Driving the Growth of Solar Energy Through Machine Learning

kWh Analytics, a data analytics start-up based in San Francisco, CA, is trying to lower the cost of solar energy through machine learning. Is it developing the right product?

Driving the Growth of Solar Energy Through Machine Learning

Can machine learning help lower the cost of solar energy? kWh Analytics, a data analytics start-up based in San Francisco, CA, believes it can and has secured $1.25 million from the Department of Energy to invest in new product development.[1] The company is seeking to build a machine learning model, which will utilize its real-world data repository on operating solar projects to quantify solar panel degradation[2] rates on an ongoing basis. The aim is to lower the costs and improve the reliability of solar energy by moving the industry away from the current practice of assuming a static annual degradation rate,[3] regardless of panel type or manufacturer, which can greatly under- or over-estimate a solar project’s energy output. The application seems promising and the company well suited to developing the new product, but has kWh Analytics set its sights on the right product?

The cost of solar photovoltaic (PV) installations fell 68% from 2010 to 2017,[4] driven largely by reductions in the cost of solar panels and other critical hardware.[5] In fact, the cost of electricity from utility-scale[6] solar PV has fallen below the cost of electricity from conventional sources such as nuclear and coal – and is cost competitive with natural gas.[7] Yet kWh Analytics saw an unaddressed opportunity to drive down the cost of solar energy from another angle – financing costs. Believing the solar PV asset class to be mispriced by financiers, kWh Analytics developed a repository with data on the operational performance of solar PV projects – now comprising 20% of American solar assets[8] – with the aim of persuading lenders and equity investors to lower their return requirements. The thesis was that an operating solar PV project with a long-term contract in place for its energy output is in fact a very low-risk asset with highly predictable revenues and expenses over its life and that robust historical data would convince financiers of this fact. Utilizing their data repository, kWh Analytics offers two products: HelioStats, a risk management platform that allows investors to benchmark their solar investments, and the Solar Revenue Put, an insurance product that guarantees solar asset performance (underwritten by an insurance company based upon kWh Analytics’ data).

Machine learning holds the promise of helping kWh Analytics develop a product that goes beyond mere data collection and enables forward-looking solar panel degradation estimates. In the near-term, kWh Analytics aims to put a price on panel reliability via lower insurance premiums on their Solar Revenue Put for the highest reliability solar panels, as well as to identify the causes of solar panel degradation, enabling manufacturers to improve reliability.[9] While the former clearly aligns with kWh Analytics’ business model (the company charges a fee for the insurance policies they arrange),[10] it is unclear how the company will monetize learnings from the latter. In the longer term, kWh Analytics aims to expand its data repository to encompass the majority of American solar assets and take its insurance product from niche to industry-standard. Currently many solar project owners in the industry are skeptical about sharing their proprietary operational data with kWh Analytics, but if the company can develop a product that pairs machine learning with the project owners’ data to provide actionable insights on forward-looking degradation rates, it would alter the value proposition and could greatly expand kWh Analytics’ customer base.

Beyond improving degradation rate predictions, there are several potentially more impactful applications for machine learning in solar energy performance estimation that kWh Analytics should consider as it expands its product suite. The most important input for predicting solar energy generation is in fact solar irradiance[11] and academic studies have shown that machine learning models can improve solar forecasting by 27%[12] to 30%[13]. As intermittent sources of electricity such as solar and wind comprise a greater share of the electricity on the grid,[14] accurately predicting performance is critical to their successful integration with conventional, dispatchable sources of electricity.[15] A more accurate, machine-learning based solar forecasting product would hold value not just for project developers and project owners, but also for grid operators who must balance electricity supply and demand. In this context, one must ask: is kWh Analytics focusing its machine learning efforts too narrowly on solar panel degradation rates? Should the company consider expanding or refocusing the product scope to include solar irradiance?

[Word Count: 722]

 

[1] “kWh Analytics Wins $1.25MM U.S. Department of Energy Award to Quantify Degradation Rates and Increase the Affordability of Solar,” press release, October 25, 2018, on New York Times website, https://markets.on.nytimes.com/research/stocks/news/press_release.asp?docTag=201810250900BIZWIRE_USPRX____BW5109&feedID=600&press_symbol=3587534, accessed November 2018.

[2] The power production potential of a solar panel gradually degrades (declines) over time. Degradation refers to the loss of power produced relative to the rated (nameplate) power of a solar panel.

