Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Automating Large-Scale Simulation Calibration to Real-World Sensor Data Richard E. Edwards Distributed Intelligence Lab Department of Electrical Engineering and Computer Science University of Tennessee, Knoxville TN, USA March 13, 2013 Funded by Whole Building & Community Integration Group, Oak Ridge National Laboratory, Oak Ridge TN, USA Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 1 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 2 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 3 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Simulation Modeling I Nuclear Power I Climate I Buildings I Physics Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 4 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Simulation Problems I Simulation Model vs Actual Structure I I I I Stated material properties vs Actual material properties Estimated duty cycle vs Actual usage Expected structure vs Built structure Simulation Limitations I I I Computer limitations Code-base maintenance Simulating new technologies Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 5 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Simulation Problems I Simulation Model vs Actual Structure I I I I Stated material properties vs Actual material properties Estimated duty cycle vs Actual usage Expected structure vs Built structure Simulation Limitations I I I Computer limitations Code-base maintenance Simulating new technologies Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 6 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Previous Simulation Calibration Methodologies I Two Overall Methodologies I I Manual procedural calibration Semi-Automated statistical calibration Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 7 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Previous Simulation Calibration Methodologies I Two Overall Methodologies I I Manual procedural calibration Semi-Automated statistical calibration Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 8 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Manual Procedural Calibration I Core Components I Physical Audit Period I I I Yoon and Lee (1999) Pendrini et al. (2002) Sensitivity Analysis I Westphal and Lamberts (2005) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 9 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Manual Procedural Calibration I Core Components I Physical Audit Period I I I Sensitivity Analysis I I Yoon and Lee (1999) Pendrini et al. (2002) Westphal and Lamberts (2005) Additional Components I Multi-phase Audit Procedure I Raftery et al. (2011) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 9 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Manual Procedural Calibration I Core Components I Physical Audit Period I I I Sensitivity Analysis I I Westphal and Lamberts (2005) Additional Components I Multi-phase Audit Procedure I I Yoon and Lee (1999) Pendrini et al. (2002) Raftery et al. (2011) Problems I I Entire process is slow Non-transferable Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 9 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Previous Simulation Calibration Methodologies I Two Overall Methodologies I I Manual procedural calibration Semi-Automated statistical calibration Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 10 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Semi-Automated Statistical Calibration I Core Components I Surrogates (Approximations) I I I I I I Ordinary Least Squares (OLS) Multivariate Adaptive Splines (MARS) Kriging Radial Basis Functions (RBF) Jin et al. 2001 Automatic Sensitivity Analysis (Feature Selection) I I OLS (Helton et al. 2006) MARS (Storlie et al. 2009) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 11 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Semi-Automated Statistical Calibration I Core Components I Surrogates (Approximations) I I I I I I Automatic Sensitivity Analysis (Feature Selection) I I I Ordinary Least Squares (OLS) Multivariate Adaptive Splines (MARS) Kriging Radial Basis Functions (RBF) Jin et al. 2001 OLS (Helton et al. 2006) MARS (Storlie et al. 2009) Problems I Scalability & Generality I I Tian and Choudhary (2012) – Macro-scale E+ simulation calibration (10 parameters, 1 output) Zeng et al. (2013) – Crop simulation calibration (100 parameters & ∼18 days ) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 11 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Problem Overview I Simulations require calibration I Manual method problems: I I I Slow Not transferable