A Multi-Objective Optimization Framework That Incorporates Interpretable CatBoost and Modified Slime Mould Algorithm to Resolve Boiler Combustion Optimization Problem Authors: Shan Gao, Yunpeng Ma Published: 2024 ABSTRACT The combustion optimization problem of the circulation fluidized bed boiler is regarded as a difficult multi-objective optimization problem that requires simultaneously improving the boiler thermal efficiency and reducing the NOx emissions concentration. In order to solve the above-mentioned problem, a new multi-objective optimization framework that incorporates an interpretable CatBoost model and modified slime mould algorithm is proposed. Firstly, the interpretable CatBoost model combined with TreeSHAP is applied to model the boiler thermal efficiency and NOx emissions concentration. Simultaneously, data correlation analysis is conducted based on the established models. Finally, a kind of modified slime mould algorithm is proposed and used to optimize the adjustable operation parameters of one 330 MW circulation fluidized bed boiler. The experimental results show that the proposed framework can effectively improve the boiler thermal efficiency and reduce the NOx emissions concentration, where the average optimization ratio for thermal efficiency reaches +0.68%, the average optimization ratio for NOx emission concentration reaches −37.55%, and the average optimization time is 6.40 s. In addition, the superiority of the proposed method is demonstrated by ten benchmark testing functions and two constrained optimization problems. Therefore, the proposed framework is an effective artificial intelligence approach for the modeling and optimization of complex systems. 1. Introduction 1.1. Background At present, thermal power generation still dominates the supply of electric energy, and coal is the main thermal power supply resource. However, as a kind of non-renewable energy source, coal is becoming more and more scarce with the massive consumption of thermal power generation. And the combustion process will be accompanied by large amounts of pollutants discharged into the atmosphere, such as NOx and SO2, which seriously affect the atmospheric environment and human living. Therefore, it is an urgent problem to improve boiler’s thermal efficiency and reduce NOx emissions for thermal power generation. The above-mentioned problem is regarded as a multi-objective optimization problem, which is called a boiler combustion optimization problem, and it is also an important research direction [1]. It is widely known that the boiler combustion system has complex characteristics with non-linearity, strong coupling and large hysteresis, making it difficult to be optimized by traditional mathematical and statistical methods. To address this problem, many researchers commonly used artificial neural networks to build the thermal efficiency model or a NOx emissions model based on the boiler’s historical operation data [2,3,4,5]. And based on the established model, all kinds of optimization algorithms were applied to adjust the running parameters for improving the thermal efficiency and reducing NOx emissions. In addition, solving the optimization problem of the boiler combustion process is beneficial to the ecological modernization of power plants [6,7]. For modeling, Zhou et al. [8,9,10] applied artificial neural networks (ANNs) and support vector machine (SVM) variants to develop the prediction models for thermal efficiency and NOx emission concentration. The experimental results showed that ANNs or SVM variants can obtain good prediction results. These research results had an important impact on the subsequent research on the boiler combustion optimization problem. Si et al. [11] applied an improved SVM to achieve a dynamic adaptive model of NOx emission concentration, which provided a new idea for online modeling. Li et al. applied the extreme learning machine (ELM) and fast learning network (FLN) to model thermal efficiency and NOx emission [12,13,14]. The experimental results showed that the ELM and FLN can obtain higher accuracy and faster computing speed than ANNs and SVMs. Ma et al. [15] proposed a hyper-parameter self-optimized broad learning system to model the thermal efficiency and emission concentrations of Nox and SO2. For optimization, researchers were keen on applying meta-heuristic optimization algorithms to solve the boiler combustion optimization problem [16,17,18]. Rahat et al. [19] applied a kind of novel multi-objective optimization algorithm to solve the contradiction between thermal efficiency and NOx emission concentration. Wang et al. [20] combined random forest (RF) with an improved flower pollination algorithm to build a NOx emission concentration model and reduce NOx emissions, separately. Hu et al. [21] used a deep neural network and multi-objective optimization algorithm to balance boiler thermal efficiency and NOx emission concentration. Three years ago, Xu et al. proposed several methods to solve the boiler combustion optimization problem, including the ICFAR-IPSO-LSTM-IMOPSO model [22], CIPSODM-LSTM algorithm framework [23], data-driven evolutionary optimization and online retrieval method [24]. In our group, Li and Ma et al. [25,26,27] used an improved artificial bee colony (ABC) and teaching and learning optimization (TLBO) to enhance the thermal efficiency and reduce the NOx emission concentration. In summary, artificial neural networks and meta-heuristic optimization algorithms can effectively solve the boiler combustion optimization problem. 