Modeling of Long Term Memory Effects in RF Power Amplifiers with

Dynamic Parameters
Ali Soltani Tehrani1, Thomas Eriksson1, Christian Fager2
1
Department of Signals and Systems, 2 Department of Microtechnology and Nanoscience,
Chalmers University of Technology, SE-41296, Göteborg, Sweden

Abstract—This paper presents a new radio frequency power and a new modeling equation is proposed. In [7], a circuit-
amplifier behavioral model that is capable of modeling long term based approach was used to construct a new model structure to
memory effects. The proposed model is derived by assuming include long-term memory effects with regards to the thermal
linear dependence of the parameters of a conventional model
to a long term memory parameter, which enables the model to filter of the PA.
better track the signal-induced changes of the power amplifier In all these works, the focus has been on identifying and
electrical behavior. The model is experimentally tested on a 100 W developing model structures that can model long term memory
Doherty power amplifier, with a signal that has a step-like change effects. In this work, we instead focus on deriving a model with
in power, representative of a realistic communication system with parameters that depend on the long term memory effect. This
bursty behavior. Results show that the proposed model is able
to improve the normalized mean squared error performance by enables us to extend most of the commonly derived behavioral
around 2–3 dB. models for PAs to include long term memory estimates with
relatively low complexity.
I. I NTRODUCTION II. M ODEL DESCRIPTION
The main propelling factor in the relatively short history A. Model formulation
of radio frequency power amplifier (PA) behavioral modeling Traditional PA behavioral models - which are linear in terms
has been the need for efficient low complex algorithms that are of parameters - can be written as
able to properly describe PAs for given input signal character-
istics. Communication signals in early generation mobile sys- y = Hx θ, (1)
tems had relatively constant amplitude, which allowed system where y is a vector of the baseband output samples of the
designers to utilize memoryless models. These systems were PA, Hx is a matrix consisting of column vectors of different
mainly designed for voice calls, and power amplifiers would nonlinear and memory of the baseband input signal x, and θ
routinely operate under steady-state temperature conditions. are the model parameters vector. Different behavioral models
As users demand for more services and higher data–rates solely differ in the proposed Hx . Volterra based models
increases, efficient spectral utilization became necessary and like the memory polynomial model [8], generalized memory
envelope varying communication signals have been employed. polynomial model [9] and others [1] belong to this group.
For these type of wideband and high peak to average signals, it In the model we propose here, instead of developing a
is noticed that memory effects in the power amplifier become new Hx we focus on including the long term effects in the
more pronounced. The output of a power amplifier not only parameters of the behavioral model θ. The new behavioral
depends on the current input sample, but on previous samples model can thus be written as
as well [1].
y = Hx θ(s), (2)
In [2] it was noted that there are mainly two categories of
memory effects that degrade communication signals, electrical where θ(s) are the parameters of the behavioral model that
memory effects and electrothermal. The former is attributed to depend on the long term memory estimate s. Intuitively, this
matching effects at the terminal impedances over the input corresponds to an amplifier whose physical parameters may
signal, and the latter to temperature drifts, biasing effects vary with the input signal, due to e.g. self-heating, biasing
and self heating which causes undesired effects on gain effects and etc. Assuming a simple first order dependence
variations and PA behavior [3]. In [4], it was shown that of the parameters with the long term behavior s, the new
for communication signals with wide modulation bandwidth, proposed model can be written as
the electrical short term memory dominated the behavioral
yLT[n] = Hx[n] (θ0 + s[n]θ1 ) , (3)
modeling performance.
As new communications signals and usage pattern emerge, where θ0 are the commonly modeled static parameters of
modeling longer term memory effects are once again gaining the behavioral model (static with respect to the long term
in importance. In [4] and [5], thermal networks are developed memory term), and θ1 are the dynamic parameters. Hx[n] are
to compensate for thermal gain variations. In [6] two tone the columns of any RF behavioral model structure linear with
measurements are used to identify long term memory effects parameters, for example the MP or Volterra models.

978-1-4673-1088-8/12/$31.00 ©2012 IEEE

 −48
N

1 2
N |x[n − k]|
k=0 −48.5
s[n]

NMSE (dB)
−49

M 
P 
 
(0) (1) p−1
θp,m + s[n]θp,m x[n − m] |x[n − m]|
x[n] p=1 m=0
ymodel [n] −49.5

−50
0 5000 10000 15000 20000 25000 30000
Window Length N (number of samples)
Fig. 1. Block diagram of the proposed model, with an MP model structure.

Fig. 2. Identification of window size, N , for the proposed behavioral model.

