Islanding Detection using PMUs and A3C

February 16, 2021

This is a part of my Bachelor’s Thesis:

  • An inbuilt IEEE 34 node bus system is procured where renewable energy sources (PV, Wind and battery) are added at the distribution end in Simulink.
  • Focus is made to work on deep reinforcement learning (A3C) classification framework to identify whether islanding has occurred or not.

Microgrid Energy Management System Optimization

July 05, 2020

Winner of the Smart India Hackathon, track set by Mathworks. Later on, the work got accepted in IEEE IEMRE’21. Renewable Energy and Demand Forecasting in integrated smart grid.

  • Designed an integrated microgrid model in Simulink.
  • Load consumption and energy generation were predicted using ensemble learning algorithms and LSTM cells of Recurrent neural networks (RNN) with R2 values of 0.95 for solar, 0.82 for wind, 0.91 for load and 0.95 for price.
  • Linear-Programming based optimization was used for the scheduling of the sources to meet the load demand, minimizing the cost of distribution and storage of power.

Autonomous Self-Driving Car Simulation

April 14, 2020

  • Graphics creation using Kivy modules in Python. A car with 3 sensors in front used in the simulation.
  • Deep Q-Learning reinforcement techniques used with specific reward policies where the taxi runs downtown to the airport and back.
  • Sand used for its simulation so that the taxi can learn through experiences stored in a batch of 100.
  • Reward = -1 given to the taxi when it crashes into the sand or reaches outskirts of the city, Reward = -0.2 given when it moves further away from the destination, Reward = +0.1 given when it approaches in correct direction of the destination, Reward =+1 given when it reaches the goal.

Seq2Seq Architecture based DL chatbot

January 20, 2020

This is a Proof-of-Concept Project on Seq2Seq Neural Architecture.

  • Trained on Movie-lens dataset having conversations between different people by building a Seq2Seq neural architecture.
  • It is trained on 100 epochs with a batch size of 64, and number of layers =3. The encoding and decoding embedding size are of 512 each.
  • Dropout regularization used to remove any kind of overfitting and the losses are optimized using Adam optimizer.
  • All the sentences in a batch, whether they are questions/ answers must have the same length. Hence ‘PAD’ tokens used.

Home Automation System

January 17, 2019

Here, the objective was to control the lighting system using a smartphone application.

  • 8051 Microcontroller acts as the controller unit of the project where Kiel u Vision 5 provided its coding interface.
  • HC05 bluetooth module establishes the communication channel between phone and controller.
  • Comparator IC LM324 converts the analog form of the signal to digital form required by the microcontroller
  • Motor driver IC L293d is the small current amplifier which takes in low current control signals from the microcontroller and converts it to higher current signal for the lighting system.