Semi-Automated method problems: I I Scalability Generality Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 12 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Select Best Sensors Simulation Parameters yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 13 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Step 1 Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Select Best Sensors Simulation Parameters yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 14 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Select Best Sensors Simulation Parameters Step 2 yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 15 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Step 3 Select Best Sensors Simulation Parameters yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 16 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Step 4 Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Select Best Sensors Simulation Parameters yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 17 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Step 5 Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Select Best Sensors Simulation Parameters yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 18 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview cument Calibration Diagram Calibration Signal Environmental Sensors Filter Sensor Data Or Relational Model Map sensors to Simulation Output Simulation Approximation Estimate Simulation Parameters Select Best Model Delay Step 6 Select Best Sensors Simulation Parameters yes no Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 19 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Summary I Step 1 — Select Best Prediction Model I Step 2 — Select Best Sensors (Step 1 result) I Step 3 — Map Sensors to Simulation Outputs (Step 2 result) I Step 4 — Build Simulation Approximation (Optional) I Step 5 — Build Simulation Relational Model I Step 6 — Calibrate Simulation (Step 3, 4, 5 results) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 20 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Contributions I Step 1 I Best predictor for hourly residential electrical consumption Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 21 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Contributions I Step 1 I I Best predictor for hourly residential electrical consumption Step 2 I I Best sensors for predicting electrical consumption Novel feature selection method Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 21 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Contributions I Step 1 I I Step 2 I I I Best predictor for hourly residential electrical consumption Best sensors for predicting electrical consumption Novel feature selection method Step 4 I Large-scale E+ residential approximation Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 21 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Contributions I Step 1 I I Step 2 I I I Best sensors for predicting electrical consumption Novel feature selection method Step 4 I I Best predictor for hourly residential electrical consumption Large-scale E+ residential approximation Step 5 I Large-scale regression structure learning Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 21 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 22 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion EnergyPlus (E+) I Requirements I I I Building model Operation schedule Weather data I Advantages I Disadvantages I I I Most general model Runtime 3∼7 minutes 300+ adjustable variables Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 23 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Sensor Data Sets I Two Subdivisions I I I Wolf Creek I I I Wolf Creek Campbell Creek approximately 250 sensors 15 minute resolution Campbell Creek I I approximately 140 sensors 15 minute resolution Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 24 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion E+ Simulation Data I Markov Order 1 (MO1) I I I Markov Order 2 (MO2) I I I Adjust parameters independently min,max adjustment Adjust two parameters together min,max adjustments Fine Grain (FG – Brute Force) I I Adjust 14 parameters Small incremental adjustments Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 25 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion E+ Simulation Data Markov Order 1 Markov Order 2 Fine Grain # Outputs 95 95 82 # Inputs 156 (151) 156 (151) 162 (14) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data # Simulations 299 29,727 11,989 Gigabytes 3.9 387.2 136.0 University of Tennessee 26 