Robust Hourly Cooling Load Prediction Based On The Historical Weather and Load Data
1
ROBUST HOURLY COOLING LOAD PREDICTION BASED ON THE HISTORICAL WEATHER AND LOAD DATA ABSTRACT This study presents a new prediction scheme for optimal control of cooling system with thermal energy storage and better energy management of a smart grid system. The proposed prediction scheme could provide up to 48 hours ahead accurate cooling load prediction. The proposed models are hourly indexed. Different sets of parameters are derived for cooling load prediction in each hour of day. The motivation for hourly indexing is based on the effect of different temperature trend and buildings utilization on the cooling load between each hour. The robust regression approach incorporates fast least trimmed squares algorithm for possible outlier detection and increased accuracy of the models. This study can enhance the prediction accuracy of cooling load for better scheduling and planning of the cooling facilities. Yin Guo 1 and Ehsan Nazarian 2 1 Facilities Management & Planning 2 Industrial & Management Systems Engineering University of Nebraska-Lincoln MOTIVATION AND BACKGROUND • An accurate cooling load prediction model is essential for - Optimal control of cooling system - Smart grid system integration • Lack of generic models for both large building complexes and individual buildings Jeonghan Ko Industrial and Operations Engineering The University of Michigan, Ann Arbor Kamlakar Rajurkar Mechanical & Materials Engineering University of Nebraska- Lincoln FACULTY ADVISORS OBJECTIVE • To develop fast and accurate robust regression based prediction scheme - incorporating historical weather characteristics and buildings occupancy and usage patterns - with outlier detection capability APPROACH • Autoregressive models with exogenous inputs models (ARX) that combine - Historical cooling load and weather data - Future weather prediction • Hourly indexing of ARX models to incorporate the similarity of load profiles of the same hour from day to day (24 ARX models, one for each hour) • Differentiation of weekend and weekday models to reflect building occupancy and usage patterns MODELING • Fast Least Trimmed Square algorithm for coefficients estimation and outlier detection - Recursive h-subset selection of n data points - Least Square estimation of coefficients - Arrangement of data points based on ascending residuals - Selection of h-subset of n data points with smallest residuals - Repetition until convergence - Global minimum sum of squared residuals for the converged model PARAMETER ESTIMATION ARX MODELS 1-HOUR AHEAD PREDICTION 24-HOUR AHEAD PREDICTION CL t : Cooling load at time t T Max : The maximum temperature of day T t : temperature at time t T Min : The minimum temperature of day RH t : Relative humidity at time t • 1-hour ahead prediction models are the same as up to 6-hour ahead prediction models ASSUMPTIONS • t-24 is defined by the day of week - Monday: 24-hour ahead refers to Friday - Tuesday to Friday: 24-hour ahead refers to the day before - Saturday: 24-hour ahead refers to the last Sunday - Sunday: 24-hour ahead refers to the day before NUMERICAL STUDIES • UNL City campus hourly cooling load and weather data: 2010-2012 • Analysis period: April-October • Learning data: 2010-2011 • Testing data: 2012 RESULTS 1-HOUR AHEAD PREDICTION FOR SEPTEMBER 2012 1 38 75 112149186223260297334371408445 0 4,000 8,000 12,000 Actual Load Conventional LS Estimate Robust fast-LTS Estimate Hour Cooling Load (Tons/hour) 24-HOUR AHEAD PREDICTION FOR SEPTEMBER 2012 1 38 75 112149186223260297334371408445 0 4,000 8,000 12,000 Actual Load Conventional LS Estimate Robust fast-LTS Estimate Hour Cooling Load (Tons/hour) CONCLUSIONS • ARX model’s integration of autoregressive and regression models enhanced the accuracy in prediction. • Robust regression approaches could detect outliers and decrease the prediction error by 7%. SPONSORS AND SUPPORTING ORGANIZATIONS • Nebraska Center for Energy Science Research • Nebraska Utility Corporation • Lincoln Electric System (LES) PARTICIPATING UNL UNITS • Industrial & Management Systems Engineering • Construction Management • Facility Management & Planning 24 24 25 , 24 24 , 2 24 23 , 24 22 , 24 21 , 2 20 , 2 19 , 2 2 18 , 2 17 , 2 2 16 , 2 15 , 2 14 , 1 13 , 1 12 , 1 1 11 , 1 10 , 2 1 9 , 1 8 , 1 7 , 6 , 5 , 2 4 , 3 , 2 , 1 , 0 , t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t Min t Max t t t RH T a RH a T a T a CL a RH RH a T T a RH T a RH a T a T a CL a RH RH a T T a RH T a RH a T a T a CL a RH T a RH a T a T a T a T a a CL 168 168 16 , 168 15 , 2 168 14 , 168 13 , 168 12 , 24 24 11 , 24 10 , 2 24 9 , 24 8 , 24 7 , 6 , 5 , 2 4 , 3 , 2 , 1 , 0 , t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t Min t Max t t t RH T a RH a T a T a CL a RH T a RH a T a T a CL a RH T a RH a T a T a T a T a a CL
description
Robust Hourly Cooling Load Prediction Based On The Historical Weather and Load Data. Participating UNL Units Industrial & Management Systems Engineering Construction Management Facility Management & Planning. Sponsors and Supporting Organizations - PowerPoint PPT Presentation
Transcript of Robust Hourly Cooling Load Prediction Based On The Historical Weather and Load Data
ROBUST HOURLY COOLING LOAD PREDICTION BASED ON THE HISTORICAL WEATHER AND LOAD
DATA
ABSTRACTThis study presents a new prediction scheme for optimal control of cooling system with thermal energy storage and better energy management of a smart grid system. The proposed prediction scheme could provide up to 48 hours ahead accurate cooling load prediction. The proposed models are hourly indexed. Different sets of parameters are derived for cooling load prediction in each hour of day. The motivation for hourly indexing is based on the effect of different temperature trend and buildings utilization on the cooling load between each hour. The robust regression approach incorporates fast least trimmed squares algorithm for possible outlier detection and increased accuracy of the models. This study can enhance the prediction accuracy of cooling load for better scheduling and planning of the cooling facilities.
