Due to the significant increase in the share of variable renewable electricity generation, the power grid flexibility must be enhanced. To achieve this while reducing CO2 emissions, improving the manoeuvrability of low-carbon sources such as nuclear power plants (NPP) is essential. Time series forecasting is a crucial task to enhance NPPs manoeuvrability. In particular, because the ability to forecast critical core operating parameters in a fast and trustworthy manner would substantially help operators perform load-following operations. Traditionally, core nuclear reactor operating parameters are estimated through neutron transport simulations, which require a significant amount of computational resources. Alternatively, machine learning (ML) techniques such as neural networks have been increasingly used to address this type of problem. In this study, we propose a novel databased ML method combining tree-based regressors (Random Forest or Gradient Boosted Trees) and time series models to forecast the normalized axial offset, a core parameter that must be monitored and controlled by the operators working in the control room. We consider a rigorous optimization procedure to determine hyperparameter values leading to high performance. Finally, we validate our approach on French 1300 MW pressurized water reactors (PWR) historical data. We show that, knowing the operators' projected commands, we can compute minute-by-minute forecasts in a few fractions of a second for the next eight hours, with suitable accuracy for the task at hand.