An Integrated Simulation Optimisation Decision Support Tool for Multi-Product Production Systems

Over the past decades, the rising energy prices and imposing environmental regulations have motivated manufacturers to improve the energy efficiency of their manufacturing processes. Manufacturers need to also consider energy efficiency in addit ion to classical performance measures. The additional performance d imension (energy-related indicators) significantly increases the complexity of classical production planning problems (e.g. scheduling), already known as NP-hard problem). To overcome the inherited complexity, an integrated simulation-optimization framework is proposed. The proposed approach tackles scheduling problem in a multi-product/multi-machine manufacturing environment and optimizes several production objectives simultaneously. A case study is presented to demonstrate the applicability of the proposed approach in a real-life industrial facility.


Introduction
In a typical manufacturing system, time dependencies and dynamics related to mu ltiple product flows affect the material flo w (Duggan, 2012) and also the energy consumption (Alvandi et al., 2016).These variab les may be associated with a single machine tool or a process chain, or even the whole factory (Herrmann at al., 2014).In other words, cycle time variations and machine job sharing stemming fro m mu ltitude product routings, can severely impact on the energy consumption of the entire system.
It has been generally acknowledged that efficiency improvements in a factory can only occur when holistic understanding of energy and resource usages is understood (Alvandi et al., 2016).There exists vast energy reducing opportunities on the system-level that advantageously do not require large capital investment.Operational method such as optimised shop floor scheduling is one examp le of such methods.Implementation of optimised shop floor scheduling is normally less costly and can easily be applied to existing systems (Fang et al., 2011).
Up to now, the focus of shop floor scheduling optimisations has been main ly on trad itional performance measures such as optimising lead times or minimising slack times on single machine/ parallel-machines.Within the context of mult i-object ive scheduling, the consideration of the energy related performance (objectives) within the entire factory has been undermined, hence hinders further research.
The focus of this research is to develop a simulat ion-optimisation framework for solving mult i-objective scheduling problem within mult i-product systems including all processes within the process chain.In the following sections, current energy related optimisation studies will be briefly explored.In order to address the gap found within the literature, a simu lation-optimisation framewo rk will be presented which will be further applied to a manufacturer of composite brake shoes and disc pads for railway industry.

Background
Most of manufacturing systems involve complex, dynamic systems which consume energy, water and raw materials.Improving energy efficiency with respect to operational methods in the area of job scheduling has been explored by researches in recent years.Mouzon et al. (2007) developed several algorith ms and a mult i-objective mathemat ical programming model to investigate the problem of scheduling jobs on a CNC machine for reducing energy consumption and total complet ion time.They pointed out that there was a significant amount of energy savings (up to 80%) when non-bottleneck mach ines were turned off and set on idle mode.This research was performed on the machine level and only considers a single mach ine environ ment and tackled single object ive problem; in further work by (Mouzon & Yild irim, 2008) a mu lti-objective optimization schedule that min imized the total energy consumption and the total tardiness of one machine was solved.Fang et al. (2011) presented a mult i-objective mixed integer linear programming formu lation for optimizing the operating schedule of a flow shop.They considered maximis ing productivity (make span) and min imising energy related (energy consumption, carbon footprint and peak power load).The presented mult i-objective model considered a simple case of scheduling 36 jobs on two machines.A Pareto frontier was established that showed the trade-off between throughput time and peak power.An optimal scheduling procedure for sequencing jobs on one process (painting process) has been proposed by (Wang et al., 2009) with the aim to reduce energy consumption in an automotive paint shop.
In order to investigate potential opportunities for energy efficiency and making trade-offs that are transparent to the decision makers, simu lation is a proven modelling technique (Alvandi et al., 2015).In line with the presented body of knowledge, some authors adopted energy oriented discrete event simulation model (Herrmann et al., 2011).The model represented the production system with all the interdependencies and dynamics of technical building services and also considered the technical, economic, and ecological evaluation of the performance of process chains.Melouk et al. (2013) developed simulation optimization-based decision support tool for steel manufacturing and conducted tests on the impact of simultaneous change of inventory levels of both slabs and coils.Within their work, they solved single objective optimisation problem of minimising inventory cost of slabs and coils.Mousavi et al. (2015) proposed a simu lation-optimisation framework to model dynamics of individual processes and the entire system.The method has been applied to a mass production system producing small variety of products with large quantity.Their work evaluated the effects of process parameters and the role of lot-sizing problems wh ile exp loring the benefits of simulation techniques for modelling the dynamic energy consumption on a system-level.However, the method falls short when it co mes to mu lti-p roduct/multi-machine environ ments and optimisation of the entire system.
In order to capture the inherited comp lexity in terms of various product routing and mach ine job sharing, a modelling framework that considers product level parameters together with machine level and system level is deemed necessary.

