Modified second-order generalized integrators with modified frequency locked loop for fast harmonics estimation of distorted single-phase signals

Modified second-order generalized integrators with modified frequency locked loop for fast harmonics estimation of distorted single-phase signals


This paper proposes modified Second-Order Generalized Integrators (mSOGIs) for a fast estimation of all harmonic components of arbitrarily distorted single-phase signals such as voltages or currents in power systems. The estimation is based on the internal model principle leading to an overall observer system consisting of parallelized mSOGIs. The observer is tuned by pole placement. For a constant fundamental frequency, the observer is capable of estimating all harmonic components with prescribed settling time by choosing the observer poles appropriately. For time-varying fundamental frequencies, the harmonic estimation is combined with a modified Frequency Locked Loop (mFLL) with gain normalization, sign-correct anti-windup and rate limitation. The estimation performances of the proposed parallelized mSOGIs with and without mFLL are illustrated and validated by measurement results. The results are compared to standard approaches such as parallelized standard SOGIs (sSOGIs) and adaptive notch filters (ANFs).

Second-Order Generalized Integrator, Frequency Locked Loop, amplitude estimation, phase estimation, frequency estimation,

pstoolNo file


***This paper has been submitted to IEEE Transactions on Power Electronics***


: natural, real, complex and rational numbers. For the following, let . : column vector (where means “is defined as” and means “transposed”). : zero vector. : vector of ones. : Euclidean norm of . : real (non-square) matrix. : diagonal matrix with diagonal entries taken from vector . : zero (non-square) matrix. : identity matrix. : block diagonal matrix with matrix entries , .

I Introduction

I-a Motivation and literature review

In view of the increasing number of decentralized generation units with power electronics based grid connection and the decreasing number of large-scale generators, the overall inertia in the grid is diminishing. This results in a faster transient response and higher harmonic distortion of physical quantities (such as currents or voltages) of the power system. Fast frequency fluctuations endanger stability of the power grid. Significant harmonic distortions of voltages and currents can deteriorate power quality and lead to damage or even destruction of grid components. To be capable of taking appropriate countermeasures such as (i) improving stability and quality and (ii) compensating for such deteriorated operation conditions, it is crucial to detect and estimate fundamental and higher harmonic components of the considered quantities in real time as fast and accurate as possible. Hence, grid state estimation became of particular interest to the research community in the last years and has been studied extensively (see e.g. [2, 3, 4, 5, 1, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23] to name a few).

It is well known that a signal with significant harmonic distortion can be decomposed and analyzed by the Fast Fourier Transformation (FFT). However, this method requires a rather long computational time and a large amount of data to be processed [24, p. 320]. Usually, several multiples of the fundamental period () are needed to estimate the harmonics with acceptable accuracy [6]; when the frequency is estimated online as well, the estimation time is even longer.

The majority of the publications deals only with the estimation of fundamental signal parameters (such as amplitude and phase) and fundamental frequency. For signals with negligible harmonic distortion, several well known and rather fast methods are available [25, Chapter 4] such as Second-Order Generalized Integrator (SOGI) or Adaptive Notch Filters (ANF) with and without Phase-Locked Loop (PLL) [20] or Frequency Locked-Loop (FLL) [2, 1]. However, if the signals to be estimated have significant harmonic distortion, these approaches fail and have to be extended by the parallelization of several SOGIs (see e.g. [1, 6, 7, 12, 23]) or several ANFs (see e.g. [3, 4]); each of those being capable of estimating the individual harmonics separately. However, the resulting estimation system is highly nonlinear (in particular in combination with FLL or PLL) and difficult to tune. The estimation speed is usually faster than those of FFT approaches but still rather slow. Other estimation approaches use Adaptive Linear Kalman Filters [26], SOGIs in combination with discrete Fourier transforms [27] or circular limit cycle oscillators [28]. A comparison of estimation speed and estimation accuracy mainly focuses on fundamental signal and frequency estimation. A comparison of all the results presented in the contributions above yields that the estimation speeds vary between . The estimation speed depends on the tuning of the estimation algorithms and the operation conditions (such as changing harmonics with varying amplitudes, phases and frequencies) during the estimation process. In particular, when the frequency is changing abruptly, the overall estimation process is drastically decelerated. The FLL can be considered as the bottleneck of grid state estimation. Moreover, the performance of the parallelized estimation of the individual harmonics is mostly not discussed and evaluated.

