DEEP LEARNING IN REFINING PROCESSES – A PRE-STUDY OF INTERNAL AND EXTERNAL VARIABLES IMPACT ON PULP PROPERTIES FREDRIK BENGTSSON*, ANDERS KARLSTRÖM* AND JAN HILL**, *CHALMERS UNIVERSITY OF TECHNOLOGY , GÖTEBORG, SWEDEN, **QUAL TECHAB, TYRINGE, SWEDEN
IS IT POSSIBLE TO CONTROL (C)TMP PROCESSES AS WELL NOW WHEN WE HAVE THE TECHNOLOGY? THE WORLD’S FIRST THOUGHT- CONTROLLED PROSTHETIC DEVICE 2019-04-11 Chalmers University of Technology 2
TOPICS IN THIS PRESENTATION • Background • Refiner control using soft sensors. Example used in this presentation: CD- refiners. • Pulp property estimation. • Future direction. 2019-04-11 Chalmers University of Technology 3
BACKGROUND(1) During the 90th we started More information about In 2015, this led to a soft a research program at refining conditions was sensor concept suitable In the old days Holmen Paper Braviken obtained from sensor arrays for refining control inside the refining zone. Temp. T g 0 0 P max 11 screw FZ FZ Y C GU 0 g 0 D T max est 22 H O FZ FZ 2 FZ T 5 T 4 C 0 g g D est 32 33 H O CD CD 2 CD P : Relates to the production screw, not Plate gap. screw D : Corresponds to dilution wa ter to each ref. zone. C : Consistency in the periphery of each ref. zone. T : Temperture maximum in the flat zone. Radial max position Dependent on plate pattern r max r 4 r 5 2019-04-11 Chalmers University of Technology 4
BACKGROUND(2) Vision: By controlling the (C)TMP-refiners, specific energy can be reduced significantly at the same time as pulp property variations are maintained at an acceptable level. To reach the vision: Introduce T max - control in the flat zone(FZ) and Consistency control in both FZ and the conical zone (CD). Keystones: 1. Manipulate the work load between the zones in an unortodox way to increase production and stabilize pulp quality. 2. Select proper operating points defined by requested pulp quality. 2019-04-11 Chalmers University of Technology 5
BACKGROUND(3) SOFT SENSORS FOR CD-REFINERS WERE INTRODUCED ON-LINE 2015 2019-04-11 Chalmers University of Technology 6
BACKGROUND(4) SOFT SENSORS FOR CD-REFINERS WERE INTRODUCED ON-LINE 2015 Flat zone CD zone Energy Mixing Energy Inlet Inlet Outlet and point and mixing mixing mixing Material between Material point zone zone balances FZ & CD balances 1,…,n 1 1,…, n 2 Mass flow (Water) Mass flow (Steam) Mass flow (Chips and/or fibers) Work related to the motor load distribution 2019-04-11 Chalmers University of Technology 7
REFINER CONTROL USING SOFT SENSORS A new set of process variables, internal states , available: • Residence time → • TCtrl – Temperature control Consistency CCtrl – Consistency control • Forces on bars → • Defibration work Increased amount of process • Thermodynamical work data to handle. • Backflowing steam → • Steam velocity Possibility to systematically • Pulp velocity analyze pulp properties, process • stability as well as different Water velocity estimation tools. • etc. 2019-04-11 Chalmers University of Technology 8
REFINER CONTROL USING SOFT SENSORS Between zones 2019-04-11 Chalmers University of Technology 9
REFINER CONTROL USING SOFT SENSORS TCtrl – in Automatic mode ; Major features Future challenge 10 2019-04-11 Chalmers University of Technology
REFINER CONTROL USING SOFT SENSORS Two data sets: 0 - 600 hr 2019-04-11 Chalmers University of Technology 11
REFINER CONTROL USING SOFT SENSORS Two data sets 0-600 hr Future Challenge – minimize variations even more by using overall MPC for specific pulp properties 2019-04-11 Chalmers University of Technology 12
PULP PROPERTY ESTIMATIONS Pulp samples in time domain ˆ x m t x m t Comment: The zero-elements in the vector x m (t) are possible to predict using piece-wise Pulp samples linear functions where residence time and/or Sp.Energy from lab. tests and consistency are used as independent variables. x s Pulp property t s time Oct Nov Dec Jan Feb Mar 2019-04-11 Chalmers University of Technology 13