Proposed Simulation Optimisation Framework
Base on (Fu et al., 2000), a general description of the proposed framework for a simu lation-optimisation is presented in Figure 1

Simulation Model
The simulation model represents the manufacturing system co mprising all the process machines, technical building services and process chains with various products flowing in between.A Discrete Event Simulat ion (DES) model is developed using in the AnyLogic® model develop ment environment.As illustrated in Figure 2, the developed simu lation model takes a holistic v iew of the factory and represents the manufacturing system and all the material and energy flows within interactive modules: unit process, process chains and TBS.
Figure 2. Simulation model framework (adopted from (Herrmann et al., 2011)) Unit Process Module: Each un it process is configured according to the corresponding machine specific parameters (e.g.energy profile, scrap rate, etc.).Basic unit process modules for the representation of machines with d ifferent operational states are created by using state charts (Alvandi et al., 2015).Following is a brief explanation on configured machine states: • Off: refers to the state when the machine is switched off.
• Ramp up: refers to the acceleration of the main drive of the machine when switched on.
• Standby: refers to the period when the machine remains ready for production.
• Preproduction: refers to the activities for preparing the production, such as loading a workpiece.
• Production: refers to the state when the machine is processing the work piece.
• Postproduction: refers to the state when the au xiliary equip ment of the machine (e.g.lubrication pu mp, chip remover) is activated.• Changeover: refers to the state when the set-up of the machine is being changed.
• Failure: refers to the state that the machine is broken down and requires maintenance Process Chain Module: The process chain of a product is defined as responsible processes and machines to transform material to product.Mult iple process chains are modelled by lin king and connecting the unit processes according to predefined product-machine routing logic.Separate modules are configured to translate the routing matrix for guiding the product passing through multitude of processes and machines (Alvandi et al., 2016).

TBS Module:
The process chain requires other resources and auxiliary services such as lighting, heating and compressed air.In the proposed framework, technical building services (TBS) are responsible to supply services such as steam, co mpressed air, and conditioned air, cooled air and air purificat ion.Similar to unit processes, the TBS models are developed as having different operational states, such as ramp up, standby, production, ramp down and off.The supply load on these devices is dependent on the total demand of associated production processes that is configured by the machine specification.Due to high flexibility of AnyLogic®, several developed TBS models can be directly connected into the proposed simulation model, such as steam generation units (Ghadimi et al., 2014), co mpressed air systems and HVA C units the (Mousavi et al., 2014).Figure 2 depicts the proposed framework for the simulation model.

Optimisation Engine: multi-objective optimisation
There are three main groups of methods for solving mult i objective optimisation problem.Methods which consider one objective at a t ime: the lexicographic method is a good example of methods in this group where all the objective functions are sorted in order of their importance (Marler & Arora 2004).Other methods such as Weighted Sum method belong to the second group that normalises a set of objectives into a single objective by mu ltip lying each objective with a user defined weight vector (Deb, 2001).The third group considers the objective functions simultaneously and try to find a set of Pareto solutions instead of one single solution.
Weighted Sum is the simplest approach and arguably the most widely used classical approach which is selected for the proposed framework.Weighted Sum solves a mult i-objective problem as a single objective.In this method, each sub-objective is solved as a single-objective problem wh ich then will be scaled and weighted depending on its relative impo rtance or weight.The mult i-objective function can be maximized or minimize (Equation 1 formulates a minimisation problem).

