Control-oriented LSTM Modeling of Indoor Temperature and Building Heating Demand: A Proposed Framework and Case Study

Energy flexible buildings serve as a mitigating factor against grid congestion, which refers to the event where the electricity transmission demand exceeds the available capacity of the electricity grid. Therefore, the company DWA aims to be able to assess this potential during a building’s design phase. This assessment can be done through the employment of model-based control algorithms, where a model of the actual building energy system is used to design the optimal control strategy based on specific objectives. The topic of this thesis is to create such a system model that can be used within a model-based controller. Specifically, in this project, we propose a novel control-oriented, Long-Short Term Memory (LSTM)-based modeling framework to predict indoor temperature and heating power demand of a complex building’s heating system. This modeling framework consists of two separate LSTM models for each output, where the power model is additionally split up into a classifier and a...Show moreEnergy flexible buildings serve as a mitigating factor against grid congestion, which refers to the event where the electricity transmission demand exceeds the available capacity of the electricity grid. Therefore, the company DWA aims to be able to assess this potential during a building’s design phase. This assessment can be done through the employment of model-based control algorithms, where a model of the actual building energy system is used to design the optimal control strategy based on specific objectives. The topic of this thesis is to create such a system model that can be used within a model-based controller. Specifically, in this project, we propose a novel control-oriented, Long-Short Term Memory (LSTM)-based modeling framework to predict indoor temperature and heating power demand of a complex building’s heating system. This modeling framework consists of two separate LSTM models for each output, where the power model is additionally split up into a classifier and a regressor. Further, the model is organized in a cascade structure, where, first, power is predicted, and is subsequently used as input to predict indoor temperature. Secondarily, apart from our proposed model, we also introduce a three-component model assessment framework in this study, which consists of error metrics, feature importance, and step response tests to support model development and assessment. The research consists of three parts. First, the proposed model was compared to simpler LSTM-based modeling structures. Second, the model structure was tuned in terms of input features and hyperparameters. Lastly, the final model was compared to a benchmark autoregressive model with exogenous inputs (ARX) in the prediction accuracy, and further evaluated in terms of feature importance, extrapolation qualities, and robustness against initializations. The results could be summarized into five key findings. Firstly, the proposed model structure outperformed conventional LSTM-based structures and the benchmark model in the predictive accuracy. Secondly, the degree to which the final model meets satisfactory prediction accuracy according to official standards remains ambiguous, due to a lack of consensus in the way these evaluation standards ought to be applied. Thirdly, the employment of a three-component assessment framework proves valuable in the development and evaluation of a control-oriented model. Fourthly, simplifying the model by the exclusion of indoor disturbance features increased model extrapolation qualities. Fifthly, the final model showed considerable extrapolation qualities with respect to outdoor disturbances and control inputs. It can be concluded that the current research provides a valuable, high potential, model framework for the control-oriented modeling of both indoor temperature and heating power load. In addition, the demonstrated value of using a three-component model evaluation framework for data-driven system models encourages further deployment of these methods.Show less

Leiden University Student Repository

Documents

In Collections

Control-oriented LSTM Modeling of Indoor Temperature and Building Heating Demand: A Proposed Framework and Case Study