Jan-2018

Plant optimisation: leveraging data analytics and modelling

More advanced modelling skills are needed at the design stage to realise the promise of better optimisation

GOPI SIVASUBRAMANIAN
Arivergy

Article Summary

Optimisation can be defined as the process, method or technique to achieve operation of a plant to produce the desired product at the lowest cost while maximising the return on investment on assets in the plant.

Today, reflecting the business climate, there is increased exposure of the various optimisation offerings, at the equipment level, sub-unit level or unit level. A spate of vendors has entered into the data mining/analytics area, asset optimisation, asset performance, and event prediction. Data analytics are some of the tools and services they offer that come under the broad umbrella of optimisation. All of them promise better utilisation of the asset. While this is true, plant engineers have to be aware of the scope and limitation of the optimisation offering they are considering.

Layers of optimisation
A process plant can potentially have optimisation at many layers. This is well described in a figure by Edgar et al (see Figure 1).¹

For the purpose of this review and discussion, the three layers closest to the process in Figure 1, namely layers 5, 4 and 3, will be referred to and analysed. An easily identifiable (with refinery operations) explanation for the three layers is given in Robert Young’s paper on real time optimisation.²

Process monitoring, analysis and control (Layer 5) is the layer closest to the process. It measures and monitors equipment and process streams via instruments. In DCS, this layer represents the information that is processed, stored and used for control and other actions.

A good example of this layer is the FIC controller maintaining a target flow rate by manipulating a control valve.

Unit management and control (Layer 4) is the model based control layer, also known as model predictive control (MPC) or advanced process control (APC). A dynamic and linear model based approach, MPC takes the DCS information as input to predict plant behaviour. Using manipulated variables, it minimises the difference between predicted behaviour and preferred behaviour. The model is a simple linear relationship and could be correlated from plant runs with step changes. A sound model based on first principles ensures success.³

A classic example of MPC is the variation of pumparound duty in a crude column when the throughput is changed.

Plantwide management and optimisation (Layer 3) is the real time optimisation (RTO) layer and is above the MPC. RTO is a steady state and non-linear model based optimisation. Optimisation calculation is based on the whole unit model including the heat and mass balance, reaction kinetics, cost functions of feeds/products, and utilities consumption and costs.

A good example is the RTO of a fluid catalytic cracking unit (FCC), wherein the severity of operation is strongly related non-linearly to heat balance and yield of products.

Optimisation today
Exploring the various offerings in the market today, one can group them into various layers (see Table 1).
The bulk of the recent push in optimisation and asset performance has been in Layer 5, especially in the area of data analytics and equipment connectivity/IIoT.

Data analytics can be useful in predicting equipment or process limitations.4 Where there is a deluge of data, organising and analysing the data to make meaningful relationships requires considerable effort. Data analytics tools provide operators with an easy path to understand the data and analyse them.

Another important area for data analytics is to predict event occurrence. Multiple vendors offer event prediction software and services. The event could be a process occurrence such as a high liquid level trip or an equipment level incident like a wet gas compressor trip due to seal failure. Predicting event or incident occurrence can buy valuable time for an operator to intervene and avoid a trip or shutdown. However, the tool will be useful only as long as the causal relationships are correctly captured by the prediction logic.

Data analytics is also useful in discovering causal relationships between a key process parameter and other variables that may not be easily visible. This is very useful for studying or analysing equipment/unit performance extremities to identify variables that cause such extreme behaviour.

IIoT has opened the doors for remote asset management for smaller enterprises and remote process units. While big entities had remote asset monitoring before the advent of IIoT, IIoT will enable wider deployment, better tools for analytics and diagnostics, open standards and falling costs.

The other development which is still relatively nascent is the merging of the MPC and RTO layers to give dynamic RTO (D-RTO) or non-linear MPC (NMPC). Developments in this field will help in the optimisation of batch and semi-batch processes.

Process model
The bedrock on which process optimisation is built is the process model. A model could be a simple linear algebraic expression or a complex simulation model. Ideally speaking, the simulation model that is used for design would be extended and used for various optimisation efforts in other layers. However, due to computing efforts and isolated platforms, over the years different models were built to address the needs of different optimisation strategies and methods. Recently, there has been a considerable move amongst vendors to have platforms ready for a model to go from the design phase through the project lifecycle into the operations and optimisation phase.

During the design phase, a simulation model is built that will be used to generate the heat and material balance, stream summaries and process information for equipment design. These models are typically steady state and sequential modular simulation models.

The models that are used in MPC/APC could be equations that are fitted to model the behaviour of the section of the unit that is controlled. The model used for MPC could take instrument measurements as inputs to predict process behaviour and affect control action to bring the operation to the desired optimum. These are models that are not reflective of the whole process or unit. Their utility value is in the ability to take instrument measurements and control the unit dynamically to meet desired operation. The importance of capturing the right first principles and the correct process or technology nuances in establishing the model is key to bringing about the desired results.³