Exceptions are the contributions [1, 6] and [3]; which explicitly discuss and show the estimation performance of the parallelized SOGIs and ANFs, respectively, for each considered harmonic component. For example, in [1], the proposed parallelized SOGIs with FLL (called MSOGI-FLL) are capable of extracting fundamental frequency and amplitudes and phases of a pre-specified number of harmonics . Local stability analysis and tuning of the parallelized SOGIs and FLL were thoroughly discussed. As outcomes of the tuning rules, the gain of the -th SOGI should be chosen to be which represents a “tradeoff between settle time, overshooting and harmonic rejection”. Simulation and measurement results were presented for three-phase signals. Six harmonics (including fundamental positive sequence) and the fundamental frequency were correctly estimated. The estimation speeds for the individual harmonics vary between . Frequency estimation takes about to return to a constant value. In [6], a similar idea is proposed. The proposed algorithm is also based on parallelized SOGIs but a FLL has not been implemented. If the frequency is known, the method is capable of estimating the harmonics in approximately 4. Only simulation results were presented for seven harmonics. No results were presented when the frequency is unknown and varying. Implementation and tuning of the parallelized SOGIs are hardly discussed. In [3], parallelized ANFs with FLL are implemented. For a constant (estimated) frequency, a complete stability proof for the parallelized structure is presented showing that stability is preserved if all gains are chosen positive. The parallelized ANFs with FLL are implemented in Matlab/Simulink to estimate a signal with six harmonic components (including fundamental). The fundamental frequency of the considered signal undergoes step-like changes of and . Frequency and harmonics estimation errors tend to zero; but the estimation speed is rather slow and varies between .

As already noted, the (parallelized) SOGIs and ANFs rely on a precise estimate of the fundamental (angular) frequency for proper functionality. If the frequency is known a priori, it can be fed directly to the parallelized systems. Otherwise, the observers must be combined with a FLL (or PLL), which allows to additionally estimate the fundamental angular frequency online. Since the FLL estimation depends on the harmonic amplitudes of the input signal, [1, 15, 13] describe a Gain Normalization (GN) which robustifies the frequency estimation. Nevertheless, due to its nonlinear and time-varying dynamics, the tuning of the overall estimator consisting of parallelized SOGIs or ANFs and FLL is a non-trivial task. As a rule of thumb (coming from the steady-state derivation of the FLL adaption law), the tuning of the FLL should be slow compared to the dynamics of the parallelized SOGIs or ANFs and, therefore, significantly degrades the settling time of the overall estimation system [8]. Apart from that, negative estimates of the angular frequency lead to instability. In this context, [10] describes a saturation of the estimated angular frequency to avoid a sign change. However, this saturation does not necessarily (i) ensure convergence of the estimation error or (ii) accelerate the transient response of the FLL. In [22], the FLL is extended by output saturation and anti-windup to avoid too large estimation values. But the proposed anti-windup strategy comes with additional feedback gain (tuning parameter), which, if not properly chosen, might lead to instability. Other approaches for frequency detection are based on Phase Locked Loops (PLLs) [17], [9, 11, 19, 20] which can be combined with SOGIs as well. PLL approaches are not considered in this paper.

