Decentralized energy generation, often in the form of industrial Combined Heat and Power (CHP) units, meets a significant part of the global energy demand. The power production from these units has to be forecast, both individually and collectively, to ensure balance and avoid congestion events on the power grid. However, despite its importance, forecasting CHP generation remains an under-studied problem. With increasing proliferation of renewable energy sources, large forecast errors for CHP generation are rapidly taking center-stage as well. In this paper, we propose a global, ensemble-based machine learning (ML) method to improve short-term forecasting accuracy. We demonstrate its efficacy using data from one hundred largest (industrial) CHP units in Belgium, showing a forecast error reduction of up to 27% compared to the baseline method. This can greatly reduce balancing needs and costs, as well as congestion issues. Our results also highlight several nuances that must be kept in mind by grid operators and market players alike, including fitting global models that leverage data from many CHPs which can lead to better forecast accuracy but limit interpretability of the results, and the fact that there is no single best forecast model for all CHPs.