Abstract:
The digital twin (DT) is one of Industry 4's enabling technologies, aiming to improve the
performance of physical entities through the use of a virtual counterpart. DT is a real-time
representation of a system or physical entity that can adapt to operational changes in real time
by using data and information gathered online to predict the physical twin's future. Despite
all the advancements in both the industrial and academic sectors, there are still challenges for
which DT adaptation requires more investigation. DT has been developed and used for three
major purposes: service, design, and manufacturing for general industrial purposes. To
develop DT to achieve all three major purposes is not realistic since different industries have
different implementation orders of DT depending on their hardware/software level, and
maturities of service, design and manufacturing. In this project, we will focus on DT virtual
entity development purposes based on their own situations to implement. Therefore, this
research will focus on developing virtual entities for the identified gaps in industry
application.
Full implementation of ‘proper’ DT requires twinning between the virtual and physical
entities. However, this is complicated by computational effort, limitation of online
measurements, and process complexity. This work introduces new approaches for processes
in which mechanistical methods alone are incapable of providing DT virtual entities. The
processes addressed are cream cheese fermentation and PCM energy storage, in the dairy and
renewable energy industries respectively. These processes pose distinct challenges,
necessitating novel approaches that combine data-driven and mechanistical methods.
Process design has advanced significantly in dairy industry as a result of the high global
demand for dairy products; however, process operations (manufacturing) require further
improvements in order to increase product throughput while maintaining consistent quality.
DT can become an important tool toward improving the online process operation in the dairy.
However, proper virtual representations of systems using DT introduces a significant
challenge in the dairy industry due to the complexity of mechanisms, raw material variations,
and the limitations of online measurements. These factors reduce the fidelity of current
industry and academic models.
Additionally, developing DT for new renewable energy technologies is one of the most
critical gaps. Online solutions can help improve the economic viability of such technologies,
by improving energy management during operation. The DT application for improving online
operations necessitates a real time solution. However, this is challenging due to the
complexity and nonlinearity of the systems involved in renewable energy technologies.
Two DT virtual entity development directions for the aforementioned industries are
investigated in this study. The results can aim in energy savings in New Zealand which is
necessary for achieving Zero Carbon Act objectives. The first is in the dairy industry, which
is vital to New Zealand's economy and consumes a significant amount of energy. A cream
cheese production unit including fermentations vats is studied. DT virtual entity needs to
optimize the energy while meeting the quality standards by providing online scheduling of
the vats. The second process involves using energy storage and solar air collectors as
renewable energy sources. This system is well suited to New Zealand's temperate climate.
However, this system's economic viability must be improved while meeting energy demands.
DT can optimise the economical design of solar collectors and energy storages. DT can also
provide an optimal control solution, allowing the system to save even more energy and cost.
In the cream cheese fermentation unit, the batch duration variation is the main challenge
of optimizing the vats scheduling since mechanistic modelling of the dynamics of the
fermentation process is difficult and includes many time-varying parameters. The other case
introduced a different challenge, the mechanistic model of the system is available for active
phase change material (PCM) systems; however, its complexity makes the optimization using
that model computationally infeasible. To provide solutions, combination of data-driven and
mathematical modelling as well as optimization were applied. The detailed virtual entity
developments for the two industries are introduced in the following paragraphs.
The first process investigated the scheduling of cream cheese fermentation vats. The
scheduling of fermentation vats was complicated by batch duration variations and the
limitations of online measurements. For the first time in this work, a DT virtual entity was
developed to schedule the cream cheese fermentation vats. The scheduling framework was
twined with the fermentation vats for filling and draining them using the information for
predicting the batch durations required for reaching the quality. The pH predictive models
were studied because pH is an important quality indicator. Novel grey and black box models
were developed for pH prediction. It was possible to create grey and black box models that
could use available online measurements to improve pH prediction. Only one fermentation
had a large error of 34.09 %, and the average network percentage relative error was less than
14 %. However, the results showed that the pH prediction at the final point matches the
experimental data in all the test fermentations.
The pH predictive models were applied in a scheduling framework. The scheduling
challenge is due to the inability to predict batch durations from the fermentation start time.
This causes batches interferences during draining, resulting in higher energy consumption
and waste. The new DT virtual entity for scheduling presented a mathematical programming
optimization that was formulated to solve this problem by assuming an initial default batch
duration at the start time and updating the batch duration later based on the pH prediction
model later when enough measurements were available. The 12 hours initial batch duration
led to the best results in reducing idle time, energy consumption, and waste.
In the second process, an energy storage was coupled with a solar air collector for
supplying heat to a hut in Auckland. The model of the entire system plays an important role
for achieving DT virtual entity. The entire system's first principles models were developed
and validated using experimental data for this purpose. The results confirmed the dynamics
model's reliability, with an average mean square error of 4°C between measured and
predicted hut temperatures over 11 days. The validated model was used for designing. For
the first time in this research, economical designing of the system was carried out for various
application scenarios in Auckland. The results showed that the optimum surface area of the
solar collector was the same at 1 square meter for all the scenarios; however, the optimum
amount of PCM mass for service, domestic, and office scenarios was 35 kg, 20 kg, and zero,
respectively.
The model was also used for developing a virtual entity for optimal control of the system.
The desired optimal control should provide online solutions despite the system complexity
coming from the nonlinearity and existence of binary variables. A novel optimal control was
developed using reinforcement learning (RL). For addressing the approach advantages, it was
compared with a model predictive control (MPC), as a classical approach. RL as a data driven
optimal control can be trained offline and provide real time solutions. A novel formulation
was proposed that allowed for the adjustment of a reward to prioritise between thermal
comfort and energy cost savings. By prioritizing the cost saving, 97% more cost was saved
compared to prioritizing thermal comfort case, however, the offset from the desired
temperature was 54% less. A balance between energy cost savings and thermal comfort was
achieved by adjusting the reward.