big data meets renewable energy: building a real-time
TRANSCRIPT
Big data meets renewable energy: Building a real-time asset management platform
Stamatis Stefanakos, D ONE
About
▪ D ONE: Based in Zurich, established in 2005
▪ Data - Data - Data
▪ International & Swiss clients
▪ Dogfooding: Investor for data-driven startups
Power generation mix (EU; 2017)
Data source: Agora Energiewende and Sandbag (2018): The European Power Sector in 2017. State of Affairs and Review of Current Developments.
Renewables are bigger than you thought?
Wind
Solar
Biomass
Hydro
11%
6%
4%
9%
Net electricity installations (in GW; EU; 2000—2017)
Data source: Wind in power 2017, WindEurope, windeurope.org
Wind and solar are leading in new investments
So, all is great! What is the need?
▪ Unreliable performance → renewables have no switch!
▪ Deregulation → more risks, less return
▪ Relative secrecy → hard access to data
Actually: Investors in renewables are having a hard time
Making renewable investment more attractive through the power of data analytics
IMPROVE PERFORMANCE&
MAXIMIZE ROI&
FIT FOR MARKET
DATA POWER&
BRAIN POWER
WIND&
PV&
HYDRO&
BIOMASS
The “True Power Platform”Integrated data platform to make renewables predictable
Performance analysis
Performance benchmark
Portfolio aggregation
Production forecast
▪ Expected power vs as is
▪ As a function of weather and technology
▪ Availability & loss analysis
▪ Per turbine, park, portfolio
▪ Weather adjusted & normalised
▪ Trends over time
▪ Support strategic decisions (invest / divest)
▪ Bottom-up▪ Dispatching based on
plant condition
How a wind turbine generates electricity
GearGenerator
Tower
Blade
Rotor and pitch system
Nacelle
Yaw drive
Controller
Nacelle: Sits on top of the tower and contains all components.
Pitch system: Rotates the blades and can take them out of the wind.
Yaw drive: Orients the whole structure to face the wind depending on wind direction.
Gearbox: increases rotational speed from the low-speed rotor to the higher speed generator
Generator: Transforms mechanical energy into AC electricity.
Controller: Starts and stops the turbine from working.
The physics behind the data: Theory vs realityWhy does production deviate from the power curve?
Power*
Wind speed
Cut-in speed
Rated speed
Cut-out speed
Rated power
* Power as a function of wind-speed for a given air-density. Typically given at sea level for a fixed temperature.
The physics behind the data: Wind
Power available: Pin = 1/2 A 3
= wind speed [m/s] = air density [kg/m3]
A = 2 = radius of windmill wings [m]
Produced power: Pout = p Pin [W = 1J /sec]
p = efficiency constant provided by the manufacturer.
Betz’s law : p < 16/27 (~60%)
Typical wind turbines achieve at peak 75% to 80% of the Betz limit.
How much power can we achieve?
Pin Pout
A
How to estimate power expected?
▪ Estimate the meteorological parameters - Correct the wind speed (typically measured behind the blades). - Estimate air pressure (which depends on altitude but also on temperature). - Still relies on having an accurate efficiency constant.
▪ Learn the power curve from the data - Using e.g. neural nets. - Assumption: you have historic data with no actionable inefficiencies (or only sporadic that will average out). - Advantage: takes the efficiency constant from the OEM out of the equation.
Setting off to build the platform
▪ Start-up mode - Use cases well understood. However: pivoting is always a possibility. - Not much data to start with.
▪ People - End clients, energy professionals, data scientists. - Analysis of data with SQL-only skills required.
Our understanding of the requirements in the beginning of the project
▪ Data - Batch to start with (aggregated to 5-min). - 100 measurements per second. - Weather from turbine or park, weather from 3rd party, yaw, pitch, rotor, shaft, gear, generator.
▪ Algorithms - MVP first. - Fast-track & EoD.
