Soft-computing models for soot-blowing optimization in coal-fired utility boilers

Abstract

Fouling and slagging are classical difficulties in pulverized fuel utility boilers, which cause important degradation and dramatic reduction of efficiency. Current cleaning devices, traditionally based on prefixed soot-blowing manoeuvres, mitigate the problem only partially and offers clear optimization possibilities. But the development of predictive control is far from easy: soot-blowing represents a significant fraction of efficiency, the effectiveness has an intrinsic level of randomness and the fouling dynamics involves nonlinear feedback loops. The complexity of the phenomenon makes not applicable theoretical models or statistical analysis. The problem requires the application of alternative methodologies as expert systems or soft-computing based techniques. The present paper aims to develop a probabilistic model to predict the effectiveness of soot-blowing based in Artificial Neural Networks and Adaptive Neuro-Fuzzy Inference Systems. The validity of the model has been illustrated in a real case-study boiler, a 350 MW_e Spanish power station. Training and test data for these models are provided by a monitoring system based on heat-flux measurements in the furnace water-walls. For evaluation and comparison purposes, the quality of prediction obtained for the mentioned algorithms is analyzed in terms of performance indices. The connection weight approach reveals the relative importance of input variables in each soft-computing model. Finally, the integration of these models into an advisory tool is discussed.

Introduction

Deposition of ash particles during combustion on heat transfer surfaces is an undesired phenomenon, which causes important losses of availability and efficiency in thermal power stations, and may involve an annual economic impact of $1.2 billion only in US [1]. In addition, these troubles are more critical in biomass combustion and they make difficult the development of efficient technologies for industrial applications.

The deposits are classified in two categories: fused deposits in the radiant zone are known as slagging, while fouling are sintered deposits in the convective passes.

On-load cleaning of hardened deposits is usually performed by means of soot-blowers of air, steam or water, located on furnace and convective passes. Traditionally, the sequencing of soot-blowing has been conducted upon a periodic schedule or under the criteria of operator, resulting in over-blowing in many cases. However, there are penalties concerning soot-blowing to be considered. On the one hand, each blowing manoeuvre implies a loss of efficiency, as it represents an energetic and economic cost. On the other hand, soot-blowing over a clean section causes tube erosion, accelerating the degradation of surface tubes and reducing the longevity of the boiler.

So, the problem is of an optimization nature, as it aims to minimize the combined cost of fouling and soot-blowing operations. This fact was recognized very early in the field of boiler design [2], along with the possibilities of automated decision controls about soot-blowing activation to substitute preset sequences [3]. Intelligent soot-blowing (ISB) must be a combination of modelling, soft-computing techniques and on-line systems to monitor the boiler performance.

Consulting companies and some boiler manufactures have developed automatic systems with the goal of optimizing soot-blowing. In general, they are based in a monitoring system of heat-flux transfer and temperature of certain points of the boiler [4], [5], [6], [7], [8]. So, the optimized strategies of soot-blowing are established according to certain dirty levels or temperature thresholds [9], [10], [11], [12], [13], [14], [15], [16]. But published information about technical details is, unfortunately, very scarce. Many studies report great advantages in real applications, without even describing the techniques used (see, for example, Ref. [17]).

Artificial Neural Networks (ANN) have shown their versatility to model the behavior of complex systems in a wide number of scientific and commercial areas, including energy systems and fouling phenomena [18], [19], [20]. The main advantage of ANN is that it does not need any mathematical model, since it learns from historical data to recognize non-evident relations and patterns in a set of input–output variables, without any prior assumption about their nature.

Fuzzy logic and fuzzy set theory, introduced in 1965 [21], embrace a wide set of inference methods that interpret the values of an input vector according to certain simple rules and assign values to the output vectors. There are two main types of fuzzy inference systems, Mamdani-type [22] and Sugeno-type [23], varied in the way outputs are determined. Mamdani’s fuzzy inference method, which is the most commonly used, was proposed as an attempt to control a steam engine and boiler combination by implementing a set of linguistic control rules extracted from operator’s experience. A combination of neural network concept and fuzzy logic of Sugeno’s type has given rise to neuro-adaptive fuzzy networks (ANFIS) [24], which are able to automatically adjust the internal parameters in a learning process similar to that in ANN.

The use of ANN in the context of utility boilers soot-blowing optimization was first suggested in Ref. [25]. Several papers report the successful application of knowledge systems [26] or neural networks combined with simple models of deposits as a basis for theoretical development [27]. A previous work of Teruel et al. [28] was the first detailed publication about neural network models to predict fouling, in which this method was applied to a 350 MW_e utility boiler furnace. ANN together with expert systems have been proposed for the optimized control of fouling in the superheater and convective passes of a 50 MW_e biomass boiler [29]. Fouling and slagging in these areas show much less randomness than the furnace, and it is not measured locally but using indirect techniques based on thermal balance models. The outcome is a control strategy that compares the energetic cost of cleaning manoeuvres with the thermal improvements and decides about the activation of soot-blowers, representing potential savings up to 12 GWh/year in the case study.

The predictive system developed in Ref. [28] was divided into three local neural networks of simple and similar architecture. Preliminary results were excellent and very promising with respect to an application to an optimal control of soot-blowers’ strategies. The present paper aims to show the robustness and adequacy for control of such procedure. The main purposes of this work are threefold: (1) to improve the prediction model presented in [28], both introducing technical modifications, and model definition changes to facilitate the integration in an intelligent control system for soot-blowing; (2) to systematically evaluate the quality of predictions as an advisory tool for the operator staff; (3) to demonstrate the robustness of the method, by comparison of the predictions for several historical records and using two different techniques, Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS).

The paper is organized as follows. Section 2 recalls the soft-computing techniques used for these tasks: ANN and ANFIS. The development of the improved predictive models based on historical data series is described in detail in Section 3. The quality of predictions for several historical data series is discussed in Section 4, including a comparison of ANN and ANFIS models. Section 5 outlines the application of the developed models as an advisory tool to recommend soot-blowing.