We present the development and validation of a Digital Twin (DT) for optimizing process parameters in the production process of wood fibreboards [1], a highly energy consuming process. The DT is designed to predict key performance indicators (KPIs) related to the energy consumption and product quality as a function of process parameters in steady-state, and it is intended to be integrated into a Decision Support System (DSS) with simulation and optimization capabilities to assist plant operators in decision-making. Our contributions to the design and validation of DTs for process parameters optimization are: (i) the adoption of a hybrid linear-nonlinear neural network architecture, tuned to be as linear as possible, thus having a more predictable behavior when extrapolating to new process parameters, (ii) the incorporation of physical domain knowledge in the form of monotonicity via a Jacobian-based regularization, (iii) the use of a validation loss function that selects models with higher sensitivity to the inputs, making them more suitable for process parameters optimization. After the initial training, the DT needs to be connected to the data stream and detect eventual changes in the process behavior to trigger model updates. In this paper, (iv) we tune the update procedure to perform robustly across several real-world conditions, using historical data. As model updates impact the workload of plant operators—since new optimal process parameters must be adopted with each update— (v) we balance the tradeoff between update frequency and model accuracy via a Pareto front. Based on data from a real fibreboard production plant, the developed DT demonstrated high sensitivity to input parameters, making it suitable for the integration into optimizers for process parameters optimization. In holdout datasets ranging from 15 to 41 days, the Percentual Median Absolute Error (PMAE) of the online DT—regularly updated with new data based on the triggers—was 5.8% ± 4.4%, compared to 12.2% ± 11.9% in the offline case, showing the effectiveness of the model updates. With the DT validated, our next goal is to validate the DSS for assisting plant operators in decision-making.