Optimisation Engine: Search Method
Metaheuristic approaches have drawn considerable attention from many researchers in the last decade.Bianchi et al. (2009) identified that these methods are a valid alternative to exact classical optimisation algorith ms as well as stochastic comb inatorial optimisation problems.In fact, metaheuristic algorith ms (such as simu lated annealing, Tabu search) have dominated the optimisation routines of simu lation software due to their flexibility and robustness (Fu, 2002).These algorith ms are known to be flexib le in dealing with any type of solution space (either discrete, continuous or a combination of them) and are capable to quickly achieve high quality solutions.
Advances in the area of metaheuristic optimization coupled with improved computing environ ments resulted in creation of general-purpose "black bo x" optimizers (Glover et al., 1999).As a matter of fact, this approach has dominated the optimisation routines of the simu lation software o wing to their flexib ility in dealing with any type of solution space (either discrete, continuous or a co mbination of them) and their ability to quickly achieve good quality solutions (Fu, 2002).In black-bo x approach, the metaheuristic optimizer selects a set of values for the decision variables and uses the responses (or objective function values) generated by the simulation model to make decisions regarding the selection of the next trial solution.This procedure is iterative until a termination condition is fulfilled which means the best solution is found and is returned as the optimal (or near optimal) solution.
OptQuest search engine combines the metaheuristics of Tabu Search, Neural Networks, and Scatter Search into a single search heuristic.A simple way to describe its search mechanism: if a candidate solution does not fit the constraints, OptQuest eliminates that solution and exp lores candidates that are more likely to be better.The efficiency of OptQuest search algorith m is very much dependant on the size of the solution space and the starting point (Kleijnen and Wan, 2007).
OptQuest uses three stopping criteria (Kelton & Law, 2000): • Run until maximu m number of configurations defined by the user is achieved (MNC).
• Run until no improvement is obtained in the value of the objective function fo r 100 consecutive configurations (Automatic Stop in OptQuest).

•
Co mbination of above rules; optimizat ion runs until non-imp roving configurations are equal to 5 percent of Vol. 11, No. 6;2017 MNC.
By using OptQuest™ to deal with the mult i-objective optimization problems, there are a few classical methods that could be applied, like weighted sum approach or goal-oriented optimization as described earlier.Although these approaches are d ifferent fro m each other, their main aim is the same, wh ich is to convert a mult i-objective optimization problem into a single-objective optimizat ion problem.A mong these approaches, the weighted sum approach is the simp lest and probably the most widely used classical approach.In order to derive theories from practice, the proposed simu lation-optimisation methodology was applied to a manufacturer of co mposite brake shoes and disc pads for the railway industry.Report any other analyses performed, including subgroup analyses and adjusted analyses, indicating those that were pre-specified and those that were exploratory (though not necessarily in the level of detail of p rimary analyses).Consider putting the detailed results of these analyses on the supplemental online archive.Discuss the implications, if any, of the ancillary analyses for statistical error rates.

Case Study
The investigated company offers around 100 different products, which generally are grouped into three product families.These products differ in terms of various material co mpositions, weight and shape.For the ease of demonstration, six products are chosen-two fro m each product family.Figure 3 is a representation of the process sequences and a reference map for the flo w of 6 product families.In general, all p roducts follo w five consecutive processes to be manufactured: weighting, mixing, biscuit pressing, moulding and curing.However, the products vary in shapes, thicknesses, and material co mpositions to suit different braking systems requirement.As a result, key variables, such as routing and processing time are product specific.
Figure 3.Representation of process sequence and sample product routings (Alvandi et al., 2016) In the first process step (weighting), there is only one weighting unit (WU) for scaling the raw materials.In mixing, three mixers (M X) are used to generate the frict ion material co mpound.The batches of specific products are mixed and stirred until they become a homogeneous mixtu re.Based on the product (batch) type and the mixer's technical specification (e.g. machine capacity) the batches are assigned to specific mixers.
The mixed co mpound is transferred to pressing machines where it is pressed into cuboid biscuit shaped units.Depending on the product and material mix, this process step is performed by different biscuit presses (BP).
There are five biscuit presses, three automatic and two that are manually operated.
Parallel to the brake pad production line, the back-plates are produced in a separate line.In the mould ing workstation, which contains ten moulding presses (MP) and heaters, the biscuits are heated and pressed against their corresponding back-plate to form the actual brake b lock or pad.Depending on the type of product and material used, either radiofrequency waves or induction is used to heat the biscuits to approximately 150°C so they can be moulded in the presses immed iately.The presses can be equipped with different types of moulds with different capacity (ranges from two to eight cavities).
Curing is the final process for most of these products, so in this workstation the moulded products are placed into one of five curing ovens including two electric (EO) and three gas ovens (GO).In this process step, depending on the type of product, only one of five programs can be selected.During the curing process, the friction materials are hardened and homogenised.