The remainder of this paper focuses on modifications of the parallelized “standard SOGIs” and the “standard FLL” as introduced in [2] and [1] which will allow to improve estimation speed and estimation accuracy significantly. Key observation which motivates the modifications is that almost all papers above, except [12], do only consider one single tuning factor (gain) for individual SOGI design. This single tuning factor limits the possible estimation performance. In [12], two gains are considered but their influence on the speed of harmonics estimation is not exploited and investigated. Therefore, this work proposes a modified (generalized) algorithm which achieves a prescribed settling time of the estimation process. It is capable of estimating amplitudes, angles and angular frequencies of all harmonic components of interest in real time. The proposed algorithm consists of parallelized modified SOGIs tuned by pole placement. The modified SOGIs come with additional feedback gains (additional tuning parameters) which provide the required degrees of freedom to ensure a desired (prescribed) settling time. Since the standard FLL was derived and is working for the standard SOGI only (as shown in [1] or [7]), also a modified FLL is proposed to guarantee functionality in combination with the parallelized modified SOGIs. The novelty of this paper is characterized by the following five main contributions:

  1. Modification (generalization) of standard SOGIs (sSOGIs) to modified SOGIs (mSOGIs) with prescribed settling time (see Sections II-B and II-C);

  2. Parallelization of the mSOGIs and their tuning by pole placement (see Section II-D);

  3. Modification (generalization) of the standard FLL to the modified FLL (mFLL) with phase-correct adaption law, sign-correct anti-windup strategy and rate limitation for enhanced functionality in combination with the proposed mSOGIs (see Section III-B);

  4. Theoretical derivation of the pole placement algorithm and the generalized adaption law for the mFLL (see Appendix A-A and A-B,respectively);

  5. Implementation and validation of the proposed estimation algorithm by simulation and measurement results and Comparison of the estimation performances of parallelized mSOGIs, sSOGIs and (ANFs) with and without FLL (see Section IV).

I-B Problem statement

Single-phase grid signals (e.g. voltages or currents) with significant and arbitrary harmonic distortion are considered. The considered signals are assumed to have the following form


with fundamental amplitude , harmonic amplitudes and angles (in ), respectively; where indicates the -th harmonic component (per definition ). Observe that does not necessarily need to be a natural number or larger than one; any rational number is admissible as well (e.g.  or ). Moreover, to consider the most general case, the phase angles

of the -th harmonic component depend on the time-varying angular fundamental frequency and the initial harmonic angle . The main goal of this paper is threefold:

  1. to propose a modified Second-Order Generalized Integrator (mSOGI) with prescribed settling time for a fast online estimation of amplitudes and angles , such that, after a specified transient phase, estimated signal (indicated by “  ”) and original signal do not differ more than a given threshold . More precisely, the following should hold


    after a prescribed (specified) settling time ;

  2. to propose a modified Frequency Locked Loop (mFLL) ensuring stable operation and fast estimation of the angular frequency in combination with the proposed parallelized mSOGIs; and

  3. to show the overall estimation performance and compare it to other standard approaches such as parallelized sSOGIs and ANFs with and without FLL.

Remark I.1.

Note that in (1), time-varying amplitudes (of each harmonic component) and time-varying angles are considered. The typical assumption (see, e.g. [25, Appendix A]) of a constant fundamental angular frequency such that is not imposed, since it is not true in general.

I-C Principle idea of proposed solution

The principle idea of the proposed solution is illustrated in Fig. 1. The depicted block diagram is fed by the input signal  to be estimated. All subsystems of the overall nonlinear observer are shown. The outputs of the block diagram are the respective estimated components of the input signal (see Sect. I-B).

Figure 1: Block diagram of the nonlinear observer (consisting of parallelized mSOGIs and modified FLL (mFLL) with gain normalization (GN)).

