Forecasting chaos: AI-enhanced prediction of indoor climate dynamics

This research explores the use of advanced deep learning models to analyse and predict indoor climate dynamics, specifically temperature and velocity, based on computational fluid dynamics (CFD) simulations in an air-conditioned room. Building on previous work that clustered sensor data into three representative zones, this study focuses on the temporal behaviour of these key positions. We compare Artificial Neural Networks (ANNs) and Long Short-Term Memory (LSTM) networks, demonstrating that ANNs can achieve prediction performance comparable to LSTMs while offering faster inference and the flexibility to operate without strictly sequential data. A key contribution of this study is the identification of semi-chaotic behaviour in indoor environments: in cases where certain variables exhibit partial order, this underlying structure can dominate over chaotic fluctuations, enabling confident and accurate prediction. This insight enhances the understanding of predictability in complex systems and has significant implications for both scientific modelling and engineering applications. By leveraging hybrid physics-informed approaches through the integration of CFD-based insights with data-driven modelling, this work offers scalable strategies for optimising intelligent sensor placement, heating, ventilation, and air conditioning (HVAC) systems, as well as predictive environmental control in smart building systems.