Time Series Forecasting with LSTM Neural Networks: An Air-Quality Case Study in Lombardy, Italy

Overview

The revolution of machine and deep learning models has powered remarkable advancements in image and text generation & identification. While these fields have been rapidly transformed, the realm of time series data, traditionally spearheaded by statistical learning, lags behind. This repository dives deep into the exploration of Long Short-Term Memory (LSTM) neural networks—a tool predominantly utilized in text classification and generation—as a viable tool for time series forecasting, focusing on an air-quality case study in Lombardy, Italy.

Highlights

• LSTM Configurations: A deep dive into multiple configurations of the LSTM model, dissecting their performance, complexity, and interpretability.

• Hyperparameter Tuning: Comprehensive exploration of the hyperparameter space across datasets from various geographic locations. This leads to insights into the best-performing hyperparameters.

• Interpretable Deep Learning: Despite the inherent black-box nature of deep learning models, this study evaluates feature importance at a granular level using Lime (local interpretable model-agnostic explanations).

• Raw Data Exploration: The data, derived from a plethora of weather and air-quality stations in Lombardy, undergoes a rigorous exploratory analysis. This includes addressing discrepancies in variables measured by different stations and handling missing values. Furthermore, an innovative approach ties weather and air-quality stations based on their geographic proximity.

Report Structure (available here)

1. Introduction: Setting the stage with the necessary theoretical foundations and methods.

2. Explorative Data Analysis: Delve into the raw data, understanding its structure, discrepancies, and missing values.

3. Modeling & Results: Detailed presentations on the LSTM configurations, hyperparameter exploration, and the final outcomes.

4. Interpretability Study: A closer look at how Lime assists in making sense of the LSTM model's predictions.

5. Conclusions & Future Work: Reflect on the insights gained and propose directions for further research and development.