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Performance Metrics I Root Mean Squared Error(RMSE): v u N u 1 X RMSE = t (yi − pi )2 N −1 I Coefficient of Variance(CV): I Mean Absolute Percentage of Error(MAPE): MAPE = i=1 i=1 I CV = RMSE × 100 ymean Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data N 1 X |yi − pi | N yi Mean Bias Error(MBE): MBE = 1 N−1 PN i=1 (yi ymean − pi ) ×100 University of Tennessee 27 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 28 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Step 1 – Select Best Prediction Model I Problem – Identify the best model for predicting hourly residential electrical consumption I Data Sets: I I I ASHRAE Great Energy Prediction Shooutout Campbell Creek Explored: I 7 Different models Residential Model I Commercial Model I I I I Selected Least Squares Support Vector Machines (LS-SVM) Selected Feed Forward Neural Network (FFNN) Results Published: I R. E. Edwards, J. New, L. E. Parker, Predicting Future Hourly Residential Electrical Consumption: A Machine Learning Case Study, Energy and Buildings, vol 49, pages 591-603, June 2012 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 29 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Step 2 – Sensor Selection I Problem – Identify the best sensors for predicting hourly residential electrical consumption I Data Sets: I I Campbell Creek Explored: I 3 Feature Selection Methods with Linear Regression Model I I I I I Stepwise Information Complexity with Inverse Fisher Information Matrix (ICOMP(IFIM)) with Genetic Algorithm (GA) Novel voting method via estimating ICOMP(IFIM) distribution Bruteforce selection up to Best 4 Conclusions: I I I ICOMP(IFIM) with GA performs well Voting method is usually better than ICOMP(IFIM) with GA Best bruteforce set performs worse than selection methods Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 30 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 31 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Overview I Problem – Determine large-scale E+ surrogate effectiveness I Data sets: I I I I Approaches: I I I MO1 (4 Gigabytes) MO2 (10 Gigabytes) FG (10 Gigabytes) Feed Forward Neural Networks (FFNN) Lasso regression with Alternating Direction Method of Multipliers Layout: I I I I FFNN learning details and results Lasso regression learning details and results Discussion Summary Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 32 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Feed Forward Neural Network Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 33 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Network Structure I Number Hidden Units I I We explored 5, 10, 15 Number of Output Units: 10 I I I Output variables were grouped based on how they were stored Similar outputs were stored next to each other within the dataset Other pairings could produce different results Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 34 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Learning Process I Batch I I Stochastic I I Uses all examples to make an update Updates network per training example We use a hybrid approach I I Data is divided into mini-batches Network is updated per mini-batch Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 35 Preliminaries Step 1 & 2 5 5 Step 4 Step 5 & 6 x 10 5 Sensible RMSE Latent RMSE Sensible MTR Latent MTR 4 RMSE 5 Fine Grain Loads with 15 Hidden Unit FFNN x 10 4 3 3 2 2 1 1 0 0 65 I Conclusion Mean Target Response Introduction 66 67 68 69 70 71 72 73 74 75 E+ Load Variables 76 77 78 79 80 FFNN with 15 and 5 hidden units fit the Fine Grain loads best Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 36 Preliminaries Step 1 & 2 5 5 Step 4 Step 5 & 6 x 10 5 Sensible RMSE Latent RMSE Sensible MTR Latent MTR 4 RMSE 5 Fine Grain Loads with 10 Hidden Unit FFNN x 10 4 3 3 2 2 1 1 0 0 65 I Conclusion Mean Target Response Introduction 66 67 68 69 70 71 72 73 74 75 E+ Load Variables 76 77 78 79 80 Does not fit variable 77 and 79 as well Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 37 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Fine Grain with 5 Hidden Unit FFNN 50 RMSE MTR 45 RMSE 40 40 35 35 30 30 25 25 20 20 15 15 10 10 5 5 0 0 I 45 4 8 12 16 20 24 28 32 36 40 44 E+ Non−Load Variables 48 52 56 60 Mean Target Response 50 0 64 FFNN with 5 hidden units has difficulty modeling non-loads Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 38 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Fine Grain with 10 Hidden Unit FFNN 50 RMSE MTR 45 RMSE 40 45 40 35 35 30 30 25 25 20 20 15 15 10 10 5 5 0 0 4 8 12 16 20 