Time Series and Machine Learning Approach for Weather Forecasting

In this project, our goal is to analyze and forecast the temperature of Pune city using time series modeling and machine learning techniques. Specifically, we aim to study the trend patterns in the weather dataset, identify seasonal variations, and fit appropriate time series models. By doing so, we intend to generate forecasts that capture both trend and seasonality. Additionally, we will compare the performance of different time series models to determine the most accurate and reliable approach for temperature prediction. This work has practical implications for urban planning, agriculture, and daily life, helping to better anticipate and respond to weather changes in Pune.

The Steps I Took

Collected historical weather data from Kaggle.
Processed and cleaned the data for analysis.
Conducted exploratory data analysis (EDA).
Created visualizations for temperatures, rainfall, correlations, ACF/PACF, and seasonal decomposition.
Fitted SARIMA, Exponential Smoothing, and ARIMA models.
Compared models using AIC, BIC, MSE, RMSE, and MAE metrics.
Determined Exponential Smoothing model as best for forecasting.

Why I Took These Steps

Ensured reliable data for analysis.
Prepared the data for accurate model training.
Identified patterns and trends in the data.
Visualized key data insights for better understanding.
Applied various statistical models for comparison.
Evaluated model performance for accuracy.
Selected the most accurate forecasting model.

Tools I Used

R Software
Python
Tableau

Challenges:

Memory Limitations: During the fitting of the SARIMA model, Python encountered a memory error due to insufficient available memory. This prevented the model from forecasting beyond a short time horizon of 48 hours, limiting the ability to project temperature trends over longer periods.
Forecasting Limitations: While ARIMA successfully generated forecasts for the entire year, the sheer volume of forecasted temperature values exceeded practical reporting limits. As a result, only a subset of 24-day forecasts could be effectively included in the report.
Reporting Constraints: The need to condense and present forecasted data within manageable limits posed a challenge. Balancing comprehensive model outputs with concise reporting required prioritizing key timeframes and findings to effectively convey insights from the forecasting exercise.

Here are some of my wonderful visuals from this project:

Bar Graph of monthly average maximum and minimum temperature

Bar Graph of Monthly Average Temperature

Bar Plot of Monthly Average Rainfall

Correlation Heatmap

Autocorrelation Function (ACF)

Partial Autocorrelation Function (PACF)

Seasonal Decomposition

Results:

The analysis of weather data revealed distinct seasonal patterns and variations:
- The average maximum temperature peaks in April and May.
- The average minimum temperature drops to its lowest in December and January.
- Rainfall follows a seasonal trend, with peak rainfall in July and lowest in January.
The Exponential Smoothing model is the optimal choice for this dataset, excelling in MSE, RMSE, and MAE metrics. It also shows superior AIC and BIC values, making it a simple yet effective model. Its ease of implementation and accurate forecasting capabilities further validate its suitability.

The blue line represents the original temperature data, and the orange line represents the forecasted values of temperature for the next 365 days (1 Year).
Using the Exponential Smoothing model, we forecast the temperature for the next 2 days (48 Hours).