Delhi Climate Time Series Forecasting

This project focuses on time-series forecasting using the Delhi Climate Dataset obtained from Kaggle, which contains daily climate data from 1st January 2013 to 24th April 2017, collected through the Weather Underground API. The dataset includes the following attributes: meantemp (mean temperature), humidity, wind_speed, and meanpressure. The project is divided into two parts: Univariate Time-Series Modeling and Multivariate Time-Series Modeling. This README provides an overview of both parts and the methodologies applied.

Table of Contents

1. Project Overview
2. Dataset
3. Part 1: Univariate Time-Series Modeling
   • Exploratory Data Analysis (EDA)
   • Modeling Approaches
4. Part 2: Multivariate Time-Series Modeling
   • Time-Series Properties
   • Modeling Approaches
5. Conclusions
6. Requirements
7. Usage

Project Overview

The aim of this project is to develop models that can forecast the meantemp attribute from the climate data. The project is split into two phases:

1. Univariate Time-Series Modeling: Focuses on forecasting the mean temperature (meantemp) using only its historical data.
2. Multivariate Time-Series Modeling: Predicts the meantemp while also taking into account additional variables like humidity, wind_speed, and meanpressure.

Dataset

The Delhi Climate Dataset contains the following attributes:

• meantemp: Daily mean temperature in °C
• humidity: Daily average humidity in percentage
• wind_speed: Daily average wind speed in km/h
• meanpressure: Daily average atmospheric pressure in hPa

Time period: January 1, 2013 - April 24, 2017
Location: Delhi, India

The dataset is available on Kaggle via this link.

Part 1: Univariate Time-Series Modeling

In this part, the focus is on predicting the meantemp attribute using univariate time-series models. The process is as follows:

1. Exploratory Data Analysis (EDA)

• Univariate analysis: Trends and seasonality in the meantemp data over time.
• Bivariate analysis: Although other attributes like humidity, wind_speed, and meanpressure were analyzed, they were not used in this phase of forecasting.

2. Modeling Approaches

• Auto-Regressive (AR) Model: Models meantemp based on its past values.
• Seasonal Auto-Regressive Integrated Moving Average (SARIMA): Captures both trend and seasonal components in the data.
• Artificial Neural Networks (ANNs): Non-linear model using a feedforward neural network.
• Long Short-Term Memory (LSTM): A type of recurrent neural network suitable for sequential data, capturing long-term dependencies.

Hyperparameter tuning for both ANNs and LSTMs was performed using grid search to optimize the models' performance.

Part 2: Multivariate Time-Series Modeling

This section extends the previous work by incorporating all the attributes (meantemp, humidity, wind_speed, and meanpressure) into the forecasting models to predict meantemp.

1. Time-Series Properties

Before applying models, the relevant time-series properties of each attribute (e.g., stationarity, seasonality) were studied.

2. Modeling Approaches

• Vector Autoregression (VAR): A multivariate model that predicts multiple time-series attributes simultaneously, including meantemp, humidity, wind_speed, and meanpressure.
• Support Vector Regression (SVR): A machine learning approach applied specifically for meantemp prediction.
• Gradient Boosting (GB): Ensemble learning method to improve predictive accuracy.
• Long Short-Term Memory (LSTM): This deep learning model was again utilized for multivariate forecasting due to its ability to model complex time dependencies.

Conclusions

• Univariate models like SARIMA and LSTM showed the potential to capture trends and seasonal patterns in meantemp.
• Multivariate models like VAR and LSTM improved forecasting accuracy by incorporating additional climate attributes.
• Hyperparameter tuning and careful model selection significantly impacted the performance of the models.