Short-horizon Prediction of Indoor Temperature using Low-Order Thermal Networks : A case study of thermal models for heat-system control applications

Sammanfattning: Optimizing and controlling the heating systems in buildings is one way to decrease their load on the power grid, as well as introduce load flexibility to be used in Demand Response (DR) applications. A requirement in occupied buildings is that the thermal comfort of the residents is guaranteed, making the optimization of heating systems a constrained problem with respect to indoor temperature. Thermal models capable of predicting indoor temperatures over short (24 hour) horizons are one way to guarantee this comfort. The accuracy and computational complexity of these models have the most significant impact on controller performance. The data requirements and the expert knowledge required for model implementation are also important factors, since they determine the development costs and, finally, whether a model is feasible to implement. First a literature study explores current modeling approaches that depend only on time-series sensor data and that are suited for control applications. A modeling type found to be fit for such purposes are grey-box models, specifically physically inspired inverse models whose parameters are estimated based on data, such as Resistance- Capacitance (RC) models. This modeling of a dynamical system approach uses prior information in the form of the assumed physical equations and has the potential to increase the performance on sparse data problems. The simple form of the model also has a low level of complexity, making it well suited for control applications. However, expert knowledge can be needed for choosing the model equations as well as initializing the parameters. Then the effects of varying RC model complexity, parameter initialization, and training data are investigated in the case study. The chosen models are 1R1C, 2R2C, and 3R2C. They are fitted using the Nelder-Mead algorithm and validated using the data collected from the RISE Research Villa. Parameter initializations are varied by two orders of magnitude and then fitted on different data sequences to avoid relying on expert knowledge in model creation. The initializations that converged with the best R2 training fit on all sequences were deemed reasonable initializations for the problem and used in the training length comparison. The training length of the models varies from 24 to 384 hours. The results showed that increased training data length correlates positively with performance up to 192 hours for all models, but further increasing it gave inconclusive results. The higher order models evaluated struggled to beat the simplest model or even the constant prediction baseline in Mean Absolute Error (MAE) performance at all training lengths, indicating either that the models selected are unsuitable or that the data features chosen are unrepresentative of the indoor temperature dynamics. Regardless, the MAE errors presented here are comparable to the outcomes of related works. This is possibly an artifact of this dataset having a low variance in temperature and thus resulting in lower errors, which underlines the importance of the data used in case-studies.