Data Collection
The studied organisation utilized Citect v. 7.20 fo r its SCADA system (supervisory control and data acquisition system) and the Oracle database for enterprise resource planning.The Oracle® ERP system provided useful input information (such as BOM, p roduction schedules) while Citect SCA DA® provided detailed machine specific values (such as scrap rate, machine status, mean time between failure (MTBF) and mean time to repair (MTTR)).Table 1 shows an example of model input for configuring the mixing processes.Every product in this system has its own batch size, which means that the batches contain a different number of products.These products are stored in front of the machine according to their types, and when a given storage reaches its batch size; all the stored products are released and assigned to the machine.Fro m a modelling perspective, there are two dynamic parameters in the process module which change depending on the product type, i.e., the batch size and production time, both of which require new value for each process module when a new product is ready for production.After a g iven product has been processed on a particular machine, the model checks the product routing to determine the relevant machine for the consequent process and then assigns the product to that machine.

Data Collection
Routing dictates the process steps that the product must go through as well as the specific machine tool for that processing step.Table 2 shows the routing and production times of each product family.Routing in the weighing and mixing workstations is not flexib le because each product family can be processed on only one machine, unlike the biscuit pressing, moulding and curing workstations where products in the same family have several alternative machines.In the simulated case, the product routing was predetermined by production planner according to product specification and machine technical specification.
Also within Tab le 2, the established production times are per unit o f product.Because the product flows in different forms and units, the production time should be scaled accordingly.In weighting, mixing, and b iscuit pressing, the time is per batch since based on the batch size and weights, production times vary.In mould ing the time is given per product because the shape of the product changes after raw materials have been pressed into blocks.Finally, production time in the curing wo rkstation is per cycle because processing does not commence until the ovens are full.The production system operates on 2-shifts/day for 5-days/week.Production starts at 6 AM and fin ishes at 10 PM.For the purpose of this study a copy of the (weekly) production plan for one month was obtained from the production planning department and then the system was simu lated.With regards to energy consumption, the amount of electricity consumed by each machine during different operational states is measured by a portable power analyser and then determined by studying the energy profile.With regards to the consumption of gas in the ovens, the nameplate values of each machine were used and further validated with the gas meter readings.

Simulation Validation
A model for the entire factory was developed in Anylogic®, Microsoft Excel® and Java® using the method described in Section 3. It involved configuring the production processes in Anylogic® and defining the production parameters in Microsoft Excel®.A snapshot of this simulation model is shown in Figure 4.The daily electricity consumed by the entire factory was simu lated for a period of one month, and then compared with the actual consumption data.The simu lation run was set to continue for five weeks.One week was dedicated (as transient phase) to warming up the system where the simulat ion results were d isregarded.Having a warm-up period ensured that the simu lation was not influenced by the initial conditions and it had reached a steady-state phase before collecting data fro m the simu lation (Banks, 1998).The results of the daily electricity consumption simu lation of the entire factory for the given month were co mpared with actual consumption data and shown in Figure 5.The simulation error is determined by calculat ing the rat io of the difference between the simulated and actual total electricity consumption to actual total electricity consumption ( see Equation 2).
% = ( − )  * 100 (1) As Figure 5suggests, the daily simu lation error is less than 10% on average.Notably, the first three Saturdays as can be seen, cause the highest difference between 14-18 %.A close investigation has identified an unplanned weekend production (over t ime work to catch up delayed orders) on the curing ovens has been carried out which was not included in the simulat ion.Furthermore, the model shows a higher accuracy for estimating the energy consumption at an aggragated level (e.g.weekly and monthly).Tab le 3shows the comparison between the simulated factory energy consumption and the factory energy measurements on weekly and monthly basis.

