Eating the Sun - 2019 Solar Eclipse

Using data science to minimise the impact of extraordinary events on national power generation.

By Mukul Pateriya, Madhura Patil, Sanand Sule, and Niladri Roy for Climate Connect Technologies 

“After the elixir of immortality was churned out of the primordial ocean, the asuras were tricked out of their share by the devas. However, the cunning asura Svarbhanu sought to still drink it and thus attain immortality. Disguising himself as a beautiful deva, he attended the divine banquet where it was to be served, sat between the Sun and the Moon, and stole a sip. But they being all-knowing, saw through his disguise, and when Vishnu approached revealed his deception. Vishnu promptly beheaded him in punishment and sent his severed head flying across the sky. Though Svarbhanu had taken a sip, the elixir had not reached the rest of his body before decapitation, and only his head gained immortality. Thus the head became celestial the body Rahu (and his headless body Ketu). Now eternally seeking vengeance, Rahu perpetually pursues the Sun and Moon across the heavens. They are sometimes caught and swallowed, and at these times the skies darken, and fear descends upon the Earth. But having no throat, nor any hands to hold them, the Sun and Moon soon reappear, and light is restored.”

- The Mahapuranas, Vishnu Purana -

Introduction

The explanation for solar eclipses in Hindu mythology is both fun, and rather graphic. In reality, they occur when the Moon passes the Sun, such that it blocks most of the Sun's light. The Moon’s diameter relative to the Earth is smaller than the Sun’s, which causes the appearance of an annulus, or ring. 

Figure 1: The obstruction of solar radiation during an eclipse phenomenon

Though a transient event of genuine wonder, it can also have real-world consequences. Notably for the energy sector as more solar generation assets become active on electricity grids. Normal machine learning (ML) algorithm outputs may not suitably constrain impacts and penalties. As such, Climate Connect had to draw upon its depth of expertise, and derive novel approaches to manage services across India through the recent eclipse event.

The team exceeded expectations, and successfully minimised the impact on customers.

2019 Solar Eclipse

An annual solar eclipse occurred on December 26, 2019. Whilst the celestial spectacle captivated India, between 50% to 90% of the usual solar energy was blocked from the Southern States, right up to the Northern States. This had a significant impact on the power generated by solar plants during the event.

Figure 2: The gradual visibility of Solar radiation across India, and reference impact on solar generation.

India’s total solar capacity is currently around 28 GW. This is already a sizable portion, and a rapid increase is expected in the coming years. So precise schedules are required to manage and balance grid operations. This is even more pertinent during exceptional events, because the normal ML-algorithm output may not suitably constrain DSM penalties. Customers were concerned given the grid balancing commitments to fulfil and mitigate for. With potentially large financial implications at stake, senior management from all customers expected us to be ready. 

The limitations of machine learning

Eclipses are very irregular, and somewhat improbable events. They do not appear frequently enough in historical data for ML-algorithms to pick them up. Nor do the underlying parameters relevant to eclipses factor in standard generation forecasting. For example, information regarding solar and moon angles. These are too complex and cost-prohibitive to code in. The portfolio of assets under management by Climate Connect did not to-date involve any eclipses. So though sophisticated, the ML algorithms had no capacity to predict generation behaviour during this time.

Humans and hackathons

Such an outlier event requires expert intervention, to derive suitable solutions. Several in-house experts – including meteorologists, operational analysts, and delivery engineers – used the classic data science tactic of a mini-hackathon. Whilst scrummed in and deeply immersed, the team simulated the conditions for every state across India containing assets. Then the response of each plant, including darkness, attenuation, and irradiance. Based on these computations, forecast schedules were derived and submitted to customers.

Successful results across the board

The figure below shows the typical performance achieved by CC across all assets and states. Time blocks were precisely forecast and kept within allowable bands to +/-15 %. The encircled area indicates the successful curve typically achieved across the event duration.

Figure 3: Down and rise ramp rate during the solar eclipse. 

Conclusion 

The team's approach of combining expertise, and scrumming together, applies to any super special event. For typical weather-related events, even ‘extreme’ weather, there is plenty of data available for algorithms to train on. But special events, such as eclipses, need special calculations and formulas, which are not readily available, and often have to be derived from scratch. Knowledge must instead be sought from other domains (e.g. Astronomy and Astrophysics).

The team's success exemplifies the best goal of machine learning, which is not total human replacement. But rather 80% automation, with 20% key human oversight.