Periodic Reporting for period 3 - FUDIPO (Future Directions of Production Planning and Optimized Energy- and Process Industries)
Berichtszeitraum: 2019-10-01 bis 2021-01-31
PROBLEM/ISSUES: In process industries it is important to get correct information from the process to perform optimal control of the processes. Different type of information is gathered and analysed and then utilized for enhancing both production rate and minimize the product quality variations. By doing this the cost per produced unit is enhanced and the prize of the sold products increased, as better quality has a higher value. The goal is also to make it possible to modify processes and implement new controls in an easy and fast way when new products are to be produced.
IMPORTANCE FOR SOCIETY: This is important for society as better performance is making EU industry competitive, which generates money that is needed to fulfill all societal obligations, including both environmental and social responsibilities.
OVERALL OBJECTIVES: The purpose of FUDIPO is to develop and implement in full-scale, and validate up to TRL 6 a system for optimization on all levels in a factory integrating the different control levels from the separate production units to mill level. FUDIPO will integrate machine learning functions on a wide scale into several critical process industries, showcasing radical improvements in energy and resource efficiency and increasing the competitiveness of European industry (pulp & paper, oil refining, waste water treatment, metals, food, manufacturing, ceramic production, steel production, etc).
IMPORTANCE FOR SOCIETY: This is important for society as better performance is making EU industry competitive, which generates money that is needed to fulfill all societal obligations, including both environmental and social responsibilities.
OVERALL OBJECTIVES: The purpose of FUDIPO is to develop and implement in full-scale, and validate up to TRL 6 a system for optimization on all levels in a factory integrating the different control levels from the separate production units to mill level. FUDIPO will integrate machine learning functions on a wide scale into several critical process industries, showcasing radical improvements in energy and resource efficiency and increasing the competitiveness of European industry (pulp & paper, oil refining, waste water treatment, metals, food, manufacturing, ceramic production, steel production, etc).
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.
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.
We are developing AI systems with machine learning at different levels mimicking the human brain.
At lowest level we sort data into "good and bad" where "bad data" is e.g. missing data, outliers and similar. These are used for diagnostics. "Good data" is normal data depending on process variations.
We use physical and statistical models for data reconciliation, where missing data is replaced by most probable from solving equation systems with larger parts of the plants beeing balanced with measured input data. From this we use reconsiled data for model based control and optimization. We also adapt models to include new experience.
As an add on at the top of this we calculate risk for failures of different process equipments, external conditions like orders, price for chemicals, energy etc as well as price of the products. Here we develop new type of "deep learning" systems to create new information on how to optimize the production with respect to long term versus short term profits and risks, and suitable trade off between these perspectives, where planning of maintenance is one major condition to consider.
At lowest level we sort data into "good and bad" where "bad data" is e.g. missing data, outliers and similar. These are used for diagnostics. "Good data" is normal data depending on process variations.
We use physical and statistical models for data reconciliation, where missing data is replaced by most probable from solving equation systems with larger parts of the plants beeing balanced with measured input data. From this we use reconsiled data for model based control and optimization. We also adapt models to include new experience.
As an add on at the top of this we calculate risk for failures of different process equipments, external conditions like orders, price for chemicals, energy etc as well as price of the products. Here we develop new type of "deep learning" systems to create new information on how to optimize the production with respect to long term versus short term profits and risks, and suitable trade off between these perspectives, where planning of maintenance is one major condition to consider.