This repository includes the code that was developed for my master's thesis, Novel Deep Learning Strategies for Time Series Forecasting, in which I explore how novel deep-learning strategies can be used for time series forecasting. The code was developed during the autumn of 2023 as a part of the course TKP4680 Chemical Engineering, Specialization Project and in the spring of 2024 for the course TKP4900 Chemical Process Technology, Master's Thesis at the Norwegian University of Science and Technology. The documents written for the specialisation project and the master's thesis are included in the Thesis folder.
The work involved understanding state-of-the-art models and developing and testing novel strategies throughout four case studies. The case studies performed during this work are the following:
- 〽️ Lotka Volterra system (Applied Mathematics): Synthetic data from an LV system is often used to analyse nonlinear, dynamic systems.
- 🐇 Hare and Lynx system (Ecology): Real-world data from the Hare and Lynx ecosystem recorded through 80 consecutive years.
- 💵 Nasdaq Composite (Finance): Real-world historical data of the Nasdaq Composite.
- 🏎️ F1 Telemetry (Engineering): Real-world data from a high-performance engineering system - specifically, the telemetry from a Formula One race
- Augmented Neural Ordinary Differential Equations using Expanding Horizon (ANODE-EH)
- Augmented Neural Ordinary Differential Equations using Multiple Shooting (ANODE-MS I + II)
- Modified Neural Ordinary Differential Equations using Multiple Shooting (MNODE-MS)
- Neural Simulation Error Method (NPEM) (Note! Named NPEM from Prediction Error, but is in fact a Simulation Error Method)
The folder structure of the repository is as follows: