AI Adoption for Steam Boiler Trip Prevention in Thermal Power Plants

Intelligent monitoring systems (IMSs) are efficient and successful tools for identifying equipment malfunctions [1]. Enhanced performance and precision are observed when adopting a fusion of multiple IMSs, like the combination of “Artificial Neural Network" (ANN) and “Genetic Algorithm” (GA), resulting in the formation of hybrid “artificial intelligent systems” (AIS). An area of notable significance for their application lies in the context of steam boilers within small to medium industries and “Thermal Power Plants” (TPP). The early detection of faults and anticipating potential undesired incidents significantly contribute to a more secure and dependable industrial operation [2]. Consequently, the detection and diagnosis of faults are imperative to uphold the safety standards of power plants.

Muselli and Ridella [3] introduced a method that melds GA with simulated annealing to generate and select a set of points within the network connection weight space. This integration accelerates the reliability and convergence of the IMS. The resulting system, aptly named “interval genetic algorithm”, was tailored for applications encompassing boilers, superheaters, and turbines. In a similar vein, Guglielmi et al. [4] harnessed “multilayer feed-forward” (MFF) and radial basis function NN to address prevailing fault detection and diagnosis challenges within an online 320 MW power plant system, inclusive of four feed heaters along the high-pressure water line.

Zhou et al. [5] presented a novel fault diagnosis methodology that intertwines GA and “Classical Probability” techniques, drawing from expert knowledge and data derived from a comprehensive 950 MW Beijing nuclear power plant. Meanwhile, Shi et al. [6] employed a combination of RBF neural network and GA to devise an automated fault diagnosis system for a nuclear power plant, thereby enhancing the accuracy and practicality of the diagnostic process, particularly in cases characterized by non-typical scenarios.

Agrawal et al. [7] introduced an online model for fault detection and diagnosis, employing “Fuzzy Logic” (FL) to evaluate residuals and a Bayesian network for troubleshooting. The findings underscore the positive impact of integrating FL with the Bayesian network in enhancing fault detection and diagnosis. Romeo and Gareta [8] devised an algorithm using NN to monitor a biomass boiler, highlighting the advantages of NN in this context. Rusinowski and Stanek [9] established an NN-based model for steam boilers, utilizing operational measurement data for material and energy balances. Alnaimi et al. [10] constructed an IMS integrating four distinct ANN algorithms. This IMS was evaluated using real-time data sourced from a Malaysian power plant, and the results were juxtaposed with corresponding “Root Mean Square” (RMS) calculated values. Notably, the optimal NN structure was identified using the “Broyden-Fletcher-Goldfarb” (BFG) “Quasi-Newton Training Algorithm”.

Fast and Palme [11] implemented an online system for monitoring and diagnosing “Combined Heat Power” (CHP) plant components, leveraging ANN. This system was smoothly incorporated into the control unit of the plant alongside a power generation and information management server. They developed a user-friendly graphical interface embedded in the CHP information management system. Nistah et al. [12] designed a boiler fault prediction model, striving for minimal misclassification rates and mean squared errors through ANN. They trained the ANN with a set of operational parameters and validated the model against real fault values from an operational plant, achieving an average accuracy of 92%.

Ismail et al. [13] devised an AIS model to predict tube leakage in steam boilers using ANN, training, and validating the model with operational data from a TPP. Their results underscored the superiority of employing a single hidden layer over a two-layer configuration in the feed-forward with backpropagation architecture. Selvi et al. [14] compared a static and dynamic ANN-based model for predicting water levels in a boiler drum. The static model, constructed with a feed-forward architecture containing a single hidden layer of 28 neurons, was juxtaposed with the dynamic model utilizing “NARX Architecture,” with a single hidden layer featuring 9 neurons. Both models were validated and assessed using real operational data obtained from an Indian plant, revealing superior prediction accuracy in the case of the dynamic model.

Panchal and Kumar [15] proposed implementing a FL methodology and formulated a model for a plant's reliability and risk analysis. The authors propose that the limitations often encountered when employing outdated “Failure Modes and Effects Analysis” (FMEA) can be surmounted using a decision support system grounded in FL. Smrekar et al. [16] constructed an ANN framework using plant-derived data to predict upcoming steam properties. The insights of plant operators influenced the original input parameters, and the final selection of input parameters was fine-tuned for optimal performance. Iliyas et al. [17] successfully crafted a predictive model for NOx and O₂ emissions using an RBF neural network, validating this model against actual data from an operational plant.

