Grid-Forming Power Converters Tuned Through Artificial Intelligence to Damp Subsynchronous Interactions in Electrical Grids

The integration of non-synchronous generation units and energy storage through power electronics is introducing new challenges in power system dynamics. Speciﬁcally, the rotor angle stability has been identiﬁed as one of the major obstacle with regards to power electronics dominated power systems. To date, conventional power system stabilizer (PSS) devices are used for damping electromechanical oscillations, which are only tuned sporadically leading to signiﬁcant deterioration in performance against the ever-changing operating conditions. In this paper, an intelligent power oscillation damper (iPOD) is proposed for grid-forming converters to attenuate electromechanical inter-area power oscillation. In particular, the iPOD is applied to a synchronous power controller (SPC) based grid-forming power converter to increases gain of the active power control loop at the oscillatory frequency. Predictions regarding the mode frequency, corresponding to the current operating points, are given by an artiﬁcial intelligence ensemble model called Random Forests. The performance of the proposed controller is veriﬁed using the two area system considering symmetrical fault for random operating points. In addition, a comparison with PSS installed in each generator reveals the individual contribution with respect to the inter-area mode damping


I. INTRODUCTION
Renewable-based Energy Systems (RES) and Energy Storage Systems (ESS) are widely adopted to reduce green house gas emissions and improve efficiency. These systems, due to their incompatibility with synchronous frequency (50/60 Hz), require power electronic devices to act as an intermediate before connection with the rest of the network is possible. Consequently, as the presence of power electronics increases, the power system dynamics are affected introducing new challenges regarding power system stability [1]. In fact, ENTSO-E has identified the angular stability of power systems as one of the key challenges related to High Penetration of Power Electronic Interfaced Power Sources (HPoPEIPS) [2].
The associate editor coordinating the review of this manuscript and approving it for publication was Xiaorong Xie .
Conventionally, the PSS has been used to damp low frequency oscillations in a system by regulating the synchronous generator's excitation voltage.Among the standardized PSS schemes, the classical PSS1 controller consists of three main blocks: gain, washout filter and phase compensation. Considering inter-area modes, the time constant of the washout filter needs to be carefully chosen, since it captures the low frequency oscillations [3]. The PSS1 uses a single input signal, typically the speed deviation ω. More sophisticated designs are the PSS2B and PSS4B controllers. In PSS2B, speed transducers are added to improve reactive power modulation during mechanical power changes and eliminate torsional forces that can harm the generator shaft. PSS4B builds on top of the PSS2B by incorporating low, intermediate and high frequency bands to broaden the range of targeted modes [4].
One major limitation of the PSS is the high number of parameters that need to be tuned for attenuating system VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ oscillations [5]. For instance, a typical PSS has 8 parameters: gain, time constant of wash out filter, and four constants for lead-lag compensator. These parameters usually are tuned at commissioning of the power plant and remain fixed despite the changes of operating conditions [6].
There is an open discussion in the literature on possible methods for adaptive PSS and optimal tuning [3]. This becomes even more relevant now, as power electronics can alter the oscillatory profile of a given system, changing or creating new modes. Currently, the number of synchronous generators is slowly phased out [7] and as a result so will the number of PSS. Yet, power electronic interfaced generation units can provide a fast response to the grid and assist the system in maintaining its stability in case of a disturbance [8]. Such grid friendly response can be produced by means of the currently wellknown grid-forming power converters. In general, the control system of a grid-forming converter can either be based on droop [9] or on virtual synchronous machine (VSM) [10]. In the former case, the contribution of the power converter is usually limited to primary frequency and voltage regulation. To add grid-supporting functionalities, e.g. inertia emulation and power oscillation damping, the VSM implementation can be adopted in the grid-forming power converter [11].
Among the various VSM strategies, reported in the literature, the Synchronous Power Controller (SPC) has been widely used mainly due its simple yet effective control structure [12]. By employing the SPC, power converters can provide virtual damping and synthetic inertia to the grid, increasing its dynamic flexibility to improve local and interarea stability [13], [14]. In fact, the SPC with multiple power loop controllers (PLC) has been presented in [15]. By combining a band-pass filter with a modified swing equation, the inter-area oscillations at a predefined frequency can be attenuated. However, to ensure the system in [15] works effectively, the center frequency of its band-pass filter has to be tuned properly to match the frequency of the inter-area oscillation mode, which is usually not accessible due to the dynamic behavior of the power system.
Recently, data-driven methods based on artificial intelligence (AI) are studied for accurately and efficiently determining the state of the power system in real time and thereby avoid instability [16]. Specifically, the relationship between system variables and stability indexes can be modeled to predict the post-fault state of the system. For instance, an ensemble approach based on extreme learning machines is adopted in [17] to predict stability after a disturbance, while a recurrent neural network with long short term memory (LSTM) cells is trained in [18] for the same task considering time dependencies in the data. Furthermore, prediction of damping state (i.e. well or poorly damped) is proposed by [19] using decision trees and by [20] using neural networks.
Driven from the above, this paper presents an intelligent Power Oscillation Damper (iPOD) that enables the SPC to adapt for maximum -oscillatory-mode attenuation. The iPOD is based on an ensemble AI predictor called Random Forests (RF), which is trained to predict the real and imaginary components of the targeted mode of oscillation as it moves according to the current operating conditions. The main contribution of the system, presented in the following, is a novel control scheme that incorporates intelligence by exploiting the ability of AI to predict the characteristic values of an oscillatory mode in real time and uses this information to adaptively tune a dedicated loop of the SPC to damp oscillations and enhance the system stability i.e. iPOD. In this work, special focus is put on describing all the steps for developing the AI models in detail. The proposed approach is verified under a symmetrical fault for random operating points as well as against conventional PSS devices using the wellknown two area system described in [21]. The AI predictor is trained for tracking the inter-area mode between the two areas, however the same principles can be applied to develop multiple AI predictors to predict more than one inter-area modes.
The remainder of this paper is structured as follows. Section II provides a detail description regarding the test case. Section III presents the proposed AI-based adaptive SPC model. Section IV analyse the results and Section V discusses the conclusions.

