Data for: Scaling-Up Grey-Box Models for Predicting Building Dynamics - Model Structure, Seasonal Variation, and Physical Interpretability

The data accompany the article Scaling-Up Grey-Box Models for Predicting Building Dynamics: Model Structure, Seasonal Variation, and Physical Interpretability, which presents an elaborate analysis of different RC model structures, different training data, different validation data, and different model selection methods. The training and validation data have been acquired by using energy modeling software TRNSYS for a building located in Zagreb. One year’s worth of data has been divided into 52 training periods, each lasting 7 days, and 182 validation periods, each lasting 2 days. Models were trained by using a combination of Latin Hypercube Sampling (LHS) and non-linear least squares. For the LHS, the number of initial parameter guesses was varied, resulting in 7 different parameter guesses densities: 2, 5, 10, 20, 50, 100, 200. Each of those guess densities was used to produce that many models for each of the training periods. That means that if the number of initial guesses is 200, there were 200 models trained for each training period, resulting in 200x52=10 400 models.

The data with a timestep of 15 minutes used for training and validating the models can be found in the folder “0. Training and validation data”. Parameter values, RMSE for training periods and RMSE for validation periods can be found in folders “1. SE zone”, “2. SW zone”, “3. NW zone”, “4. NE zone”. The platform in which the user can interact with RC models to increase and decrease model parameters can be found in the folder “5. Platform for RC Model Parameter Analysis”. Each folder contains a description file regarding the data.

Any additional inquiry about the data set and the RC model platform can be sent to the corresponding author, Nikola Badun ([e-mail]).