LCO
BBs
Slurry
PPs
1400
600
1200
800
1200
500
1000
1000
400
600
800
800
300
600
400
600
200
400
400
200
100
200
200
0
0
0
0
-0.3
-0.2
-0.1
0.0
0.050 0.075 0.100 0.125
-0.2
-0.1
0.0
0.075 0.050 0.025 0.000
Cat gas
Dry gas
Cat gas octane
600
700
500
500
600
400
500
400
400
300
300
300
200
200
200
100
100
100
0
0
0
-0.05
-0.00
-0.05
0.10
10 12 14 16
0.100 0.025 0.050 0.075
Figure 3 Bimodal CLNN gains between ROT and FCC product yields illustrating the learned shift in product yields with a new catalyst formulation during CLNN process model training
predictions of control variables to then proactively use key handles to increase the objective function based on the refiner’s defined price deck. Decisions are sometimes different from what is defined by traditional steady-state approaches as the CLNN understand non-linearity and changes in the process state by continuously leveraging their knowledge of the FCC
dynamics. For example, the CLNN chose to cut feed rather than maximise feed in the given state, making room for an ROT increase while respecting the key operating limits to increase the objective function significantly (see Figure 4 ). Prior to CLNN, operators and process engineers were forced to change bounds in APC to conservatively stay within operating limits as the process response to
MV: Total feed rate
Objective function
+$40k/day
MV: ROT
CV: Regen slide valve dP
Upper bound
MV: Resid feed
Reactor-regen dP
Lower bound
CLNN engaged
PV signal
MV setpoint
Bound
Figure 4 Prior to CLNN, APC bounds and MV targets were manually adjusted to maintain targets and constraints. When engaged, CLNN drives the MVs dynamically to increase the objective function while respecting operational constraints
88
PTQ Q3 2022
www.digitalrefining.com
Powered by FlippingBook