Unit load prediction method based on weighted just-in-time learning with spatio-temporal characteristics for gas boiler power generation process

Abstract

Load management of gas boiler generator units in metallurgical captive power plants heavily relies on operators’ coordination and scheduling. The operating load fluctuates frequently due to varying electricity demands across production processes and its sensitivity to multiple operational parameters. To accurately predict unit load, promptly reflect load changes, and ensure stable operation, we propose a unit load prediction model utilizing a weighted just-in-time learning algorithm with consideration of key parameter spatio-temporal characteristics (WJITL-ST) and a long short-term memory network with temporal pattern attention (TPA-LSTM) mechanism. First, we thoroughly analyze the process mechanism and use the maximum information coefficient to identify and select the variables most relevant to unit load as model inputs. Next, the WJITL-ST method, based on data segment retrieval, selects historical data most similar to the retrieved segments for online local modeling. The TPA-LSTM algorithm is then used to model and predict unit load. Finally, experiments using actual production data from a 150MW gas boiler generator unit in a metallurgical captive power plant show that the proposed method achieves higher prediction accuracy, and demonstrates superior performance under fluctuating operating conditions.