24 28 32 36 40 44 E+ Non−Load Variables 48 52 56 60 Mean Target Response 50 0 64 I Fits non-loads better than the 5 hidden unit model I The 15 hidden unit model is very similar to the 10 hidden unit Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 39 Preliminaries Step 1 & 2 5 5 Step 4 Step 5 & 6 x 10 5 Sensible RMSE Latent RMSE Sensible MTR Latent MTR 4 RMSE 5 Order 1 Loads with 10 Hidden Unit FFNN x 10 4 3 3 2 2 1 1 0 0 74 I Conclusion Mean Target Response Introduction 75 76 77 78 79 80 81 82 83 84 E+ Load Variables 85 86 87 88 89 MO1 results are similar to FG results Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 40 Preliminaries Step 1 & 2 5 5 Step 4 Step 5 & 6 x 10 5 Sensible RMSE Latent RMSE Sensible MTR Latent MTR 4 RMSE 5 Order 1 Loads with 15 Hidden Unit FFNN x 10 Conclusion 4 3 3 2 2 1 1 0 Mean Target Response Introduction 0 74 75 76 77 78 79 80 81 82 83 84 E+ Load Variables I Does not fit variable 82 and 86 as well I 5 or 10 hidden units is best Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 85 86 87 88 89 University of Tennessee 41 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Order 1 with 15 Hidden Unit FFNN 50 RMSE MTR 45 RMSE 40 45 40 35 35 30 30 25 25 20 20 15 15 10 10 5 5 0 Mean Target Response 50 0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 E+ Non−Load Variables I Best non-load model I variables 28 to 36 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 42 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Order 1 with 5 Hidden Unit FFNN 50 RMSE MTR 45 RMSE 40 45 40 35 35 30 30 25 25 20 20 15 15 10 10 5 5 0 Mean Target Response 50 0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 E+ Non−Load Variables I Slightly worse than 10 and 15 I variables 28 to 36 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 43 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion FFNN Result Summary I Loads are difficult to estimate for both datasets I FFNN perform better on the Fine Grain dataset Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 44 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Lasso regression I Model: y = hw , xi + b I Optimization Criteria: N 1X (yi − hw , xi i)2 + λ|w | 2 i=1 I Exact-Model Training Time: O(n3 ) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 45 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Optimization Method I Alternating Direction of Method of Multipliers (ADMM) I I Boyd et al. 2010 Benefits: I Completely decentralized Defines methods for splitting across examples and features Can solve Lasso, SVM, and Ridge Regression I Guaranteed to converge I I I By changing a single function Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 46 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion ADMM I I Data splitting across examples: A0 A1 A= . .. An b0 b1 b=. .. bn Lasso optimization process I I xik+1 = argminxi ( 12 ||Ai xi − bi ||22 + ρ2 ||xi − z k + uik ||22 ) z k+1 = S λ (x̄ k+1 + ū k ) ρN I I uik+1 = uik + xik+1 − z k+1 First step is the only heavy lifting I xik+1 = (ATi Ai + ρI )−1 (ATi bi + ρ(z k − uik )) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 47 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion FG Results x 10 5 Sensible RMSE Latent RMSE Sensible MTR Latent MTR 4 RMSE 5 Fine Grain Loads with Lasso Regression x 10 4 3 3 2 2 1 1 0 Mean Target Response 5 5 0 65 66 67 68 69 70 71 72 73 74 75 E+ Load Variables 76 77 I Does not estimate FG loads as well as FFNN I Based on variable 65 and 67 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 78 79 80 University of Tennessee 48 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion FG Results Fine Grain with Lasso Regression 50 RMSE MTR 45 RMSE 40 40 35 35 30 30 25 25 20 20 15 15 10 10 5 5 0 0 I 45 4 8 12 16 20 24 28 32 36 40 44 E+ Non−Load Variables 48 52 56 60 Mean Target Response 50 0 64 Estimates non-load variables worse than FFNN Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 49 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion MO1 Results x 10 5 Sensible RMSE Latent RMSE Sensible MTR Latent MTR 4 RMSE 5 Order 1 Loads with Lasso Regression x 10 4 3 3 2 2 1 1 0 Mean Target Response 5 5 0 74 75 76 77 78 79 80 81 82 83 84 E+ Load Variables 85 I Estimates MO1 loads better than FG loads I Worse than FFNN Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 86 87 88 89 University of Tennessee 50 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion MO1 Results Order 1 with Lasso Regression 50 RMSE MTR 45 RMSE 40 40 35 35 30 30 25 25 20 20 15 15 10 10 5 5 0 0 0 I 45 Mean Target Response 50 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 