2.1. The Thermal Resistance Model of a Low-Temperature Superheater
As shown in
Figure 2, heat balance and the heat transfer calculation of heating surfaces are based on the basic theory of boiler thermal balance and performed one by one along opposite directions of the flue gas flow from the outlet of the air preheater. The inlet working-fluid temperature of the low-temperature superheater is calculated according to the heat balance principle—Equations (1) and (2) for the working fluid side and the flue gas side based on the known inlet and outlet gas temperatures of each heating surface, the working fluid outlet parameters and the pipe arrangement structure of each heating surface. Heat transfer Equation (7) can then be applied to calculate the heat transfer coefficient of the boiler’s low-temperature superheater.
Heat absorption on working fluid side:
Heat release on flue gas side:
where
is the convection heat absorption on the working fluid side;
is the convection heat release on the flue gas side;
is the quantity of the working fluid flow;
denotes the calculating fuel quantity;
and
represent the enthalpy of the inlet working fluid and outlet working fluid, respectively;
and
are the enthalpy of the inlet flue gas and outlet flue gas, respectively;
is the reduced value of steam enthalpy in the desuperheater;
is the heat retention coefficient;
is the air leakage coefficient; and
is the theoretical cold air enthalpy.
Physical parameters such as
,
,
,
,
,
and so on are collected in the distributed control system (DCS), and
,
,
and so on can be obtained through manual boiler design and thermal calculation. According to the heat balance equation
, the enthalpy of the inlet working fluid at the low-temperature superheater is as follows:
According to the enthalpy temperature table of superheated steam, as shown in
Figure 3, the corresponding inlet working fluid temperature
is expressed as follows:
where
is superheated steam enthalpy, and
is superheated steam pressure.
Next, the logarithmic mean temperature difference
is obtained from the inlet and outlet working fluid temperatures and the inlet and outlet flue gas temperatures, which can be described as follows:
where
is the difference between the inlet temperature of the flue gas side and the inlet temperature of the working fluid side of the low-temperature superheater, and
is the difference between the outlet temperature of the flue gas side and the outlet temperature of the working fluid side of the low-temperature superheater.
When the maximum temperature difference
and the minimum temperature difference
are satisfied using
, the logarithmic mean temperature difference
can be simplified as follows:
Finally, the actual heat transfer coefficient of the heating surface should be
where
is the heat transfer area of the low-temperature superheater.
Taking the low-temperature superheater as an example,
Figure 4 depicts the heat transfer in a single superheater tube. The tubular convective heating surface is regarded as a heat transfer model of multi-layer cylindrical wall, and the heat transfer coefficient
K is calculated as follows:
where
is the heat release coefficient of the flue gas side;
is the heat absorption coefficient for the working fluid side;
is the ash-layer thermal conductivity;
is the metal tube thermal conductivity;
is the scale-layer thermal conductivity;
is the inner radius of the scale layer;
is the inner radius of the metal tube;
is the outer radius of the metal tube;
is the outer radius of the ash layer;
is the length of the metal tube;
is the thermal resistance of the ash layer;
is the thermal resistance of the metal tube; and
is the thermal resistance of the scale layer.
Since the influence of
is small, the effect of
is ignored. Before the raw water is replenished into the boiler, the power plant must treat the boiler feed water to remove the salts, impurities and gases, so that the quality of the supply water meets certain requirements. Therefore, the thermal resistance of the scale layer
is small and can be ignored to simplify the calculation of
. Thus, the thermal resistance of the ash layer
is calculated as follows, according to the Equation (8):
where the heat release coefficient of the flue gas side
is the convective heat transfer coefficient
and the radiation heat release coefficient
, and is calculated as:
For the low-temperature convection heating surface, the radiation heat release coefficient
. When tube bundles on the convection heating surface are arranged in parallel, the heat release coefficient on the flue gas side becomes
When the tube bundles on the convection heating surface are arranged in staggered rows, the convective heat transfer coefficient on the flue gas side is calculated as
The heat absorption coefficient
on the working fluid side is given by
where
,
,
,
are the correction factors determined by the structural dimensions of tube bundles, and represent the correction factor related to pipe pitch, the correction factor associated with the vertical tube row number, the correction factor related to airflow and wall temperature and the relative length correction factor, respectively;
is the outer diameter of the low-temperature superheater tube;
is the inner diameter of the low-temperature superheater tube;
and
are the flow rate of the working fluid and flue gas at the average temperature, respectively;
and
are the thermal conductivity of the working fluid and flue gas at the average temperature, respectively;
and
are the kinematic viscosity of the working fluid and flue gas at the average temperature, respectively;
and
are the dynamic viscosity of the working fluid and flue gas at the average temperature, respectively; and
and
are the constant-pressure specific heat of the working fluid and flue gas at the average temperature.
The thermal resistance of the ash layer is used to characterize the fouling degree of the convection heating surface. Generally, the larger the value is, the more serious the fouling of the heating surface is.
2.2. Wavelet Decomposition Model
In 1988, Mallat proposed a concept of multi-resolution with a fast algorithm for wavelet decomposition and reconstruction [
29,
30]. The original signal
is a continuous wavelet. Then, the inner product of
and
, which is called the continuous wavelet transform, is described as
However, in actual engineering calculations, the signals are discrete. As a result, it is necessary to discretize the wavelet transform. The binary discrete wavelet transform (DWT) can be expressed as
The basic principle is as follows. A signal
in space
can be represented by basic functions in two orthogonal subspaces,
and
, which can be determined using Equation (16). According to Equation (17), the first level decomposes
into a low-frequency part
and a high-frequency part
, the second level decomposes
into a low-frequency part
and a high-frequency part
and so on, until the multi-resolution decomposition of signals can be realized.
In the scale metric space , the coefficient is decomposed to two wavelet coefficients and in the scale metric spaces and . Similarly, the two wavelet coefficients and can be used for reconstruction to get . The reconstruction algorithm and the decomposition algorithm are corresponding and mutually inverse.
2.4. SVR Theory
An SVR [
32,
33,
34] method is usually realized through regression and prediction. The principle of SVR is to learn a function
, so that the function value is as close as possible to the real value. Given the training samples
, with
for the input,
for the target output,
as the number of training samples. The regression model equation is calculated as
where
,
;
,
is kernel function. SVR transforms low-dimensional nonlinear problems into high-dimensional linear problems by introducing kernel functions. Solving the dual Lagrangian problem of SVR:
The above process must satisfy Karush–Kuhn–Tucker (KKT) conditions:
Eventually, the solution to SVR is
where
is the kernel function.
The kernel function can map the nonlinear problem of low-dimensional space to high-dimensional space, which will then become a linear problem. However, constructing the kernel function
is a significant problem. The most crucial step is to determine the mapping of input space to feature space, which can only be achieved if we know the distribution of data within the input space. However, in most cases, we do not know the specific distribution of data being processed. Therefore, it is generally challenging to construct a kernel function that conforms precisely to the input space, and as a result this paper uses the radial basis function (RBF) instead of rebuilding a new kernel function.
The Gaussian radial basis function is one of the most widely used kernel functions. It offers a better performance regardless of whether a sample is large or small, and it has fewer parameters than the polynomial kernel function. Therefore, the Gaussian kernel function is preferred in most cases.
Vapnike’s [
35] research demonstrates that the kernel parameter
and penalty factor
are the key factors affecting the performance of SVM. A larger
will reduce the training error, but will also result in over-fitting at the same time, which will increase the test error. When the kernel parameter
is small, the regression prediction has better accuracy. However, if the kernel parameter
is much lower, the accuracy of the model will drop significantly.