A long short-term memory networks-informer based prediction model in coal management of thermal units

AbstractAs clean energy penetration increases, thermal power units must transition into flexible regulating sources. To mitigate the economic impact of frequent operational adjustments, developing an accurate coal consumption prediction model becomes essential. Leveraging the vast real-time operational data generated by digital transformation in power units, this study proposes a LSTM denoise-Informer prediction model, designed for hourly coal consumption forecasting in thermal units. The framework first applied LSTM’s gating mechanism to denoise raw operational data along with relevant influencing factors. Subsequently, it utilized Informer with ProbSparse self-attention to extract features from extensive time-series data and establish precise input-output mappings. Validated against two years of operational data from a thermal unit, our model demonstrates superior performance compared to standalone models such as LSTM, Informer, PatchTST, and DLinear, with significantly reduced mean absolute error (MAE) and mean square error (MSE). Besides, this study incorporated a feature-importance ablation experiment to elucidate the marginal contributions of features beyond the dominant parameter. This advancement improves hourly coal consumption prediction accuracy, enabling optimized fuel scheduling, energy efficiency gains, and economic benefits for power units, while offering a novel methodology for industrial time-series forecasting.