Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond explaining: XAI-based Adaptive Learning with SHAP Clustering for Energy Consumption Prediction

Tobias Clement, Hung Truong Thanh Nguyen, Nils Kemmerzell, Mohamed Abdelaal, Davor Stjelja

TL;DR

This work addresses robust energy consumption prediction for buildings under distribution shifts by integrating explainable AI with adaptive learning. The core method, SCAL, combines SHAP-based explanations, dimensionality reduction, and clustering to form an explanation space, from which clustering characteristics guide iterative model refinement. Three building blocks—SHAP clustering in explanation space, extraction of clustering characteristics, and adaptive model refinement—enable interpretable, Sharper-generalizing models across regression and classification tasks. Empirical results on building energy data, plus transferability tests on Financial Distress and Power datasets, show improved test performance and richer explanations, with the approach offering practical benefits for robustness and anomaly detection in real-world deployments.

Abstract

This paper presents an approach integrating explainable artificial intelligence (XAI) techniques with adaptive learning to enhance energy consumption prediction models, with a focus on handling data distribution shifts. Leveraging SHAP clustering, our method provides interpretable explanations for model predictions and uses these insights to adaptively refine the model, balancing model complexity with predictive performance. We introduce a three-stage process: (1) obtaining SHAP values to explain model predictions, (2) clustering SHAP values to identify distinct patterns and outliers, and (3) refining the model based on the derived SHAP clustering characteristics. Our approach mitigates overfitting and ensures robustness in handling data distribution shifts. We evaluate our method on a comprehensive dataset comprising energy consumption records of buildings, as well as two additional datasets to assess the transferability of our approach to other domains, regression, and classification problems. Our experiments demonstrate the effectiveness of our approach in both task types, resulting in improved predictive performance and interpretable model explanations.

Beyond explaining: XAI-based Adaptive Learning with SHAP Clustering for Energy Consumption Prediction

TL;DR

This work addresses robust energy consumption prediction for buildings under distribution shifts by integrating explainable AI with adaptive learning. The core method, SCAL, combines SHAP-based explanations, dimensionality reduction, and clustering to form an explanation space, from which clustering characteristics guide iterative model refinement. Three building blocks—SHAP clustering in explanation space, extraction of clustering characteristics, and adaptive model refinement—enable interpretable, Sharper-generalizing models across regression and classification tasks. Empirical results on building energy data, plus transferability tests on Financial Distress and Power datasets, show improved test performance and richer explanations, with the approach offering practical benefits for robustness and anomaly detection in real-world deployments.

Abstract

This paper presents an approach integrating explainable artificial intelligence (XAI) techniques with adaptive learning to enhance energy consumption prediction models, with a focus on handling data distribution shifts. Leveraging SHAP clustering, our method provides interpretable explanations for model predictions and uses these insights to adaptively refine the model, balancing model complexity with predictive performance. We introduce a three-stage process: (1) obtaining SHAP values to explain model predictions, (2) clustering SHAP values to identify distinct patterns and outliers, and (3) refining the model based on the derived SHAP clustering characteristics. Our approach mitigates overfitting and ensures robustness in handling data distribution shifts. We evaluate our method on a comprehensive dataset comprising energy consumption records of buildings, as well as two additional datasets to assess the transferability of our approach to other domains, regression, and classification problems. Our experiments demonstrate the effectiveness of our approach in both task types, resulting in improved predictive performance and interpretable model explanations.
Paper Structure (22 sections, 8 figures, 3 tables, 1 algorithm)

This paper contains 22 sections, 8 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Energy consumption prediction
  4. XAI-based model improvement
  5. SHapley Additive exPlanations (SHAP)
  6. Methodology
  7. Building Block 1: SHAP Clustering in Explanation Space
  8. Building Block 2: Extraction of SHAP Clustering Characteristics
  9. Building Block 3: Adaptive Model Refinement Based on SHAP Clustering Characteristics
  10. Experimental Setup and Data Set
  11. Results
  12. SCAL Performance
  13. Cluster Analysis in Explanation Space
  14. Model improvement cases
  15. Data error detection case
  16. ...and 7 more sections

Figures (8)

  • Figure 1: SCAL Pipeline: Adaptive Learning via SHAP Clustering in Three Building Blocks.
  • Figure 2: Overview of the three quality metrics: number of clusters ($M$), intra-inter distances between clusters, and the presence of noise cluster.
  • Figure 3: The distribution of energy consumption of a general educational institution's training and test set (building 1) reflects the significant growth in energy consumption after the COVID-19 pandemic.
  • Figure 4: The categories of buildings based on their similarity in the explanation space. Five buildings categories are clustered.
  • Figure 5: SHAP clusters embedding of AHT and SCAL on the four different buildings' training set. Each cluster goes with its index, where -1 indicates the noise cluster. SCAL improves the cluster's quality with a higher silhouette score (SS) and the presence of the noise cluster, where the noise cluster presents ($\textcolor{teal}{✓}$) and does not present (✗).
  • ...and 3 more figures