Weather Prediction using LSTM and GRU

This project focuses on predicting weather conditions using historical weather data. The workflow follows a structured pipeline that includes data preprocessing, exploratory data analysis (EDA), model building, and evaluation of LSTM and GRU models.

Workflow Overview

1. Importing Necessary Libraries

The notebook begins by importing essential libraries for data manipulation, visualization, deep learning, and evaluation. The PyTorch library is used for building and training neural network models.

2. Data Loading and Exploration

• The dataset (weatherHistory.csv) is loaded using Pandas.
• Initial exploration includes displaying the first few rows, checking data types, and understanding missing values.
• Missing values are handled using forward fill to maintain temporal consistency.
• Redundant columns (e.g., Daily Summary) are removed to avoid duplication of information.

3. Data Preprocessing and Feature Engineering

• Features are selected and transformed to prepare the dataset for deep learning models.
• Normalization is performed using MinMaxScaler to scale numerical features.
• Categorical variables (e.g., Precip Type) are encoded using OneHotEncoder.
• The dataset is split into training and testing sets using train_test_split.

4. Model Preparation and Training

• The dataset is converted into TensorDataset and loaded into PyTorch DataLoader.
• Two models are defined and implemented:
  • LSTM Model: A Long Short-Term Memory network for capturing temporal dependencies.
  • GRU Model: A Gated Recurrent Unit network as an alternative to LSTM.
• Both models are trained using the Adam optimizer and Mean Squared Error (MSE) loss function.
• Training involves multiple epochs, and performance is monitored using loss curves.

5. Model Evaluation and Comparison

• The trained models are evaluated using standard metrics:
  • Mean Squared Error (MSE)
  • Mean Absolute Error (MAE)
  • R-squared Score (R²)
• Performance comparison between LSTM and GRU models is conducted to determine the better-performing model.

Results and Observations

• The notebook provides a visual representation of loss curves for both models.
• Evaluation metrics help in assessing the accuracy and generalization of LSTM and GRU for weather prediction.
• The better model is selected based on its ability to minimize error and maximize predictive performance.

Requirements

To run this notebook, install the required dependencies:

pip install pandas numpy matplotlib seaborn torch scikit-learn

Conclusion

This project demonstrates the application of deep learning in time-series forecasting, particularly using LSTM and GRU networks. By preprocessing data effectively and leveraging sequential models, we can make meaningful weather predictions.