Rainfall Prediction using LSTM for Darjeeling District

Project Overview

This project aims to develop a rainfall prediction model for the Darjeeling district in West Bengal, India. The model leverages the power of Long Short-Term Memory (LSTM), an advanced Recurrent Neural Network (RNN) architecture. By utilizing a decade's worth of daily rainfall data sourced from the India Water Resources Information System (India WRIS) website, the model undergoes training and testing to accurately predict future rainfall patterns.

Data

• Source: India WRIS website
• Scope: Daily rainfall data for Darjeeling district, spanning 10 years.
• Preprocessing: Data cleaning, checking for null values, stripping trailing white spaces ans un necessary rows in the dataset and visualization of rows.

Methodology

1. Data Acquisition: Download and preprocess historical rainfall data.
2. Model Development:
   • Build an LSTM model architecture.
   • Experiment with different hyperparameters for optimal performance.
3. Model Training: Train the LSTM model on the historical dataset.
4. Model Evaluation: Evaluate the model's accuracy using appropriate metrics (e.g., Mean Squared Error, Mean Absolute Error).

Dependencies

• Python
• TensorFlow/Keras
• Pandas
• NumPy
• Matplotlib (for visualization)

Getting Started

Data Preparation:

• Download rainfall data from India WRIS for the specified period.
• Preprocess the data as needed (cleaning, handling missing values, visualization).

Model Development:

• Implement the LSTM architecture in the preferred deep learning framework (here LSTM)
• Experiment with different hyperparameters (number of layers, neurons, dropout, etc.) to optimize performance.

Model Training and Evaluation:

• Train the model on the prepared dataset.
• Evaluate the model's performance using suitable metrics.

Future Work

• Explore other deep learning architectures (e.g., GRU, CNN-LSTM) for potential enhancements.
• Develop a web application or API for real-time rainfall predictions.

Setbacks

• There's a risk of overfitting, especially when the model is trained on a limited dataset, which might not generalize well to unseen data.
• The model's performance is highly dependent on the quality of data preprocessing, which can be time- consuming and requires careful handling.
• The model's predictions are heavily dependent on historical data, which may not always accurately predict future rain conditions due to unforeseen events​