II. TEST CASE
A popular power system to analyse the behavior of an interarea mode is the two-area system [21], depicted in Fig. 1. The system consists of two areas with one load (L07, L09), one shunt capacitor (C7, C9) and two generators (G1-G2, G3-G4), respectively. Each generator is equipped with IEEE Type AC4A Excitation System (AVR), steam turbine governors (GOV) and PSS type 1. These areas are connected through a weak tie-line where 400MW are flowing from Area 1 to Area 2. Modal analysis reveals two local modes with frequencies of 1.05Hz and 1.08Hz in Area 1 and Area2, respectively, as well as an inter-area mode with frequency of approximately 0.55Hz.
The specific design of this system not only allows for studying the behaviour of electromechanical oscillations between two areas, but also on demonstrating the effectiveness of control methods (such as the iPOD) for attenuating them. To demonstrate the impact of the iPOD, a virtualsynchronous generation unit equiped with an SPC is connected at bus 7, through two step-up transformers. The rated apparent power of the aforementioned unit is 100MVA, equal  to 3% of the total apparent power of the four synchronous generators together.

A. CONTROL STRUCTURE
For mitigating inter-area oscillations, a conventional SPC is equipped with an iPOD as shown in Fig. 2. Generally, the SPC consists of four main control blocks, namely, the current controller, the virtual admittance, the reactive controller, and the PLC. In the SPC, the reference for the current control loop is generated by means of a virtual admittance whose transfer function is where R and L are the virtual resistance and the virtual inductance, respectively. As shown in Fig. 2, the virtual electromotive force e emf is generated by combining the outputs of the active and reactive control loops through a voltagecontrolled oscillator (VCO) as In (2), the magnitude of the electromotive force is regulated through the reactive power controller as follows: (3) where k p and k i are the proportional and integral gains of the reactive power controller, and V based is the nominal voltage. Likewise, the phase angle θ used in (2) is produced by the active power loop controller (PLC). In this implementation, although any other controller structure could be used [8], the PLC is implemented through the conventional swing equation whose transfer function is where H and D are the inertia constant and the damping coefficient respectively, while ω base is the grid nominal frequency, which acts as a feed-forward term to improve the dynamic performance of the SPC. Aiming to increase damping to inter-area oscillation modes, the PLC is incorporated with the proposed iPOD as  shown in Fig. 3. Essentially, the iPOD includes two proportional gains k 1 and k 2 , and a second-order band-pass filter. The gain k 1 is set to −1 for removing the targeted frequency component from the signal before acted on the PLC, whereas k 2 defines the damping factor provided by the iPOD at the specified frequency. The reason behind the choice of a bandpass filter is that the frequency of the oscillation is considered to be a known parameter. Indeed, the oscillation frequency is predicted by the AI predictor which will be explained in the following subsections. The transfer function of the band-pass filter can be given as where ω c is centre frequency and B is passing band. From (4) and (5), the augmented transfer function of the active power controller can be derived as follows: From (6), the combined frequency response of the PLC and the iPOD is obtained, as depicted in Fig. 4 for different VOLUME 8, 2020 values of gain k 2 . As shown, the iPOD adds a significant amplification to the active power control loop at the tuned frequency. Such a high gain implies that the proposed PLC with iPOD control structure can provide higher damping to low-frequency inter-area oscillation.
It is worth noting that the augmented transfer function (6) has two parts. While the first term is conventional PLC, the second term reflects added effects of the iPOD. In fact, the second term can be rewritten in a form of a lead-lag compensator as follows: Equation (7) reveals that the two gains k 1 and k 2 make it possible to adjust not only gain but also phase lag of the active power control loop at the specified frequency. In practice, the inter-area modes depends on the operating condition of the electrical grid. Therefore, in order to predict the oscillation modes in real-time, the AI-based algorithm called Random Forest is employed.

