The unprecedented digitization of present-day health care systems has resulted in an exponential growth in the quantity and variety of medical data, including electronic health records (EHRs), medical imaging, and output from wearable sensors and genomics dataset. These data sources offer a strong potential for enhancing clinical decision-making, personalized medicine, and predictive analytics.

Given the rising global demand as eventually growing complexity of modern power systems, accurate energy consumption forecasting has become one of fundamental requirements some efficient way to manage energy. Many of the complex, nonlinear and dynamic mechanisms that influence energy usage patterns on increasing levels can be simply due to the weather phenomenon, human behavior or economic activities (e.g. one needs predictive rules about electricity consumption), which Old forecasting approaches based on statistical and rule-based models alone cannot adequately capture [5].

Artificial Intelligence (AI) systems are increasingly deployed in dynamic real-world environments where data distributions evolve over time. This phenomenon, commonly referred to as data shift or dataset drift, poses a significant challenge to the reliability and robustness of machine learning models. When the training data distribution differs from the deployment environment, model performance can degrade, leading to inaccurate predictions and reduced trust in AI systems.

Concept drift is one of the biggest challenges in deploying deep learning systems for dynamic, real-world environments where data distributions evolve as time passes. Conventional deep learning models typically operate under the assumption of stationary distributions, meaning that they will perform poorly in instances where patterns change over time, such as shifts in user behavior, environmental changes due to noise or lighting conditions, and dynamics of system robustness.

Hierarchical Temporal Learning (HTL) has been at the forefront of a new paradigm for modelling complex time series comprised of multi-scale entangled temporal dependencies. Historical methods such as traditional machine learning and deep learning fail to characterise long-range dependencies or hierarchical structures that occur in real-world data (such as climate systems, financial markets, healthcare monitoring, smart infrastructure etc.

Indeed, as you may be aware, over the past few years neural networks have been applied to many different fields with such success that they heavily contribute to healthcare diagnostics, financial forecasting, autonomous systems and climate modelling. Nonetheless, the deployment of such models in high-variance environments — where data distributions are not only noisy and dynamic but often also unpredictable — poses serious threats to the reliability or trustworthiness of their predictions.