E+ Non−Load Variables Estimates non-load variables as well as FFNN Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 51 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Discussion Points I Prediction time I Data Clustering via Error I MO2 contains unobserved behavior I HVAC Operating Features Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 52 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Prediction time Model FFNN 5 FFNN 10 FFNN 15 Lasso Training Time (Hr) ∼2 ∼8 ∼24 ∼0.2833 Total Training Time (Hr) ∼18 ∼72 ∼216 ∼25.50 I Lasso regression method scales best I FFNN is only competitive when run in parallel Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data Prediction Time (sec) ∼2.70 ∼2.85 ∼2.93 ∼2.90 University of Tennessee 53 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Error Clustering – FFNN FG Variable 65’s CV Error Clustering 40 Group 1 Group 2 Group 3 CV Error Metric 38 36 34 32 30 0 I 100 200 300 400 500 Simulation 600 700 800 Suggests multiple models would be best Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 54 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Error Clustering – Lasso FG Variable 52’s CV Error Clustering FG Variable 63’s CV Error Not Clustering 2 Group 1 Group 2 Group 3 4 CV Error Metric CV Error Metric 5 3 2 1.5 1 0.5 1 0 100 200 300 Simulation 400 500 600 0 0 100 200 300 Simulation I Variables with linear relationships cluster I Variables with non-linear relationships do not cluster Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 400 500 600 University of Tennessee 55 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Error Clustering – MO1 Order 1 Variable 75’s CV Error Group 44 Group Unkown CV Error Metric 42 40 38 36 34 0 50 100 150 200 Simulation 250 300 350 I Contains a single grouping I Implies sampling resolution directly impacts the number of groups Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 56 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion MO2 Scale Shifts Dryer’s Total Heat Gain Rate 2000 Nominal Abnormal Watts 1500 1000 500 0 Jan Feb Mar Apr May Jun Jul Aug Time Sep Oct Nov Dec I Reason behind variable 10’s high variance (MO1 exp) I 3 Simulations in sampled MO2 test data match the Abnormal shift. Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 57 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion MO1 HVAC ON/OFF vs Living Room (LR) Latent Load HVAC Heating ON/OFF vs MO1 LR Latent Heating HVAC Cooling ON/OFF vs MO1 LR Latent Cooling 2 2 ON OFF Latent 1.5 1.5 1 1 0.5 0.5 0 0 Jan Feb Mar Apr May Jun I ON OFF Latent Jul Aug Sep Oct Nov Dec Time Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Time The ON/OFF state temporally matches well Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 58 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion FG HVAC ON/OFF vs Living Room (LR) Latent Load HVAC Heating ON/OFF vs FG LR Latent Heating HVAC Heating ON/OFF vs FG LR Latent Heating 2 2 ON OFF Latent 1.5 1 1 0.5 0.5 0 ON OFF Latent 1.5 0 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Time Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Time I The ON/OFF state does not temporally match as well I FG Living Room Latent Heating distributed throughout the year Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 59 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion HVAC Results I FG HVAC Experiments I I I I Improved performance – 6 loads Diminished performance – 8 loads Unchanged performance – 2 loads MO1 HVAC Experimetns I I I Improved performance – 4 loads Diminished performance – 6 loads Unchanged performance – 6 loads I Latent loads are still difficult to estimate I Some sensible loads are equally difficult Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 60 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Summary I Both models approximate non-load variables well I FFNN is best over all I Lasso trains much faster, and scales better I Multiple models are most likely needed I Generalization may be hindered by scale shifts Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 61 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 62 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Overview I Problem – Estimate building parameters via their relationships I Data sets: I I I I Approaches: I I I I MO1 (4 Gigabytes) MO2 (10 Gigabytes) FG (10 Gigabytes) Direct Lasso regression structure learning Bayesian Lasso regression structure learning Both require Alternating Direction Method of Multipliers Layout: I I I I I I I Problem forumlization Structure learning review Direct Lasso structure learning