Data-driven forecasting with model uncertainty of utility-scale air-cooled condenser performance using ensemble encoder-decoder mixture-density recurrent neural networks

In the current work, an encoder-decoder mixture-density network (MDN) is developed using recurrent neural networks (RNN) for the prediction of utility-scale air-cooled condenser backpressure. Back pressure has a direct impact on thermal power plant generated load. The current study compares deterministic models (standard RNNs) to probabilistic models (MDN-RNNs). This is done using three datasets with increasing complexity to understand how significant the effects of plant operating parameters and ambient conditions are on the ACC back pressure. A hyperparameter search was performed to find the best encoder-decoder RNN architecture. An MDN layer was then attached to the selected architectures, to develop models capable of predicting the uncertainty. A two-layer encoder-decoder MDN-RNN model was selected and then trained with and without early stopping regularization active. The resultant models were combined using an ensemble approach. It was found that the ensemble model was able to achieve a better prediction of outliers without overfitting the data. The basic model, a standard RNN with the smallest input dimensionality achieved an average RMSE of 5.66 kPa whereas the end-to-end ensemble model (meta-MDN model trained using the highest dimensionality input) achieved an RMSE of 3.14 kPa which translate into a model accuracy increase from 68% to 82%. (c) 2021 Elsevier Ltd. All rights reserved.