Energy Penalty Prediction for Post-Combustion Carbon Capture Driven by Process Measurements Using a Neural Hawkes Transformer Model
Authors
Andrei Popescu, Rareș Ciobanu, Cătălin EnacheFiles
Abstract
Develop a Neural Hawkes Transformer model in this paper to predict the energy penalty of post-combustion carbon capture based on process data. The aim is to find high-energy operating periods where asynchronous changes in flue-gas load, inlet CO2 concentration, solvent circulation, absorber temperature, reboiler steam pressure, analyser delay and capture-rate targets reduce efficiency. First, normalise the process records, add load-ramp and solvent-regeneration intervals, and set up a measurement-event sequence to save both the continuous operating trend and irregular disturbance time. A Transformer branch models long-range dependencies among capture load, solvent state, heat duty, cooling demand, and product-gas quality, and a Neural Hawkes branch estimates the intensity of penalty-relevant process disturbances, such as delayed analyser response, sudden steam fluctuation, and abnormal solvent recovery. The fused representation predicts a short-term energy penalty and outputs an operator-review warning score with interpretable process evidence. A simulated dataset of 21,600 hourly records from 10 post-combustion capture units includes stable capture, load ramp, solvent deviation, steam disturbance, cooling mismatch, and sensor-noise states. Compared with LSTM, temporal convolution, Transformer-only and Hawkes-only baselines, the proposed model has reduced MAE by 0.0231 percentage points, improved high-penalty recall to 0.872, and kept the false-alarm rate at 5.7%. Based on the above experiments, combining long-range sequence representation with event-intensity model can better identify inefficient capture intervals. The model will prioritise alarms, support operator review and process diagnosis, but will not automatically change the control.