Model predictive control (MPC) is an advanced control technique. It has been deployed to harness the energy flexibility of a building. MPC requires a dynamic model of the building to achieve such an objective. However, developing a suitable predictive model is the main challenge in MPC implementation for flexibility activation. This study focuses on the application of key performance indicators (KPIs) to evaluate the suitability of MPC models via feature selection. To this end, multiple models were developed for two houses. A feature selection method was developed to select an appropriate feature space to train the models. These predictive models were then quantified based on one-step ahead prediction error (OSPE), a standard KPI used in multiple studies, and a less-often KPI: multi-step ahead prediction error (MSPE). An MPC workflow was designed where different models can serve as the predictive model. Findings showed that MSPE better demonstrates the performance of predictive models used for flexibility activation. Results revealed that up to 57% of the flexibility potential and 48% of energy use reduction are not exploited if MSPE is not minimized while developing a predictive model.

模型预测控制（MPC）是一种先进的控制技术。它已被部署来利用建筑物的能源灵活性。MPC 需要建筑物的动态模型来实现这一目标。然而，开发合适的预测模型是灵活性激活 MPC 实施中的主要挑战。本研究重点关注关键绩效指标（KPI）的应用，通过特征选择来评估 MPC 模型的适用性。为此，为两栋房屋开发了多种模型。开发了一种特征选择方法来选择合适的特征空间来训练模型。然后，根据一步提前预测误差 (OSPE)（多项研究中使用的标准 KPI）和不太常见的 KPI：多步提前预测误差 (MSPE) 对这些预测模型进行量化。设计了 MPC 工作流程，其中不同的模型可以用作预测模型。研究结果表明，MSPE 更好地展示了用于灵活性激活的预测模型的性能。结果显示，如果在开发预测模型时未将 MSPE 最小化，则高达 57% 的灵活性潜力和 48% 的能源消耗减少不会得到利用。