Solving Multi-Objective Optimisation
In this case study, each product family has several feasible alternative machines for carry ing out required operations.Product routing problem in this environment consists of assignment of each operation to a machine.When considerable routing flexib ility exists in a production system, changing product routing may significantly affect system's throughput and work-in-p rocess inventory (Calabrese & Haus man, 1991).On these basis, the main objectives to be considered for this case study are maximis ing product throughput, minimising factory energy consumption while minimising lead time of products.In the following section, the decision variables and objective functions will be discussed in more details.

Model Settings
As for arranging and setting OptQuest® parameters for the optimisation, new parameters are introduced to the simu lation model.These new parameters act as representation of the routing and for configuring those, discrete numbers are selected with defined M in/Max and step size.The Min/Max numbers are reflect ive of the number of the actual machine under each process.For examp le the decision variable set up for Product A for presentation Vol. 11, No. 6;2017 purposes is shown in Table 4.For the rest of the product types, same optimisation parameter setting is performed but due to space limitation, all tables are not presented.The production is simulated for a two week period where the first week is considered as a warm up period thus the system perfo rmance related to this period was discarded.Therefore, the objective functions were only measured for the second week.Real-world problems mostly contain computationally expensive objective functions that may result in optimizat ion runs that take several days.Stopping criteria are needed to terminate the execution of optimization algorithms.The automat ic stop is chosen as a stopping criterion with in OptQuest which terminates the optimisation when the algorithm determined that new solutions are not likely to produce a better objective value.

Decision Variables
The aim is to find the best choice of machines under each process step (Mixing, Biscuit pressing, Moulding, and Curing) for ten products.This study only focuses on finding the best routing for the ten selected products, representing all three product families.For selecting the ten products, the production order book for a particular month was studied.The order quantities for each product were ranked and the result showed that 10 of the products were responsible for 90% of production load.
The choice of machines to be used for each process for a particular product follows a set of heuristic ru les which considers technical capability of the machine fo r processing a particular product.Some o f the heuristic rules in assigning the product type to the machine include all mixers and all biscuit pressing machines, because there is no limitation in terms of technical co mpatibility between the machines and the products.Nevertheless, utilisation of moulding machines on product types has limits.Fro m technical point of v iew, product in group I (as shown in Table 2) can only be processed on MP9 and MP10 whereas the rest of the products can go to any mould machines.Also not all curing ovens are suitable for all product types.

Objective Functions
This study takes significant interest in understanding the impact of different product routing on energy, throughput and lead time.Throughput is the number of products that are produced by the production system within a certain period of time.Energy is total amount of electricity consumed by the system during the period under consideration.Lead Time is defined as the average time that the product spends in the shop floor passing through each process step till it gets through the last process.The ultimate goal of the company is to choose the best possible routing for each product type so that the energy is min imised, throughput is maximised and the lead time is minimised.As explained earlier, in Weighted Sum method, each sub-objective is solved as a single-objective problem which then will be scaled and weighted depending on its relative importance.

Energy Minimisation
As a first scenario, the aforementioned decision variables are configured and fed into OptQuest® with the objective function set on Minimisation of Energy.Optimisation was run with the Energy as the main objective function to be minimised with automat ic stop.As the problem was a determin istic problem, only one rep licat ion was performed.Table 5 shows, total run (the total number of observed solutions), the best run (the number of observed solutions before reaching the best solution) and the values of objective function are given.It should be noted that in optimising the single object ive (energy consumption), the other t wo objectives (throughput and lead time) are not priority for OptQuest® to optimise.Observing the recommended variable setting for the optimised energy as illustrated in Table 6, the following trend is seen: • For almost all products, biscuit pressing machine number 5 was chosen.
• For most products the mixing number 3, moulding number 10 and curing oven number 5 were chosen.
OptQuest® selection of these particular machines can be explained as the fo llowing: In mixing process, MX3 is recommended by the optimiser engine fo r almost all products because it consumes less energy than others on stand-by mode.The same reasoning applies for selection of BP5 as well as MP10 and Curing nu mber 5 which uses gas.It is noted that the chosen machines are in fact rated as low energy consumers among other machines.

