EP4035246A1

EP4035246A1 - Active power filter controlled by machine learning and method thereof

Info

Publication number: EP4035246A1
Application number: EP20866990.3A
Authority: EP
Inventors: Ebrahim BALOUJI; Karl BÄCKSTRÖM
Original assignee: Eneryield AB
Current assignee: Eneryield AB
Priority date: 2019-09-25
Filing date: 2020-09-23
Publication date: 2022-08-03
Also published as: EP4035246A4; WO2021061040A1; SE1951080A1; SE544845C2

Abstract

The disclosure relates to a method (300) for controlling and monitoring an active power filter device, comprising the steps of: measuring (301) current and previous states of current and voltage signals across at least a part of a power supply system and store these in a memory: pre-processing (302) said current and voltage signals, allowing for extraction and detection of features; wherein the machine learning component further predicts (303) future current and voltage levels; provide the result from the above processing steps to an active power filter device, allowing for the desired output to be obtained. The disclosure further relates to an active power filter device and a computer readable instruction set for controlling the output of an active power filter device.

Description

ACTIVE POWER FILTER CONTROLLED BY MACHINE LEARNING AND METHOD THEREOF

Technical field

The present invention relates to system, methods and arrangements for active power filters.

Background

Power electronic devices are applications, developed to control electricity energy flow in interaction with power supply systems. In a broad sense, all controllable power electronic devices are configured from solid-state circuits and control components circuit. It is common in industrial applications that power electronic devices are digitized to allow activities and interactions in power supply systems to be controlled by computing devices. As a result, large amounts of data is collected, enabling development of data driven methods, such as machine learning, as a next step. Methods for controlling and processing substantial amounts of signals in power electronic device is a relevant industrial and academic problem, currently being actively researched.

A power supply system, such as an electricity grid, comprises power generation, transmission and distribution subsystems, and Power Quality (PQ) is measured for each subsystem for long-term periods at certain sampling intervals. Therefore, especially for the problematic parts of the electricity grid, which are supplying highly nonlinear loads such as electric arc furnaces (EAF) or induction melting furnaces, huge amount of power quality event data is to be investigated, so that necessary countermeasures can be taken and compensation techniques can be developed specific to those parts. Power Quality filters and compensators are types of power electronic devices designed to compensate power quality events such as voltage dip, sag, interruptions and detect and filter out power quality variations such as harmonics and inter-harmonics. It is common in the art to provide compensators such as active power filters to power supply systems in order to compensate for power quality events and variations occurring due to non linear loads. In recent years there is a grown interest in providing power supply systems that are more efficient with the help of control and communications technologies such as machine learning. Machine learning typically refers to computational structures that can be programmed automatically by a learning algorithm.

The present art of power supply systems hasn’t explored the domain of power engineering by focusing on machine learning to improve the quality of compensators in an efficient manner. There is specifically a lack in the present art of how to improve the efficiency of active power filters by predicting signal disturbances with the help of machine learning. Accordingly, there is a need for improvements in the art to efficiently reduce the response and reaction time of active filters and further to improve the monitoring of the same.

In Chinese patent application publication, CN100461580,is disclosed a method for compensating by prediction of harmonic currents by measuring the loading current instantaneous value in real time and low passes the filtering, and gets the phase corner of the electric-net voltage by phase-locked loop at the same time; predicts the current volume of the sampling time currently according to the electricity-net current phase corner and the fitting parameter vector; adjusts the fitting parameter vector according to the predicted error; then, according to the adjusted fitting parameter vector, the electricity-net voltage phase corner and the requested compensatory controlling time- lapse, it can predict the loading current volume of the tk+DeltaT time; and predicts the request compensatory harmonic current signal of the tk+DeltaT at last.

Thus, even though the known solution fulfills certain requirements related to improving the efficiency of active power filters, there is still a need for further improvements. Summary

It is an object of the present invention to mitigate, alleviate or eliminate one or more of the above-identified deficiencies and disadvantages. The disclosure provides a method for controlling and monitoring an active power filter, comprising the steps of: measuring current and previous states of current and voltage signals across at least a part of a power supply system and store these in a memory; pre-processing the current and volt age allowing for extraction and detection of features; predicting future current and volt age levels using a machine learning component; wherein the predicted future current and voltage levels are provided to a pulse width modulation controller in said active power filter device, allowing for a desired output to be obtained.

