industries (FUDIPO) Erik Dahlquist, FUDIPO, Malardalen University - PowerPoint PPT Presentation

FUture DIrections of production Planning and Optimized energy- and process industries (FUDIPO) Erik Dahlquist, FUDIPO, Malardalen University Frankfuhrt, 18 October 2018

Partner distribution 2

At a glance sorting “good data” used for tuning from “bad data” used for fault detection 3

FUDIPO structure Sales - income Order plan Costs - expences Optimization Production plan Modified Production plan Risk of failure Maintenance OD Machine learning Decision support Soft sensors Deviation (sim- meas) Fault diagnostics Statistical models Model based control Model adaptation Physical Models Data preprocessing 4 Processes

1. Demonstrator Background: Mälarenergi AB, Block 6 High temp corrosion sensor

Results BN at CFB boiler 5 p=predict Furnace status m= measured 100% 100% Tp-Tm 90% 90% Moisture 80% 80% Tp-Tm Unballanced right Temp Unballanced left NIR 70% 70% High combustion Normal 2 Sensor 10 false 60% 60% Temp MC in Sensor 11 false Probability Sensor 12 false 50% 50% Sensor 13 false 1 fuel Sensor 14 false 0.3 Sensor 15 false 40% 40% Sensor 16 false Sensor 17 false 0.5 30% 30% 0.2 Sensor 20 false Sensor 21 false Sensor 22 false 20% 20% Sensor 24 false Sensor 25 false Temp 10% 10% 0% 0% in 0 30000 60000 90000 120000 150000 180000 210000 240000 270000 300000 Time [s] Furnace status 2011-09-10 to 2011-09-18 Cyclone 100% 100% 90% 90% 80% 80% Unballanced right Sensor Unballanced left 70% 70% High combustion fault Normal Sensor 10 false 60% 60% Corrosion Sensor 11 false Probability Sensor 12 false 50% 50% Sensor 13 false Sensor 14 false 40% 40% Sensor 15 false Sensor 16 false Sensor 17 false 30% 30% Sensor 20 false Sensor 21 false Diagnostics and decision support 20% 20% Sensor 22 false Sensor 24 false Sensor 25 false 10% 10% 0% 0% 6000 102000 108000 114000 120000 126000 132000 138000 144000 150000 156000 162000 168000 174000 180000 186000 192000 198000 204000 210000 216000 222000 228000 234000 240000 246000 252000 258000 264000 270000 276000 282000 288000 294000 300000 306000 312000 318000 324000 330000 336000 342000 6 0 12000 18000 24000 30000 36000 42000 48000 54000 60000 66000 72000 78000 84000 90000 96000 Time [s]

Learning system in a fiberline – updating models semi on-line Residual alkali Kappa number 7 Billerud-Korsnäs

Temperature was higher than predicted. Indicate channeling. -Temperature in the extraction flow during channelling: - yellow curve = measured process value - violet line = predicted value from simulation

Oil refinery at Tupras - Connection of the Physical Models Determine feed comp by NIR Optimize use of feed

Overall scheme of the WWTP at Mälarenergi (Sweden) BN and MPC Minimize electricity Reduce: NO3, NH4, BOD, PO4 Maximize biogas prod

Micro gas turbines, mCHP Fleet management of mGT plants Decision Physical and support statistical models Maintenance Diagnostics on demand Data pretreatment Measurements 11

Conclusions ● Goal to integrate different functions from low level to high level ● Build learning systems, that are self adapting ● Develop data structures that can make this posible ● Make supervised AI systems for process industries 12

Erik Dahlquist FUDIPO – Malardalen University www.fudipo.eu The projects leading to this application have received funding from the European Union’s Horizon 2020 research and innovation program