Control of safety-critical interconnected systems has recently attracted a lot of attention since these systems can depict various infrastructure systems that form the foundation of our society and economy. Model predictive control (MPC) shows significant potential as a viable option to control this class of systems. A primary downside of MPC is the necessity of a precise model which could be hard to acquire and/or complex to utilize. Thus, simplified models are typically employed together with a description of the uncertainty within the MPC framework. In this talk, we present an adaptive learning-based MPC framework comprising a learning phase and an adaptation phase. In the learning phase, a better description of the uncertainty is continually learnt using data collected online. In the adaptation phase, the MPC ingredients are adapted based on the updated uncertainty description. We also show that this scheme yields a better trade-off between closed-loop performance and computation cost when compared to existing schemes including robust, adaptive and learning-based MPC.