Bayesian Lasso structure learning Results Discussion Summary Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 63 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Variable Relationship Learning (Step 5) I Relational Learning I All Energy Plus inputs and outputs are random variables Z = {X1 , X2 , . . . , Xn } I Then PDF, P(Z ), describes interactions P(Z ) = P({X1 , X2 , . . . , Xn }) I Objective: I Learn/Approximate P(Z ) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 64 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Variable Inference (Step 6) I Both are MAP Inference problems I Calibration P(Z ) = N Y P(X |Yi )P(X ) i=1 I I I X = {Building Parameters} Yi = {Weather Data, Operation Schedule, Sensor Data} Approximation P(Z ) = N Y P(Yi |Xi )P(Yi ) i=1 I I Xi = {Building Parameters, Weather Data, Operation Schedule} Yi = {Simulation Output} Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 65 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Variable Relationship Learning I Computing P(Z ) is intractable in the general case I Assume P(Z ) factorizes into simpler components I i.e. P(Z ) = P(X1 |X2 )P(X2 |X3 ) . . . P(XN ) I Factorizations are best represented by a Probabilistic Graphical Model (PGM), G I G = {V , E } I I I V are random variables E , edges, represents conditional dependencies Find G that best approximates P(Z ) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 66 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion PGM Structure Learning I Classical Approaches: I I I Score and Search Constraint Based Less Traditional: I Regression Based Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 67 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Score and Search I I Search through possible G Saving the G that best approximates P(Z ) I I Best G minimizes a criteria function Most common criteria: BIC(Data; G) = L(Data; G , θ) + I log M ∗ Dim[G ] 2 Common Algorithms I Greed-Hill Climbing Sparse Candidate Algorithm (SCA) I Max-Min Hill-Climb (MMPC) I I I Friedman et al. 1999 Tsamardinos et al. 2006 I Advantage I Disadvantage I I I I Global Criteria Function Applicable only to Bayesian Networks (DAG) Too expensive for Markov Network (Undirected Graph). Search space size Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 68 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Constraint Based I Searches through variable dependencies, via statistical test I Dependent variables are connected I Common Algorithms I SGS (named after authors) I Grow and Shrink I I I Margaritis et al. 1999 Advantage I I Spirtes et al. 1993 Applicable to Bayesian and Markov Networks Disadvantage I I Scalability - not always guaranteed Hard to apply to continuous random variables Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 69 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Regression Based I Assumes functional relationship between random variables I Uses variable selection methods to determine dependencies I Methods I Sparse Graphical Model for exploring gene expression data I Structure learning with large sparse undirected graphs I Modified Grow and Shrink I I I (Li, Fan. 2007) (Jean-Philippe Pellet and Andre Elisseeff 2008) Advantage I I I (Dobra, A., et al., 2004) I Applicable to Bayesian and Markov Networks Highly Scalable Disadvantage I Existing methods assume the joint probability P(Z ) is Gaussian Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 70 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Gaussian Graphical Model (GGM) I Assumes PDF, P(Z ), is Gaussian P(Z ) = I n 1 1 exp(− (x − µ)T Ω(x − µ)) 2 Requirements: I I I I 1 (2π) 2 |Σ| 2 estimating Σ estimating µ computing Ω, Σ−1 Exact Inference: p(x) ∝ exp(x T Jx + hT x) I I J=Ω h=µ Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 71 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Learning Gaussian Graphical Model I Approach I I I Learn a Bayesian Network Extract the undirected GGM Estimating Ω: Ω = (1 − Γ)T Ψ−1 (1 − Γ) I Γ Bayesian Network weights I I I Ψ−1 I I I Upper triangle Zero Diagonals Diagonal matrix Each entry is the estimated variance Estimating h I I Not required, if data is centered Estimate centering parameters instead Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 72 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Directly Estimating Γ and Ψ−1 I Approach I I Impose an ordering over the variables Perform N − 1 regressions I I I Γi = βi+1,n I I Response is Xi Candidate Parents are {Xi+1 , . . . , Xn } βi+1,n computed via Lasso regression Ψ−1 = diag (σ1 , · · · , σn )−1 I σi is the estimated response