In order to be able to include long term memory effects

without going to very high model order – which is to be
avoided for both identification and run-time complexity rea- the communications signal is a WiMAX-like signal consisting
sons – or by pruning – which may not be an ideal solution – of two low power high power segments, each around 2
an estimate of the long term state variable has to be created. ms, repeated. The high power segment has 10 dB higher
Since in this work it is assumed that changes in these long average power than the low power segment. In this work,
term effects are relatively slow (the power does not switch two commonly used behavioral model structures are used for
between high and low quickly), the average input power over comparison, the memory polynomial model proposed in [8],
a finite window is used as a metric for modeling the long term and the generalized memory polynomial model [9].
memory changes. The first step is to identify the length of the window that
n
estimates the long term memory effect, N , from equation
1  (4). In this work, 50000 samples are used in the offline
s[n] = |x[k]|2 (4)
N identification stage and a full search is done which is then
k=n−N +1
used for all further analysis. The result of this full search
N is the size of the window of the finite impulse response with M = 2 and P = 5 for the MP model can be seen in
estimate. The larger the window size, the more the instan- Fig. 2. A window size of around N = 11000 samples, which
taneous power will be averaged. The block diagram for this corresponds to around 0.35 ms, was found to be the optimal
model with an MP model structure is shown in Fig. 1. setting. This value relates to the thermal time constants as well
as the parameters of the active bias circuit in the PA used.
B. Model Identification
In the first experiment, different model orders are used
In order to identify the parameters of this model, it can to evaluate the in-band performance – with the normalized
be noticed that once s[n] is calculated from (4), with some mean square error (NMSE) – and the out-of-band performance
re-writing of terms, equation (3) can be rewritten as – with the adjacent channel error power ratio (ACEPR) as
yLT = [Hx s ⊗ Hx ][θ0 θ1 ]T , (5) defined in [10]. The data is captured at a 6 dB backoff
from peak operating power. It is noticed that the proposed
where Hx is any of the commonly proposed behavioral models model has around 2.5 dB better in-band performance and
in the literature, and s ⊗ Hx is the Kronecker product of the around 3–4 dB better out-of-band performance for the different
elements of the column vector s with the columns of Hx . configurations. This can be explained by the capability of
It can be noticed that with a known s[n], the model is linear the proposed model to track the long term state (the change
with respect to the parameters θ. Therefore the unknown pa- in input amplitude level and corresponding PA self-heating),
rameters [θ0 θ1 ]T can be calculated with normal least squares while the normal memory polynomial model has to average
technique and written as the effect of the low and high power input segments, although
 −1 mainly dominated by the higher power segments errors.
[θˆ0 θˆ1 ]T = [Hx s ⊗ Hx ]T [Hx s ⊗ Hx ] [Hx s ⊗ Hx ]T y.
In Fig. 3, the instantaneous NMSE computed over blocks
(6)
of 2000 samples for the models is shown. This experiment
From equation (4), for the identification of s[n], it can be is done by first identifying the parameters of the models
noticed that the only unknown coefficient is N , the size of using the entire data set and then evaluating them blockwise
the window used to calculate the average power. The optimal on a separately measured set of data. The proposed model
value for N can be found with a full search. shows a consistent 2-4 dB better modeling performance than
conventional models. The NMSE improvement is especially
III. R ESULTS AND ANALYSIS higher in the transitions between low and high power segments
A 2.65 GHz 100 W LDMOS Doherty PA is used to test of the data, where PA behavior drifts are normally highest, and
the performance of the proposed behavioral model. In order the proposed model has around 5 dB better NMSE. This can be
to mimic bursty usage patterns in future generation systems, explained because for identification of conventional models, all

978-1-4673-1088-8/12/$31.00 ©2012 IEEE

 TABLE I
C OMPARISON OF THE PROPOSED MODEL AND MP AND GMP FOR −38
DIFFERENT MODEL ORDERS . M IS THE MEMORY DEPTH .
GMP
−40 LT−GMP
MP
(a) MP (values in dB)
−42 LT−MP
Nonlinear MP Proposed model

NMSE (dB)
order NMSE ACEPR NMSE ACEPR −44
M=4 M=2
P=5 -46.8 -59.9 -49.4 -64.2 −46
M=8 M=4
P=5 -46.9 -59.9 -49.6 -64.4 −48
M=4 M=2
−50
P=7 -47.9 -61.5 -49.7 -64.5
−52
(b) GMP (values in dB). G in all models is 1 0 10 20 30 40
Number of parameters
Nonlinear GMP Proposed model
order NMSE ACEPR NMSE ACEPR
Fig. 4. The accuracy/complexity tradeoff for both models.
M=4 M=2
P=5 -48.4 -62.4 -50.1 -65.6
M=8 M=4
P=5 -48.5 -62.7 -50.2 -65.6 the parameters with respect to a long term memory term like
M=4 M=2 average power, it is possible to accurately track abrupt changes
P=7 -48.8 -62.7 -50.5 -65.9
in input signals, and the modeling accuracy is improved by
around 2.5 dB. The ability of the proposed model to track
signal characteristic changes can be important for linearizing
1 −42
MP algorithms and digital predistortion, and can help lessen the
0.8 GMP load on parameter adaptation algorithms.
0.6 LT−MP −44
0.4 LT−GMP ACKNOWLEDGEMENT
Amplitude of x

|x[n]|
NMSE (dB)