B. ENSEMBLE AI PREDICTOR
The structure of the proposed iPOD requires up to date information regarding the targeted mode's frequency to provide additional damping. Therefore, its imperative to develop a model that will be able to relate the systems operating condition to the specific movement of the targeted mode. Although this can be a challenging task, following a purely mathematical approach, data-driven methods based on AI and machine learning are able to approximate this relationship without any knowledge about the inner mechanics of the system. Among the plethora of machine learning models, such as the neural network in [22] to improve damping control for Static Synchronous Series Compensator, the decision tree (DT) stands out due to its robustness against feature scales or types, meaning that it is possible to process patterns with metric and non-metric data at the same time [23]. In addition, the decision making process of the DT is highly interpretable, e.g. as in Fig. 5, and is able to remove redundant system variables, thus performing an intrinsic feature selection.
The basic principle of DTs is to classify patterns based on a series of ''questions'', where the given answer determines the next ''question'' being asked and so on. This top-down process can be displayed in the form of a tree, consisting of nodes and branches as illustrated in Fig. 5. The process begins at the root node and it ends at the leaf nodes, where the process stops. The ''question'', more formally known as the split criterion of the node, resembles an ''if-then-else'' statement that describes a specific rule. For instance, the root node of the DT depicted in Fig. 5, reveals a threshold value for the Voltage magnitude (V m ) at bus 09. Depending on the answer the subsequent question can be either about the Power Factor (PF) at bus 10 or at bus 9, as in (8). This rule based approach is the core of the high level of transparency, which is not found in other machine learning models.
Generally, the purpose of the splitting criterion is to divide the feature space into as pure as possible sub-spaces. Hence, for every candidate split criterion the resulted impurity is calculated as the mean squared error, see (9), between the true value y (i) and predicted valueŷ t for patterns in subset S t averaged over all training patterns N t at node t. The predicted value at node t is calculated according to (10).
The efficiency of a split criterion is further validated by calculating the drop of impurity at the ancestor nodes, as in (11). Here, t Y and t N are the two ancestor nodes corresponding to the Yes and No branches, while N tY /N t and N tN /N t are the portion of samples at each of the these two nodes. Ultimately, at each node, the final splitting rule is chosen to be the one that reduces the impurity at the ancestor nodes the most [24].
Despite their straightforward interpretation and intuitiveness, DTs tend to overfit the data demonstrating poor generalization. Moreover, errors propagate from higher level nodes towards lower level nodes [24] resulting in a different DT if changes to the training data occur.
To overcome the avove limitations, an ensemble AI model called Random Forest (RF) is adopted, which integrates multiple DTs into a single predictor, as in Fig. 6. The individual DTs are trained, based on the principle describe above, for the same task but separately from each other, using random subsets of patterns (i.e. system snapshots) and/or features (i.e. system variables). The random sampling is widely known as 93372 VOLUME 8, 2020 bootstrap aggregation (or bagging for short [23]) when the subsets are not unique. The ensemble paradigm demonstrates higher performance both in terms of accuracy and generalization: a direct consequence of the diversity promoted by the randomness of the different subsets that are used for developing each of the individual prediction models.
Frequently, the final output of the ensemble is given through a majority voting scheme [25] or by averaging the probabilistic prediction (i.e. soft voting). Nevertheless, the core of RF are the individual DTs where the development process is the same, albeit the variations of the feature-pattern space.

