Page 10 - North American Clean Energy January/February 2019 Issue

1
•
1
0
0
/w
wind power
Data Engineering for Wind Power Optimization
The promise and challenges
J
//
JA
AN
NU
UA
A
R
RY
Y
F
•
FE
E
B
BR
R
U
UA
A
R
R
Y
Y2
m
20
0
1
19
9/
/
/
ww
by Gareth Brown
Arti cial Intelligence. Deep Learning. Neural Networks. Machine Learning. Blockchain. Big data. We are surrounded by these technology buzz words and their promise of software panaceas. e unspoken assumption is to collect a lot of data, apply these new data tools, and watch value magically appear. Many industrial businesses spend millions of dollars implementing broad-based data collection, in the blind hope of a data solution.
ww
w.
.
n
na
ac
cl
l
e
e
a
a
n
ne
e
n
ne
er
Unfortunately, these new data tools are limited. Much like a hammer, drill, or screwdriver, only when the tool is applied by specialists will its true value be unlocked.
Without a clear vision of the value they hope to extract, many rms are unwittingly engaging in poor data engineering. is severely limits the value available to them from these new tools. Ultimately, where and when the latest data science should be used will be driven by the value proposition.
e wind industry faces at least six distinct data engineering challenges that must addressed by any wind farm owner (or performance analyst) looking to take advantage of the latest and greatest data science techniques.
The precise in ow conditions are not known.
Uniquely among turbine power generation, wind turbine generators face an unknown input resource. For hydro or steam turbine generators (whether nuclear, molten-salt, gas, or coal) the enclosed intake enables in ow conditions to be controlled and well understood. From this, operators and analysts can generate clean performance signals to assess any turbine speci c challenges.
In fact, when it comes to assessing the turbine aspect of performance, nuclear power is a lot simpler than assessing hundreds of wind turbines across a hillside. Each turbine with limited instrumentation, di erent, unknown, and constantly varying in ow conditions, can have diverse input resource.
Wind in ow conditions - driven by wakes, forestry, terrain, and atmospheric stability - can create performance variation higher than 30 percent within each wind speed bin below rated power: trying
to detect an underperformance issue of 1 to 2 percent becomes a herculean task, especially if you don’t get your data model correct (regardless of the algorithm or machine intelligence applied).
The assumptions used
in wind turbine design and data collection
have not kept pace with advancements in turbine engineering or software engineering.
Turbines with larger blades, farms
with more turbines, and farms built in more varied global locations, challenge longstanding industry assumptions. SCADA and other software systems are often generations behind those of the analysis tools, even if the turbine has the latest manufacturer software upgrades.
is has led to problems both in terms of mechanical failures, and owner-operators’ available strategies for data collection. ese severely limit the e ectiveness of any modern data science techniques, without signi cant and conscious up-front data engineering e ort.
Data collection structures re ect an industry obsession with availability as the only important metric.
e wind industry’s data collection structures have been built around answering the question, why has the wind turbine stopped spinning (availability)?
r
g
gy
y.
.c
c
o
om