A benefit of this method is that it allows for accurately predicting and compensating signal disturbances with the help of machine learning.

The pre-processing may further comprise the steps of; extracting and detecting features by using a kernel-based method; processing said features using filters such as low-pass filter and exponential smoothing.

By using this method in the pre-processing, the sample processing time and the harmonic error is kept at a minimum.

The pre-processing may further comprise; performing a discrete Fourier Transform, fast fourier transform, recursive discrete fourier transform or any combination thereof. The extracted and detected features may be at least one of individual harmonics, inter harmonics, fundamental frequency, flicker, and/or any other disturbances such as power quality events or variations in the signals.

The machine learning component may predict future voltage and current by predicting the amplitude and/or phases of at least one of fundamental component, harmonics, inter-harmonics at different frequencies as well as other power quality variations and events for at least one future cycle.

A benefit of predicting amplitudes and/or phases for at least one future cycle is that it allows for the active power filter to compensate before any power quality events or variations occur.

The machine learning component may comprise of a computational structure which is optimized using a learning algorithm, such as Artificial Neural Networks, Support Vector Machines, Decision Tree, Bayes Classifier or any combination thereof.

The proposed methods allow for a more robust and accurate prediction of current and voltage levels.

At least a part of said power supply system may supply electric arc furnaces.

A benefit of applying the method on power supply systems supplying non-linear loads is that it allows for mitigation of power quality variations and events occurring from said loads in an efficient manner. The method may comprise the step of storing and processing data related to controlling and monitoring the active power filter in a cloud-based repository, wherein the machine learning component can be updated by downloading data stored in said cloud-based repository.

A benefit of having a cloud based repository is that it allows for the method to be continuously updated and tuned by being connected to other devices in the cloud.

Hence, it allows for the active power filter to adapt to previous events and use previous data from other connected devices to learn from previous power quality variations and events. However, the method may be continuously updated and tuned even without a cloud based repository from the machine learning component. There may further be provided an active power filter, comprising a processor, at least memory, at least one input 106 and at least one output, wherein the processor is arranged to execute instruction sets for operating the method. Further a computer readable instruction set for controlling the output of an active power filter may be provided, the instruction set is arranged to perform the method.

The present invention will become apparent from the detailed description given below. The detailed description and specific examples disclose preferred embodiments of the invention byway of illustration only. Those skilled in the art understand from guidance in the detailed description that changes and modifications may be made within the scope of the invention.

Hence, it is to be understood that the herein disclosed invention is not limited to the particular component parts of the device described or steps of the methods described since such device and method may vary. It is also to be understood that the terminology used herein is for purpose of describing particular embodiments only, and is not intended to be limiting.

Brief description of the drawings

In the following the invention will be described in a non-limiting way and in more detail with reference to exemplary embodiments illustrated in the enclosed drawings, in which: Figure 1 is a schematic block diagram illustrating an example of an active power filter device;

Figure 2 is schematic block diagram illustrating an exemplary processing and control unit for an active power filter;

Figure 3 is schematic block diagram illustrating an exemplary method for power quality events and variation compensation;

Figure 4 is schematic block diagram illustrating an exemplary method for power disturbance compensation;

Figure 5 is a schematic diagram of a machine learning component; and Figure 6 is a schematic diagram of self-programmable power electronic device.

Detailed description

In the following detailed description, some embodiments of the present disclosure will be described. However, it is to be understood that features of the different embodiments are exchangeable between the embodiments and may be combined in different ways, unless anything else is specifically indicated. Even though in the following description, numerous specific details are set forth to provide a more thorough understanding of the provided device and method, it will be apparent to one skilled in the art that the device and method may be realized without these details. In other instances, well known constructions or functions are not described in detail, so as not to obscure the present disclosure.

The term “power quality” refers to the ability of electrical equipment to consume the energy being supplied to it. There are several factors to power quality events and variations such as voltage sags, momentary power interruptions, electrical noise, ground loops, high-speed transients, flickers, inter-harmonics and harmonics. The power quality is influenced by the electric appliances connected to a power supply system such as a power grid.

The term “harmonics” refers to a waveform whose frequency is a multiple of the fundamental frequency. By having a non-linear load connected to the power supply system, the current shape will be distorted and harmonics may be introduced. Examples of non-linear loads are uninterrupted power suppliers, adjustable speed drives and electric arc furnaces. This may cause overheating or performance problems for equipment and loads in power systems. The distortion signals of harmonics may be compensated by injecting a compensation current with respect to the harmonics of the load. This may be performed by for instance an active power filter.

