Advancements in time series analysis using deep learning

This doctoral thesis investigates advanced methodologies in deep learning for time series analysis, addressing challenges inherent to univariate and multivariate time series data across dynamic sectors. In response to the rapid growth of data generated by Internet-of-Things devices and sensors, this research introduces an innovative Adaptive Embedding approach tailored to dynamically adjust embedding dimensions and architecture configurations based on the characteristics of the input data. This technique significantly improves the predictability of complex time series data by optimizing the neural network’s structure for various forecasting horizons and data scenarios. This novel embedding procedure enhances model accuracy in predicting energy generation and demand within smart grids, addressing challenges in photovoltaic energy forecasting, and grid management. The research further explores embedding techniques that transform time series data into image formats, offering substantial improvements over traditional time series analysis methods. By leveraging convolutional and recurrent neural networks, comprehensive validation across several energy-focused applications is provided, including electric grid anomaly detection, energy theft identification, and community-level energy optimization. Results indicate that the proposed methods outperform conventional approaches in predictive performance and computational efficiency, offering scalable solutions for real-time applications. Overall, this thesis lays foundational advancements in time series embedding and provides practical insights for deploying deep learning in energy systems.