LAUSR

Abstract With the increase in renewable energy penetration of electrical grids, coal power stations will be required to operate flexibly rather than functioning as baseload units. During flexible operation of conventional coal-fired stations, thermal stresses are induced in reheaters which could lead to tube ruptures and unplanned plant downtime. The current study sets out to develop a data-driven sequence-to-sequence recurrent neural network model capable of predicting future reheater metal temperatures using plant operational data. The best-performing network and training algorithm configuration was found by implementing a coarse grid search of hyperparameter combinations. The proposed model architecture uses stacked encoder and decoder sections with GRU cells and 512 hidden units per layer. An input sequence length of 8 min was used to predict an output sequence of 5 min, with sequence intervals of 1 min. The results indicate that the encoder-decoder GRU network has adequate accuracy. The mean absolute percentage error for the test dataset was below 1% which corresponds to a root-mean-squared error in predicted metal temperatures of 6.2 °C.