INDUSTRIAL USE CASES:
OIL REFINING PLANT (Tupras) crude oil is bought in different qualities. If
the properties are better predicted, the production planning can be optimized to
utilize the oil available in the best possible way to meet the consumer needs.
PULP AND PAPER (Billerudkorsnäs) paper qualities depends on the properties of
fibers, which if are predicted by NIR measurement combined with lab
measurements for tuning a suitable mix of wood, chips with different quality can be
chosen. By matching demand from the paper mill, backward calculation can be
made on demand for fibers and how to cook the fibers, where and how much to store at
different vessels along the fiber line, etc.
HEAT AND POWER PLANTS (MTT, Mälarenergi) the focus will first be placed on stabilizing
the temperature in the boiler and steam system (or turbine entry).
BIOLOGICAL WASTE-WATER TREATMENT PLANTS (ABB) combining algae with microorganisms may
eliminate the aeration demand (energy costs) when using activated sludge
processes, giving much more biomass and biogas.
STATUS after 18 months: The focus is on implementing learning systems. We first have developed physical simulation models of the processes. In these models we have parameters which are to be adapted to follow process and sensor changes and deterioration. The models are used to correlate different variables in the processes to each other. By comparing simulated data which represent the "normal operations" to measured data we can determine process deviations and sensor faults. From this we can calculate most probable values for variables at different positions in the plant (data reconciliation) and detect faults of different kind. From reliable data we can plan for optimal production using model based control in different time perspectives. We also can build soft sensors where e.g. NIR spectra are correlated to different raw material properties. In e.g. the pulp mill at Billerud-Korsnäs we have analysed different type of wood chips used in the process with respect to lignin content and type. Concentration of chemicals, temperature and residence time in the cooking zone are simulated using the physical model. From this we predict how much lignin should be dissolved and then compare to what we actually measure after the digester using the kappa number. Deviation is used to tune reaction parameters in the model and the results is also used to tune correlation between lignin concentration and original content to the NIR spectra. This is done continuously and the system becomes better by time, as it is a "learning system". We control the process depending on the wood quality using MPC, model predictive control, and diagnose also different type of process faults like hang-ups and channeling in the digester, which is 60 meter high and up to 10 meter in diameter. Sensor faults, screen clogging, sticking valves, poor performing pumps etc is analysed and information given on status to operators as decision support and for planning of maintenance on demand. All these functions are integrated using different objects representing the different functions to make it possible to use all or part of the functions. For the pulp fiber-line case study we are up and running with he physical model, measurement of NIR spectra, kappa number and a lot of other variables and also communicating between the data base and the different functions and measurements on-line. The goal is also to try to extract information from the different functions to generate knew process knowledge and understanding going beyond what is possible today. Here we are trying to implement a new type of "supervised deep learning" function that should be feasible for process industries. For the other cases we are at different stages and implementing from the beginning parts of the full scope of functions, like Bayesian nets for diagnostics for the waste water treatment plant, remote control and diagnostics for the MTT micro-CPH case and boiler control of a waste CHP plant at Mälarenergi where different waste types may cause problem like different types of plastic, glass, ash and others in the Circulating Fluidized bed.
Different methods are being tested in a FUDIPO PhD research school during spring 2018.
STATUS after 36 months: During the last period from 18 to 36 month we have finalized both statistical and physical models for all processes and also tuned them with real process data. Also the signal flow has been developed and signals now flow from sensors over data basis into different analytic and control functions like diagnostics, fault detection, decision support, performance monitoring, decision support, Model predictive control and production planning. It has been quite challenging to get the signal flow continuously due to high demand for data security to avoid that hackers get into the systems.