Throughput Maximisation
As a second scenario, the objective function on OptQuest® is set on maximisation of throughput.OptQuest® was run with automatic stop.The result of optimised throughput shows that OptQuest® arrived to the optimu m 15041 (pcs) after 389 total runs as shown in Table 7. Fro m the solution space (decision variable) point of v iew, the reco mmended machines under each process are different to the Energy optimisation problem as illustrated in Table 8.Here due to objective function being Throughput Maximisation, the OptQuest® search algorithm is trying to achieve the highest possible product yield by means of job sharing and maximu m utilisation.
It is of no surprise to see for each process step, almost all of the machines are utilised.Contrary to the recommended machines of Energy optimisation exercise, this time OptQuest® strategy to maximise throughput is by distributing the products to as many machines as possible.

Lead Time Minimisation
As a final scenario, the minimisation of the lead time is the third single objective problem to solve by OptQuest®.Similar to the previous optimisation, the OptQuest® will stop when it finds the optimu m point based on its built in algorith m.After 361 runs as shown in Table 9, it arrives at the Lead Time with the value of 4780.51 minutes as the best possible result.

Multi-objective Minimisation
To solve the problem o f optimisation of three objectives simu ltaneously, all three objectives need to be combined to one objective which then similarly can be solved within OptQuest® simulat ion-optimisation platform.As discussed earlier, the proposed optimisation engine ut ilised the weighted sum method to comb ine several objective functions.Table 11 shows the best and worst values of each objective function generated from the single optimisation problems solved above.For the ease of the reader to locate the optimu m results of each objective, the * sign is placed next to the optimum values.The min/ max values for each objective functions are fed into the OptQuest® as well as the single objective function f(x) which is co mbination of all three single objectives (see Equation 1).The weighted sum method, the optimisation engine also needs the relative importance of each objective function or weights.Since achieving optimu m results for all of the objectives at the same time is the main goal of the co mpany, each of the weights (w1, w2 and w3) is set at equal weighing (e.g.1/3 of total importance which means 0.33).After setting the OptQuest® with the new objective function and weights for each objective, OptQuest® started its search and eventually stopped at 321st run.Table 12 su mmarises the results for each objective for when the optimu m was reached.The OptQuest® result suggests, the optimu m energy to produce the highest possible number of products (13450 pcs) is 65265 kWh within 4897 unit of optimum time.For understanding whether the optimised results are actually any better, a baseline scenario is simulated with the current system setting for the whole production system.The duration of the simulation was also for two weeks with the first week discarded as a warm up period.
In order to demonstrate the variation between t wo sets of three d imensional results, the results on energy and lead time are plotted separately against throughput as in t wo d imensional graphs.Figure 6 presents the variation between the optimu m result and the baseline scenario on energy versus throughput.As it shows the optimal solution led to 4515 pcs in throughput and 5% decrease in total energy consumed by the system.There is a significant 51% improvement on the throughput.In Figure 7, the throughput results are plotted against the lead time results.The minimum lead time of 4897 minutes has been improved about 15% from the baseline lead time.To explain the significant improvement on throughput and lead time and the reduced energy, the solution space (optimal product routing) is studied.Table 13 presents the OptQuest® choice of optimu m routing for each product.Co mparing this setting (mult i optimisation) and the energy min imisation problem (single optimisation), the choice of mach ines, in particular mould ing machines are much more diverse.Obviously utilising variety of available machines result in lead-time reduction and increase in throughput but adversely affects the energy consumption.

