Power Plant Boiler Reheater Metal Temperature Forecasting Using Autoencoder

The fluctuating pattern of power plant operation can cause a negative impact on a specific area of the boiler. Such as overheating in the boiler reheater. Various methods, especially data-based modeling, have been used to optimize power plant operation to lessen the impact. Architectures such as CNN and RNN and their combinations become the main foundation in the data-based modeling process. This study is intended to reproduce previous research with the actual datasets of a boiler from a power plant in Indonesia. Because even with a similar model, datasets from different systems will produce different results. In this study, forecasting for the reheat metal temperature is carried out using logged data. The model used is a stacked-layer LSTM-based autoencoder. Various hidden unit numbers are used for each encoder-decoder to find the optimum number. The test results obtained show the accuracy of the prediction model with an average MAPE value below 0.6%. It is also concluded that the addition of hidden units has no significant impact on model performance. The model with 64-64 hidden layers is sufficient for forecasting reheat metal temperature because shorter training time with affordable performance.