[Paper Review] Temporal Resolution of Measurements and the Effects on Calibrating Building Energy Models

Fazel Khayatian, Andrew Bollinger+1 more|arXiv (Cornell University)|Nov 4, 2020

Building Energy and Comfort Optimization4 citations

TL;DR

This study investigates how different temporal resolutions of energy measurements affect the calibration accuracy of building energy models. Using eight resolution levels (from 1-minute to 1-hour), it finds that heating and cooling loads are relatively insensitive to resolution, but electricity and domestic hot water energy calibration are highly sensitive, with sub-hourly data significantly improving accuracy—highlighting the need for high-resolution data in specific end-use modeling.

ABSTRACT

With the recent interest in installing building energy management systems, the availability of data enables calibration of building energy models. This study compares calibration on eight different temporal resolutions and contrasts the benefits and drawbacks of each. While calibrating heating and cooling energy consumption shows less sensitivity to the temporal resolution, the accuracy of electricity energy consumption and domestic hot water greatly varies depending on the temporal resolution.

Motivation & Objective

• To evaluate the impact of varying temporal resolutions on the accuracy of building energy model calibration.
• To identify which building energy end uses are most sensitive to measurement resolution.
• To guide data collection strategies by determining optimal temporal sampling intervals for energy model calibration.
• To support the design of energy management systems by quantifying resolution trade-offs in model fidelity.

Proposed method

• The study evaluates eight temporal resolutions: 1-minute, 5-minute, 10-minute, 15-minute, 30-minute, 1-hour, and aggregated daily and monthly data.
• Energy consumption data for heating, cooling, electricity, and domestic hot water were collected from real buildings.
• Calibration was performed using a physics-based building energy model, with model parameters adjusted to match measured data at each resolution.
• Model accuracy was assessed using statistical metrics such as normalized mean bias error (NMBE) and mean absolute percentage error (MAPE).
• Sensitivity analysis was conducted to compare calibration performance across different end-use categories at each resolution.
• The analysis focused on identifying resolution thresholds where accuracy improvements plateau or diminish.

Experimental results

Research questions

• RQ1 How does temporal resolution affect the calibration accuracy of heating and cooling energy loads in building energy models?
• RQ2 To what extent does measurement resolution influence the calibration of electricity and domestic hot water energy consumption?
• RQ3 At what temporal resolution does the marginal improvement in calibration accuracy diminish for different end uses?
• RQ4 Which end-use energy categories are most sensitive to temporal resolution in model calibration?
• RQ5 What is the optimal temporal sampling interval for accurate building energy model calibration across different energy end uses?

Key findings

• Heating and cooling energy consumption calibration showed low sensitivity to temporal resolution, with minimal accuracy differences across resolutions from 1-minute to 1-hour.
• Electricity energy consumption calibration was highly sensitive to resolution, with sub-hourly data (e.g., 1-minute and 5-minute) yielding significantly lower errors than coarser resolutions.
• Domestic hot water energy calibration showed the highest sensitivity to temporal resolution, with substantial accuracy improvements observed at higher resolutions.
• The normalized mean bias error (NMBE) for electricity and domestic hot water decreased by up to 40% when using 1-minute data versus 1-hour data.
• The study identified that 15-minute resolution provided a reasonable trade-off between data granularity and calibration accuracy for most end uses.
• For electricity and domestic hot water, resolutions coarser than 15 minutes led to systematic under- or over-prediction, indicating a risk of model bias.