Sensitivity Analysis
Sensitivity analysis is a natural step of robust optimisation that follows a systematic approach to changing the value of model decision variab les over some range of interest in order to observe how the changes affect model behaviour (Balci, 1998).It also can identify those input variables which the values of objective functions are very sensitive.
Ult imately, the validity o f the model can be enhanced by assuring that those values are determined with sufficient accuracy.In this section, how sensitive the value of the objective functions were to the weight of the objective functions was investigated by defining 19 more scenarios in addition to the three single objective scenarios.For each scenario, a different co mb ination of weights is selected for the objectives; see Table 14.The sensitivity of the objective weights was analysed by comparing 22 scenarios with a different co mbination of weights.The values for each object ive vary according to weight change and show they are sensitive to changing weights.For examp le, when scenario S9 is co mpared with S0 (a 40% reduction in the throughput weight) the throughput objective was reduced by around 8.3% (fro m 15041 pcs in S0 to 14115 pcs in S9) whereas in a non-sensitive model, more reduction was expected (around 40% reduction in throughput).
Table 14.Optimisation results for 21 scenarios absolute and relative deviation between the baseline and other scenarios.The scenarios can be co mpared with each other to quantify the impact each weight combination has on the objectives.
As explained before, in solving this multi-object ive problem, all objectives are important to decision maker; the equal share of 0.33 fo r each objective signifies scenario 3 (S3) of the presented table.However, it is always possible that the given priority on operation objectives would change according to different business strategies.This highlights the applicability of the sensitivity analysis in understanding the impact that each priority setting (weights) will have on the objectives.It is already expressed earlier that the equal weights of 0.33, S3 achieved desirable result and could optimise each of the objectives (throughput, energy and lead-time) by 51%, 5% and 15% accordingly.

Conclusion and Outlook
Most of efficiency improvement decisions are mult i-objective problems in which management needs to handle the challenges of conflicting objectives.Due to complexit ies and uncertainties exist in real-world problems, and a simulation-optimisation was considered as the most practical platform.
There are many examples in literature for solving scheduling problems on a single machine or parallel-machine for minimising for instance tardiness or maximising machine utilisation.However instances that apply mu lti-objective optimisation, considering energy and traditional business objectives are rarely found within the research field.Th is paper exp loits the use of simulation-optimisation framework for energy efficient production planning and control.
The proposed framework utilises the simu lation model for evaluation of performance of the system under different scenarios.As for optimisation part of the framework, several decision variables are evaluated simu ltaneously, searching the solution space to find optimal or near optimal solution.As presented, improvement opportunities are evaluated and optimised considering all relevant parameters (machine, process chain, factory and product).
Fro m the system perspective, the inclusion of product view in the optimisation of energy consumption for the whole factory shows a great leap towards holistic system improvement.Through optimal product routing, it is possible to achieve highest possible throughput and lowest possible energy consumption.The proposed optimal wh ich is co mposed of two segments: Simu lation model is the first segment and the second segment is an optimisation engine (optimiser) which interacts with the simu lation model with its exp loratory algorithm.The simulation and optimisation co mmunicate closely with each other.Firstly, optimisation engine provides trial solutions to the simulation model.Simulat ion model then runs those solutions and returns the values of objective function to the optimisation engine.Optimisation engine takes advantage of these output results to improve its search for selecting new trial solutions.

Figure 4 .
Figure 4. Snapshot of the simulation model

Figure 5 .
Figure 5. Daily electricity consumption of the entire factory; Actual data vs.Simulation

Figure 6 .Figure 7 .
Figure 6.Throughput vs. Energy for baseline and optimal solution

Table 2 .
Routing and production times of different product families

Table 3 .
Comparison between actual consumption and simulation

Table 4 .
List of decision variables for product A for optimisation

Table 5 .
Objective values when minimising total energy consumption

Table 6 .
Optimal product routing for the products when minimising total energy consumption

Table 7 .
Objective values when maximising throughput

Table 8 .
Optimal product routing for the selected products when maximising throughput

Table 9 .
Objective values when maximising throughputAs for the decision variable( as in Table10), it is noticed that the OptQuest® algorith m uses the same search strategy as it did for minimising lead time and proposes job sharing and line balancing to other machines.

Table 10 .
Optimal product routing for the selected products when maximising throughput

Table 11 .
The best and the worst of each objective during single optimisation runs

Table 12 .
Optimal value of objectives with equal importance

Table 13 .
Optimal product routing for the selected products

Table 15 .
Absolut and relative deviation between baseline and scenarios