penetration of PV impacts the capacity generation considerations for utilities and balancing authorities. In addition, it does this inconsistently across hours of the day or day of the year. This is because PV generation is variable by nature across multiple time domains: the production of a fleet of PV systems may not produce the same amount of energy today as it did yesterday.

In California, where nearly 40% (1) of the nationwide solar generation occurs, the California Independent System Operator (ISO) and state utilities will need to balance the swings in “net load” caused by the variability in solar generation by scheduling reserve resources with increasingly steep ramp rates and increasing associated costs. The net load profile presents an additional ramp-down in the morning and a bigger, steeper ramp-up before peak load in the evening. The magnitude of these additional ramps will only increase as additional BTM PV systems are installed. The absolute accuracy of current load prediction methods will continue to decrease as overall PV capacity grows within a balancing area. This will have a direct impact on the California ISO’s ability to manage and purchase reserve capacity.

Forecasting production of BTM PV systems, however, is not completely uncertain and can be predicted using both cloud motion vector (CMV) satellite and numerical weather prediction (NWP)-based forecasts. Fundamentally, PV production forecasts are heavily reliant on the ability to predict cloud cover. Accurate prediction requires a good understanding of spatial cloud correlation and movement and generation, which must be factored into operational PV production forecasts.

The California ISO is charged with maintaining the reliability and accessibility of operation for the region's electricity grids. Its balancing area covers zones serviced by three of the largest US investor-owned utilities (PG&E, SCE and SDG&E), all of which are experiencing rapidly increasing numbers of BTM PV systems. Figure 1 illustrates the extensive coverage of BTM PV generation which California ISO load estimates are subject to.

This paper will discuss results of recent Department of Energy Solar Utility Networks: Replicable Innovations in Solar Energy (SUNRISE) and California Energy Commission Electric Program Investment Charge (EPIC) funded research in which Clean Power Research partnered with the California ISO and Itron to integrate fleet BTM PV forecasts, for all PV systems within the state of California, into their Short Term Forecasting Group’s Automated Load Forecasting System (ALFS).

Supplied by Itron, ALFS utilizes an artificial neural network methodology that incorporates forecasted weather, such as minimum and maximum temperature, and conditions including type of day and business hour to determine the California ISO forecast of regional electricity demand. The system “learns” how to improve its load forecasting accuracy with increased experience based on historical observations. Itron’s ALFS currently supplies load forecasts to over 60% of North America’s ISO and Balancing Area Authorities.

Overall, ALFS forecasts with BTM PV input (ALFS-BTM) hour ahead (HA) and day ahead (DA) forecasts performed better than baseline ALFS forecasts when compared to actual load data. Cloudy day ALFS-BTM forecast improvement was substantial and reached as high as 30% for cloudy hours observed in this analysis. Specifically, ALFS-BTM DA forecasts were observed to have the largest reduction of error during the afternoon on cloudy days as noted in a five-day case study period (not shown). However, DA ALFS-BTM forecasts showed little or no improvement during early morning hours due to the higher morning load variability associated with several CAISO forecast zones.

Inconsistency in overall ALFS-BTM forecast results led Itron to evaluate various ways of incorporating the forecast results into their ALFS framework. Figure 2 shows ALFS forecasts with and without BTM forecast input utilizing two different ALFS neural network training methods. Here we see that one method (#1) responds better at shorter forecast time horizons while the other (#2) responds better at the longer forecast time horizons. These results indicate that an optimized hybrid ALFS forecast solution would be the best way to incorporate BTM PV fleet forecasts into its framework.

This paper will expand on these results including further analysis and results BTM PV fleet forecasts have on California ISO ALFS load forecasts broken down by forecast zone, time of day and year.