🎓 Bachelor's Thesis - Time series forecasting as image reconstruction using U-Net

Thesis Title: Forecasting come Image Reconstruction: un approccio basato sull'architettura U-Net
Degree: BSc in Statistical Sciences for Big Data University: Università degli Studi di Salerno Academic Year: 2023/2024

---

📌 Project Overview

This thesis proposes a novel approach to time series forecasting by reframing it as an image reconstruction task using deep learning. Instead of traditional numeric-based forecasting models, this work leverages image-based methods by converting time series data into 2D representations.

---

🔬 Methodology

1. Data Transformation
   Time series are converted into images using:

   • GASF (Gramian Angular Summation Field)
   • GADF (Gramian Angular Difference Field)
   • MTF (Markov Transition Field)

2. Model Architecture

   • Custom U-Net models were trained for each image representation.
   • Models were designed to perform forecasting as image reconstruction.

3. Comparison Baseline

   • A traditional autoregressive (AR) statistical model was implemented as a benchmark.

---

📊 Evaluation Metrics

• Forecasting Accuracy:
  • Mean Squared Error (MSE)
  • Mean Absolute Error (MAE)
• Image Quality:
  • Peak Signal-to-Noise Ratio (PSNR)
  • Structural Similarity Index Measure (SSIM)

---

🏆 Key Results

• The GASF-based U-Net model outperformed other deep learning models in terms of image quality.
• The AR model was superior in pure forecasting accuracy (lower MSE and MAE).
• The best model depends on the target: visual quality vs. numerical accuracy.

---

🚀 Future Work

The thesis outlines several directions for improving deep learning performance:

• Enlarging the dataset
• Reducing prediction windows
• Exploring advanced architectures like Transformers and Diffusion Models

---