Latest Projects
Project | 01
.png)
Digital Twin - IfM @ Cambridge
Forecasting household-level electricity demand
​​
Intro: A digital twin is a digital representation of a physical asset or a system. Recognizing the importance of data for the UK economy, the UK National Infrastructure Commission published a report in which it was suggested to develop a "National Digital Twin".
Personal Contribution: Our purpose at IfM is to define a dynamic digital twin of the West Cambridge campus. My specific task is to develop, try and test various deep-learning (and other) methodologies for forecasting household-level electricity demand. These methodologies include but are not limited to: LSTMs, GRUs, CNNs and Random Forests. In order to be able to utilize arbitrarily big datasets in the modeling process, I have to use a multiple parallel input and multi-step output version of the traditional techniques, a fact that requires a lot of careful and proper data engineering.
Project | 02
.png)
Credit Risk - Amex @ NYC
Developing semi-supervised GANs
​​
Intro: Credit card issuers face a constant decision-making problem: whether or not to extend credit to people who are applying for that. In an ideal setting, they would know who is going to default and who is not and they would make their decisions based on this. In a real setting though, they have to quantify the risk associated with any incoming customer.
​
Personal Contribution: Building models for existing customers is a fairly easy problem. However, my specific task was to come up with ways to infer default tags for declined population segments. In order to do so, I expanded and tested semi-supervised GAN architectures, which are able to get both labeled and unlabeled input. The results were really promising, as I achieved nearly identical performance (measured in capture rate) with the current super-expensive benchmark models.
Project | 03
Expansion Planning - Rutgers @ NJ
Designing RL frameworks for energy planning
​
Intro: Expansion planning problems refer to the monetary investment need to be made in the central grid or in a microgrid to meet the constantly growing demand for reliable electricity. The criticality of formulating systematic, analytical and novel methodologies to tackle these problems can be easily justified by considering the technological advancements continuously being made in the area, and the high operating costs incurred in these processes.
​
Personal Contribution: My goal in this project is to to formulate a highly detailed long-term expansion planning problem in a microgrid setting and solve it using advanced artificial intelligence techniques. Towards this direction, I utilize novel reinforcement learning algorithms in order to develop a unified dynamic optimization framework and use it to derive expansion strategies. I also identified the overestimation bias problem of the Q-Learning algorithm in my setting and developed a simulation-based technique to mitigate it.