130 likes | 323 Views
A Parallel Statistical Learning Approach to the Prediction of Building Energy Consumption Based on Large Datasets. Hai Xiang ZHAO, Phd candidate Frédéric MAGOULÈS, Full Prof. in ECP. Outline. Background SVM theory Obtain historical data Data analysis Experiments and results
E N D
A Parallel Statistical Learning Approach to the Prediction of Building EnergyConsumption Based on Large Datasets Hai Xiang ZHAO, Phd candidate Frédéric MAGOULÈS, Full Prof. in ECP
Outline • Background • SVM theory • Obtain historical data • Data analysis • Experiments and results • Conclusion
Background • Energy efficiency for buildings. • Complex --- many influence factors: • Ambient weather conditions • Building construction and materials • Occupants and their behaviors • Inner facilities • ... • Approaches: • Engineering, simulation, statistical models...
A complex system with many factors involved • Heat gain: • Outside: • Solar radiation • △T (through walls) • Inside: • Electrical plants • Occupants • Heat loss: • Infiltration • Ventilation — Ambient weather conditions — Building construction, materials — Occupants and their behaviors — Inner facilities — ...
Background • Energy efficiency for buildings. • Complex --- many influence factors: • Ambient weather conditions • Building construction and materials • Occupants and their behaviors • Inner facilities • ... • Approaches: • Engineering, simulation, statistical method...
Support vector machine (SVM) Samples: Decision function: Loss function: Maximize: Constraints: ,
Kernel: Radial Basis Function (RBF) Parallel approach (D. Brugger 2007) Sequential minimal optimization (SMO) Kernel evaluation Distributed storage of kernel rows Performance evaluation Mean squared error (MSE) Squared correlation coefficient (SCC)
Obtaining historical data— Simulation in EnergyPlus Table 1: Input parameters for a single building. Fig 1: Flow chat of simulating multiple buildings.
Data analysis • Analyzing steps – Fig 2 • Estimation of — — 5-fold cross validation • Experimental environment: • — Two nodes, CPU 8 * 2.5GHz • 1333MHz FSB and 4G memory • — Gigabit Ethernet Fig 2: Flow chat of the data analyzing
Results Fig 4: Measured and predicted district heating demand for the last building in heating season. Fig 3: Dry bulb temperature in the first 20 days of January and July.
Fig 5: Running time of the training process using a parallel implementation of SVMs. Fig 6: Comparison of the speedup with a theoretical optimal linear speedup. • Table 2: Comparison of parallel and sequential implementations. • SVs — The number of Support Vectors • MSE — Mean Squared Error • SCC— Squared Correlation Coefficient
Conclusion • A simulation approach to collect enough historical time series data for multiple buildings • A statistical learning method is then applied to predict the energy behavior in a completely new building • A parallel implementation of support vector regression with RBF kernel is applied to analyze large amounts of energy consumption data.