In Fig. 1, all components (subsystems) of the nonlinear observer are explicitly shown. One can summarize: For , the overall nonlinear observer consists of the following subsystems:

  • a parallelization of modified Second-Order Generalized Integrators (mSOGIs) to estimate amplitude and phase of each of the harmonic components of the input signal . The -th mSOGI will output the estimated state vector

    compromising estimates of in-phase and quadrature signals of the -th harmonic component, i.e.  and , respectively. All estimated signal vectors are merged into the overall estimation vector


    The output signal represents the estimate of the input signal and is established by the sum of all estimates of the in-phase signals of the mSOGIs;

  • a modified Frequency Locked Loop (mFLL) with gain normalization, generalized frequency adaption law, sign-correct anti-windup strategy and rate limitation to obtain the estimate of the fundamental angular frequency . The mFLL is tuned by an adaptive gain which depends on estimation input error , estimation vector and estimated angular frequency ;

Section II and Section III introduce the different subsystems (i.e. mSOGIs and mFLL) illustrated in Fig. 1 and explain in more detail their contribution to the proposed solution for real-time estimation of amplitudes and phases of all harmonics of the input signal as in (1) as well as the fundamental frequency .

Ii Second-Order Generalized Integrators (SOGIs): In-phase and quadrature signal estimation

The key tool to estimate in-phase and quadrature signals of a measured sinusoidal signal is a Second-Order Generalized Integrator (SOGI) [25, App. A]. Their parallelization in combination with a FLL (see Sect. III) allows to detect all harmonic components and the fundamental frequency. First, a standard SOGI (as e.g. discussed in [1]) for the -th harmonic is revisited. After that, the proposed modification to it is introduced to obtain the modified SOGI with prescribed settling time. It is shown that the modified SOGI is actually a generalization of the standard SOGI. Next, their estimation performances are compared. Finally, to be capable of estimating all harmonics, the proposed modified (or standard) SOGIs are parallelized to obtain the overall observer system. Throughout this paper, the more powerful state space representation will be used, since the considered parallelized SOGIs with FLL represent a nonlinear system and transfer functions are not applicable.

Ii-a Standard SOGI (sSOGI) for the -th harmonic component [1]

For now, let and consider only the -th harmonic component . If the estimate of the -th harmonic frequency is known, the implementation of a sSOGI for the signal allows to obtain online estimates and of in-phase and quadrature signal, respectively. A sSOGI for the -th harmonic component is depicted in Fig. 2 (a). Its dynamics are given by the following time-varying differential equation


with arbitrary initial value (mostly set to zero), gain (single tuning factor) and estimate (possibly time-varying) of the fundamental angular frequency . The gain only allows for a limited tuning of the dynamic response of the sSOGI. For constant only, characteristic equation and poles of the -th sSOGI are given as follows5


The respective root locus is shown in Fig. 2 (b). Hence, stability is guaranteed for all . However, since the pole closest to the imaginary axis determines the settling time of the system, the smallest settling time is obtained for which clearly limits the tuning of the transient performance of the sSOGI. Moreover, this choice leads to two real poles at and, hence, the sSOGI is not capable of oscillating by itself. Therefore, common tunings are  [1] or  [3].

(a) Block diagram of -th sSOGI with one single tuning parameter .
(b) Root locus of -th sSOGI for (unstable for ).
Figure 2: Standard Second-Order Generalized Integrator (sSOGI[1]: (a) Block diagram and (b) root locus of -th sSOGI.

Ii-B Modified SOGI (mSOGI) for the -th harmonic component

To overcome the problem of the limited tuning without the possibility to prescribe the settling time, the modified SOGI (mSOGI) with additional gain is introduced. The resulting block diagram of the -th mSOGI is illustrated in Fig. 3 (a). Note that the additional gain does not impair functionality but gives the necessary degree of freedom to enhance the transient performance as will be shown in the next subsection.

(a) Block diagram of -th mSOGI with two tuning parameters and .
(b) Root locus of -th mSOGI for and (in general unstable for or ).
Figure 3: Modified Second-Order Generalized Integrator (mSOGI): (a) Block diagram and (b) root locus of -th mSOGI.

