Machine Learning and Simulations for Asset Management at Alliander

Transcript

1. Machine Learning and Simulations for Asset Management March 13, 2020, NL-RSE Meetup Gijs Molenaar
2. pythonic.nl • Me! • Freelance Software/Research Engineer • MSc AI (UvA) • PhD astronomy (South Africa)
3. • Public Utility company • 5.7 million client connections • 5700 FTE
6. The problem • Fixed budget • How to optimally distribute budget over departments • Balance between long term and short term goals
7. Amsterdam • Go fully electric by 2030 • 900.000 people • 230.000 private cars • 10 years • 2610 working days • 1000 charging stations a day?
8. Our solution • Build agent based simulation of company • Run millions of simulations up to 40 years in the future • Let Reinforcement Learning agent explore and optimize problem space
9. What do we need? • A physical model • A bunch of agents manipulating this model
12. Reinforcement Learning
13. Reinforcement Learning st st+1 a
14. Deep-Q learning Simulation agent != Reinforcent learning Agent
15. State • Number of connection issues Reward • function of • Budget spend • Average waiting time for new connection • Average Waiting time for fix of broken connection Action Decision of budget split
16. Project structure • 3 teams: • Data Engineers • Data Scientists • Software Engineers
18. Evaluating 2 strategies • Compute in neo4j • Compute in NumPy + NetworkIt
20. Rpy2 R has NA values in int arrays, numpy has NaN in float arrays
21. Status report • Still working on simulations • Underestimated the effort required to get this right • Maybe too ambitious? • Useful nonetheless
22. Research Engineer in Corporate environment • Different setting • Requires different mindset • More budget • A lot of politics
23. pythonic.nl available again! • 1 June 2020 • First getting married in Namibia • If we don’t get Corona banned
24. Questions?