top of page
This site was designed with the
.com
website builder. Create your website today.Start Now

Intro

I took my B.Tech. Project - I (BTP -I) in the department of computer science, IIT Kgp. I completed my BTP - I under the expert guidance of Prof. Sudeshna Sarkar. The project was titled forecasting of sea wave height using deep learning architecture. 

This project provided me with an opportunity to dive into the fundamentals of Recurrent neural networks and LSTM. My project was rewarded with an A grade for the research. 

IIT_Kharagpur_Logo_edited.png

Wave height forecasting using deep learning

July - November 2019

About my project

How I did it

This is my first B.Tech. Project and first inter-departmental project. I got this opportunity to complete my BTP - I in CS department under guidance by Prof. Sudeshna Sarkar (HoD, School of AI). I worked on the dataset retrieved from UK national weather service
‘Met Office’. In this project I used time-series analysis with LSTM. Under this project I experimented with different types of LSTMs and variables which may affect the results. Total number of variables were 18 including wind speed, wind direction etc. Out of them I experimented with different variables for the best results. 

​

The problem statement: 'Prepare a model for wave height forecasting using deep learning.'

​

The project can be broken down into two major parts: 

  • Preparation of the dataset 

  • Experimenting with different models

​

This research was conducted at IIT Khargapur, India and is in the continuation for my summer research project in 2018. 

time_series.png

Skills and Tools

  • Data analysis

  • Fuzzy logic and membership function

  • OpenCV

  • Djikstra algorithm

Month.png
Preparation of the dataset

  • The dataset has 18 variables with 174,532 readings. For better visualization, I plotted scatter plot in MATLAB of wave height against months and year. 

  • I noticed an interesting pattern in the time series of wave height when plotted against month (shown in the figure below) 

  • Further normalized the dataset so that data can be fed into the model. Then I divided the dataset for testing training and validation in the ration of 7:2:1.

time_series.png
Experimenting with different models

  • We used broadly three types of LSTMs for prediction:

    • Single-layer LSTM​s (Total trials: 4)

    • Multiple layer LSTMs (Total trials: 2)

    • Bi-directional LSTMs (Total trials: 2)

  • I experimented with different batch size and prediction period like 5 days, one week, 15 days. 

Result

After experimenting with different types of LSTMs, different batch size and different forecasting period, our result showed that the best performing model was Bidirectional LSTM with a batch size of 3 months and 15 days.

The model showed the accuracy of 94.6% for the prediction of up to 15 days in advance.

Click here for Letter of recommendation. 

bottom of page