The state space representation of the -th mSOGI is given by the following time-varying differential equation:


with arbitrary initial value and estimate of . The gains and now allow (theoretically6) for a limitless tuning of the dynamic response of the mSOGI. The tiny but crucial difference between the mSOGI in (6) and the sSOGI in (4) is the additional gain in the system matrix and the vector . For a constant frequency , the characteristic equation and the poles of the -th mSOGI are given by


The special choice of the additional gain in (7) gives the key feature of the mSOGI: For any , the real parts of the poles in (7) can be chosen arbitrarily; whereas the capability of the mSOGI to oscillate with angular frequency is preserved (see imaginary parts of ). The root locus of the -th mSOGI is depicted in Fig. 3 (b). The mSOGI is stable for any ; and, the larger is chosen, the faster is its transient response.

Remark II.1 (Generaliziation of the sSOGI).

The introduction of the additional gain for the mSOGI in (6) represents actually a generalization of the sSOGI in (4). Clearly, for , the mSOGI simplifies to the sSOGI. In other words, only now, the term “second-order generalized integrator” is really appropriate.

Ii-C Comparison of the estimation performances of sSOGI and mSOGI

If , both SOGIs are capable of estimating in-phase signal and quadrature signal of the -th harmonic signal . The estimated amplitude is given by the norm of the estimated signal and its quadrature signal. The estimated phase angle is given by . Hence, the parameters and of the -th harmonic can be detected online.

In Fig. 4, the transient responses of sSOGI and mSOGI are shown in cyan and blue, respectively, for the first harmonic (i.e. , see Fig. 4(a)) and for second harmonic (i.e. , see Fig. 4(b)). Four tunings of the gain are implemented and illustrated by different line types:  (dotted), (dashed), (dash-dotted) and (solid). The larger is chosen, the faster is the transient response of the mSOGI. Moreover, for (dash-dotted) or (solid), settling times of e.g.  and can be guaranteed for the fundamental signal, respectively. For the second harmonic, the transient response is twice as fast as for the fundamental signal. For the sSOGI, a prescribed settling time cannot be ensured, since one pole approaches the imaginary axis for large choices of (see also Fig. 2). In particular, the estimation of the quadrature component is slow (see in Fig. 4) which degrades the estimation speed of positive, negative and zero sequences in three-phase systems (not considered in this paper).

(a) Estimation of fundamental component of first harmonic ().
(b) Estimation of second harmonic component ().
Figure 4: Comparison of estimation performances of -th sSOGI (, , , ) and -th mSOGI (, , , ) for four different tunings of gain , respectively. Signals shown in (a) for and in (b) for are from top to bottom: input signal and its estimate , estimation in-phase error , quadrature signal and its estimate and estimation quadrature error

Ii-D Parallelization of the mSOGIs

This far, the presented SOGIs (sSOGIs and mSOGIs) can only estimate in-phase signal and quadrature signal of the -th harmonic signal . By parallelizing of the mSOGIs or sSOGIs (see Fig. 1), it is possible to extract in-phase and quadrature signal of each harmonic component for all . For the parallelized sSOGIs, stability is preserved for a positive choice of all gains, i.e.  for all  [3]. Stability for the parallelized mSOGIs will be guaranteed by pole placement. Moreover, the settling time can only be pre-specified by the parallelized mSOGIs.

The idea of the parallelization can be motivated by recalling the internal model principle which states that “[e]very good regulator [or observer] must incorporate a model of the outside world being capable to reduplicate the dynamic structure of the exogenous signals which the regulator [or observer] is required to process.” [30]. In the considered case, the exogenous signal as in (1) can be reduplicated by the parallelization of sinusoidal internal models [31, Chapter 20], which have the overall dynamics




The initial values of the internal model in (8) allow to determine amplitude and angle of the -th harmonic.

