LSTM-Based Temperature Forecasting with STM32 Deployment

Train a multivariate LSTM to predict temperature using weather data, and deploy it on resource-constrained STM32 microcontrollers.

Project goal – Use 10 years of hourly weather data to train a deep learning model that can predict temperature without requiring a physical sensor. The final model is optimized and converted to TensorFlow Lite format for embedded deployment.

---

🧠 Model Overview

This project uses an LSTM (Long Short-Term Memory) neural network trained on 8 weather features (e.g., humidity, pressure, wind) to predict temperature one hour into the future. Input is based on a one-week rolling window (168 time steps × 8 features).

---

📁 Folder structure — top level

.
├── fetch_data.py             ← fetches Meteostat data and saves as CSV
├── temperature_data.csv      ← full 10-year dataset (2015–2025, Hyderabad)
├── process_data.py           ← normalizes and splits data, saves scaler
├── scaler_params.txt         ← saved MinMaxScaler values
├── lstm_model.py             ← builds, trains, and evaluates the LSTM model
├── convert_model.py          ← converts model to TensorFlow Lite format
├── LICENSE                   ← MIT License             
├── README.md                 ← this file

---

🔧 Installation

git clone https://github.com/barnav1/PS-I
cd PS-I
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

Or install manually:

pip install pandas numpy matplotlib scikit-learn tensorflow meteostat

Python 3.8+ recommended.

---

🚀 Running the Pipeline

Step 1: Fetch Historical Weather Data

python fetch_data.py

Downloads hourly weather data from Meteostat API for Hyderabad (2015–2025).

Step 2: Preprocess the Data

python process_data.py

• Filters and normalizes the 8 numerical features
• Saves scaler parameters to scaler_params.txt

Step 3: Train the LSTM Model

python lstm_model.py

• Splits data into train/val/test
• Trains an LSTM model with early stopping
• Plots loss curves and saves main.keras

Step 4: Convert to TensorFlow Lite

python convert_model.py

• Converts .keras to .tflite
• Saves final model as converted_model.tflite

---

📉 Model Performance

• NRMSD: ~2% on the test set
• Model captures trends over a 168-hour window with high precision
• Inference is lightweight and suitable for CPU-only deployment

---

🧪 Features Used

• temp: air temperature (target)
• dwpt: dew point
• rhum: relative humidity
• prcp: precipitation
• wdir: wind direction
• wspd: wind speed
• pres: pressure
• coco: weather condition code

---

📊 Dependencies

• tensorflow
• pandas
• numpy
• matplotlib
• scikit-learn
• meteostat

---

📝 Acknowledgments

• Meteostat API for free access to historical weather data
• TensorFlow & Keras for model building
• STM32Cube.AI for ML-on-MCU deployment tools

MIT License – see LICENSE for details.

---