variance I σi = P|D| n=1 (xni − PN j=i+1 βj xnj )2 |D|−2 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 73 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Variable Order I Order matters for Direct Method I I A greedy algorithm can select good orders Dobra et al., 2004 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 74 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Variable Order I Order matters for Direct Method I I I Order does not matter, proved by (Li, Fan. 2007) I I A greedy algorithm can select good orders Dobra et al., 2004 Dependent upon applying Wishart Prior to Ω Prior provides iterative Lasso Regression method I Iteratively estimates Γ and Ψ−1 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 74 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Bayesian Γ and Ψ−1 Estimation I Wishart prior W (δ, T ) I I I I I Γi = βi+1,n I I I δ defines shape T defines prior covariance hyperparameter T = diag (θi , · · · , θn ) i ) P(θi ) = γ2 exp( −γθ 2 βi+1,n√computed via Lasso regression λ = γψi Ψ−1 = diag (ψ1−1 , · · · , ψn−1 ) I ψi−1 = PN j=i+1 I δ−1+N−2i+|D| P|D| PN 2 n=1 (xni − j=i+1 βij xnj ) P −1 M 1 P(θj ) j=1 ψ̂j M βij2 θi−1 +θi−1 + E[ψi−1 |θi , βi , D] = Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 75 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Small Scale Bayesian MO1 Result Bayes Error Variables 27 to 52 1.2 1 1 Difference in Absolute Error Difference in Absolute Error Bayes Error Variables 1 to 26 1.2 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 5 10 15 Variables 20 0 26 25 31 1.2 1 1 0.8 0.6 0.4 0.2 0 52 62 67 Variables 73 98 103 149 154 0.6 0.4 0 78 78 83 88 93 Variables Bayes Error Variables 131 to 151 1.2 1 Difference in Absolute Error 1 Difference in Absolute Error 51 0.2 57 1.2 0.8 0.6 0.4 0.2 I 46 0.8 Bayes Error Variables 105 to 130 0 104 41 Variables Bayes Error Variables 79 to 104 1.2 Difference in Absolute Error Difference in Absolute Error Bayes Error Variables 53 to 78 36 0.8 0.6 0.4 0.2 109 114 119 Variables 124 129 0 130 135 139 144 Variables 139 good estimates, 8 poor estimates, 4 moderate estimates Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 76 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Small Scale Direct MO1 Result Direct Error Variables 27 to 52 1.2 1 1 Difference in Absolute Error Difference in Absolute Error Direct Error Variables 1 to 26 1.2 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 5 10 15 Variables 20 0 26 25 31 1.2 1 1 0.8 0.6 0.4 0.2 0 52 62 67 Variables 73 98 103 149 154 0.6 0.4 0 78 78 83 88 93 Variables Direct Error Variables 131 to 151 1.2 1 Difference in Absolute Error 1 Difference in Absolute Error 51 0.2 57 1.2 0.8 0.6 0.4 0.2 I 46 0.8 Direct Error Variables 105 to 130 0 104 41 Variables Direct Error Variables 79 to 104 1.2 Difference in Absolute Error Difference in Absolute Error Direct Error Variables 53 to 78 36 0.8 0.6 0.4 0.2 109 114 119 Variables 124 129 0 130 135 139 144 Variables 139 good estimates, 13 poor estimates, 0 moderate estimates Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 77 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Small Scale Random MO1 Result Bayes Error vs Rand Error Variables 27 to 52 1 0.5 0.5 Difference in Absolute Error Difference in Absolute Error Bayes Error vs Rand Error Variables 1 to 26 1 0 −0.5 −1 0 5 10 15 Variables 20 0 −0.5 −1 26 25 31 1 1 0.5 0.5 0 −0.5 −1 52 57 62 67 Variables 73 −1 78 78 0.5 I Difference in Absolute Error Difference in Absolute Error 0.5 0 −0.5 119 Variables 124 51 83 88 93 Variables 98 103 Bayes Error vs Rand Error Variables 131 to 151 1 114 46 0 Bayes Error vs Rand Error Variables 105 to 130 109 41 Variables −0.5 1 −1 104 36 Bayes Error vs Rand Error Variables 79 to 104 Difference in Absolute Error Difference in Absolute Error Bayes Error vs Rand Error Variables 53 to 78 129 0 −0.5 −1 130 135 139 144 Variables 149 154 Sampling [0, 1] uniform distribution is competitive Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 78 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion FG Results Rand FG Error 1 0.8 0.8 Difference in Absolute Error Difference in Absolute Error Bayes FG Error 1 0.6 0.4 0.2 0.6 0.4 0.2 0 0 9 27 28 29 30 66 67 68 69 157 158 159 160 181 Variables 9 27 28 29 30 66 67 68 69 157 158 159 160 181 Variables I Bayesian is statistically better – 4.05±0.98 vs 5.07±1.01 I Statistically different variables 27, 28, 68, 157, 158, 159, 160, and 181 Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 79 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion FG Results 0.8 0.6 0.4 0.2 Estimate Actual 