Note that, for constant and differing harmonics for all , the overall internal model (8) is completely state observable, since


Now, by substituting estimate for , the observer is obtained and consists of the parallelized mSOGIs (as introduced in (6) for the -th harmonic). The observer dynamics are nonlinear and given by


where observer state vector and observer gain vector


merge the individual state vectors and gain vectors of the mSOGIs as in (6). The observer will be tuned by pole placement and, hence, the gains in can be determined by comparing the coefficients of the characteristic polynomial


of the system matrix in (11) and the coefficients of a desired polynomial


with prescribed stable roots (poles) , , in the negative complex half-plane. The detailed derivation of the analytical solution of the pole placement algorithm is presented in Appendix A-A. The resulting feedback gain vector is obtain as follows






It can be shown that, for any positive (but possibly time-varying) angular frequency estimate for all , the closed-loop observer system (11) is bounded-input bounded-output (BIBO) stable and input-to-state stable (ISS). Moreover, if , then the estimation state error decays exponentially to zero. [For the reviewers: To not further lengthen the manuscript, the proofs are omitted. They can be provided upon request.]

Remark II.2 (place command in Matlab versus analytical expression in (15)).

For small (e.g. ), the Matlab command place can be used to compute . For large , place might not provide a proper result. Moreover, place cannot place poles with multiplicity greater than . That is why, the analytical expression in (15) has been derived. It can be used to achieve pole placement for arbitrarily large .

Iii Frequency-Locked Loop (FLL): Frequency estimation

As mentioned above, a correct estimate of the fundamental angular frequency is essential for a proper functionality of the parallelized SOGIs and the harmonics detection. The following subsections motivate and discuss the necessary modifications of the FLL to ensure its functionality also with the parallelized mSOGIs.

Iii-a Standard FLL (sFLL) [1]

In this subsection, the standard FLL is re-visited. Its block diagram is shown in Fig. 5. Adaption law and gain normalization are briefly explained.

Figure 5: Block diagram of standard Frequency Locked Loop (sFLL) with gain normalization [3, 1, 15].

Adaption law

As shown in Fig. 1, any of the parallelized mSOGIs requires an estimate of the fundamental angular frequency . The estimate is the output of the FLL. The nonlinear adaption law of the sFLL is given by [1]


where is a positive but non-constant adaptive gain,  [3, 1] is a constant “selection” vector (to extract only the fundamental estimate from ), is the estimation vector of the parallelized sSOGIs and is the estimation error (difference between input and estimated input ). A proper choice of the initial value, e.g. , of the sFLL adaption law is beneficial for functionality and adaption speed.

Remark III.1 (Impact of negative estimates of the angular frequency).

Note that, in view of the adaption law in (18), the estimated angular frequency might also become negative, i.e.  for some time instant . However, a negative will result in instability of the parallelized sSOGIs and all estimated states will diverge.

Gain Normalization (GN)

The FLL should be robustified to work for signals with arbitrary fundamental amplitudes and angular frequencies (see [1, 15]). This can be achieved by introducing the following adaptive sFLL gain


which depends on gain , frequency estimate and norm of the fundamental estimation vector leading to a “normalized” FLL adaption law. The gain is a constant tuning factor of the FLL.

Remark III.2 (Avoiding division by zero).

Depending on the initial values and the time evolution of estimation process, the denominator in (19) might become zero for certain time instants . This must and can easily be avoided by introducing a minimal positive value for the denominator by substituting for in (19) where is a small positive constant.

Iii-B Modified FLL (mFLL)

The FLL is the weakest subsystem (bottleneck) of the overall grid estimation system; in particular, its tuning endangers system stability, estimation accuracy and estimation speed. Only if the frequency is detected correctly, the mSOGIs or sSOGIs work properly. Therefore, to improve stability and performance of the estimation process a modified FLL is proposed. The block diagram of the proposed mFLL is depicted in Fig. 6. Remarks III.1 and III.2 have already been considered in the block diagram. In addition, the mFLL is equipped with a generalized adaption law, a sign-correct anti-windup strategy and a rate limitation. All three modifications enhance performance and stability of the mFLL. The generalized adaption law increases adaption speed. The anti-windup strategy guarantees that the estimated angular frequency remains bounded and positive for all time and the rate limitation prevents too fast adaption speeds which might endanger stability. Details will be explained in the next subsections.

Figure 6: Modified Frequency Locked Loop (mFLL) with gain normalization, sign-correct anti-windup and rate limitation.