C. ENSEMBLE AI PREDICTOR TRAINING AND TUNING 1) DATABASE DESCRIPTION
To generate a sufficient amount of measurements, an automated simulation process is designed to parse through a wide range of probable operating scenarios, and to store the variables of interest. In other words, this automated process will formulate a database D ∈ R M ×N , where M ∈ Z + is the total number of patterns and N ∈ Z + is the total number of features.
DigSilent's PowerFactory is the selected platform to run the simulations and Python 3.7 for executing Algorithm 1. The loading of the system (line 3 of Algorithm 1) is chosen from (12), where r 1,2 ∼ U (0, 1) are random coefficients drawn from a uniform distribution between the open set 0 and 1, µ is the distribution mean and σ is the standard deviation. The processing time on a machine with 20 processors and a single Powerfactory licence requires around 50 minutes per 1000 simulations. Our final dataset consists of about 22000 patterns and 122 features, i.e. D ∈ R 22000×122 . A summary of the system variables constituting our feature vector can be found in Table 1.
For each pattern in the dataset there exists a set of eigenvalues that corresponds to a random operating point. From the set of eigenvalues the most dominant mode is extracted, in this case the inter-are mode. Hence, each pattern is labeled with the target variable comprised of the real and imaginary component i.e. y ∈ R 22000×2 .
The behaviour of dominant inter-area mode for the random operating points is revealed in Fig. 7. From the figure it becomes apparent that there are many instances were the dominant inter-area mode moved to the, unstable, right hand side of the real axis. In addition, the frequency of the dominant inter-area mode varies significantly, as the conditions change, from 0.45Hz up to 0.76Hz. Interestingly, from Fig. 8 it appears to exist a relationship between voltage magnitude and voltage angle that can be used by the RF to map system variables to the inter-area mode. For example, almost every voltage magnitude that corresponds to the unstable inter-area mode is less than 1 p.u. Similar patterns can be seen in other locations of the system not presented here for brevity.

2) PREPROCESSING
The simulated environment assumes full observability of the system, that is, measurements of system variables are recorded in every location of the system. However, in reality this can be costly and impractical. In addition correlated features can impair the ability of a prediction model to make accurate predictions. Therefore, several methods are used for reducing the size of the feature vector to include only relevant and uncorrelated features. In general, the objective of data preprocessing is mainly two-fold; to improve the accuracy of the model and its training speed [26].    The typical workflow followed for processing a dataset consists of three steps: filter, select and re-scale as in Fig. 9. However, due to the robustness of DTs to features of different scale and type last step can be omitted, as indicated by the dashed lines in Fig. 9.
The highly correlated features are filtered out by using a minimum acceptance threshold and Pearson's correlation coefficient, as in (13), where {j, p ∈ [1, . . . , N ] | j = p} are feature indexes while the bar above the variable denotes the sample mean. The uncorrelated features comprising the feature vector reduced from 122 to just 36 features.
The inherent ability of DTs and RFs to perform feature elimination is exploited to analyse the individual contribution of each feature in the prediction process. Typically, greedy search algorithms like the Sequential Feature Selection (SFS) are preferred as they require less time and computational resources to eliminate features [26]. Fortunately, the size of the feature vector in hand allows the implementation of exhaustive search algorithms such as the Recursive Feature Elimination (REF), which guarantee optimal feature elimination. In this study, the REF algorithm is used alongside the RFs to decrease the size of the feautere vector even further to only 22.
The RFs are trained in python 3.7 using the scikit-learn package [27]. Although it is possible to predict both real and imaginary parts of eigenvalues with a single RF, here it is preferred to train a separate RF for each component. Therefore, if performance drops for one of the two target variables, the problematic model can be still retrained without losing the predictions for the other variable.
The optimal parameters, valid for both RF models, are estimated using grid search and summarized in Table 2. Notice that three out of five parameters refer to the development of the individual DT. Specifically, the minimum patterns per split define a lower bound to the number of patterns needed to make a split, the maximum features define the number of features to be considered for the split criterion at each node, while the maximum depth sets an upper bound to the DT growth. Finally, the Bootstrap implies feature and pattern sampling selection is done with replacement. The RF is evaluated using a 10-fold cross validation, recording both the R2 score, in (14) and Mean Absolute Error (MAE), in (15). The former measures how well the features explain the target variable with values ranging from 0 (worse) to 1 (best).