Azure: Components for λ architectureMix & match: Multiple options even within one platform
File
Stream
DB
Batch
Stream
Ingestion Storage Processing
Ingestion Processing
Code Synchronization
Serving
Source: https://www.analytix-unboxed.com/mapping-microsoft-azure-analytics-components-to-lambda-architecture/
Data Factory
Sqoop
Kafka IoT Hub
Event Hub
Blob
HDFS
Data Lake Store
Data Lake Analytics
Batch
Hive | Spark SQL DW
Stream Analytics
StormSpark Streaming
Event Processor Host
SQL Azure
SQL DW
Document DB
Analytics Services
Redis Cache
Search
HBase
System architecture: λ (or φ?)Platform is built on Azure with three processing lines
Input processor & storage
Front-end layer
User application
Visuals
Streaming line
Fast-track batch
EoD batch
SQL ServerWeb application w/
embedded Power BI & R visuals
Custom made queuing(python)
PV, WT, weather measured & FC
File
OPC Stream
Azure Blob storage
A closer look: Analytical batchData vault and analytics in the core of the platform
Azure Power BI Embedded SQL Server + RAzure Blob Storage
Data LakeRawVault
BusinessVault
Power BI ServingAnalytics
Inbound over
polybase
data movement within DB
Deep Dive: Azure Storage Options
▪ Unstructured data with frequent and fast retrieval → Azure Blob Storage
▪ Analytics on stored data → Azure Data Lake
▪ Both frequent and fast data retrieval + analytics → Blob + SQL server (winji)
Azure Blob Storage vs Azure Data Lake Store
Azure Data Lake Store Azure Blob Storage
Purpose: Optimized for parallel big data analytics Not optimized for analytics
Access: Access via Azure Active Directory at file & folder level Access via Access Keys
Price: High storage, low operation price Low storage, high operation price
Suggestion
PolybaseWhy move data?
Azure Power BI Embedded SQL Server + RAzure Blob Storage
Data LakeRawVault
BusinessVault
Power BI ServingAnalytics
Inbound over
polybase
▪ Simplicity Reduced complexity : - No need for external tools to access data - Access data using T-Sql
▪ Productivity Analysis without moving data before implementation
Rationale
Analytics with SQL Server + RAnalytics, all you can eat
Azure Power BI Embedded SQL Server + RAzure Blob Storage
Data LakeRawVault
BusinessVault
Power BI ServingAnalytics
▪ Flexibility Broad analytical options
▪ Simplicity Ready to go R services
▪ Productivity Integrated R services allows fast ETL
Rationale
Deep dive: Data vaultEndless scalability & auditability
Azure Power BI Embedded SQL Server + RAzure Blob Storage
Data LakeRawVault
BusinessVault
Power BI ServingAnalytics
Rationale
▪ Flexibility Parallel workloads and data ingestion allowing automation
▪ Scalability Extendable model without re-engineering
Deep dive: Data vaultInmon, Kimball, and now Linstedt
Associations
Details
Business Keys
Farm
Sat
Sat
WT
MeteoLink
Sat
Owner
Sat
Sat
LinkLink
Sat
WT Type
Sat
Data Vault
Dimensional
OwnerFarm
Meteo
WT Fact
WT Type
3NF
Farm
WT Fact
WT LI
Owner
MeteoWT Type
Deep dive: Data vault“A repeatable modelling technique”
LINK_FARM_WT
PK Hub_Farm_WT_Key Char(32)
UF Hub_Farm_Key Char(32)
UF Hub_WT_Key Char(32)
Hub_Load_Dts Datetime
Hub_Record_Src Varchar(50)
Farm_Name Varchar(50)
WT_Code Varchar(25)
HUB_WT
PK Hub_WT_Key Char(32)
U WT_code Varchar(25)
Hub_Load_Dts Datetime
Hub_Record_Src Varchar(50)
HUB_FARM
PK Hub_Farm_Key Char(32)
U Farm_Name Varchar(50)
Hub_Load_Dts Datetime
Hub_Record_Src Varchar(50)
SAT_FARM
PF Hub_Farm_Key Char(32)
P Hub_Load_Dts Datetime
Hub_Record_Src Varchar(50)
HASH_DIFF Char(32)
Farm_Local_Id Varchar(50)
Manager Varchar(50)
Installation_date Datetime
SAT_WT
PF Hub_WT_Key Char(32)
P Hub_Load_Dts Datetime
Hub_Record_Src Varchar(50)
HASH_DIFF Char(32)
WT_name Varchar(50)
Installation_date Datetime
- Increase in the number of data objects- Complexity is shifted downstream- Contra-trend?
Hubs : Unique list of Business Keys
Links : Unique List of Associations / Transactions
Satellites : Descriptive Data for Hubs and Links (Type 2 with history)
Critics
Delivering an IoT project
▪ Data Not a given in win-ji’s case: need partners and early projects to develop solution.
▪ Technology & Architecture Important to have flexibility (e.g., Azure: pick & play, multiple solutions for different parts of the puzzle). But: Finding the best architecture can be complicated (tools, compatibilities) - this should not be delaying you.
▪ Team You need the right mix with data science, system engineering, and most importantly: subject-matter expertise aiming high in the value chain.
Learnings
Delivering an IoT project
▪ Top management buy-in is crucial A given in win-ji’s case.
▪ From “blind” to data-driven: Short time to value 5-10% increase in power output in most cases realistic without any sophisticated ML.
▪ Are you in a fast changing environment? Be prepared to adapt. In the case of renewables the market is rapidly changing due to deregulation.
Learnings (cont’d)