Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From RNNs to Foundation Models: An Empirical Study on Commercial Building Energy Consumption

Shourya Bose, Yijiang Li, Amy Van Sant, Yu Zhang, Kibaek Kim

TL;DR

The results show that dataset heterogeneity and model architecture have a greater impact on post-training forecasting performance than the parameter count, and despite the higher computational cost, fine-tuned FMs demonstrate competitive performance compared to base models trained from scratch.

Abstract

Accurate short-term energy consumption forecasting for commercial buildings is crucial for smart grid operations. While smart meters and deep learning models enable forecasting using past data from multiple buildings, data heterogeneity from diverse buildings can reduce model performance. The impact of increasing dataset heterogeneity in time series forecasting, while keeping size and model constant, is understudied. We tackle this issue using the ComStock dataset, which provides synthetic energy consumption data for U.S. commercial buildings. Two curated subsets, identical in size and region but differing in building type diversity, are used to assess the performance of various time series forecasting models, including fine-tuned open-source foundation models (FMs). The results show that dataset heterogeneity and model architecture have a greater impact on post-training forecasting performance than the parameter count. Moreover, despite the higher computational cost, fine-tuned FMs demonstrate competitive performance compared to base models trained from scratch.

From RNNs to Foundation Models: An Empirical Study on Commercial Building Energy Consumption

TL;DR

The results show that dataset heterogeneity and model architecture have a greater impact on post-training forecasting performance than the parameter count, and despite the higher computational cost, fine-tuned FMs demonstrate competitive performance compared to base models trained from scratch.

Abstract

Accurate short-term energy consumption forecasting for commercial buildings is crucial for smart grid operations. While smart meters and deep learning models enable forecasting using past data from multiple buildings, data heterogeneity from diverse buildings can reduce model performance. The impact of increasing dataset heterogeneity in time series forecasting, while keeping size and model constant, is understudied. We tackle this issue using the ComStock dataset, which provides synthetic energy consumption data for U.S. commercial buildings. Two curated subsets, identical in size and region but differing in building type diversity, are used to assess the performance of various time series forecasting models, including fine-tuned open-source foundation models (FMs). The results show that dataset heterogeneity and model architecture have a greater impact on post-training forecasting performance than the parameter count. Moreover, despite the higher computational cost, fine-tuned FMs demonstrate competitive performance compared to base models trained from scratch.

Paper Structure

This paper contains 10 sections, 1 equation, 5 figures, 12 tables.

Table of Contents

  1. Introduction
  2. ComStock Dataset
  3. Experiment Design
  4. Experiment Results
  5. Concluding Remarks
  6. Analysis of IL-HET and IL-HOM
  7. Correlation Analysis for Feature Selection
  8. Model Architectures
  9. Hyperparameter Tuning
  10. Predictions

Figures (5)

  • Figure 1: (a) Distribution of building type in IL-HET. (b) Boxplot of the spread of load values for each building type in IL-HET. (c) Load shapes from two buildings each from IL-HET and IL-HOM.
  • Figure 2: Test set performance of base models for $L=512$ and (a) $T=4$, (b) $T=48$, (c) $T=96$.
  • Figure 3: (a) Distribution of building type in IL-HOM. (b) Boxplot of the spread of values in the time series of each building type in IL-HOM.
  • Figure 4: Training instability in Transformer training on IL-HET compared to IL-HOM.
  • Figure 5: Predictions of the test set loads for two selected buildings from IL-HOM and IL-HET datasets for $T=512, L=48$, respectively, for the best models reported in Table \ref{['table:features']}.