Harveston Climate Forecasting

Project Overview

This focuses on predicting critical environmental variables for Harveston, a self-sufficient agricultural region experiencing climate shifts. Our models forecast five key climate variables to help farmers make informed decisions about planting cycles, resource allocation, and preparation for weather extremes.

Forecasting Targets

1. Average Temperature (°C)
2. Radiation (W/m²)
3. Rain Amount (mm)
4. Wind Speed (km/h)
5. Wind Direction (°)

Repository Structure

/climate-forecasting
├── data/                           
│   ├── sample_submission.csv       
│   ├── test.csv                    
│   └── train.csv                   
│
├── Notebooks_and_Scripts/
│   ├── 01_EDA_and_Analysis.ipynb   # Exploratory analysis
│   ├── 02_model_training.ipynb     # Model development
│   ├── utils.py
│   └── plots/                      # Generated visualizations
│
├── final_submission.csv            
├── technical_report.pdf            
├── README.md                      
└── requirements.txt                                

Environment Setup

Requirements

The required libraries for this project are listed in requirements.txt. The main dependencies include:

• numpy
• pandas
• scikit-learn
• lightgbm
• optuna
• matplotlib
• seaborn

Installation

1. Clone this repository:

git clone https://github.com/FouetteBytes/climate-forecasting.git
cd Data_Crunch_106

2. Create and activate a virtual environment (optional):

python -m venv env
source env/bin/activate  # On Windows: env\Scripts\activate

3. Install required packages:

pip install -r requirements.txt

Solution Approach

Our solution employs a comprehensive approach to time series forecasting, combining advanced feature engineering with ensemble modeling techniques:

Data Preprocessing

• Temperature unit standardization (Kelvin to Celsius)
• Missing value imputation using hierarchical methods
• Outlier detection and treatment
• Geographic clustering of kingdoms

Feature Engineering

• Temporal features with cyclical encoding
• Lagged variables (1-30 days)
• Rolling window statistics (multiple window sizes)
• Exponentially weighted moving averages
• Differencing features for trend removal
• Cross-feature interactions

Modeling

• Primary model: LightGBM with optimized hyperparameters
• Multi-seed ensemble approach (3 seeds)
• Target-specific feature selection
• Time-series cross-validation
• Specialized handling for directional data (Wind Direction)

Running the Code

Exploratory Data Analysis

jupyter notebook 01_EDA_and_Analysis.ipynb

This notebook performs comprehensive data exploration, visualizing distributions, temporal patterns, and correlations between variables.

Model Training and Prediction

jupyter notebook 02_model_training.ipynb

This notebook implements the complete modeling pipeline:

1. Data preprocessing
2. Feature engineering
3. Model training with hyperparameter optimization
4. Ensemble prediction
5. Submission file generation

License

This project is licensed under the MIT License - see the LICENSE file for details.