It is often difficult to determine the cause of upset or detrimental boiler conditions because the large number of boiler input and output parameters create a complex problem to troubleshoot.
Such was the case in this study, where a radiant superheater and internal desuperheater were retrofitted in an existing industrial water-tube boiler at a South African sugar mill. After installation, the steam temperature in the new radiant superheater was consistently 35% higher than the design value. Excessive final steam temperatures can lead to thermal fatigue of the superheater material and therefore boiler downtime; continuous opening of the superheater vent valve, which leads to a reduced boiler efficiency; and tripping of the boiler or turbines due to excessive inlet steam temperatures.
It was suspected that the higher-than-expected steam temperatures could be attributed to combustion dynamics, including the flame temperature and position of the flame in the furnace. However, this was difficult to model due to the sheer number of factors affecting combustion, the variances in these factors, and the degree to which these variances differed. The objective was therefore to develop a predictive model to determine the cause(s) of the excessive final steam temperatures encountered. This was done by using three machine learning models and varying the combinations of boiler input parameters to predict the final steam temperature response.
Machine learning is a type of artificial intelligence (AI). It entails supplying an algorithm with a set of data that it uses to teach itself about the underlying relationships within the data. A well-trained machine learning algorithm is therefore able to make predictions or decisions without being explicitly programmed to perform the task. In this root cause analysis, different machine learning models were compared, all of which have previously been used for various applications in the power-generation sector. These models are artificial neural networks, random forests and support vector regression.
This article focuses on what is likely the most well-known and powerful method, namely neural networks. An artificial neural network is a computing system inspired by the biological neural networks of the brain. It uses mathematical functions to simulate the working of the brain’s neurons, including dendrites, axons and synapses, and can consist of anything from a single neuron up to tens or even hundreds of millions of neurons.
Boiler and fuel information
The boiler in this study has an evaporation rate of 140 tph and final steam conditions of 30 bar(g) and 400°C, while co-firing sugarcane bagasse and furfural residue. The boiler has a three-pass evaporator bank configuration, with a two-stage superheater with interstage attemperation by means of an integral indirect contact desuperheater. The primary superheater is screened and has a drainable horizontal configuration, and the secondary is a radiant pendant superheater.
It is noteworthy that furfural residue has about 10% higher calorific content than bagasse. Another big difference is found in the particle size of the fuels; the bulk density of furfural residue is three times higher than that of bagasse. As a result, a flame fueled by furfural residue burns higher in a boiler’s furnace than would be the case with a bagasse-fueled flame.
From April to December 2016, data logs were collected for the boiler’s steam temperature and flow, flue gas oxygen content, fuel feeder speeds, induced and forced draughts, over-fire air pressures, fuel density, and fuel moisture. Although data on other parameters were collected, they were not utilised in the formulation of the learning models.
From the final steam temperature, boiler evaporation rate and flue gas oxygen data, some general trends were noted, but none that represented an obvious correlation between any one parameter and steam temperature. With regard to feeder speed vs. boiler load, a large variance in feeder speeds was found, from about 300 to about 1000 rpm when running at load.
When looking at final steam temperature vs. boiler load, it was found that the temperature varies between about 390 and 410°C for loads roughly between 80 and 140 tph – again, no obvious correlation between boiler load and steam temperature was found. After training the neural network and performing sensitivity analysis on the AI model of the boiler by varying input parameters, it was determined that the factors with the greatest influence on steam temperature were fuel quality (particle size and moisture) and furnace draught.
With the trained neural network and keeping all inputs except fuel moisture and chute density fixed at nominal values, it was found that low fuel moistures (indicative of a large percentage of furfural residue) coupled with high chute density result in high steam temperatures. Conversely, low fuel moistures and low chute densities result in low temperatures. To study the effect of furnace draught, all inputs were fixed on nominal values except forced draught (FD) damper position and chute density.
It was found that high steam temperature resulted from a large FD fan damper position and high chute density, again indicative of a large portion of furfural residue. When comparing steam temperatures calculated from inputs for test data with the actual steam temperatures experienced, the average error was found to be less than 1%, which, in this case, is less than 4°C. The other two machine learning models utilised (random forests and support vector regression) came to the same conclusion as the neural network model.
Using machine learning techniques, an accurate model of the boiler with an error of less than 1% could be produced. This model was then used to complement existing modelling techniques to redesign the superheater in question to obtain the design final steam temperature.
Machine learning has no knowledge of how a boiler works or the underlying physics at play, i.e. it has no preconceived notions of which factors affect a boiler’s operation. It can therefore objectively predict suspected correlations or detect correlations that were not previously considered. As a snapshot of the boiler in 2016, this machine learning model can be used in future as a reference to detect gradual changes in the boiler’s operation. This model furthermore opens up new opportunities for the predictive maintenance of industrial water-tube boilers.
This article was fist published in the June 2019 edition of Watts Watt and is republished here with permission.
Contact Debby Riddle, Actom, Tel 011 820-5239, firstname.lastname@example.org
The post Application of machine learning algorithms in boiler plant root cause analysis appeared first on EE Publishers.
Source: EE plublishers