PULP PROPERTY ESTIMATIONS Specific energy External variables Consistency in FZ MPC based on Machine Pulp Soft Sensor parameters Consistency in CD Quality pulp & vector Refiner segment (non-linear & parameters Residence time in FZ handsheet model based on handsheet Measured internal first principals) models estimations Variables (Temp) Residence time in CD Non-linear Linear Karlström, A. and Hill J. ” CTMP Process Optimization Part III: On the Modeling of Scott-Bond, Z-strength and Tensile index ” . Submitted for publication in NPPRJ 2017. 2019-04-11 Chalmers University of Technology 14
PULP PROPERTY ESTIMATIONS Multivariate modeling approach Dynamic modeling using a System identification approach ˆ θ θ θ f ( x ) x x x b m 1 , 2 , ,q m m m 1 m 1 m 2 m 2 mk mk m Model parameters - Case I {θ ,…, θ Properties Intercept Cons.(FZ) Cons.(CD) R.time(FZ) R.time(CD) Spec.E CSF Freeness 767.4522 6.5706 -5.3285 -320.0951 1430.6061 Sheet density 311.3555 -13.5518 7.6204 100.8509 55.8056 Tensile strength -3.5278 -0.1770 0.0831 -3.6486 62.7926 Tensile index 40.2999 -2.1925 1.0245 -16.5077 200.9831 Elongation to rupture 1.6649 -0.0341 0.0143 -0.7405 6.0259 Tensile energy absorption 80.0754 -4.2263 2.2176 -38.5411 314.1703 Tensile energy absorption index 0.4283 -0.0111 0.0046 0.7856 -4.6984 Tensile stiffness -451.7375 -5.3301 1.3169 1757.0678 -6261.3361 Tensile stiffness index 5.2630 -0.0675 -0.0026 0.5462 -1.9310 Tear strength 6435.5391 -14.0063 0.2833 811.0374 -25020.3831 Tear index 53.1587 -0.1836 0.0703 -5.6864 -140.9302 Short-span compressive test index 24.1904 -0.6072 0.3068 -1.8931 10.6965 ISO brightness 88.6031 -0.7198 0.0572 -57.3180 331.3528 Scott-Bond 155.0146 -7.2287 3.5924 120.7834 -508.4503 Z-strength 144.0442 -6.0570 3.1333 -129.1267 890.5317 Shives(>=0.3mm) -1787.1167 -53.8174 94.7427 -1459.5470 10070.7017 Long fibers 8.0583 -0.0944 -0.0102 -8.7292 47.3025 Fines 34.3848 0.2466 -0.6137 70.0445 -240.6360 Karlström, A. and Hill J. ” CTMP Process Optimization Part III: On the Modeling of Scott-Bond, Z-strength and Tensile index ” . Submitted for publication in NPPRJ 2017. 2019-04-11 Chalmers University of Technology 15
PULP PROPERTY ESTIMATIONS Machine learning is learning from data, system identification can be described as learning dynamic models from data. Predictors used: Specific Energy, Consistency in FZ and CD, inlet consistency 2019-04-11 Chalmers University of Technology 16
PULP PROPERTY ESTIMATIONS In the same way as the multivariate regr. models The models seem to be useful! Approx. 0.5 hr 1) Build models based on DC-gains and refiner dynamics! Introduce learning algorithms. 2) Use delay/filter to foresee future changes after latency! 2019-04-11 Chalmers University of Technology 17
RESULTS CCtrl and TCtrl – in Automatic mode Control concepts provides a possibility to deside where in the operating window to run the refiners. This gives an opportunity to: • minimize process variances; • stabilize the pulp quality; • analyze performance of refining segments; • optimize machine performance; • introduce an overall MPC-concept for good enough pulp (and handsheet) property control; • optimize the specific energy; • plan production. 2019-04-11 Chalmers University of Technology 18
RESULTS Control concepts provides a possibility to deside where in the operating window to run the refiners. This gives an opportunity to: • minimize process variances; • stabilize the pulp quality; • analyze performance of refining segments; • optimize machine performance; • introduce an overall MPC-concept for good enough pulp (and handsheet) property control; • optimize the specific energy; • plan production. 2019-04-11 Chalmers University of Technology 19
PULP PROPERTY ESTIMATIONS Next step! Introduce machine learning algorithms on-line for ”model tuning”! 2019-04-11 Chalmers University of Technology 20
Recommend
More recommend