De et al. [18] detailed a “feed-forward with backpropagation” (FFBP) ANN model designed for biomass and coal-fired CHP plants. This model was trained using data from the plant to predict its performance. Moreover, in their comprehensive review paper, Nistah et al. [19] analysed current techniques in developing IMS for “Coal-fired Power Plants”. Their conclusion endorsed the viability of incorporating remote accessibility and seamless interaction between plant operators and the control system interface, highlighting the IMS's accurate trip prediction capabilities.

Victor et al. [20] embarked on the development of an AI model for a “Pressurized Water Reactor” (PWR) within a “Nuclear Power Plant”.

They fashioned an FL controller for the PWR pressurizer, the parameters of which were modelled through ANN. Data from a 2785 MW thermal Westinghouse 3-loop PWR simulator was employed to validate the pressurizer ANN model and the FL controllers. Simulation outcomes demonstrated a reasonable agreement between the developed ANN model's responses and the simulated power plant, showcasing the FL controllers' enhanced performance compared to conventional ones.

Mayadevi et al. [21] had recently provided a comprehensive survey encompassing diverse applications of expert systems within power generation plants. The publication highlights the notable technological progression of expert systems and their harmonization with contemporary methodologies, including FL, NN, machine learning, and computerized data acquisition systems. The study affirms that expert systems possess substantial potential to alleviate the operational burden on plant operators and experts, effectively serving as fault diagnosis and maintenance experts for the plant. Furthermore, the integration and fusion of various intelligent systems have expanded the problem-solving capabilities of expert systems, contributing to an enriched scope of solutions.

It can be deduced that prior research has predominantly employed a one-hidden layer (1HL) architecture for ANNs in the context of fault detection and diagnosis. Additionally, the training of IMSs has largely relied on simulation data. Some studies have incorporated the insights of plant operators as inputs for NNs. However, a comprehensive hybrid intelligent system framework combining NN and GA has not been previously formulated for diagnosing steam boiler trips within TPPs. It is noted that a standardized approach for real data preparation in IMSs, elucidating the nature of boiler operational variables, has not yet been established. Furthermore, the topologies of NN-based IMSs have received little attention.

The primary objective of the current study is to devise two IMSs to diagnose steam boiler trips within a three-unit 700 MW coal-fired thermal power plant (CFTPP). The case study encompassed six instances of steam boiler trips. The initial AIS, designated as IMS-I, employs a pure ANN configuration with influential variables selected based on the expertise of plant operators. The second AIS, referred to as IMS-II, adopts a hybrid structure combining ANN and GA. In IMS-II, GA substitutes the plant operators' decision-making process to identify the most influential variables for each of the six trips. Real operational data associated with the six steam boiler trips and relevant operational variables were collected to train and validate the two developed IMSs. The research also outlines the data acquisition and processing procedure instrumental in training and validating both pure and hybrid AIS.

The article is organized into four distinct sections. The first section conducts a review to identify gaps and delineate the problem statement concerning boiler trip diagnosis. The second section delves into problem identification and presents specifications relevant to the power plant. Section 3 comprehensively elaborates on the methodology and procedures adopted, encompassing data acquisition, manipulation, and the creation of the two IMSs. Section 4 showcases and analyses the diagnostic outcomes and compares the performance of the developed IMSs. Finally, the article concludes in Section 5, summarizing the insights gained and accomplishments derived from the case study.

The primary objective of this case study is to assess the performance of the “pure ANN (IMS-I)”, “hybrid ANN-GA (IMS-II)”, and the corresponding IMS-I and IMS-II systems in diagnosing various types of steam boiler trips. The underlying hypothesis posits that the hybrid AIS developed would exhibit superior sophistication to the pure AIS for these diagnostic tasks. This automated approach aims to surmount challenges intrinsic to human intervention within the pure AIS, such as entrapment in local minima or inadequate search exploration.

The methodology employed In the research Is visually delineated in Figure 2, commencing from acquiring data within the TPP and extending through four distinct phases to culminate in the generated outcomes. The code development for modelling the power plant’s boiler is executed across these four sequential phases, as graphically depicted in Figure 2.

2.png

Figure 2. Phases of methodological execution

3.1 Phase 1: Data acquisition and manning

Due to the significance of plant data preparation, a structured framework has been introduced for six prevalent boiler trips. These specific trips, identified by experienced plant engineers as the most recurrent culprits behind the occurrences of trips, are associated with relevant operational variables. The schematic representation of the modelling procedure is provided in Figure 3. The comprehensive interpretation of plant data involves the following procedural steps:

Step 1: Data Identification for Preliminary Analysis – The operational variables of the boiler are pinpointed and collected for each distinct boiler trip, as listed in Table A1.

