Over the past decade or so, the use of short term forecasting has been key to successfully integrating wind and solar power. All North American ISOs, as well as most other utilities and system operators with any meaningful wind penetration, use wind forecasting systems. Many are also procuring solar forecasts, and with penetration rates increasing rapidly, solar forecasting is becoming more widely used. In Europe, many Transmission and Distribution System Operators (TSOs and DSOs) have been using solar forecasts for a number of years. Here, I’ll highlight some notable aspects of solar forecasting, based on interactions with utilities/ISOs, forecast providers and other researchers, as well as ongoing EPRI R&D projects.
Forecasting of solar PV is similar to wind power forecasting techniques in many ways – first you forecast the weather in the area for the time horizon desired, then translate this into an estimate of power output based on what you know about the resource (e.g. panel and inverter characteristics, layout and availability). But some differences do exist. For one, solar has a defined diurnal characteristic based on the position of the sun; wind has diurnal patterns in most places, but these are more a general pattern of when output is higher or lower. The underlying shape of solar can be anticipated; the main challenges involve understanding the variance from this shape due to cloud cover, air particles, and temperature. We can more easily see many of the things that impact solar output (namely, clouds) and so can use additional techniques not as relevant for wind, such as satellite data and sky imaging; much of the interesting R&D involves these techniques. Whereas wind plants, particularly in the US, tend to be large – 10s or 100s of MWs – solar is often distributed, such as with rooftop PV. The system operator does not have visibility of the output of this resource, making it harder to develop forecasts and verify performance. At the same time, the distributed nature may also work to smooth out forecast errors.
One of the more interesting questions in this area is how to approach this area of distributed PV forecasting. ‘Top down’ approaches estimate PV output across an area based very little information – maybe just the installed PV and a few data points about the weather. ‘Bottom up’ approaches, on the other hand consider the detailed specifications of all PV (rooftop, commercial and utility-scale), forecast for each panel and then aggregate. As with most good engineering questions, the answer to which approach to use is, ‘it depends’. For one, it depends on the information available to the forecaster – some operators have access to a lot of metadata (static data about the installed PV), and live or recent output on a high temporal and spatial resolution.
There’s also a question of need – depending on the penetration level, particular application being considered and the flexibility of the system, different solutions may be required. If there is a lot of solar in a small area with relatively low amount of balancing resources and significant transmission constraints, for example, then you would want to have a pretty granular view of solar output – this might require detailed modeling approaches based on a lot of data. DSOs managing active networks may also want highly granular forecasts. On the other hand, larger areas with little or no transmission constraints and sufficient balancing resources, may only need a system-level estimation. Many system operators see the output of PV as a load modifier, and thus may mainly need to know the overall impact on load. This area is something we as an industry will need to focus on in the next few years, and I’m sure lessons learned from our current experiences in the US, Europe and elsewhere will allow us to better understand how to do this well. From a regulatory viewpoint, enough data will need be available during and after interconnection to ensure the forecasts perform at a sufficiently high level, while the closer link between transmission and distribution networks will require increased TSO/DSO coordination and information sharing.
The question of how much detail is required gets to the final topic I’ll mention, one that is often brought up by operators – the value of a forecast. It is important, as we deploy instrumentation and invest in R&D, to understand the value of improving forecasts. While it’s clear that forecasts provide value when used in operations, answering the question of how much value can prove quite tricky. From what we’ve seen, it depends quite closely on market design and/or operational practices, where certain design approaches will put more of a premium on improved forecasts than others; as an example, making decisions more frequently and closer to real time reduces the exposure to forecast error for longer time horizons, but may increase the value of a very short range forecast. Using the probabilistic information often available with wind and solar forecasts allows us to manage risk exposure, and a number of ongoing research efforts are looking at the practical use of such information. Thus, as we consider the issue of value, we should consider first and foremost how the improvement in forecasting methods allow us to better manage risk, from both an economic and reliability perspective.
Summing up – solar forecasting is now mainstream. A large number of vendors provide solar forecasts, using different methodologies, different data sources, and with varying degrees of granularity and accuracy. As we continue to look to use forecasts in a more optimal manner, we should push for the development of forecasting methods that are closely linked with system value, and design operational practices around the capabilities of forecasts. This involves close interaction between forecasters, system operators, generators and researchers, and deployment of techniques and technologies that allow us to successfully integrate solar in the power system.
Aidan Tuohy
Senior Project Manager
Electric Power Research Institute
Leave a Reply