LSTM vs GRU — Time-Series Forecasting
A reproducible deep-learning experiment comparing LSTM and GRU recurrent neural networks for daily temperature time-series forecasting, with modular training, evaluation, testing, and documented results.
Problem
Sequential forecasting requires models that can learn temporal dependencies from previous observations. This project examines how two gated recurrent architectures—LSTM and GRU—perform on the same univariate temperature-forecasting task under a common experimental setup.
Approach
The project transforms daily temperature observations into 10-day lookback sequences, applies MinMax scaling, and trains comparable LSTM and GRU models using TensorFlow/Keras. Both models are evaluated with Mean Squared Error using the same train/test split and training configuration, with the experiment organized into modular data, model, training, evaluation, and visualization components.
Outcome / Learning
Both recurrent architectures captured the temporal pattern in the dataset, with the GRU producing a slightly lower test MSE in the recorded experiment: 0.007073 versus 0.007380 for the LSTM. The project provided hands-on experience with sequence construction, recurrent neural networks, preprocessing, controlled model comparison, evaluation, and structuring an academic notebook experiment as a reproducible ML repository.
Key Features
- LSTM and GRU recurrent neural-network comparison
- Daily temperature time-series forecasting
- 10-day sliding lookback sequence generation
- MinMax feature scaling
- 80/20 sequential train-test split
- Common training configuration for model comparison
- MSE-based train and test evaluation
- Training and validation loss analysis
- Modular Python ML pipeline
- Automated data tests and CI workflow
- Reproducible experiment documentation