3) PERFORMANCE ANALYSIS
Learning curves are showing the performance of a machine learning model as a function of the training set's size. Particularly, Fig. 10 depicts the learning curve of the RF model developed to predict the {λ} component of interarea mode. The figure clearly shows that after 8000 training patterns there is no significant gain in the performance of the RF model. Hence, by using fewer training patterns we can limit the time spent for fitting the model to just 30s. Furthermore, the cross-validation score seems to converge to a certain value indicating that further addition of training data cannot improve the model. Similar behaviour can be viewed for the corresponding model of the {λ} component not shown here due to space limitations. In addition to the metrics described in the previous section, both training and testing CPU processing times are considered. The cross validated accuracy and speed of -RF and -RF are listed in Table 3. The tables prove the high accuracy of RF to predict both the real and imaginary values of the inter-area mode. The error of the -RF is slight lower than the one calculated for -RF however the difference is quite small. On average a single RF model needs approximately 50s for training and 0.7s on a high performance computer using 32 CPUs. Overall, both RFs demonstrate high accuracy on predicting the inter-area mode using only 22 system variables. Concretely, RFs can provide accurate information about the interarea mode such as damping, frequency and damping ratio. Although the iPOD is designed for tracking and attenuating the inter-area mode between the two areas, the same principles can be applied to develop the iPOD for more than one inter-area modes. This implies that the iPOD will be modified by integrating multiple AI predictors (or a single multi-output AI predictor) and multiple band pass filters [15] as dictated by the number of inter-area modes in the system. Multiple interarea mode attenuation is however, out of the scope of this paper. The next step is to verify that this information allows the iPOD to realize mode attenuation and enhance system stability.

IV. VERIFICATION
The intelligence of the iPOD scheme, proposed in this paper, for the SPC is based on the premises that a machine learning model, like RF, can predict in real time the dominant mode real and imaginary components. Specifically, the SPC can utilize the prediction about the mode's frequency to re-tune the iPOD band-pass filter to adapt in the new operating conditions. The gains of the iPOD are chosen as k 1 = −1 and k 2 = 0.018. This section, aims to verify that the iPOD can improve the attenuation of power oscillations in the system under contingencies given a random operating point (see Section III-C1), and compare it to the conventional PSS performance in the well-known two area system used as a reference [21].

A. SYSTEM RESPONSE TO DISTURBANCE
To excite the inter-area mode a 3-phase fault is defined in the middle of the upper transmission line that connects bus 7 and 8 i.e. L 1 7−8 . The fault is cleared by opening the breakers at both ends and re-closing them after 100ms. As mentioned, loads in areas 1 and 2 are set randomly to move the system at a new operating point. The snapshot of the system at time t m (before fault occurrence) is fed into RF, which makes a prediction about the inter-area to tune the SPC.
Note that in this study communication delays have not been considered, hence all measurements at t m are in sync. In reality, however, this might not be the case. Typically, a data storing and preprocessing module handles the information flow before they are passed on for further processing [28].Communication protocols, such as the well known IEC 61850, which support timestamping are used for ensuring that measurements can be sorted according to their corresponding time tags [29], [30]. Illustrated in Fig. 12 is the frequency recorded at buses 6 and 10 of areas 1 and 2, respectively. As a comparison the system response with the proposed SPC equipped with the iPOD is plotted against the base case (i.e. without SPC) and the SPC-only case (i.e. without online tuning and attenuation of low-frequency power oscillations). The inherent capability of the regular SPC to damp oscillations is revealed in the frequency response of both areas [31]. Yet, by predicting the mode frequency and re-tune the iPOD parameters the damping can be enhanced even further when the SPC+iPOD is used, as it is shown in Fig. 12, forcing oscillations to decay faster.
Concretely, the stochastic variations of loads cause the inter-area mode to vary in both frequency and damping, therefore, it is vital to be able to monitor these changes. The SPC can significantly improve damping of inter-area modes as shown in Fig. 13 depicting the results of modal analysis for random operating points. It is worth noting that the effectiveness of the iPOD scheme reflects the accuracy of the RF predictions.

