[Paper Review] Utilizing Domain Knowledge: Robust Machine Learning for Building Energy Prediction with Small, Inconsistent Datasets

Xia Chen, Manav Mahan Sing+1 more|arXiv (Cornell University)|Jan 23, 2023

Energy Efficiency and Management4 citations

TL;DR

This paper proposes Component-Based Machine Learning (CBML), a knowledge-embedded approach that integrates domain-specific building energy principles into machine learning models to enhance robustness and data efficiency. By encoding structural and physical knowledge into model architecture, CBML achieves superior performance on small, inconsistent datasets—reducing training time, improving interpretability, and enabling effective use of sparse data with up to 98.75% less data than standard methods.

ABSTRACT

The demand for a huge amount of data for machine learning (ML) applications is currently a bottleneck in an empirically dominated field. We propose a method to combine prior knowledge with data-driven methods to significantly reduce their data dependency. In this study, component-based machine learning (CBML) as the knowledge-encoded data-driven method is examined in the context of energy-efficient building engineering. It encodes the abstraction of building structural knowledge as semantic information in the model organization. We design a case experiment to understand the efficacy of knowledge-encoded ML in sparse data input (1% - 0.0125% sampling rate). The result reveals its three advanced features compared with pure ML methods: 1. Significant improvement in the robustness of ML to extremely small-size and inconsistent datasets; 2. Efficient data utilization from different entities' record collections; 3. Characteristics of accepting incomplete data with high interpretability and reduced training time. All these features provide a promising path to alleviating the deployment bottleneck of data-intensive methods and contribute to efficient real-world data usage. Moreover, four necessary prerequisites are summarized in this study that ensures the target scenario benefits by combining prior knowledge and ML generalization.

Motivation & Objective

• To address the critical bottleneck of data dependency in machine learning for building energy prediction.
• To develop a machine learning framework that remains robust under extremely small and inconsistent datasets.
• To enable efficient utilization of fragmented, incomplete data from diverse building sources.
• To improve model interpretability and reduce training time through embedded domain knowledge.
• To identify and validate the prerequisites for successful integration of domain knowledge with data-driven models.

Proposed method

• Component-Based Machine Learning (CBML) is introduced as a knowledge-encoded ML framework that embeds building energy system abstractions into model structure.
• Domain knowledge—such as thermal dynamics, HVAC behavior, and building envelope properties—is encoded as semantic constraints within the model architecture.
• The method uses a modular design where each component represents a physical building system, with predefined relationships and constraints.
• CBML leverages a hybrid training strategy: supervised learning on available data combined with physics-based regularization to enforce domain consistency.
• The model is trained on highly subsampled datasets (1% to 0.0125% of full data) to simulate real-world data scarcity.
• Model performance is evaluated using standard regression metrics (e.g., RMSE, R²) under extreme data sparsity and inconsistency.

Experimental results

Research questions

• RQ1 Can a machine learning model trained on 1% or less of typical building energy data achieve reliable predictions?
• RQ2 How does embedding domain knowledge improve model robustness to inconsistent and sparse data?
• RQ3 To what extent does knowledge integration reduce training time and improve interpretability in energy prediction models?
• RQ4 What are the essential prerequisites for successful fusion of domain knowledge and machine learning in building energy applications?
• RQ5 How does CBML compare to pure data-driven models in terms of generalization and data efficiency?

Key findings

• CBML achieved significant improvements in prediction robustness on datasets sampled at 0.0125% of full size, outperforming pure ML models under extreme data scarcity.
• The model demonstrated efficient data utilization across heterogeneous data sources, maintaining high performance even with incomplete or inconsistent records.
• Training time was substantially reduced due to constrained search space from embedded domain knowledge, enabling faster convergence.
• The model exhibited high interpretability, as each component’s behavior aligned with known physical principles, enabling traceable decision-making.
• Four key prerequisites were identified: (1) modular model design, (2) semantic encoding of domain rules, (3) data heterogeneity tolerance, and (4) consistent physical consistency enforcement.
• The method maintained strong predictive performance (e.g., low RMSE) even when data were highly sparse, demonstrating resilience to data quality issues.