1.2. Motivation In this paper, a new multi-objective optimization framework is proposed to solve the boiler combustion optimization problem. The proposed framework combines the interpretable CatBoost model by TreeSHAP and a modified slime mould algorithm (MSMA). The interpretable CatBoost model is firstly used to build the synthetic model of boiler thermal efficiency and NOx emissions based on the historical operation data of one 330 MW circulation fluidized bed boiler (CFBB). Based on the established synthetic model, the modified slime mould algorithm is used to optimize the CFBB’s adjustment operation parameters for improved thermal efficiency and reduced NOx emissions. The experimental results show that the interpretable CatBoost model can obtain better model performance than RF [28,29], SVMs [30,31] and MLP [32,33,34], and the MSMA can improve boiler thermal efficiency and reduce NOx emissions. Now, the new proposed framework needs to be explained. For the model method, CatBoost [35,36,37] is a kind of novel ensemble learning model that was first proposed by Yandex in 2017, and TreeSHAP [38,39] is an interpretable method for black box models that was proposed by Lundberg in 2017. Combining CatBoost with TreeSHAP can enhance its interpretability, which is called the interpretable CatBoost model in this paper. For the optimization algorithm, a slime mould algorithm (SMA) [40] is a kind of novel meta-heuristic algorithm that was first proposed by Lee et al. in 2020, which mainly simulates the changes of behavior and morphology of slime mould in nature during foraging. The SMA has several advantages, including a simple principle, few parameters and good global search capability. In order to further enhance the convergence accuracy and convergence speed of the original SMA, two mechanisms are proposed and introduced in the SMA. Firstly, iterative chaotic mapping with infinite collapse (ICMIC) [41] is used to initialize the population individuals for increasing population diversity and enhancing global search capability. Secondly, a reverse interpolation strategy (RIS) is proposed on the basis of opposition-based learning (OBL) [42,43] to update the positions of the population individuals obtained from each iteration, which helps to improve the quality of the solution and strengthen the local search ability. Compared with five other state-of-the-art meta-heuristic algorithms (including the butterfly optimization algorithm (BOA) [44], whale optimization algorithm (WOA) [45], seagull optimization algorithm (SOA) [46] and sparrow search algorithm (SSA) [47]) on 10 benchmark testing functions and two classical constrained optimization problems, the proposed MSMA outperforms on most testing functions. Moreover, compared to the original SMA, the MSMA is able to obtain better or equivalent convergence accuracy with less convergence time without increasing the calculation burden. 1.3. Contributions In order to solve the boiler combustion optimization problem, a new multi-objective optimization framework is firstly designed in this paper. The contributions of this paper are summarized as follows:A kind of novel multi-objective optimization framework is proposed, which combines interpretable CatBoost by TreeSHAP and a modified slime mould algorithm. Particularly, this framework has model interpretability and adaptivity.A kind of modified slime mould algorithm is proposed to solve benchmark testing functions and constrained optimization problems. A kind of novel multi-objective optimization framework is proposed, which combines interpretable CatBoost by TreeSHAP and a modified slime mould algorithm. Particularly, this framework has model interpretability and adaptivity. A kind of novel multi-objective optimization framework is proposed, which combines interpretable CatBoost by TreeSHAP and a modified slime mould algorithm. Particularly, this framework has model interpretability and adaptivity. A kind of modified slime mould algorithm is proposed to solve benchmark testing functions and constrained optimization problems. A kind of modified slime mould algorithm is proposed to solve benchmark testing functions and constrained optimization problems. The structure of this paper is as follows: the study object and basic methods are introduced in Section 2; the modified SMA is presented in Section 3; Section 4 contains the application of the proposed framework to the boiler combustion optimization problem; and the conclusions are presented in Section 5. 2. Related Work 2.1. Study Object The study object of this paper is a 330 MW circulating fluidized bed boiler (CFBB). The main study focus of this paper is to maximize the thermal efficiency (TE) while minimizing the NOx emissions concentration of the CFBB. To achieve the optimization goal, a kind of new multi-objective optimization framework integrating the interpretable CatBoost model and MSMA is proposed in this paper. The flow chart of the proposed framework is shown in Figure 1, and the core steps are summarized as follows: Step 1. The interpretable CatBoost model combined with TreeSHAP is used to build the synthetic model of boiler thermal efficiency and NOx emissions based on historical operation data. Note that the detailed descriptions of the interpretable CatBoost and TreeSHAP are given in the literature [35,36,37,38,39], separately. Step 2. Based on the established model in step 1, the value ranges of adjustable parameters of the CFBB with positive impact on thermal efficiency and negative impact on NOx emissions are obtained by further analyzing the features affecting the thermal efficiency and NOx emission concentration. Step 3. Based on the established model and selected parameter features, the boiler’s adjustment parameters are optimized by a modified SMA with the fitness function of maximizing the thermal efficiency and minimizing the NOx emission concentration of the CFBB. And the constraints for the value ranges of the adjustable parameters of the CFBB are obtained in step 2. 2.2. Slime Mould Algorithm (SMA) In this subsection, a detailed descriptions of the slime mould algorithm is given. The specific content of the SMA is explained below. In order to understand the following method, the symbols used in this section are explained in Table 1. The slime mould algorithm (SMA) is a kind of novel meta-heuristic algorithm that was first proposed by Lee et al. in 2020, which is inspired by foraging behavior and the morphology of slime mould. Slime mould is a eukaryotic organism that lives in a humid and cold environment and takes its nutrients mainly from external organic matter. When slime mould approaches its food source, its bio-oscillator generates a propagation wave through the vein to increase the flow of cytoplasm. The higher the concentration of food, the stronger the propagation wave generated by the bio-oscillator and the faster the cytoplasmic flow. The SMA has a simple principle, few adjustment parameters and strong global search capability. Therefore, the SMA is used to solve the boiler combustion optimization problem. Its basic principle is as follows. (1) The stage of approaching food In the approaching food phase, the slime mould approaches the food through the odor in the air, and the expression of its approaching behavior is shown in Equation (1). (1)X(t+1)=Xbest(t)+vb⁡W×XAt−XBt, r