Generalized adaption law

The presented adaption law (18) of the sFLL does not work properly for the mSOGIs. It does not guarantee a a sign-correct adaption for all time. Therefore, the adaption law must be generalized to fit to the parallelized mSOGIs. It is clear that for a sign-correct adaption of the estimated angular frequency , the generalized adaption law must ensure that the following conditions hold


To illustrate the intuition behind these conditions, assume that the input signal has a constant fundamental angular frequency and that the parallelized mSOGIs are fed by an arbitrary positive but constant estimate . Then, in steady state, the system states with their characteristic amplitude and phase responses can be used to analyze whether and are in-phase or counter-phase. In Appendix A-B3, it is shown that this sign-correct adaption is guaranteed when the selection vector is chosen as follows


This choice of can be used for sSOGI and mSOGI as well. It is actually a generalization of the standard choice with for all (recall (18)). Finally, note that the sign-correct adaption was derived based on a steady state analysis (see Appendix A-B3). This implies that the mFLL (and sFLL) dynamics should be slow compared to the dynamics of the parallelized mSOGIs (which can be achieved by an adequate choice of ).

Sign-correct anti-windup strategy

Usually, the grid frequency should not exceed a certain interval. This physically motivated limitation can be exploited for the frequency estimation. The principle idea of the proposed sign-correct anti-windup strategy is illustrated in Fig. 7. More precisely, the adaption of the estimated angular frequency shall be stopped (i.e. ), when

  • the estimated angular frequency leaves the admissible interval, i.e. with lower and upper limit , respectively (see Fig. 7); and

  • the right hand side of the adaption law (20) has wrong sign (otherwise the estimation gets stuck at one of the limits).

Figure 7: Illustration of the principle idea of the proposed sign-correct anti-windup strategy.

This yields to the following sign-correct anti-windup decision function


where is proportional to the time derivative of the estimated angular frequency as can be seen when decision function and frequency adaption law are combined as follows


The consequences of this adaption law with sign-correct anti-windup are that the estimated angular frequency is positive and remains bounded for all time, i.e.  for all . Moreover, once within the admissible interval , the estimated angular frequency will remain inside this interval. Clearly, if the initial value of the frequency estimate starts outside of , it will approach the interval due to the sign-correct frequency adaption (as illustrated in Fig. 7 for ). Note that the proposed anti-windup strategy does not require tuning of an additional feedback gain as in [22]. Instability can not occur, since the proposed sign-correct anti-windup strategy is based on the simple idea of conditional integration [31, Section 10.4.1].

Rate limitation

Recall that the overall observer (6) is nonlinear. Considering the estimated angular frequency as time-varying parameter, the observer becomes a time-varying linear system. If the time derivative is limited (rate limitation), the observer can be considered as slowly time-varying system [32] (which simplifies stability analysis). For this rate limitation of the adaption law, the admissible rate of the estimated angular frequency must be bounded, i.e.  where and are desired lower and upper thresholds, respectively. The idea of the rate limitation is illustrated in the graph shown in Fig. 8.

Figure 8: Illustration of the rate limitation of .

Usually, the rate limitation leads to a smoother adaption and is a meaningful choice. Reasonable rate thresholds were found out to be per . The rate limitation can be ensured by introducing an additional saturation function to the adaption law (23) leading to the generalized adaption law for the mFLL as shown next.

Iii-C Generalized adaption law

Finally, combining rate limitation, sign-correct anti-windup strategy and sFLL with gain normalization, the generalized adaption law of the mFLL with as in (21) can be introduced. It is given by


which guarantees that (i) the derivative of the estimated angular frequency is bounded, i.e. , and (ii) the estimated angular frequency is positive and bounded from below and above, i.e.  for all .

Iv Implementation and measurement results

To validate the proposed algorithms, measurements at a laboratory setup are carried out. [For the reviewers: To not unnecessarily lengthen the paper further, simulation results are not presented; but could easily be provided as well. The authors believe that measurement results are more important and more convincing than simulation results.] The laboratory setup is shown in Fig. 9.