Step 2: Data Preprocessing – During this phase, the data is subjected to preprocessing, encompassing the filtration of noisy and defective data points, followed by their normalization within the range of one to zero. Manual intervention was employed to identify and remove any data points deviating beyond the ± 0.8 thresholds, regarded as noise. It's important to note that numerous instances exhibited no scattering of signals beyond the ± 80% range of the sensor’s mean value.

Step 3: Data Post-analysis – This phase involves segmenting the data into two separate sets for each trip: Data-set-A and Data-set-B. Data set A constitutes 70% of the complete data set for each trip, utilized for the initial training of the developed ANN. On the other hand, Data Set B, comprising 30% of the total trip data, is reserved for validation purposes. It is pertinent to mention that the chosen percentage allocation results from iterative experimentation. Notably, every phase of the plant data preparation incorporates distinct inter-phases, contributing to the overall process.

3.png

Figure 3. Data preparation scheme

3.2 Phase 2: Development of IMS-I (Pure ANN)

The fundamental training approach employed in this study is the FFBP Training Algorithm. This multidimensional minimization algorithm has undergone several adaptations to reduce errors effectively.

3.2.1 IMS-I structure

As elucidated in the preceding section, the thirty-two variables of steam boilers are designated as the inputs for the ANN. Each individual ANN model encompasses two outputs, where “0” signifies normal operation, and “1” denotes a malfunctioning state (an operational trip). Throughout the training and validation of the pure IMS, the outputs manifest as continuous values ranging between 0 and 1. Consequently, based on the outcomes yielded by the ANN model, a decision-making process is employed to determine the optimal meaningful threshold value that distinguishes between faulty and normal operations.

Numerous ANN configurations were investigated, encompassing 1 and 2 hidden layers, 1 to 10 neurons for each hidden layer, 3 distinct activation functions, and 4 training algorithms. To ascertain the accuracy of IMS-I, the “Root Mean Square Error” (RMSE) of the NN outputs (an indicator of steam boiler trips) is scrutinized against a novel dataset that hasn’t been employed in the NN model’s training. This procedure called the validation of the developed IMS-I, ensures the reliability of the model’s performance. The training of IMS-I was conducted utilizing the tools available in MATLAB. The structure of IMS-I, incorporating its input and output parameters, is visually depicted in Figure 4.

41.png

42.png

3.2.2 Training algorithms

Within the context of multidimensional minimization algorithms aimed at error reduction, the backpropagation training algorithm has undergone several adaptations.

Table 2. Convergence times for nine widely recognized Back Propagation accelerated training algorithms

Table 2 presents the convergence times observed for nine widely recognized Back Propagation accelerated training algorithms addressing a specific problem.

Only four types of minimization algorithms, as indicated in bold in Table 2, have been considered in this study. The selection process was informed by the computational times reported by Demuth et al. [22], as detailed in the subsequent sub-sections. It is worth mentioning that Mayadevi et al. [21] endorsed Rprop due to its expedited nature compared to standard steepest descent optimization methods.

3.3 Phase 3: Development of hybrid IMS-II (ANN+GA)

Gas operates as search methods inspired by the principles of natural genetics. Over time, gas has gained substantial popularity as an optimization technique, largely due to its demonstrated success in identifying optimal solutions, often outperforming more conventional optimization algorithms. According to Sivanandam and Deepa [23], typical operational characteristics of a genetic algorithm encompass:

3.3.1 Structure of the IMS-II

IMS-II represents a hybrid system that integrates GA selection for determining influential variables within the ANN, contrasting the conventional approach of relying on plant operator decisions. Numerous instances have demonstrated the robust search capabilities inherent in Gas. The process employed by Gas involves generating new populations through selection, recombination, and mutation operators.

The GA Selection Process initiates the creation of a breeding pool composed of strings with above-average fitness from the current population. This pool forms the foundation for generating new population members. Specifically, the Tournament Selection method is adopted, wherein the population is repeatedly subdivided into random tournaments, each comprising two population members. This study includes the fittest individual from each tournament in the reproducing pool. This process continues until the size of the reproducing pool matches that of the population. A notable advantage of Tournament Selection is its independence from the range of fitness values, and its concurrent implementation is straightforward.