The term “interharmonics” refer to frequencies that can be observed between the harmonics of the power frequency voltage and current, further frequencies which are not an integer of the fundamental. They can appear as discrete frequencies or as a wide band spectrum.

The term “active power filter” refers to a device having the purpose to compensate for power quality variations and events by injecting active power with the same frequency but reverse phase to cancel out the harmonics. Active power filters (APF) are developed for the compensation of time- and frequency-varying harmonic and interharmonic com ponents of loads with nonlinear and stochastic characteristics. These devices attempt to continuously detect the amplitudes and phases of the undesired frequency components and filter them out by supplying currents or voltages with the same amplitudes. In the case of power quality events and/or variations, the phases may be shifted by at least half-cycle for each component. Hence, each generated current or voltage component should be ideally out-of-phase and exactly at the same amplitude so that perfect com pensation is achieved. Therefore, an important key factor in these types of methods is fast and accurate detection of the harmonics and interharmonics since misdetection may lead to amplification of the undesired current and voltage components. The active power filter may be directly connected to a grid or locally connected.

The term “Machine learning” typically refers to computational structures that can be pro grammed automatically by a learning algorithm. An instance of such a structure with de fined and fixed parameters is referred to as a model.

Machine learning is typically data-driven, meaning that it is achieved through the obser vation of data samples. In supervised learning each data sample is a pair of the form (x; y) where x is the input and y is the desired output of the model, known as the label. Dur ing the learning process, the model is altered in away so that the output will be more similar to the labels when given the corresponding inputs.

The term “pre-processing” refers to a method of processing or restructuring the data in order to extract features for further processing or prediction.

The term “modified synchronous reference frame (MSRF)” refers to a method of analysis that may be used in the pre-processing step to obtain both positive and negative se quences of d, q, and zero components of harmonics and inter-harmonics, as well as the fundamental component.

The term “exponential smoothing” refers to a method used to estimate the DC variation of the data to obtain a low-pass filtered form of the positive and negative sequences of the d and q components of the frequency components. The disclosure relates to an active power filter device 100, comprising a processing and control unit 201 , at least one memory 202, at least one input 106 and at least one output 107, where the processing and control unit 201 is arranged to execute instruction sets for operating a method. Figure 1 discloses an active power filter device 100 comprising a processing and control unit 201. The active power filter device 100 may be in connection to a load 101 being at least partially supplied by a power supply system. Further, the active power filter device 100 may be connected to a cloud-based repository 102 allowing for all processed and measured data to be stored in a centralized external server 105 in connection to a database 101. This may allow for all measured and processed data on an active power filter device 100 to be stored and used for large- scale optimisation to solve general high computational problems that benefit multiple devices. Accordingly, the cloud based repository 102 may further also transmit stored or processed data and other information to individual active power filter devices 100 connected to the cloud based repository 102. As seen in figure 1 the active power filter device 100 may be connected to an external device located remotely to the device, e.g. in a global network or a cloud-based repository 102 allowing for updates of instructions/instruction sets and/or data. The active power filter device 100 may further comprise an actuating device allowing for actuation of switching strategies.

In Figure 2, the processing and control unit 201 of the active power filter device 100 is shown in more detail. The processing and control unit 201 may comprise at least one processor, at least one memory 202, at least one input interface 205, at least one output interface 206, and at least one communication interface 207. The processing and control unit 201 may comprise at least one pre-processing module 250, at least one machine learning component/module 260, and at least one actuate control module 270. The pre processing module 250 may process and perform the step of pre-processing 302 which will be further discussed herein. The at least one actuate control module 270 may control an actuating device allowing for execution of switching strategies in the active power filter 100.

