Abstract Model-based Predictive Control (MPC) is widely used in Industry due to its ability to handle multivariate systems subject to input and output constraints. Two phases can be distinguished in an MPC implementation: identification and control. The purpose of this thesis is twofold: make contributions in identification for MPC, and propose a new MPC control methodology. Closed-loop performance in an MPC implementation relies heavily on predictive performance of the model, then model identification is a crucial point in MPC and the part that often demands most of the time of the project. This thesis deals first with identification for MPC. Model identification aims at approximating a process, and models are often fit for purpose. Provided the purpose of the model in MPC is to perform multi-step ahead predictions, the identification strategy needs take multi-step ahead prediction errors into account. This identification strategy is often denoted MRI (Model Predictive Control Relevant Identification). In this thesis, identification for MPC covers three main topics. First, MRI and different approaches to it are defined. Second, fitting a multiple input multiple output model is compared to fitting several multiple input single output models in terms of MRI concluding the approach to obtain a single multiple input multiple output model is preferable in MRI for a sufficiently large prediction horizon. Finally a PLS-based (Partial Least Squares) line search numerical optimization approach denoted PLS-PH that deals with parametric MRI in the case of collinearity in the identification data set is proposed. An example shows PLS-PH can outperform conventional MRI parametric approaches if there is collinearity in the identification data set. Once the model to perform multi-step ahead predictions is ready, the controller can be formulated. A model-based predictive control methodology in the space of the latent variables for continuous processes is proposed in this thesis and denoted LV-MPC. LV-MPC takes the dynamic matrix approach for MRI and uses PLS to obtain the latent variable space in which the decisions of the controller are made. Implementing identification and control in the latent variable space: eases identification in the case of correlation in the identification data set, acts as a prefilter reducing the effect of noisy data, reduces computational complexity, and provides tools that ensure the predictor is used in the region in which it has been identified, hence improving closed-loop performance. Several examples show how LV-MPC can outperform traditional MPC in terms of computational complexity and closed-loop performance whilst it is easy to tune. LV-MPC for continuous processes is a novel approach that combines the best of two methodologies widely used in Industry: MPC and latent variable methods. This thesis describes LV-MPC and tackles with many of the challenges in MPC, however, there is room to add more functionalities to the methodology, then LV-MPC starts what can become a new trend in MPC.