N-BEATS neural network structure for sub-critical boiler data-driven modelling

Carbon Capture and Storage (CCS) processes can play an important role in reducing the effects of global warming in the coming years. When coupled with a thermal power plant, it has the ability to prevent a large amount of CO2 from being released into the atmosphere. However, this technology is associated with a reduction in the efficiency of the power plant it is paired with. As a general rule, the higher the capture rate of CO2, the less effective the plant is. As it is a recent application, it is still possible to reduce the penalty associated with its use. Therefore, it becomes very important to model the thermal power plant associated with this process as accurately and quickly as possible. This work focuses on the boiler of such a plant. Based on experimental data, it aims to assess the feasibility of modeling the boiler of a CCS-coupled coal-fired power plant using a traditional statistical time-series method (ARIMA), a state-of-the-art artificial neural network (ANN) structure, and a hybrid model that combines the previous algorithms. The ANN hyperparameters are optimized using a Tree-structured Parzen Estimators approach. Using different optimization intervals, different sets of hyperparameters are chosen to produce different models. The results showed that the ANN outperformed the ARIMA method by 18%, however, the optimization time outweighed the positive improvement. The ARIMA-based forecasts had an error of only about 1% sMAPE, which is a satisfying value. The combination of the ANN and ARIMA methods using a simple average produced a model that was 18 % better than the best “single” algorithm used.