[3] The industry typically assumes 0.5% degradation per year for crystalline silicon panels.

[4] International Renewable Energy Agency, Renewable Power Generation Costs in 2017, p. 11, www.irena.org/-/media/Files/IRENA/Agency/Publication/2018/Jan/IRENA_2017_Power_Costs_2018_summary.pdf?la=en&hash=6A74B8D3F7931DEF00AB88BD3B339CAE180D11C3, accessed November 2018.

[5] E.g. inverters, which convert Direct Current (DC) electricity produced by solar panels to Alternating Current (AC) electricity used by the electric grid, and trackers, which rotate solar panels over the course of a day to keep them oriented towards the sun in order to maximize electricity production.

[6] Utility-scale is typically defined as a solar PV plant with a capacity greater than 5 megawatts AC.

[7] Lazard, “Lazard’s Levelized Cost of Energy Analysis—Version 12.0,” November 8, 2018, p. 2, www.lazard.com/media/450773/lazards-levelized-cost-of-energy-version-120-vfinal.pdf, accessed November 2018.

[8] kWh Analytics, “Solar Revenue Put” (PDF file), downloaded from kWh Analytics website, www.kwhanalytics.com/wp-content/uploads/2018/04/Solar-Revenue-Put-Two-Pager-1.pdf, accessed November 12, 2018.

[9] “kWh Analytics Wins $1.25MM U.S. Department of Energy Award to Quantify Degradation Rates and Increase the Affordability of Solar,” press release, October 25, 2018, on New York Times website, https://markets.on.nytimes.com/research/stocks/news/press_release.asp?docTag=201810250900BIZWIRE_USPRX____BW5109&feedID=600&press_symbol=3587534, accessed November 2018.

[10] Jason Kaminsky, Chief Operating Officer, kWh Analytics, in-person conversation.

[11] Solar irradiance is the amount of electromagnetic radiation received from the sun per unit area (typically square meters).

[12] Navin Sharma et al., “Predicting solar generation from weather forecasts using machine learning,” October 17-20, 2011, 2011 IEEE International Conference on Smart Grid Communications, https://ieeexplore-ieee-org.ezp-prod1.hul.harvard.edu/stamp/stamp.jsp?tp=&arnumber=6102379&tag=1, accessed November 2018.

[13] Siyuan Lu et al., “Machine learning based multi-physical-model blending for enhancing renewable energy forecast – improvement via situation dependent error correction” July 15-17, 2015, 2015 European Control Conference (ECC), https://ieeexplore-ieee-org.ezp-prod1.hul.harvard.edu/stamp/stamp.jsp?tp=&arnumber=7330558, accessed November 2018.

[14] The electric grid is the network of transmission and distribution lines that connects producers (electrical generation facilities including solar PV plants and natural gas plants) with consumers (people and businesses).

[15] Conventional sources such as coal and natural gas plants can be dispatched (i.e. produce electricity) as needed provided they have fuel on site; they are not dependent on the presence of intermittent natural resources like sunlight or wind.

Previous:

User Centric AI: Can Samsung be a game changer for the AI industry?

Next:

Project Maven: Machine Learning in the Military Target Selection Process

3 thoughts on “Driving the Growth of Solar Energy Through Machine Learning

  1. Trevor – very interesting article. I found it interesting to learn how much opportunity there is to better forecast solar degradation and how variable degradation rates across panels are likely unknown and driving hidden performance gaps long-term. My initial reaction to the pursuit of solar irradiance forecasting is that it requires a completely different set of skills, data, and algorithms to solve which poses a massive barrier; however, I see their insurance products being their core business model long-term and cracking the solar irradiance forecasting would likely help them offer a broader and more robust suite of insurance products. Will be interesting to watch them in the coming years!

  2. Very interesting piece! It’s interesting to consider machine learning and the monetization of data collection in an industrial context since we see data proven to be a source of advantage and revenue for consumer facing companies, but in this context there are no advertisers to sell to or product offerings to optimize. In a business of efficiency, it will be interesting to see how they leverage their existing data to perhaps help producers, or if they pivot to a different model as you consider in your piece.

  3. Great read! I think the biggest challenge with solar developments is the lack of multiple complete cycles especially given a lot of growth has been seen in the recent years, and this industry has long project life. It is very interesting to see how the use of data can further lower the risk perception of investors, especially in developing markets, and hence lead to large scale deployments across the globe!

Leave a comment