The Center for Data Innovation spoke with Michael Herzig, founder and CEO of Locus Energy. Herzig discussed how Locus’ innovative use of data analytics for solar photovoltaic systems helps solar PV fleet owners and managers increase performance and decrease costs.
Joshua New: Could you introduce Locus Energy?
Michael Herzig: Locus Energy is a solar monitoring and data analytics platform provider for the PV market, spanning the residential, commercial and utility sectors with more than 50,000 systems deployed in North America. We draw data from solar PV installations, as well as from other public and private sources, to develop a proprietary solar data analytics platform, the PVIQ suite, which helps fleet operators reduce O&M costs and drive best practices throughout the solar PV asset lifecycle.
PVIQ consists of two main components, the Virtual Irradiance (VI) tool and the Performance Waterfall report. VI uses satellite and weather data, along with data from Locus’ more than 50,000 installations, to provide highly accurate information on how much sunlight is hitting the ground at a specific location every 15 minutes. The Performance Waterfall report, PVIQ’s second main element, assesses the drivers behind underperformance, including factors such as inverter malfunction, shading, and soiling.
New: Your analytics platform often relies on virtual models instead of physical sensors. Can you explain what feeds into these models and why this is beneficial?
Herzig: The cost of installing physical sensors often doesn’t make sense for smaller projects because the hardware is so expensive. Virtual models, on the other hand, are far cheaper and are applicable to all solar PV projects, regardless of size. Sensors have also been known to fail or to provide inaccurate data for a variety of reasons, from weather to maintenance. Virtual models provide a consistently accurate assessment of irradiance, bypassing physical factors that may affect the readings. In addition, they serve to validate data from physical sensors. Furthermore, sensors only collect data for a specific location, whereas the virtual model can calculate irradiance for the entire country with a spatial resolution of one square kilometer. All this said, we still incorporate data from physical sensors on the vast majority of our larger PV sites, and incorporate their data into our analytics models.
New: You anticipate incorporating machine-learning techniques to improve these models and increase efficiency. What would this accomplish, and what are some of the challenges in implementing these techniques?
Herzig: More sophisticated modeling, such as using machine-learning techniques, can yield improved model results. The key challenges with these approaches tend to be understanding precisely where to apply one of the numerous available approaches, as well as avoiding over-fitting the training data available. Machine learning techniques tend to require a substantial training data set to work well, and one can easily over-fit the data if you are not careful. Over-fitting is a problem where you think you have improved the model results because the model performs better on the training data set, but in reality you have created a less-general model that will perform less well on new data in the future. Large data sets help alleviate this problem, so we see our large data set as an advantage for algorithm development.
New: With large scale, distributed deployments of solar panels, what are the challenges of ensuring that all data collected is accurate?
Herzig: Currently, system performance expectations, modeling, and design often use blunt data inputs that are derived from hardware performance in test labs and other datasets that do not represent real-world performance. Locus uses real-world performance data, as well as striving for the most granular level of data available. For instance, the standard spatial resolution for VI is one square kilometer, and the standard temporal resolution for VI is 15 minutes. Also, VI can be integrated with real-time data from Locus’ 50,000 installations. In other words, VI is delivered at scale in short time intervals and can be integrated with real-time solar output data, allowing it to drive more meaningful insights. Locus collects more than 50 million data points daily, with a cumulative total of more than 30 billion data points collected to date, making it one of the most diverse data set in the industry. Our access to this enormous — and growing — dataset allows us to continually refine our algorithms and platforms to improve the accuracy of VI and Performance Waterfall.
In the case of irradiance, one issue faced by the industry is the potential inaccuracy of onsite sensors, which are used with larger-scale systems. The data provided by such sensors can become skewed due to soiling, miscalibration and other factors. In addition to serving as a substitute for sensors for smaller systems for which sensors are prohibitively expensive, Locus’ VI tool also functions as a check and balance in situations where onsite sensors are already in place.
New: Locus claims to have the deepest intellectual property portfolio in the industry. What are some innovative uses of data that Locus has pioneered?
Herzig: Locus has seven patents granted and nine pending for its solar PV analytics. The innovative uses of data that Locus has pioneered include the Virtual Irradiance (VI) tool and the Performance Waterfall report, which are part of Locus’ proprietary PVIQ suite of software analytics. The VI tool draws from private and public historical and real-time data from weather stations, satellite imagery, federal agencies and other sources to provide highly accurate data on how much sunlight is hitting the ground at a specific time and location. As the “fuel” for solar PV systems, solar irradiance is the most valuable data for assessing the performance of a solar PV system. Used with other modeling data, such as the type of panel or inverter, VI allows fleet owners and operators to assess if a system is meeting performance expectations.
The Performance Waterfall report remotely determines the root causes of a solar PV system’s failure to meet performance expectations. These may include weather uncertainty, snow downtime, shading, equipment downtime, equipment degradation and inverter problems. The prescriptive analysis provided by the Performance Waterfall report, which is unique in the industry, enables a specific, detailed understanding of which factors most affect the performance of a solar PV system and how the causes of the underperformance can best be addressed. By increasing yield and reducing O&M costs, analytics tools such as VI and the Performance Waterfall decrease the cost of solar, thus helping to speed adoption. The data also has applications for other aspects of the solar PV value chain, including insurance and warranty pricing and design and installation best practices.
