Challenges in Deep Learning Based Forecasting of Time Series with Calendar-driven Periodicities

Throughout history, forecasting has been crucial across various civilizations, influencing imagination and critical decisions. In contemporary times, it has become indispensable across diverse domains, where accurate predictions drive decision-making --- e.g., in the operation of the electrical grid or traffic planning. However, forecasting is still associated with various challenges. Thus, the present thesis delves into neural network-based time series forecasting challenges, focusing on four aspects: missing data, missing scenarios, concept drifts, and periodicities.

This thesis aims to solve these challenges using conditional generative models or profile-based methods. Conditional generative models are the main component of the proposed approaches to face the missing scenario challenge and the challenge of a small training data size. Moreover, a conditional generative model is also the main component of the proposed forecaster used for energy peak load forecasting in the BigDEAL challenge.
The proposed profile-based methods are promising in handling concept drift and support neural networks in time series forecasting. In particular, they can be supportive since they cover specific aspects of the time series that need not be captured by the machine learning model anymore if the profiles are integrated with the machine learning models. ... mehrConsequently, the proposed Profile Neural Network (PNN) and Probabilistic Profile Neural Network (ProbPNN) can beat state-of-the-art forecasting models on two datasets regarding the forecast quality and the computational time.