Ruchi Choudhary

Zack Xuereb Conti, Ruchi Choudhary, Luca Magri

State-of-the-art machine-learning-based models are a popular choice for modeling and forecasting energy behavior in buildings because given enough data, they are good at finding spatiotemporal patterns and structures even in scenarios where the complexity prohibits analytical descriptions. However, their architecture typically does not hold physical correspondence to mechanistic structures linked with governing physical phenomena. As a result, their ability to successfully generalize for unobserved timesteps depends on the representativeness of the dynamics underlying the observed system in the data, which is difficult to guarantee in real-world engineering problems such as control and energy management in digital twins. In response, we present a framework that combines lumped-parameter models in the form of linear time-invariant (LTI) state-space models (SSMs) with unsupervised reduced-order modeling in a subspace-based domain adaptation (SDA) framework. SDA is a type of transfer-learning (TL) technique, typically adopted for exploiting labeled data from one domain to predict in a different but related target domain for which labeled data is limited. We introduce a novel SDA approach where instead of labeled data, we leverage the geometric structure of the LTI SSM governed by well-known heat transfer ordinary differential equations to forecast for unobserved timesteps beyond observed measurement data. Fundamentally, our approach geometrically aligns the physics-derived and data-derived embedded subspaces closer together. In this initial exploration, we evaluate the physics-based SDA framework on a demonstrative heat conduction scenario by varying the thermophysical properties of the source and target systems to demonstrate the transferability of mechanistic models from a physics-based domain to a data domain.

Zack Xuereb Conti, Rebecca Ward, Ruchi Choudhary

This paper presents an approach based on higher order dynamic mode decomposition (HODMD) to model, analyse, and forecast energy behaviour in an urban agriculture farm situated in a retrofitted London underground tunnel, where observed measurements are influenced by noisy and occasionally transient conditions. HODMD is a data-driven reduced order modelling method typically used to analyse and predict highly noisy and complex flows in fluid dynamics or any type of complex data from dynamical systems. HODMD is a recent extension of the classical dynamic mode decomposition method (DMD), customised to handle scenarios where the spectral complexity underlying the measurement data is higher than its spatial complexity, such as is the environmental behaviour of the farm. HODMD decomposes temporal data as a linear expansion of physically-meaningful DMD-modes in a semi-automatic approach, using a time-delay embedded approach. We apply HODMD to three seasonal scenarios using real data measured by sensors located at at the cross-sectional centre of the the underground farm. Through the study we revealed three physically-interpretable mode pairs that govern the environmental behaviour at the centre of the farm, consistently across environmental scenarios. Subsequently, we demonstrate how we can reconstruct the fundamental structure of the observed time-series using only these modes, and forecast for three days ahead, as one, compact and interpretable reduced-order model. We find HODMD to serve as a robust, semi-automatic modelling alternative for predictive modelling in Digital Twins.

Vijja Wichitwechkarn, Charles Fox, Ruchi Choudhary

Foundation models for natural language processing have many coherent definitions of hallucination and methods for its detection and mitigation. However, analogous definitions and methods do not exist for multi-variate time-series (MVTS) foundation models. We propose new definitions for MVTS hallucination, along with new detection and mitigation methods using a diffusion model to estimate hallucination levels. We derive relational datasets from popular time-series datasets to benchmark these relational hallucination levels. Using these definitions and models, we find that open-source pre-trained MVTS imputation foundation models relationally hallucinate on average up to 59.5% as much as a weak baseline. The proposed mitigation method reduces this by up to 47.7% for these models. The definition and methods may improve adoption and safe usage of MVTS foundation models.

Max Langtry, Vijja Wichitwechkarn, Rebecca Ward et al.

Data is required to develop forecasting models for use in Model Predictive Control (MPC) schemes in building energy systems. However, data is costly to both collect and exploit. Determining cost optimal data usage strategies requires understanding of the forecast accuracy and resulting MPC operational performance it enables. This study investigates the performance of both simple and state-of-the-art machine learning prediction models for MPC in multi-building energy systems using a simulated case study with historic building energy data. The impact on forecast accuracy of measures to improve model data efficiency are quantified, specifically for: reuse of prediction models, reduction of training data duration, reduction of model data features, and online model training. A simple linear multi-layer perceptron model is shown to provide equivalent forecast accuracy to state-of-the-art models, with greater data efficiency and generalisability. The use of more than 2 years of training data for load prediction models provided no significant improvement in forecast accuracy. Forecast accuracy and data efficiency were improved simultaneously by using change-point analysis to screen training data. Reused models and those trained with 3 months of data had on average 10% higher error than baseline, indicating that deploying MPC systems without prior data collection may be economic.

Fabian Raisch, Max Langtry, Felix Koch et al.

Transfer Learning (TL) is currently the most effective approach for modeling building thermal dynamics when only limited data are available. TL uses a pretrained model that is fine-tuned to a specific target building. However, it remains unclear how to proceed after initial fine-tuning, as more operational measurement data are collected over time. This challenge becomes even more complex when the dynamics of the building change, for example, after a retrofit or a change in occupancy. In Machine Learning literature, Continual Learning (CL) methods are used to update models of changing systems. TL approaches can also address this challenge by reusing the pretrained model at each update step and fine-tuning it with new measurement data. A comprehensive study on how to incorporate new measurement data over time to improve prediction accuracy and address the challenges of concept drifts (changes in dynamics) for building thermal dynamics is still missing. Therefore, this study compares several CL and TL strategies, as well as a model trained from scratch, for thermal dynamics modeling during building operation. The methods are evaluated using 5--7 years of simulated data representative of single-family houses in Central Europe, including scenarios with concept drifts from retrofits and changes in occupancy. We propose a CL strategy (Seasonal Memory Learning) that provides greater accuracy improvements than existing CL and TL methods, while maintaining low computational effort. SML outperformed the benchmark of initial fine-tuning by 28.1\% without concept drifts and 34.9\% with concept drifts.