Hi,
I am a budding researcher in the area of data science, statistical computing, and machine learning with a background in computational fluid-thermal science. I am currently working with Dr. Pitchumani at Virginia Tech to develop a physics-informed machine learning tool for probabilistic solar forecasting and to integrate that into power system operation and planning. In the long run, I yearn to spearhead the revolution of AI/ML to disrupt the energy and the automotive sectors and to provide a feasible AI-human interface in those sectors.
Skills:
Language: Python, Matlab, R, SQL
ML and Data Analysis Libraries & Framework: Numpy, Panda, Scikit-Learn, Matplotlib, Plotly, Keras, Pytorch
Machine Learning skills: Clustering, Time-Series Forecasting, Ensemble methods, Decision Trees, Uncertainty Quantification, Bayesian Inference, Classification, Predictive modeling, PCA, SVM
CFD solver & Tools: Fluent, CFX, ICEM-CFD, Turbogrid, CFD post
CAD modeling: Siemens NX, Solidworks
Operating System: Windows, Linux
Misc. : Latex, Git version control, HPC computing, Working with Big Data, MS Word, MS Excel, Mathcad, Maple,
Professional website: https://sbhavsar0.wixsite.com/mysite
-
Experience
Since last year, Sujal has been working as a Graduate Research Assistant in Dr. Pitchumani’s research group at Virginia Tech. His principal role is to provide a solution to challenges posed by the increased penetration of renewable energy into existing power-grid through data-driven machine-learning-based mathematical models.
Graduate Research Assistant (ML/AI)- Advanced Material Technology Lab., Virginia Tech
Sep. 2019- May-2021
Project: Operational Probabilistic Tool for Solar Uncertainty (OPTSUN)
Funding: Department of Energy
Collaboration: Electric Power Research Institute
• Uncertainty quantification in renewable energy generation.
• Probabilistic forecast using ensemble machine learning technique
• Integrate forecasting tool into a system operation and planning
Skills and Tools: Python Programming, Unsupervised Machine Learning, Statistics, Time-series Forecasting
Project Associate - Computational Combustion and Energy Conversion Lab., Indian Institute of Technology Kanpur, India
• Achieved 3% reduction in the size of the micro gas turbine by proposing a robust design of pipe diffuser through CFD simulation
• Implemented Design of Experiment (DOE) based Bayesian optimization to optimize the design of the gas turbines.
• Created MATLAB script to handle 5 million data sets to automate the post-processing activity.
Skills and Tools: Data Handling, Computation, Programming, Simulation, Design optimization
-
Projects
Apart from the primary job, Sujal has tried to maximize the medium through which he could get an exposure to the trend in Data Science and Machine Learning. As a part of this, he has done a couple of exciting projects and competitions.
1. M5 Forecasting – Accuracy (Kaggle Competition): Sujal’s team stood at 56th out of 5558 competitors and secured the place in the top 1%, and secured the silver medal. He dealt with a hierarchical time-series of sales data from Walmart and additional explanatory variables. A challenge in this competition is to predict a daily sale at each level of hierarchy over a month’s horizon. Sujal and his team tried several approaches, ranging from traditional methods such as ARIMA, SARIM, to machine learning and deep learning methods such as Gradient Boosting Decision Trees, Recurrent Neural Networks, Wavelet, etc. as a part of figuring out the best modeling approach for the given dataset. Our team proposed an ensemble of Light Gradient Boosting with Facebook’s prophet for base forecasting at each level. The final coherency amongst different levels has been achieved through appropriate post-processing. This project demanded a heavy data-wrangling, data-management, understanding of feature engineering techniques that, too, combined with the machine learning models. In the end, our predicted forecast of the number of sales came out to be on par with actuals, and we ended up getting a silver medal on the leaderboard.
2. Customer recommendation system for the adoption of solar PV: In this project, Sujal used a consumer survey data of rural Virginia to determine adopters' decision behavior of solar PV. He developed a data-driven model using Decision Tree approach with Light Gradient Boosting framework, which takes a survey response of responder as input and predict the tendency of that consumer toward solar PV adoption. Customer recommendation system for the adoption of solar PV: In this project, Sujal used a consumer survey data of rural Virginia to determine adopters' decision behavior of solar PV. Sujal has deployed the latest variate of Generative Adversarial Network to augment the original sample with the synthetic sample. Sujal has utilized a Bayesian optimization technique to tune the hyper-parameter of the Gradient Boosting model. The proposed model was able to identify highly under-represented class out of the cluster of dominated non-adopter in a fairly accurate way.