(599a) Performing Solution of Market-Driven Optimization for Large-Scale Corporates

Performing Solution of Market-driven Optimization for Large-scale Corporates

Matteo D'Isanto¹, Flavio Manenti¹, Maria Grazia
Grottoli¹, Sauro Pierucci¹, Guido Buzzi-Ferraris¹,
Marcello Altavilla², Ornella Martinelli³, Davide
Jurissevich⁴, Roberto Di Marco³

¹Politecnico di Milano, Dipartimento di Chimica,
Materiali e Ingegneria Chimica ?Giulio Natta?, Piazza Leonardo da Vinci 32,
20133 Milano, Italy

²Linde Gas Italia, Terni Plant, Viale Benedetto Brin
214, 05100, Terni, Italy

³Linde Gas Italia, Headquarter, Via Guido Rossa 3, 20010,
Arluno, Italy

⁴Linde Gas Italia,Trieste Plant, Via Di Servola 1, 34145,
Trieste, Italy

Many projects to rationalize and optimize the
supply chain of complex corporates are ongoing in United States and Europe. The
leitmotif is the high potential margins that may come from a high level
optimization of decentralized productions and distribution networks, but also
the need to raise the decision-making level from plantwide's purpose to the
enterprise wide level. Actually, the global crisis of these last years is
strongly pushing towards global optimization since the single production sites,
although optimized, present relevant economic losses and the net operating
margins are very small with respect to several years ago, especially in
chemical and process industries. Thus, the corporate optimum, which almost
never corresponds to the optimum of the single production sites, is a promising
way to significantly improve the profits as well as to debottleneck
inefficiencies highlighting and exploiting hidden potentialities.

The main problem is that such an approach
unavoidably means to optimize in an integrated way all the operations of the
corporate and, hence, to account simultaneously for billions of variables and
to elaborate a solution in real-time for the optimal decision support.

Science fiction for the large industries, at
least until computers will not be enough powerful, the parallel computing
enough exploited, and the algorithms enough robust and efficient at the same
time.

The present paper discusses (and applies by the
field ? Linde's air separation units) certain novel methodologies to solve
separately several problems behind the real-time market-driven corporate
optimization, looking forwards to their integration for a performing solution
of the overall problem. Specifically, the following topics will be discussed:

·        
The
analysis of very large data sets. New methods for the analysis of very large
(industrial) data sets have been developed and validated. Such methods are able
to promptly detect the outliers that could affect the data acquired by the
field and the DCS.

·        
Prevision
of the future market demand. A combination of weighted averages based on the
historian and linear short-term extrapolations can be successful to foresee the
incoming market demand for each final customer.

·        
Allocation
of available commodities to match the demand. A Gaussian-based approach can be
useful and very performing to assign the right source of commodity to the final
customer so as to satisfy it by accounting for the distance, the availability
of cryogenic trucks, and the availability of product.

·        
KPIs
for final customers supply. Each society monitors its production and
distribution by means of specific key performance indicators (KPIs). Some of
the most important ones will be introduced and the item of the cryogenic truck
is optimized basing on them and clustering the customers.

·        
Just
in time production and minimum inventory. Being one of the most relevant cost
items, the liquid levels of the cryogenic tanks must be minimized. To do so, it
is necessary to foresee the market demand with a certain reliability and to
change the operational sets of the production plants accordingly. If the former
bullets are mainly related to mixed-integer linear programming, this point is
strongly related to nonlinear programming.

·        
Optimal
production. Basing on nonlinear models for air separation units, the production
and the energy consumptions are both optimized in a predictive way.

What is in common among all these topics is that
the solvers are all based on the novel Attic method and the robust optimizer
(exploiting parallel computing for shared memory machines) belonging to BzzMath
library. The application to Linde Gas Italia Corporate is also provided as
validation case.