Predicting power demand of building heating systems is a challenging task due to the high variability of their energy profiles. Power demand is characterized by different heating cycles including sequences of various transient and steady-state phases. To effectively perform the predictive task by exploiting the huge amount of fine-grained energy-related data collected through Internet of Things (IoT) devices, innovative and scalable solutions should be devised. This paper presents PHi-CiB, a scalable full-stack distributed engine, addressing all tasks from energy-related data collection, to their integration, storage, analysis, and modeling. Heterogeneous data measurements (e.g., power consumption in buildings, meteorological conditions) are collected through multiple hardware (e.g., IoT devices) and software (e.g., web services) entities. Such data are integrated and analyzed to predict the average power demand of each building for different time horizons. First, the transient and steady-state phases characterizing the heating cycle of each building are automatically identified; then the power-level forecasting is performed for each phase. To this aim, PHi-CiB relies on a pipeline of three algorithms: the Exponentially Weighted Moving Average, the Multivariate Adaptive Regression Spline, and the Linear Regression with Stochastic Gradient Descent. PHi-CiB’s current implementation exploits Apache Spark and MongoDB and supports parallel and scalable processing and analytical tasks. Experimental results, performed on energy-related data collected in a real-world system show the effectiveness of PHi-CiB in predicting heating power consumption of buildings with a limited prediction error and an optimal horizontal scalability.