Yin Guo1 and Ehsan Nazarian2
1Facilities Management & Planning2Industrial & Management Systems Engineering
University of Nebraska-Lincoln
MOTIVATION AND BACKGROUND• An accurate cooling load prediction model is essential for
- Optimal control of cooling system
- Smart grid system integration
• Lack of generic models for both large building complexes and individual buildings
Jeonghan KoIndustrial and Operations
Engineering The University of Michigan,
Ann Arbor
Kamlakar RajurkarMechanical & Materials
EngineeringUniversity of Nebraska-Lincoln
FACULTY ADVISORS
OBJECTIVE• To develop fast and accurate robust regression based prediction scheme
- incorporating historical weather characteristics and buildings occupancy and usage patterns
- with outlier detection capability
APPROACH
• Autoregressive models with exogenous inputs models (ARX) that combine
- Historical cooling load and weather data
- Future weather prediction
• Hourly indexing of ARX models to incorporate the similarity of load profiles of the same hour from day to day (24 ARX models, one for each hour)
• Differentiation of weekend and weekday models to reflect building occupancy and usage patterns
MODELING
• Fast Least Trimmed Square algorithm for coefficients estimation and outlier detection
- Recursive h-subset selection of n data points
- Least Square estimation of coefficients
- Arrangement of data points based on ascending residuals
- Selection of h-subset of n data points with smallest residuals
- Repetition until convergence
- Global minimum sum of squared residuals for the converged model
PARAMETER ESTIMATION
ARX MODELS
1-HOUR AHEAD PREDICTION
24-HOUR AHEAD PREDICTION
242425,2424,22423,2422,2421,
220,219,2218,217,
2216,215,214,113,112,
1111,110,219,18,17,
6,5,2
4,3,2,1,0,
ttttttttttt
ttttttttttt
tttttttttttt
ttttttttttt
tttttttttMintMaxttt
RHTaRHaTaTaCLa
RHRHaTTaRHTaRHa
TaTaCLaRHRHaTTa
RHTaRHaTaTaCLa
RHTaRHaTaTaTaTaaCL
16816816,16815,216814,16813,16812,
242411,2410,2249,248,247,
6,5,2
4,3,2,1,0,
ttttttttttt
ttttttttttt
tttttttttMintMaxttt
RHTaRHaTaTaCLa
RHTaRHaTaTaCLa
RHTaRHaTaTaTaTaaCL
CLt: Cooling load at time t TMax: The maximum temperature of dayTt: temperature at time t TMin: The minimum temperature of day RHt: Relative humidity at time t
• 1-hour ahead prediction models are the same as up to 6-hour ahead prediction models
ASSUMPTIONS• t-24 is defined by the day of week
- Monday: 24-hour ahead refers to Friday
- Tuesday to Friday: 24-hour ahead refers to the day before
- Saturday: 24-hour ahead refers to the last Sunday
- Sunday: 24-hour ahead refers to the day before
NUMERICAL STUDIES
• UNL City campus hourly cooling load and weather data: 2010-2012
• Analysis period: April-October
• Learning data: 2010-2011
• Testing data: 2012
RESULTS
1-HOUR AHEAD PREDICTION FOR SEPTEMBER 2012
1 30 59 88 1171461752042332622913203493784074364650
2,000
4,000
6,000
8,000
10,000
12,000
14,000
Actual Load Conventional LS EstimateRobust fast-LTS Estimate
Hour
Cool
ing
Load
(Ton
s/ho
ur)
24-HOUR AHEAD PREDICTION FOR SEPTEMBER 2012
1 30 59 88 1171461752042332622913203493784074364650
2,000
4,000
6,000
8,000
10,000
12,000
14,000
Actual Load Conventional LS EstimateRobust fast-LTS Estimate
Hour
Cool
ing
Load
(Ton
s/ho
ur)
CONCLUSIONS• ARX model’s integration of autoregressive and regression models
enhanced the accuracy in prediction.
• Robust regression approaches could detect outliers and decrease the prediction error by 7%.
SPONSORS AND SUPPORTING ORGANIZATIONS
• Nebraska Center for Energy Science Research• Nebraska Utility Corporation• Lincoln Electric System (LES)
PARTICIPATING UNL UNITS• Industrial & Management Systems Engineering• Construction Management• Facility Management & Planning
