16 | New Scientist | 3 August 2019
MORE than half a billion years
ago, this stalk-eyed minibeast
(Isoxys auritus) cruised the seas
in what is now Yunnan, China.
This remarkable fossil is one of
55 from the Chengjiang deposit
on loan to the Oxford University
Museum of Natural History, most
of which have never been seen
outside China before.
The fossils date back more
than 500 million years to
the Cambrian explosion, an
evolutionary big bang that gave
birth to modern ecosystems and
all the basic types of animal we see
today. The Chengjiang deposit has
become famous for its exquisite
specimens, many of which are
changing our understanding of
how animals evolved.
I. auritus was related to insects
and crustaceans, and was probably
a predator. This fossil is on display
as part of the museum’s First
Animals exhibition. ❚
Fossils
ARTIFICIAL intelligence is being
used to make the UK’s forecasts
for solar power generation more
accurate. This could lower energy
bills for consumers as well as
carbon emissions.
The country’s energy system
is becoming more reliant on
renewable sources of electricity
that tend to have a variable
output, like solar. About 36 per
cent of the UK’s electricity was
generated from renewables in
the first quarter of 2019.
“The growth in solar was much,
much more fast-paced than
anyone anticipated,” says Cian
McLeavey-Reville at National
Grid Electricity System Operator
(ESO), which balances supply
and demand in England,
Scotland and Wales.
But solar panels are connected
to local distribution networks
rather than the national network,
making it difficult for National
Grid ESO to monitor their activity.
Combined with the trickiness
of forecasting the weather,
that makes forecasting supply
from solar panels difficult, says
McLeavey-Reville.
Now National Grid ESO has
teamed up with the UK’s Alan
Turing Institute to make more
accurate solar power forecasts.
Previously, these were based on
two pieces of data: installed solar
power capacity and the amount
of the sun’s energy that hits Earth.
“While this simple model can
work relatively well for forecasts
at very short times ahead, the
accuracy degenerated quite
rapidly,” says Andrew Duncan
at the institute.
Duncan helped build a machine
learning model that trained itself
on scores of variables, including
temperature and technical
specifications of the solar panels.
These were fed into an algorithm
to produce forecasts of solar power
entering the grid.
Forecasts for seven days’ time
instantly became about 10 per
cent more accurate. Combining
this with further modelling
made those forecasts 33 per cent
more accurate.
This is good news on two fronts.
Balancing rapid changes in supply
is expensive because reserve
power stations suddenly need
to be called on. Such balancing
costs up to £1 billion a year,
which is ultimately paid for
through household energy bills.
Better forecasting could reduce
that by as much as £50 million.
It should also lower carbon
emissions, since most reserve
power plants burn fossil fuels. ❚
Machine learning
AI could lower household energy bills
Graham Lawton
Ancient predator on show
Exquisite Chinese fossils on display in the UK
OXFORD UNIVERSITY MUSEUM OF NATURAL HISTORY
News
“ Solar power forecasts for
seven days’ time instantly
became about 10 per cent
more accurate”
Adam Vaughan