Exploration of Machine Learning Methods for Predicting the Operation Schedule of a Combined Heat and Power Plant

摘要

In the context of sustainable energies, precise monitoring, simulation and forecasting methods are required for ensuring a sustainable and economical operation of energy systems. For combined heat and power (CHP) plants these tasks are typically performed with simulation models that model the physical behavior. With Machine Learning (ML), a data-driven modeling approach is available that can develop a reliable model without knowing detailed component characteristics. This paper compares the performance of different ML methods such as artificial neural networks, support vectors, and decision trees (including ensembles) for predicting the operation schedule of a heat-led CHP plant with heat storage. This includes the design of recursive and multiple output strategies. The experiments show that especially ensembles are well suited to solve the regression problem. They achieve an accuracy of over 98% for the forecast of a one-hour time horizon and still achieve 91% accuracy for a forecast of full days.