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.8 0.6 0.4 0.2 0 9 27 28 29 30 66 67 68 69 157 158 159 160 181 Variables 0 9 27 28 29 30 66 67 68 69 157 158 159 160 181 Variables I Random works best on 0.5 mean variables I Bayesian tracks means better I Appears to infer building parameters well Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 1 Parameter Values 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 Parameter Estimates Random Parameter Estimation Estimate Actual Parameter Values Parameter Estimates Bayesian Parameter Estimation University of Tennessee 80 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion MO2 Results 10 15 Variables 20 25 26 31 1.5 1 0.5 0.5 83 88 93 Variables 98 103 Parameter Estimates 1 78 41 Variables 46 51 52 0.5 57 1.2 1 0.8 0.6 0.4 0.2 1.2 1 0.8 0.6 0.4 0.2 109 114 119 Variables 124 129 I Appears to infer building parameters well I Tracks means well Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 62 67 Variables 73 78 Bayesian Parameter Estimates Variables 131 to 151 Estimate Actual 104 1.5 1 0.5 Bayesian Parameter Estimates Variables 105 to 130 Estimate Actual Parameter Values Parameter Estimates Bayesian Parameter Estimates Variables 79 to 104 1.5 36 Parameter Estimates 5 0.5 Parameter Values 0 0.5 Estimate Actual 1 Parameter Values 0.5 1 1.5 Estimate Actual 0.7 0.6 0.5 0.4 0.3 130 0.7 0.6 0.5 0.4 0.3 135 139 144 Variables 149 Parameter Values 0.5 1 1.5 Parameter Estimates 1 Bayesian Parameter Estimates Variables 53 to 78 Estimate Actual 1.5 Parameter Values 1 1.5 Parameter Estimates Bayesian Parameter Estimates Variables 27 to 52 Estimate Actual Parameter Values Parameter Estimates Bayesian Parameter Estimates Variables 1 to 26 1.5 154 University of Tennessee 81 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Discussion I Inference optimization methods I Estimating Distant Values Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 82 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion GA vs Gradient FG Building Parameter 2 Parameter Values 1.5 Actual Gradient−Est GA−Est 1 0.5 0 −0.5 0 I I I 5 10 15 20 25 30 Simulations 35 40 45 50 Gradient estimates near the mean often GA introduces more variance Gradient better for large parameter inference I I Variance scales with number of parameters MO1 and FG used GA, MO2 used Gradient Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 83 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Estimating Distant Values MO2 Building Parameter 3 Parameter Value 2 One−Act Zero−Act Est Zero−Est 1.5 1 0.5 0 0 50 100 150 Simulation 200 I Values concentrate on the mean closely I Distant values hard to estimate I Changing γ and δ may introduce more variance Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data 250 300 University of Tennessee 84 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Summary I Bayesian GGM estimates building parameters well I Best when using gradient optimization I However, all estimates always concentrate towards mean I Model needs more variability Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 85 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Outline Introduction Preliminaries Calibration Signal Estimation & Sensor Selection Simulation Approximation Learning Simulation Variable Relationships Conclusion Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 86 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Solution Overview I Step 1 — Select best Prediction Model I Step 2 — Select best Sensors (Step 1 result) I Step 3 — Map Sensors to Simulation Outputs (Step 2 result) I Step 4 — Build Simulation Approximation (Optional) I Step 5 — Build Simulation Relational Model I Step 6 — Calibrate Simulation (Step 3, 4, 5 results) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 87 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Addressed Steps I Step 1 — Selected LS-SVM model I Step 2 — Selected best sensors from Campbell Creek I Step 4 — Built large-scale E+ approximation I Step 5 — Built large-scale E+ relational model I Step 6 — Partially demonstrated (E+ Ground Truth Simulations) Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 88 Introduction Preliminaries Step 1 & 2 Step 4 Step 5 & 6 Conclusion Contributions I Step 1 I I Step 2 I I I Best sensors for predicting electrical consumption Novel feature selection method Step 4 I I Best predictor for hourly residential electrical consumption Large-scale E+ residential approximation Step 5 I Large-scale regression structure learning Richard E. Edwards Automating Large-Scale Simulation Calibration to Real-World Sensor Data University of Tennessee 89