Hybrid CF-CNN-BiLSTM hypertuned by Bayesian optimization for thermal power generation and decarbonization forecasting
William Gouvêa Buratto, Rafael Ninno Muniz, Ademir Nied, Carlos Frederico de Oliveira Barros, Erlon Cristian Finardi, Gabriel Villarrubia González
Abstract
Accurate thermal power generation forecasting is essential for grid stability, operational cost reduction, and decarbonization (via CO 2 emission management), but is challenged by noisy, non-linear, temporally dependent data. In this paper, a novel hybrid model integrates a Christiano–Fitzgerald (CF) random walk filter to decompose time series into trend, cyclical, and irregular components for denoising. A convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM) architecture extracts spatial features and captures temporal dependencies. Bayesian Optimization automates hyperparameter tuning. The model is evaluated on real data from Brazil’s Candiota III coal-fired power plant. The proposed hypertuned CF-CNN-BiLSTM significantly outperformed 16 benchmarks. For 6-hour-ahead forecasting, the proposed model achieved root mean squared error (RMSE) of 1.42 × 1 0 − 3 , mean absolute percentage error (MAPE) of 0.207%, and mean squared logarithmic error (MSLE) of 1.11 × 1 0 − 6 . Statistical analysis over 50 runs confirmed robustness (low mean errors, narrow interquartile ranges), though occasional outliers occurred. The integration of CF filtering, CNN-BiLSTM, and Bayesian optimization delivers high-precision thermal generation forecasts, directly supporting emission reduction strategies. Future work should incorporate exogenous variables like weather and fuel prices.