0.2 −46 This research has been carried out in GigaHertz Centre in a
0 joint project financed by the Swedish Governmental Agency
−48 for Innovation Systems (VINNOVA), Chalmers University of
Technology, Ericsson AB, Infineon Technologies Austria AG,
−50 and NXP Semiconductors BV.
R EFERENCES
−52
0 2 4 6 8 [1] J. C. Pedro and S. A. Maas, “A comparative overview of microwave and
Time (ms)
wireless power-amplifier behavioral modeling approaches,” IEEE Trans.
Microw. Theory Tech., vol. 53, no. 4, pp. 1150–1163, 2005.
Fig. 3. Amplitude of input signal used (left axis), and comparison of model [2] J. H. K. Vuolevi, T. Rahkonen, and M. J. P. A., “Measurement technique
performance vs time (right axis). Model orders are chosen for relatively similar for characterizing memory effecs in RF power amplifiers,” IEEE Trans.
complexity of each model. Mircowave Theory Tech., vol. 49, pp. 1381–1389, Aug. 2001.
[3] H. Ku and K. J. S., “Behavioral modeling of nonlinear RF power
amplifiers considering memory effects,” IEEE Trans. Mircowave Theory
Tech., vol. 51, pp. 2495–2504, Dec. 2003.
the different characteristic changes in the PA are averaged. It [4] S. Boumaiza and G. F. M, “Thermal memory effects modeling and
is also interesting to note that the performance of the proposed compensation in RF power amplifiers and predistortion linearizers,”
model for the beginning of the cycle is slightly worse. This IEEE Trans. Mircowave Theory Tech., vol. 51, pp. 2427–2433, 2003.
[5] J. Mazeau, R. Sommet, D. Caban-Chastas, E. Gatard, R. Quere, and
is because the long term memory estimate is initialized at Y. Mancuso, “Behavioral thermal modeling for microwave power ampli-
zero, and takes some time to ramp up to the correct values. fier design,” IEEE Trans. Microw. Theory Tech., vol. 55, pp. 2290–2297,
This effect is not noticed in the second cycle as the long term Nov. 2007.
[6] E. Ngoya, C. Quindroit, and J. Nebus, “Improvements on long term
memory estimate is consistent now. memory modeling in power amplifiers,” in Proc. IEEE MTT-S Int. Micr.
Fig. 4 shows the NMSE vs number of parameters tradeoff Symp. Dig., 2009, pp. 1357–1360.
for the models as proposed in [10]. It is noticed that except for [7] C. Crespo-Cadenas, J. Reina-Tosina, and M. J. Madero-Ayora, “An
equivalent circuit-based approach to behavioral modeling of long-term
the low parameter region, the proposed model has better mod- memory effects in wideband amplifiers,” Microw. Opt. Techn. Letters,
eling accuracy by around 2 dB compared to their traditional vol. 53, no. 10, pp. 2278–2281, 2011.
counterparts. [8] J. Kim and K. Konstantinou, “Digital predistortion of wideband signals
based on power amplifier model with memory,” Electron. Lett., vol. 27,
IV. C ONCLUSIONS pp. 1417–1418, 2001.
[9] D. Morgan, Z. Ma, J. Kim, M. Zierdt, and J. Pastalan, “A generalized
In this work a new power amplifier behavioral model memory polynomial model for digital predistortion of rf power ampli-
capable of modeling long term memory effects was presented. fiers,” IEEE Trans. Sig. Proc., vol. 54, pp. 3852 –3860, oct. 2006.
[10] A. S.Tehrani, H. Cao, S. Afsardoost, T. Eriksson, M. Isaksson, and
In particular, we demonstrated that it is well suited to handle C. Fager, “A comparative analysis of the complexity/accuracy tradeoff in
bursty data which result as communication traffic shifts from power amplifier behavioral models,” IEEE Trans. Microw. Theory Tech.,
voice to packet based data. The results show that by linearizing vol. 58, pp. 1510–1520, 2010.

978-1-4673-1088-8/12/$31.00 ©2012 IEEE