Figure 9: Laboratory setup: (A) Cinergia grid emulator, (B) Host-PC and (C) dSPACE real-time system .

For measurements, the voltage is produced by the grid emulator. These voltages are measured by a LEM DVL 500 voltage sensor, analogue-to-digital converted by the dSPACE A/D card DS2004 and internally filtered by a low pass filter with cut-off frequency to suppress high frequency noise. The implementation is done via Matlab/Simulink R2017a on the Host-PC. The executable observers are downloaded via LAN to the dSPACE Processor Board DS1007 and run in real time. The measurement data is captured and analyzed on the Host-PC after the experiment. The implementation data of the conducted measurements is listed in Tab. I.

sampling time
low-pass filter (for measurements)
parallelized mSOGIs
observer gains as in (15) []
mFLL ,
Anti-windup ,
Rate limitation ,
parallelized sSOGIs [1]
observer gains
sFLL , (avoidance of division by zero added)
parallelized ANFs [3]
filter gains
sFLL (without gain normalization)
Scenario (S) with constant fundamental frequency
initial values of observer (11)
initial values of sFLL (19) and mFLL (24) (and for all )
Scenario (S) with time-varying fundamental frequency
initial values of observer (11)
initial values of sFLL (19) and mFLL (24)
Table I: Implementation and tuning data for simulations and measurements.

Three estimation methods are implemented and their estimation performances are compared for the following two scenarios:

  • Estimation of an input signal with constant fundamental angular frequency (i.e.  for all ) and ten harmonics exhibiting amplitude jumps at , and . The following three estimation methods are implemented and compared:

    • parallelized mSOGIs without FLL, i.e. observer (11) with as in (15);

    • parallelized sSOGIs without FLL [1], i.e. observer (11) with (and for all ); and

    • parallelized Adaptive Notch Filters (ANFs) without FLL [3], i.e. observer (11) with .

  • Estimation of an input signal with time-varying fundamental angular frequency (i.e. ) and ten harmonics exhibiting frequency jumps at and and amplitude jumps at and . The following three estimation methods are implemented and compared:

    • parallelized mSOGIs with mFLL, i.e. (11) with as in (15) and (24);

    • parallelized sSOGIs with sFLL [1], i.e. observer (11) with ( for all ) and (19); and

    • parallelized ANFs with sFLL [3], i.e. observer (11) with and sFLL (18) (without gain normalization7).

For both scenarios, the considered input signals have a significant harmonic distortion. The parameters of the individual harmonics (amplitudes and frequencies ) of the signal for Scenario (S) and for Scenario (S) are collected in Tab. II. Within the considered time interval , three jump-like changes in amplitudes and/or fundamental frequency occur at , and . Hence, the input signal and its harmonic content changes abruptly and requires the observers to “restart” their estimation process for each step-like change.

For a fair comparison, all three estimation methods are tuned in such a way that the best feasible estimation performance was achieved within their respective tuning and capability limits. The measurement results for Scenario (S) are shown in the Figures 10, 11 and 12. The results for Scenario (S) are depicted in Figures 13, 14 and 15. These results will be discussed in more detail in the next subsections.

Table II: Data of considered input signal for both scenarios (S) and (S): Amplitudes and frequencies () of the ten harmonics.

Iv-a Discussion of the measurement results obtained for Scenario (S)

For Scenario (S), the FLL were implemented but the adaption was turned off. Fundamental and estimated angular frequency are identical for this scenario. Therefore, the estimation performances purely according to the respective observer tunings can be compared. The harmonic content of the input signal undergoes step-like changes at the time instants , and , respectively (see Tab. II and Fig. 10).

Figure 10: Measurement results for Scenario (S): Comparison of the estimation performances of parallelized mSOGIs (), sSOGIs () and ANFs () without FLL. Signals shown from top to bottom are: Input signal () & its estimate