Flexible turnaround planning for integrated chemical sites Sreekanth Rajagopalan, Nick Sahinidis, Satya Amaran, Scott Bury Enterprise-Wide Optimization (EWO) Meeting, Fall 26 September 2-22, 26
Turnaround rescheduling Motivation Respond to peak demands when demands are forecast to increase during turnaround period Unfavorable market conditions foreseeable supply/demand variations Additional resource constraints availability of skilled workforce or process specific technical experts Problem statement Given an integrated chemical sites network with a base turnaround schedule over the next 6-9 months, Benefit of moving Unit 4 turnaround from March to July? Risk of loss in rescheduling? Unit 4 Unit 2 S Unit 9 Continue operations performance exceeding expectations (catalyst activity, HEX fouling), changes in turnaround frequency time period estimate S2 S3 Unit 5 Unit 6 Unit 7 Upstream Downstream Storage Tanks 2
Potential benefits Key factors Time value of money on costs and revenue Time to plan production and inventory better Integration effect Unit 9 Unit 7 Unit 6 Unit 5 Base Unit 4 % Cases (N=25) 9 8 7 6 5 4 3 2 5+% 55+% 6+% 65+% 7+% 75+% 8+% 85+% 9+% 95+% Demand Range 9 8 7 6 5 4 3 2 Additional days worth production Unit 2 Jan Feb Mar Apr May Jun Jul Aug Sep Unit 4 turnaround in March Alternative Unit 9 Unit 7 Unit 6 Unit 5 Unit 4 % cases Alternative profitable % cases Alternative profitable (4+ (>=%) days) Unit 2 Avg improvement Avg - S.D. Improvment Jan Feb Mar Apr May Jun Jul Aug Sep Avg + S.D. improvement Unit 4 turnaround in July 3
Planning under uncertainty: unplanned outages Probability.3.2. Unplanned Outage Severity March April May June July No outage Pit-stop 6 No outage Pit-stop Full TA.4 8 7.3.9 Corrective maintenance decision policy: pit-stop (minor) and turnaround (major) Reactive planning (sequential LPs) vs. anticipative planning (stochastic programming) models tt = tt = Pit-stop (CM) Turnaround (CM) Turnaround (PM) No outage Full TA 5. tt = 2 Pit-stop 4.3 tt = 3 No outage Full TA 3. tt = 4 Pit-stop Full TA 2.9. Full TA.49 9 tt = 5 2 3 4 5 6 7 8 9 Mar Apr May Jun Jul Scenarios 4
Stochastic programming model Objective: max NPV Flow balance Demand constraints Turnaround constraints Capacity constraints : expected value of all scenario profits : material balance + stream ratio requirements : upper bound (deterministic, monthly timescale) : unit up or down for maintenance : flow and storage tank bounds Example: nodes, 6 arcs, 3() turnarounds, 9 month horizon, stages weekly discretization, planning semi-weekly 72 time periods 33 stages (4-month reschedule window) 2 time periods per stage Model tractable to reliably reschedule single turnaround: stages ~ 2T Nonanticipativity constraints : time-consistency and implementable decisions Solution time for deterministic equivalent of SP is < 2 sec Base Alternative Reactive Anticipative variables,48 48 27,648 54,736 constraints,987,987 38,435 434,35 non-zeros 5,44 5,44 5,28 96,53 5
Potential reschedulable instances NPV improvement: Anticipative vs Base Risk profiles: cumulative distribution of profits % Cases (N=25) 9 8 7 6 5 4 3 2 5+% 55+% 6+% 65+% 7+% 75+% 8+% 85+% 9+% 95+% Demand Range % cases Alternative planning profitable % cases Anticipative planning profitable % cases Anticipative planning pl. profitable profitable (>=%) (4+ days) Avg improvement 7 6 5 4 3 2 Additional days worth production Probability of NPV <= X.9.8.7.6.5.4.3.2. Base 299 Alternative 38 Reactive 34 Anticipative 36 288 29 292 294 296 298 3 32 34 36 38 3 Reactive planning NPV (scaled cost units) Anticipative planning -3% instances can be reliably rescheduled More than 4 days worth production recovery in about 5% cases Anticipative planning model provides a flexible production and inventory plan that is less risky 5% chance of loss vs. % from reactive planning 6
