Many countries are attempting to adopt energy retrofit measures for their existing building stock to achieve low or net zero energy goals. In this regard, Building Performance Simulation (BPS) tools are often used to test and inform different retrofit measures. However, these tools can be time-consuming and computationally intensive, which limits their application in complex building retrofit and optimization applications. Machine learning (ML) algorithms used as surrogates to BPS models can supplement their capabilities and overcome some of their computational limitations. Recent studies have successfully trained and validated ML-based surrogate BPS models to predict building performance for different applications at low computational costs. However, most studies are limited to predicting total building energy consumption over large time windows (e.g., yearly), which is inadequate for applications that require more granular prediction windows (e.g., daily), such as net zero energy building design. The primary objective of this study is to develop and explore the capabilities of ML-based BPS surrogate models to support applications requiring high-resolution energy predictions. This is achieved by (i) developing a BPS model of an archetype office building in Canada, (ii) generating a dataset for ML surrogate model training, validation, and testing, (iii) applying and contrasting the capabilities of different ML algorithms at predicting energy loads using different prediction windows (e.g., annually, monthly, daily, and hourly). The results demonstrate that hourly, daily, and monthly ML models achieved competitive predictive accuracies, confirmed by adjusted R2 values exceeding 0.9 and CV(RMSE) values within industry standards. In contrast, the annual ML models yielded lower adjusted R2 values, which could be attributed to smaller sample sizes and a lack of granular weather information (e.g., hourly air temperature) provided as inputs to the models. The findings confirm the ability of ML-based surrogate models to mimic BPS models’ performance for applications requiring high-resolution energy predictions.