In the context of binary-encoded Gas, the typical recombination operator is crossover. A “crossover probability” (Pc) is applied, usually between 0.6 and 0.9. This operator involves taking two strings from the reproducing pool and exchanging contiguous segments of their structures to generate two offspring. Various techniques exist for executing the crossover operation. A single-point crossover method is employed for the GA at hand, with the site along the string selected randomly.

Within the framework of the binary-encoded GA, the mutation operator introduces randomness by occasionally replacing a “Bit” with another bit through a minor mutation probability. Each string within the population might share the same value in each Bit. Consequently, crossover might not introduce any deviation from these predefined values. The mutation operation safeguards, ensuring that lost values are reintroduced, thereby maintaining diversity within the GA population.

3.3.2 Process of the hybrid system

Chromosomes serve as the representation of primary ANN topologies and the boiler's operational variables slated for optimization. These chromosomes are encoded as binary strings referred to as “genes”. Consequently, a chromosome contains as many genes as the number of ANN topologies and boiler operational variables earmarked for optimization. Figure 5(a) illustrates the Binary representation characteristic of the hybrid IMS-II.

The gene optimization process entails the creation of an initial population comprising a specified count of chromosomes, each assigned “1” or “0” to its genes at random. Within the context of selecting ANN topologies and boiler operational variables, a gene encompasses a single-bit string indicating the presence or absence of a specific plant variable.

The design of the fitness function plays a pivotal role in the application of the GA, as it dictates the optimization objective of the GA. In this study, the initial GA population is generated randomly, except for one chromosome, configured to encompass all ANN topologies and boiler operational variables. Subsequently, the generated chromosomes are assessed using the Fitness Function. This evaluation begins with encoding the chromosomes into NNs, where “1” signifies utilizing a specific boiler operational variable, and “0” signifies its omission. These NNs are then subjected to training and validation with fresh datasets. In this context, the fitness of each string is gauged by the root mean square error (RMSE), culminating in the ultimate fitness value calculation:

$(\mathrm{RMSE})_{\mathrm{ps}}=\sqrt{\sum_{\mathrm{i}}^{\mathrm{n}} \frac{\left(\mathrm{d}_{\mathrm{i}}-\mathrm{o}_{\mathrm{i}}\right)^2}{\mathrm{n}}}$          (1)

where, I = 1, 2, 3, …., n, ps is the number of training data set, and d_i and o_i are the desired and predicated outputs of node i.

Furthermore, an integral aspect of the evolutionary system involves encoding the various potential configurations of NN topologies and boiler operation variables into distinct genotypes. These NN topologies encompass factors like activation functions within hidden layers and output nodes. The training process employs multidimensional algorithms in conjunction with the backpropagation technique.

51.png

52.png

53.png

Figure 5. (a) GA binary representation, (b) WSR encompasses four essential aspects of ANN design, (c) proposed individual, comprising a total of 46 strings

A genotype is a binary sequence comprising either 0s or 1s and uniquely corresponding to a phenotype. In this context, a phenotype encompasses a set of thirty-two boiler operation variables. The “Weak Specification Scheme” (WSR), innovated and implemented within this study, employs specific associations between designated binary strings and predefined network architectures, as stipulated by the user. WSR encompasses four essential aspects of ANN design illustrated in Figure 5(b). Each genotype within the GA distinctly represents a particular phenotype, encompassing both NN topologies and training parameters. Figure 5(c) presents and explains the proposed individual, comprising a total of 46 strings.

This study developed and evaluated two artificial intelligent monitoring systems designed to detect boiler trips within a coal-fired thermal power plant. The first system, referred to as IMS-1, employs a pure ANN approach, while the second system, named IMS-2, hybridize GA techniques with ANN. These systems' efficacy was assessed using operational data obtained from six distinct trip events within the thermal power plant.

The outcomes of the investigation underscore the capability of the feed forward NN methodology to systematically explore optimal NN topology configurations for each trip, guided by the NN performance indicator, RMSE. Additionally, the integration of GA successfully facilitated the selection of superior NN topologies and specific boiler operational variables for individual boiler trips, effectively replacing subjective human judgment with a structured optimization process.

Comparing the two systems, IMS-II, which relies on optimization properties, emerges as a preferable option in many instances due to its automated approach, as opposed to IMS-I. Notably, both IMS-I and IMS-II exhibit enhanced trip detection performance compared to the actual plant control unit. However, the pure AI system, IMS-I, showcased marginally higher ability to detect all six boiler trips prior to the intervention of the plant control system. Consequently, IMS-I holds potential as an online, dependable monitoring system for TPPs.