Machine Learning in the IoT with Apache NiFi
Download as PPTX, PDF7 likes2,939 views
The document discusses the integration of machine learning in IoT using Apache NiFi, emphasizing its advantages such as data provenance and ease of use for handling telemetry and forecasting data. It outlines various approaches for implementing dynamic demand response and the limitations of using NiFi for complex logic flows. Future steps include improving error handling and model validation through enhanced architecture.
Machine Learning in the IoT with Apache NiFi
- 1. Machine Learning in the IoT with Apache Nifi Michael Bironneau April 2017 @OpenEnergi
- 2. What problem are we solving?
- 3. Image from Wiki Commons https://en.wikipedia.org/wiki/Pearl_Street_Station
- 4. 0 5 10 15 20 25 30 35 Installed Capacity (GW) Generation (GW)
- 6. Our Solution
- 9. 0 2 4 6 8 10 12 14 16 18 20 0:00 2:30 5:00 7:30 10:00 12:30 15:00 17:30 20:00 22:30 MW Total Power Average upwards flex – 120% Average downwards flex – 35%
- 12. Our Data • ~20k telemetry messages/second • ~5k messages/second report a change of state that requires secondary processing (eg. validating forecast) • Most messages require aggregation for reporting purposes
- 13. Why Apache Nifi? • Data provenance • Built-in mechanism for backpressure and fault handling • Easy to use • Built-in processors for Azure services • Easy to extend • Performance not our main concern, but nice to know that it scales
- 14. Downsides • Source control of flows – possible but diffs not very readable • Automated flow testing and CI still remain difficult • Script components not easy maintain • Not all processors work in clustered mode
- 15. Examples
- 16. Computing Response After Dispatch 0 2 4 6 8 10 12 0 5 10 15 20 25 30 Response(kW) Time Elapsed (s) Dynamic Demand Response -2 0 2 4 6 8 10 12 0 5 10 15 20 25 30 ActivePower(kW) Time Elapsed (s) Connected Power Consumption Response baseline Duration of request
- 17. Extract JSON properties Lookup previous state and cache current state Compute and publish state change metrics
- 18. Dynamic Demand Response Forecasting -2 0 2 4 6 8 10 12 0 5 10 15 20 25 30 ActivePower(kW) Time (s) Forecasted Response Before After? Dispatch Request Forecast
- 19. Extract properties from JSON Metadata/state lookups and caching Score model (Python Script) First approach - pure Nifi solution
- 20. Observations • Fun example, but not practical • Nifi scripting is not easy to test or maintain • Long, messy flows are not easy to troubleshoot
- 21. Extract JSON properties Filter Get Forecast 2nd approach – Use Nifi as Orchestrator
- 22. Observations • As practical/maintainable as the HTTP service • Where did all the logic go? This is boring! • Why use Nifi at all? – Traditional stream processing (eg. Storm) – Serverless (eg. Azure Function)
- 23. 0% 5% 10% 15% 20% 25% 30% 35% 40% 135 140 145 150 155 160 165 7:12:00 PM 12:00:00 AM 4:48:00 AM 9:36:00 AM 2:24:00 PM 7:12:00 PM 12:00:00 AM 4:48:00 AM 9:36:00 AM 2:24:00 PM MeanSqErrorOverDay BitumenTankSetpoint(DegC) Date Forecasting Error Setpoint Mean sq Error Real-time Model Validation Setpoint Change Invalid Model
- 24. Forecast Receive data Observe Difference Increment Accumulated Square Error Fit model parameters Y N Acceptable Error?
- 25. Extract JSON properties Filter Get and cache forecast Validate and re-fit if required Real-time Validation with Nifi
- 26. Next step • Store the errors in a max-heap and use these to retrain in a priority order • Better reporting
- 27. Architecture
- 29. Model Registry/Proxy Model 1 Model 1 Persistence Enrichment/ Aggregation Forecasting/ Optimisation
- 32. Thank you for listening
Editor's Notes
- #4: We are trying to improve the efficiency in power networks. About 20% at the time of the first power station – still only about 25% now.
- #5: The under-utilisation is even worse for renewables! Consumers end up footing the bill for this chronic inefficacy.
- #6: Why? Because we think we can’t control demand, so we have to over-supply in case of spikes…
- #7: Let’s control demand!
- #8: The gateway contains hard-coded information on the assets it controls, sensors that help it tell when the asset has stored energy and constraints (such as peak tariff avoidance), enabling it to dispatch them when grid frequency is too low or too high.
- #9: The aggregation means that each asset need not be a proportional control to grid frequency, but remains free to perform operational duties 94% of the time – our service is invisible to the end customer (except for the monthly checks).
- #10: Dynamic Demand can deliver approx £85,000 per MW/Yr FCDM / Static FFR £22,000 - £26,000 per MW/Yr STOR - £10,000 - £15,000 per MW/Yr
- #11: - Open Energi is turning the energy system on it’s head, so that instead of supply adjusting to meet demand, demand adjusts to meet supply By harnessing small amounts of flexible energy demand from energy-intensive equipment we can create a virtual power station and displace fossil-fuelled peaking power stations This is enabling a user-led transformation in how our energy system works, so that businesses and consumers are not only making it happen, but also seeing the benefits It’s a vital part of our transition to a zero carbon economy because we cannot maximise our use of renewables unless our demand for energy becomes more responsive
- #12: Basically, we’re 20x cheaper than building a new power station because we just tap into existing infrastructure.
- #13: This is not huge data on its own, but Low latency requirement for aggregations One message can feed into multiple streams
- #15: There are ongoing discussions to improve flow testing, CI and source control.
- #20: This is only one half of the flow!
- #23: To the third point, using Nifi gives better traceability, instantaneous feedback on pipeline health (i.e. metrics) and a simple UI.
- #25: Timeframe for all this – minutes to hours.
- #26: As an output of the PostHTTP processor we get not only the forecast but also expectation of error. We keep track of the accumulated square error, so that we can have a single “reduceable” map key in the distributed cache.
- #29: Nifi cluster – 5 nodes, 28 flows Flink cluster – 4 nodes, 16 jobs Persistence – Azure
- #31: In blue – tools used primarily by data science team. In grey – tools used primarily by software team. Others – shared infrastructure.
- #32: Nifi Auditability Shallow learning curve (easy to use) Nice UI Flink Ultimate control Windowing Steeper learning curve