B. OSCILLATION DAMPING IMPACT OF iPOD TO SYNCHRONOUS GENERATORS
In the two area system, the load L09 in area 2 is approximately 1.5 times larger than L07 of area 1. As a consequence, area 2 relies on imported power to maintain power balance. The power dependence of area 2 can be seen by analysing the power oscillations in each generator. Specifically, these responses during the symmetrical fault in L 1 7−8 for a random operating point are depicted in Fig. 14 for area 1 and in Fig. 15 for area 2.   As expected, the generators in area 2 are oscillating with higher amplitudes than those in area 1. Inter-area modes can result in power fluctuations, which can stress the structural integrity of generator units. By integrating the iPOD in the conventional SPC, the system power oscillations decay at a higher rate reducing the stress on generator shafts. This becomes even more obvious in area 2 where a 100MW oscillation can damage generators 3 and 4. Furthermore, the output power excursions during the on-fault period are reduced for all generators but generator 2, where there exists a slight increase of peak power as listed in Table 4. Nevertheless, it is   worth emphasizing that the improved damping from the SPC based on iPOD is achieved with only 1/7 of the synchronous generator apparent power.

C. iPOD COMPARISON WITH PSS
Commonly, the PSS is used to provide damping of critical modes in a system. This section compares the individual contribution of each PSS device with the corresponding contribution provided by the SPC augmented by the iPOD. For the purpose of this study, the PSS parameters are considered fixed and their values are set following the methodology described in [32] and listed in Table 5. Concretely, random points are generated following the aforementioned process and for each control scenario modal analysis of the system is performed. The results, averaged over all random points, are listed in Table 6. Clearly, PSS1 and PSS3, on average, possess the highest contribution towards inter-area power oscillation damping. However, damping provided by PSS1 and PSS3 varies significantly between each case. Particularly, PSS1 fails to provide damping for the inter-area mode for a random operating as revealed in Fig. 16. As opposed, the SPC augmented by iPOD seems to provide less damping than the previous two PSS devices. However, as Fig. 16 shows, this amount does not fluctuate much highlighting the damping capabilities of the SPC with iPOD. Furthermore, it also dominates over the contribution of both PSS2 and PSS4, which due to poor tuning are unable to damp the inter-area mode corresponding to the random operating points. Note that the nominal power of the SPC is considerably lower to the nominal power of the synchronous generators.

V. CONCLUSION
The connection of energy systems through power electronic devices is adding to the complexity of power systems by altering the overall behaviour of power systems. Rotor angle stability is of primary concern as the presence of power electronics in power systems is rising. Conventional PSS are performing acceptably for attenuating oscillatory modes, however, only for a narrow range of operating conditions. Power converters controlled through grid-forming techniques, e.g. a SPC augmented by an iPOD as in this paper, can provide an additional source for damping power oscillations with the use of data and AI.
Random Forest is a powerful machine learning model, which exploit the inherent diversity of combing individual decisions trees trained with random subsets of the original dataset. In this paper, two random forests have been developed for predicting the real and imaginary component of the inter-area mode in the two are system. These predictions are used to tune the iPOD added to SPC for improving damping of the targeted inter-area mode. As it has been proven from thorough analyses, the SPC with iPOD can enhance the stability of a power system by increasing the damping of the interarea mode. In the test presented in this paper, the response of recorded frequency demonstrated significant improvement when SPC is augmented by the iPOD scheme. In contrast to conventional PSS, the iPOD does not have a phase lag and VOLUME 8, 2020 only the frequency of the oscillation is required. Therefore, reducing the number of parameters needed to be tuned for attenuating the inter-area mode.
Furthermore, the proposed scheme is shown to be robust against network variations, performing equally and even better than the conventional PSS in the system. However, the proposed control scheme should not be considered as an adversary of PSS but rather as a complementary ally for improving the performance of the power systems of the future, which will be dominated by power electronics. The increased utilization of grid-forming converters for additional functionalities other than their primary purpose will increase flexibility and robustness in the ever-changing power system environment without the necessity for additional investments.