AI Weather Forecasting System

Overview

This project is a production-grade, end-to-end machine learning pipeline that forecasts 7-day weather conditions for cities worldwide. It integrates modern MLOps tools, cloud infrastructure, deep learning architectures, a scalable preprocessing pipeline, and a conversational LLM interface. The final product is exposed via a web application that accepts user input and returns a human-readable forecast summary.

Access the live application here:
👉 https://ai-weather-predictor.streamlit.app/

---

Project Objective

The objective of this project is to demonstrate the full machine learning lifecycle — from data collection and model development to cloud deployment and user interface — within a real-world forecasting problem. The system is designed for scalability, reproducibility, and ease of use, fulfilling all course and industry-standard best practices.

---

Problem Statement

Accurate weather forecasting is critical for planning in agriculture, logistics, travel, and disaster management. Traditional weather models often lack city-level personalization and interpretability for the general public. This project aims to fill that gap by:

• Aggregating historical weather data from multiple global stations
• Building a deep learning model capable of multi-day, multi-feature forecasting
• Enabling real-time, on-demand predictions for any global city
• Providing human-readable summaries using large language models (LLMs)

---

Key Features

Component	Description
Data Source	Historical weather data (2010–2024) via Meteostat API
Preprocessing	Time enrichment, missing value imputation, engineered weather metrics
Model	Conv1D + GRU deep learning sequence-to-sequence model
Experiment Tracking	MLflow with parameter logging, metric tracking, and artifact registry
Deployment	Streamlit web application with real-time inference and LLM integration
Monitoring	Google Sheets logging of search history, evaluation metrics visualization
Cloud Integration	Terraform-provisioned GCP infrastructure (GCS + BigQuery)

---

Project Architecture

User Input
    ↓
Streamlit Web App
    ↓
City Geolocation + Meteostat Data Fetch
    ↓
Preprocessing (feature + sequence generation)
    ↓
Model Inference (Keras)
    ↓
GPT-4o-mini Summary Generation
    ↓
Forecast Table + Text Output
    ↓
Google Sheets Logging

---

Technologies Used

Category	Tools / Libraries
ML/DL	TensorFlow, Keras, NumPy, Scikit-learn
Data Collection	Meteostat, Geopy, OpenCage API
Feature Engineering	Custom Python scripts
Experiment Tracking	MLflow
Web Interface	Streamlit
LLM Summarization	OpenAI GPT-4o-mini
Infrastructure as Code	Terraform
Cloud Platform	Google Cloud Platform (GCS, BigQuery)
Logging	Google Sheets API (gspread)
Deployment Tools	Streamlit Cloud, Docker (future scope)

---

Setup Instructions

1. Environment Setup

conda create -n weather-forecast python=3.10
conda activate weather-forecast
pip install -r requirements.txt

2. Download Raw Data

Use the built-in script to download historical weather data:

from data_collection import data_downloader
data = data_downloader()
data.to_csv("data/raw_data.csv", index=False)

3. Preprocess Data

from preprocessing import main
import pandas as pd

raw_data = pd.read_csv("data/raw_data.csv")
X, y = main(raw_data)

4. Train Model

from model_training import train_and_save_model
train_and_save_model()

5. Evaluate Model

from evaluation import evaluate_model
evaluate_model()

6. Deploy App

streamlit run app.py

---

Reproducibility

• All model files, training scripts, and preprocessing steps are versioned
• Dependencies are documented in requirements.txt
• MLflow logs all metrics and model artifacts
• Terraform scripts provision identical GCP environments
• Data saved in GCS for persistent access

---

Cloud Infrastructure

Using Terraform, the following components are provisioned on Google Cloud:

• A Cloud Storage Bucket with folders for:

  • raw_data/
  • features/
  • target/

• A BigQuery Dataset for future analytics workflows

• Configurable via variables.tf and securely authenticated using a service account JSON key

---

Model Performance

• Input shape: [30, 21] → Output shape: [7, 6]

• Forecast features: tavg, tmin, tmax, wspd, prcp, snow

• Metrics:

  • Mean Squared Error (MSE)
  • Mean Absolute Error (MAE)

• Visual comparison of ground truth vs prediction available in plots

---

Experiment Tracking

• Managed using MLflow (/kaggle/working/mlflow)
• Parameters: epochs, batch size, input/output shapes
• Metrics: loss, val_loss, mae, val_mae
• Model artifacts saved and zipped for portability

---

Web Application Highlights

• Users input a city name
• System fetches real-time weather data and makes 7-day predictions
• GPT-4o-mini provides an intuitive weekly forecast summary
• City searches are logged via Google Sheets for tracking and usage analysis
• Access the live application here: 👉 https://ai-weather-predictor.streamlit.app/

---

CI/CD & Best Practices

• Terraform for infrastructure-as-code
• MLflow for experiment reproducibility
• Modular code design (separate scripts for each pipeline stage)
• Logging of model performance and user queries
• Compatible with Docker and CI/CD tools (future extension)

---

Future Work

• Integrate CI/CD with GitHub Actions
• Enable Docker containerization and deployment to GCP App Engine
• Add automated retraining if model drift detected
• Expand monitoring with tools like Evidently or WhyLogs

---

License

This project is licensed under the MIT License.

---