Application of machine learning to develop a real-time air-cooled condenser monitoring platform using thermofluid simulation data

Abstract

A data-driven surrogate model is proposed for a 64-cell air-cooled condenser system at a power plant. The surrogate model was developed using thermofluid simulation data from an existing detailed 1-D thermofluid network simulation model. The thermofluid network model requires a minimum of 20 min to solve for a single set of inputs. With operating conditions fluctuating constantly, performance predictions are required in shorter intervals, leading to the development of a surrogate model. Simulation data covered three operating scopes across a range of ambient air temperatures, inlet steam mass flow rates, number of operating cells, and wind speeds. The surrogate model uses multi-layer perceptron deep neural networks in the form of a binary classifier network to avoid extrapolation from the simulation dataset, and a regression network to provide performance predictions, including the steady-state backpressure, heat rejections, air mass flowrates, and fan motor powers on a system level. The integrated surrogate model had an average relative error of 0.3% on the test set, while the binary classifier had a 99.85% classification accuracy, indicating sufficient generalisation. The surrogate model was validated using site-data covering 10 days of operation for the case-study ACC system, providing backpressure predictions for all 1967 input samples within a few seconds of compute time. Approximately 93.5% of backpressure predictions were within ±6% of the recorded backpressures, indicating sufficient accuracy of the surrogate model with a significant decrease in compute time.