Additional inventory to hedge against uncertainties Inventory with time for tanks 2 and 3 for the (final) scenario corresponding to no outages Tank 2 inventory level (%) 9 8 7 6 5 4 3 2 Tank 2 Tank 3 Jan Feb Mar Apr May Jun Jul Aug Sep Time Tank 3 inventory level (%) 9 8 7 6 5 4 3 2 Jan Feb Mar Apr May Jun Jul Aug Sep Time Base Reactive Anticipative Base Reactive Anticipative Anticipative plan recommends more inventory to hedge against future uncertainties A small premium of.5 scaled cost units is incurred in the form of additional holding cost 7
Sensitivity to outage probabilities What if outage probabilities data are underestimated? Dark to light: by 5% months, 2, 3, 4, by 5% for all months, + severity by %, outage + severity by % [7 profiles + nominal (black) + alternative planning(dashed)] Probability of NPV <= X Anticipative.9.8.7.6.5.4.3.2. 288 29 292 294 296 298 3 32 34 36 38 3 NPV (scaled cost units).9.8.7.6.5.4.3.2. Reactive 288 29 292 294 296 298 3 32 34 36 38 3 NPV (scaled cost units) For small target profit values (near base NPV of 299), the change in probability of loss is within 5% Even for 5%-% underestimation in outage probabilities, CDFs of profits from anticipative planning model values reschedule more than reactive plan 52
Sensitivity to reschedule time window Effect of reschedule time ( to 4 months) for decreasing demands Probability of NPV <= X.9.8.7.6.5.4.3.2. 255 26 265 27 275 28 285 NPV (scaled cost units) AP-4 AP-3 AP-2 AP- RP-4 RP-3 RP-2 RP- Rescheduling by a short window (-2 months) is risky since demands are not sufficiently low to save on loss from sales revenue Longer window (4 month) is also risky here due to corrective maintenance costs from potential outages 9
Sensitivity to demands Demand as % max capacity 95 9 85 8 75 7 65 6 55 5 Jan Feb Mar Apr May Jun Jul Aug Sep Expected additional NPV relative to Base NPV 4. 3.5 3. 2.5 2..5..5. 5% variation.. 2. 3. 4. 5. AP NPV on successful reschedule relative to Base NPV..9.8.7.6.5.4.3.2.. Expected additional NPV (RP) Expected additional NPV (AP) Premium AP Premium Unit 5 Unit 6 Unit 7 Unit 9 Variation within 5% is still profitable Anticipative better than reactive planning Premium is roughly the same doesn t cost more Expected additional NPV relative to Base NPV 6. % variation. 5. 4..8 3. 2..6...4 -..2-2. -3.. -2.. 2. 4. 6. 8. AP NPV on successful reschedule relative to Base NPV AP Premium 53
Conclusions Planning turnaround reschedules as multistage stochastic programming model hedges against uncertainties due to outages at a small premium Rescheduling turnarounds offer production recovery as high as -2 days depending on demands and integration effect Timing of reschedule as well as performance condition of the unit affects potential cost benefits and risk of loss Future work Optimal turnaround reschedule time window Simultaneous condition-based and risk-based turnaround planning Practical-scale networks
Data for case studies Reactive TA s are ~35% longer and ~2% more expensive PS s are ~25% of TA costs Costs were time value adjusted for some test cases Production rate degradation assumed 2-6% every month Time value of money: % per annum Duration (days) 3 25 2 5 5 March April May June July PS duration TA duration 3 4 Probability (x) 25 2 5 5 Cost 2 8 6 4 2 March April May June July March April May June July Unplanned outage Severity PS cost TA cost 64