This work presents the results of using the electrochemical impedance to analyse the behaviour of a BASF Celtec
P2100 MEA operated under varying operating conditions with different temperatures and gas concentrations.
Figure 1 shows the experimental setup used for these measurements.
Electrochemical impedance measurements have resulted in a full map of the impedance of a
BASF Celtec P2100 HTPEM fuel cell at different temperatures, cell current, CO and CO
2 values.
The algorithms defining the mappings to CO are derived using polynomials of the order 4 and 4, which has the following form:
CO est
(t)[%] = p
00
+ p
10 x(t) + p
01 y(t) + p
20 x(t) 2 + p
11 x(t)y(t) + p
02 y(t) 2 + p
30 x(t) 3 + p
21 x(t) 2 y(t) + p
12 x(t)y(t) 2 + p
03 y(t) 3 + p
40 x(t) 4 + p
31 x(t) 3 y(t) + p
22 x(t) 2 y(t) 2 + p
13 x(t)y(t) 3 + p
04 y(t) 4
The polynomial fits are done for measurements with i cell
(t) at the range 10 to 17.5[A], T cell
(t) in the range 155 to 175 [˚C] and CO
2 in the range 0 to 30[%]. The constant frequency is set at 105[Hz], which was placed in the middle of the desired range of interest between 100 and 110[Hz], which isn’t too slow to be uncertain, due to changing system states such as temperature and gas concentrations, but neither to fast to require fast electronics. If the fuel cell system is delivering power to a 50 [Hz] grid through a single phase inverter the frequency of the dc-link voltage will be twice of the grid frequency, i.e. 100 [Hz]. This means that by designing the dc-link capacitor properly in order to have an appropriate dc-link voltage ripple the CO-concentration can be estimated from the inherent small ac-signal of the inverter. It is therefore not necessary to have any additional hardware in order to superimpose an ac-signal in the fuel cell current. However, analyzing the impedance due to a single phase grid-inverter is left for future work as the purpose for this research is to demonstrate the relationship between the CO-concentration, temperature and impedance.
Figure 1: Single cell setup capable of mixing different concentrations of H
2
, CO and CO
2 to simulate reformate gas performance .
The 3 surface mappings derived using the possible combinations of Z
Re
(t), Z
Im
(t) and T cell
(t) are illustrated in figures 4, 5 and 6.
The setup uses two separate
Labview Real Time systems, one for controlling the fuel cell mass flows and temperature, and one for conditioning the current and interpreting the fuel cell voltage response to the galvanostatic impedance measurements.
The impedance measurement results are interpreted and can be converted into Nyquist plots, as shown in figure 2.
Figure 4: Design of the CO estimator, with the inputs: Z
Re
(t) and T cell
(t).
Figure 5: Design of the CO estimator, with the inputs: Z im
(t) and T cell
(t).
Figure 2: Nyquist plot of the shown electrical circuit. Subscripts referring to high, intermedate or low frequency (HF,IF and LF), and Ω to the ohmic losses of the fuel cell.
The plots require several measurements in different operating points, as each point in the curve requires their own operating point.
This CO estimation approach don’t require the frequency response over the spectrum, but rather one frequency. This still require several different operating points like the one presented in figure 3.
1
2
3
CO
Fit
#
Input Variables
1 st 2 nd R 2
Fit Results
Adjusted
R 2
RMSE
Z
Re
Z
Z
Re
Im
(t) Z
Im
(t) 0.800151 0.7996319 0.3825344
(t) T
(t) T cell
(t) 0.890836 0.8905519 0.2827225
cell
(t) 0.748713 0.7480604 0.4289478
Figure 6: Design of the CO estimator, with the inputs: Z
Re
(t) and Z
Im
(t).
Table 1: Table for polynomial plane error values.
The best surface mapping is according to the polynomial fit errors presented in table 1, the polynomial with the Z
Re
(t) and T cell
(t) inputs by at least 35[%]. It is followed by the polynomial with the Z
Re
(t) and Z
Im
(t) inputs, which is an improvement by around 12[%] over the last polynomial fit.
The polynomial with the Z
Re
(t) and Z
Im
(t) inputs have measurement points that are clustered in an arc, which is expected from impedance results (illustrated in figure 2), but as they don’t span the whole surface and only a part of it, then there is less robustness in the mapped surface plane.
The polynomial with the Z
Re
(t) and Z
Im
(t) inputs also have larger higher order constants than the other two. This is also visible in the figures of the polynomials as the mapped surface reach higher CO values than the other two polynomials by a factor of at least 300[%]. This indicates less robustness for the polynomial with the Z
Re
(t) and Z
Im
(t) inputs, but it doesn’t guarantee it.
Using electrochemical impedance, a BASF P2100 has been characterized in different operating points with varying temperature and gas concentrations, but same frequency. The developed experimental setup is designed such that measurements are consistent and reproducible. Increasing the number of measurements in each operating point increases the reliability of the conclusions drawn from these measurements.
Figure 12: Nyquist plot of different measurements, with 2 consecutive measurements in each operating point.
Figure 9: Plot of the CO estimation error. There are one sub graph for each CO estimation method and the order is the same as in table 1. They sub graphs are divided into areas depending on operating temperature.
The figure 9 illustrates the CO estimators 1 and 3 both fail to follow the CO value at lower temperatures, while there is no such problem with the second CO estimator. This indicate a strong relation between Z
Re
(t), T cell
(t), and CO concentration.
The presented CO estimators deliver acceptable results, but there are differences between their performances. There isn’t one algorithm that is better than others for all situations, but there is one that overall has better performance and that is the second CO estimator, which provides acceptable results in most situations, but is surpassed at lower
CO values.
The first CO estimator provides good CO estimation at 0[%], and lower CO values at increasing temperatures. This however doesn’t stop the CO estimator from giving the worst estimations at the other operating points.
The main improvement in the presented paper is the runtime CO estimation methods, which prove promising results. The other alternatives at the present time are either a gas analyzer, as known
CO sensors are cross gas sensitive, or methods under research, e.g. impedance spectroscopy.
The next step is to implement the approach on a different setup with either a reformer or a fuel cell stack. An experimental setup with a reformer can be used to test the desired improved performance, which should come with the knowledge about CO contents. Whereas a setup with a fuel cell stack rather than a experimental setup would prove the functionality of the approach on a real setup with appropriate measurement values and noise.
Figure 3: Table of the different operating points examined in this work, using different combinations of DC currents, gas concentrations and temperatures.
Figure 7: Plot of how the fitted equivalent circuit parameters vary with changing temperatures.
Figure 8: Plot of how the mean filtered (with window length of 3 measurements) fitted equivalent circuit parameters vary with changing temperatures.
In figures 7 and 8 the CO estimation results on a new set of measurements are illustrated. The methods presented follow the same order as table 1.