Boodhoo, Khivishta
(2024)
Machine learning methodologies for real-world resilient energy systems.
EngD thesis, University of Nottingham.
Abstract
With the growing global population, energy demand and consumption have increased. Higher fossil fuel consumption has led to increased greenhouse gas emissions and global warming. However, the 4th Industrial Revolution, driven by digital, biological, and physical innovations, has shown an increased focus on environmental protection and clean energy use. Technologies such as Artificial Intelligence can help meet the rising energy requirements sustainably. Artificial Intelligence mimics human intelligence to automate tasks, optimise processes, and enable predictive analytics for better decision-making. Machine learning, which is a branch of Artificial Intelligence, uses data driven algorithms to solve problems. Its applications have been widely investigated in data-rich sectors, including the energy sector. Despite its promising benefits, such as cost savings, improved efficiency, grid maintenance, and reliability, there are still multiple areas of well-needed research. For instance, it has been found that energy systems are complex, everchanging, and may need more robust algorithms owing to their dynamic and intermittent nature. The constant updates, upgrades, replacement, and integration of renewable energies in such systems require further research to improve their scalability and resilience, and to allow the full harness of the potential of machine learning in addressing those pressing energy challenges.
In this context, the thesis explores three case studies to investigate the importance of distinct steps in machine learning approaches when applied to energy systems, to improve the overall process. For a proper representation, case studies targeting data-rich and diverse energy systems, such as a wind turbine acting as an energy-producing system, an oil platform as an energy-consuming system, and vanadium redox flow batteries as an energy storage system, were selected. Case study 1 focused on increasing the energy efficiency of a wind turbine, whereby scheduling maintenance when the predicted energy output is low, increased the overall power production. In this case study, two machine learning models were developed, known as artificial neural networks, which were compared with a baseline regression model. The effects of the model’s inputs, preprocessing versus none, different granularities, and quantity of data equal to one month and one year were examined. The best performing results showed that specific preprocessing steps, such as outlier detection and removal of any manual intervention, are essential. Moreover, having a granularity of 4 minutes instead of 1 second significantly decreased the computational efficiency while maintaining the model accuracy. Ultimately, the model performance metrics were shown to have an R2 of 0.98, RMSE of 283 on the test set, and MAE of 1462 when testing on an independent dataset. These results revealed the huge applicability of developing such a model by applying it to an independent dataset consisting of weather forecasts for future power predictions.
For Case Study 2, the focus was on reducing the daily diesel consumption of diesel generators on an oil platform. Mapping diesel flow across the system was necessary to determine the model’s inputs. Preprocessing steps were conducted to improve feature-label correlations such as matching the granularity of the power loads (features) to diesel usage (label) daily. Various machine learning models, from baseline regression to tree ensembles and artificial neural networks, were explored. The baseline regression model and the artificial neural network had comparable performance metrics, with R2 values of 0.80 and 0.81, respectively. However, the artificial neural network was chosen for its ability to uncover nonlinear patterns, and its low complexity and reasonable training times which considering all made it a practical choice for this analysis especially with ongoing data collection and the addition of a diesel generator, which remained broken during the study. Daily power loads were then generated via the Cartesian product and the best combinations were used to predict the lowest diesel consumption that can be achieved daily. These specific daily power load combinations can help to save energy on oil platforms.
Case Study 3 involved the early detection of severe faults in vanadium redox flow batteries to prevent their complete shutdown; thus improving their energy efficiency. A baseline isolation forest model was developed on each of the three available sites. Another isolation forest model was developed on a single site using the dataset containing primarily normal functioning data. This model was then applied to the other clusters and sites for anomaly detection. Similarly, a hybrid model consisting of an autoencoder, and an isolation forest was developed. It was found that this model had better performance in terms of applicability across sites and fewer false positives along with false negatives than the single isolation forest model. Its performance was comparable to the three baseline Isolation Forest models developed on each site. The performance was rated according to the few known anomalies, as indicated by the domain experts and their relative percentages in the different datasets, with the hybrid model showing the lowest percentage of anomalies in the normal dataset along with the three baseline isolation forest models both identifying all the known anomalies correctly.
Through the case studies, the relative importance of each step in the application of machine learning to energy systems has been explored. For example, the most important step revealed was the need for relevant data in sufficient quantities, which has been highlighted in Case Study 3, where an unlabelled dataset (having no recorded anomalies) led us to assumptions grounded on domain experts, which may not be as precise as data collected through sensors. It was also concluded that the amount of data greatly affected model performance metrics, as shown in Case Study 1. Then, Case Study 2 indicated that focusing on the system's working principle helped identify factors affecting its operation and map the data sources and distribution across its components. Along with data exploratory analysis, basic insights about the next possible preprocessing steps arose, leading to increased data quality. Next, in line of importance, is the model development process. The selection of appropriate algorithms for each problem followed a systematic approach: we began with simpler models and progressively moved to more complex ones across the case studies. This gradual increase in model complexity allowed us to establish baselines and better understand the trade-offs between model simplicity and performance. The simplest model can suggest possible correlations and act as a benchmark. One specific step, noted as missing in various past research would be the hyperparameter tuning, which needs to be specified for reproducibility.
Then, exploring the granularity of the data resulted in downsampling and insufficient samples being observed in Case Studies 1 and 2 respectively. Moreover, matching the granularity of the features to the label was consequential, as shown in Case Study 2. Subsequently, the applicability of the developed machine learning model was further investigated when an independent dataset was used to deploy the model and assess its capabilities, as revealed in Case Studies 1 and 2. This step was found to be frequently overlooked in past studies. The importance of deployment is further highlighted in this thesis by the visualisation tools crafted for real world usage in each case study. The developed model, made accessible in the right format for the end user is crucial for bridging the gap between research and industry. The monitoring and feedback of the deployed model are also believed to be essential for its proper usage.
Moving on to the next important step, it was found through the extensive literature review of this thesis that diverse model performance metrics are often used in multiple studies to assess the performance of a model on the test set, thus rendering comparisons challenging. Conversely, as demonstrated, using a baseline model can enable an adequate evaluation of the model and using multiple common model performance metrics in one study would allow a more valuable comparison with the literature. Finally, the relevance of seasonal variations in each case study was also determined. Because this factor did not greatly affect the case studies in this thesis, this step came last in our order of importance. However, further work needs to be conducted to explore this step in the process of developing an ultimate and all-rounded machine-learning approach for energy systems.
| Item Type: |
Thesis (University of Nottingham only)
(EngD)
|
| Supervisors: |
Watson, Nicholas
Meredith, William
Triguero, Isaac
Plumbly, Josh |
| Keywords: |
Machine learning; Energy systems; Energy efficiency; Wind turbine; Diesel generators; Vanadium redox flow batteries |
| Subjects: |
T Technology > TJ Mechanical engineering and machinery |
| Faculties/Schools: |
UK Campuses > Faculty of Engineering > Department of Chemical and Environmental Engineering |
| Item ID: |
79870 |
| Depositing User: |
Boodhoo, Khivishta
|
| Date Deposited: |
10 Dec 2024 04:40 |
| Last Modified: |
10 Dec 2024 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/79870 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[thumbnail of modified thesis with corrections]](https://eprints.nottingham.ac.uk/style/images/fileicons/application_pdf.png "modified thesis with corrections")