Domain Consistent Industrial Decarbonisation of Global Coal Power Plants

TL;DR

This paper tackles the challenge of deploying machine learning–based optimisation for industrial decarbonisation in coal power plants by embedding domain expertise through a human-in-the-loop (HITL) data-driven constraint within a multi-objective optimisation framework (HITL-MLOPT). It builds predictive models (XGBoost with Hyperopt) for thermal efficiency and turbine heat rate using plant data, and uses SHAP and inductive conformal prediction for interpretability and uncertainty quantification. The HITL constraint enforces domain-consistent relationships among correlated operating variables, producing solutions that align with control-system realities. Validation on a 660 MW plant shows a TE gain of 0.64 percentage points and a THR reduction of 93 kJ/kWh, and extrapolation to global plants suggests up to 156.4 Mt of CO2 could be mitigated over their lifetimes, illustrating a scalable path to industrial decarbonisation.

Abstract

Machine learning and optimisation techniques (MLOPT) hold significant potential to accelerate the decarbonisation of industrial systems by enabling data-driven operational improvements. However, the practical application of MLOPT in industrial settings is often hindered by a lack of domain compliance and system-specific consistency, resulting in suboptimal solutions with limited real-world applicability. To address this challenge, we propose a novel human-in-the-loop (HITL) constraint-based optimisation framework that integrates domain expertise with data-driven methods, ensuring solutions are both technically sound and operationally feasible. We demonstrate the efficacy of this framework through a case study focused on enhancing the thermal efficiency and reducing the turbine heat rate of a 660 MW supercritical coal-fired power plant. By embedding domain knowledge as constraints within the optimisation process, our approach yields solutions that align with the plant's operational patterns and are seamlessly integrated into its control systems. Empirical validation confirms a mean improvement in thermal efficiency of 0.64\% and a mean reduction in turbine heat rate of 93 kJ/kWh. Scaling our analysis to 59 global coal power plants with comparable capacity and fuel type, we estimate a cumulative lifetime reduction of 156.4 million tons of carbon emissions. These results underscore the transformative potential of our HITL-MLOPT framework in delivering domain-compliant, implementable solutions for industrial decarbonisation, offering a scalable pathway to mitigate the environmental impact of coal-based power generation worldwide.

Figures (6)

• Figure 1: Design parameters of global coal power plants. (a) Major coal-based power plants operate on sub-critical technology with a median energy efficiency of 29.8% SpGlobalcoal. Main steam pressure and main steam flow rate are the two important operating parameters and the contour plots plotted against two performance parameters (b) Power (MW) and (c) Turbine Heat Rate (kJ/kWh) represent the highly nonlinear function space for optimising the power generation operation.
• Figure 2: The operating behaviour of a 660 MW supercritical coal power plant. (a) The KDE curves for input-output features show that CFR, TAF, MSP, MSF, FWT and Power have nearly similar distribution profiles, meaning that these features are changed simultaneously during the plant operation. The ECDF profiles for the identified input-correlated features are close to each other, demonstrating the synchronised behaviour of the features. (b) The $PCC$ based heat map for the input-output variables quantifies the linear dependence between the pair of variables. The power generation operation is synchronised with the state of the variables in the control layer, resulting in the large number of correlated variables in the operating variables. The low $PCC$ value for the pair of variables shows that variables can be set at different set- points during the plant operation and are independent in nature.
• Figure 3: Data-driven modelling and insight into power generation operation. (a) XGBoost models are trained under rigorous hyperparameter tuning to predict TE and THR. The predictive performance of the models is evaluated in the training and testing datasets, and the prediction intervals are also constructed around the point-predictions of the models. (b) SHAP-based model interpretability analysis is carried out to compute the feature contribution towards the model-based predictions for TE and THR. A feature with a high feature contribution value is significant towards the model-based predictions.
• Figure 4: Comparison of multi-objective optimisation analysis that attempts to maximise TE and minimise THR for coal power plant by MLOPT framework. (a)(i) The ECDF profiles and scatter plot for two most significant correlated features for TE (MSP & Power) and THR (MSP & TAF) are plotted. The solution mapping for two correlated variables after solving the optimisation problem (5) is presented in (ii). The optimisation problem (6) is solved for a tolerance value of (iii) 0.05, (iv) 0.10, (v) 0.15, (vi) 0.20, (vii) 0.25, and (viii) 0.30. The optimal values of TE and THR estimated on the quantile value of the initial guesses are presented in (b) for (i) TE and (ii) THR.
• Figure 5: Verification of the HITL-MLOPT framework-driven estimated solutions on the power plant operation. The actual values of the operating vaiables (a) MSP (MPa), (b) Power (MW), (c) TAF (t/h) and (d) CFR (t/h) are plotted corresponding to the optimal values estimated from HITL-based MLOPT analytics. (e) The optimal values of TE and THR are mapped against the actual values of TE and THR, maintained during the plant operation. The confidence intervals are drawn around the actual values with a 95% confidence level.