The processing and control unit 201 may for instance comprise a microprocessor, digital signal processor (DSP), graphical processing unit (GPU), embedded processor, field programmable gate array (FPGA), or ASIC (Application specific integrated circuit), or a combination of these. The memory (computer readable storage medium) 202 may be a non-transitory or transitory computer readable memory and arranged to store instructions or instruction sets for execution by the processing and control unit 201 and to store data. Instruction sets are preferably stored in a non-transitory memory such as solid state (SSD), magnetic disk drive storage, and optical storage such as CD, DVD, or Blu-ray, or persistent solid state memory technology such as flash memory or memory card. The storage unit may also comprise a combination of storage types. The method may be realized in a computer program product and/or stored in a computer-readable storage medium. The memory 202 may be non-transitory or transitory computer- readable storage medium which stores one or more programs configured to be executed by one or more processing and control units 201 of an electronic device with or without a display apparatus and one or more input devices. The computer program product may be pre-installed in the device or delivered on a storage medium such as for instance SSD, magnetic storage, optical storage, or delivered on a network connection as a signal with a suitable protocol, for instance Ethernet using Internet Protocol (IP) or wirelessly with suitable radio protocol such as cellular technologies or short or medium range local area network technologies for installation in the device.

Further, there may be a computer readable instruction set for controlling the output 107 of an active power filter device 100, the instruction set arranged to perform a method for controlling and monitoring an active power filter device 100 as will be described herein. The disclosure further relates to a method 300 for controlling and monitoring an active power filter device 100, comprising the steps of:

- measuring 301 current and previous states of current and voltage signals across at least a part of a power supply system and store these in a memory 202;

- pre-processing 302 said current and voltage, allowing for extraction and detection of features;

- wherein a machine learning component further predicts 303 future current and voltage levels 104;

- wherein the predicted future current and voltage levels are provided 305 to the active power filter device 100, allowing for a desired output 107 to be obtained.

The desired output 107 refers to the output wherein the power disturbances such as harmonics or interharmonics are at least partially compensated. The prediction 303 of future current and voltage levels allows for future power quality variations and to be de tected and mitigated with a fast response-time.

The method shown in figure 3 may accurately predict 303 power quality variations and events within a significant time window on a power signal recorded from highly nonlinear high-voltage systems. Additionally, there is an improved compensation 305 accuracy compared to other predictive methods. The compensation step 305 of figure 3 may include the step of providing the predicted future current and voltage levels to the active power filter device 100, allowing for the desired output 107 to be obtained.

It may be possible to predict power quality events and variations with high accuracy up to one cycle ahead in time. This is may allow for pre-processing 302 of the measured signal and computation of the prediction, which allows an active power filter device to commence the compensation exactly when the disturbance is predicted to occur. This may effectively reduce the response and reaction time of active power filter devices to zero, while maintaining very high compensation accuracy.

Figure 3 discloses the method of controlling and monitoring an active power filter 100. Figure 3 discloses the steps of measuring 301 the signal i.e. it discloses the step of extracting and detecting current and voltage levels from a non-linear load 101. A voltage interface unit and a current interface unit may be utilized during the measurement step 301. The voltage and current levels may be measured in a predetermined measuring point. As disclosed in Figure 3 the measurement signals achieved in the measuring step 301 may further be pre-processed 302 and predicted 303. Further, the predicted values may be reconstructed 304 i.e. an original waveform may be reconstructed 304 in order to be prepared to be fed/provided to the active power filter device 100 for injecting a compensation 305 as shown in the last step of the method disclosed in figure 3.

A data acquisition (DAQ) unit may transform the values from analog to digital and may allow the current and voltage values to be digitized and used in the machine learning parts in the measurement step 301 of the method disclosed in figure 3.

The machine learning component 260 may be arranged to detect power quality events and variations. Furthermore, the machine learning component 260 may predict 303 future values before the active power filter device 100 compensates 305 the harmonics and/or interharmonics. Also the active power filter device 100 may compensate 305 other power quality events and variations.

The methods of power quality event and variation detection performed by a machine learning component 260 after pre-processing 302 may be at least one of time-domain or frequency-domain analysis. In frequency-domain analysis, power quality variations such as harmonics and inter-harmonics, and events, may be analysed by Discrete Fourier

Transform (DFT), Fast Fourier Transform (FFT) and/or Recursive Discrete Fourier

Transform (RDFT). The pre-processing 302 may further comprise; performing a FFT, RDFT, DFT or any combination thereof. Further, sliding Discrete Fourier Transform may be used in order to detect harmonics and synchronize to act as an adaptive-harmonic compensator.

In the pre-processing step 302, there may be utilized a Time-domain (TD) approach which may predict 303 the future samples of the original load current waveform before estimating power quality events and variation samples. The TD approach may predict 303 samples with a prediction error less than 1%.

The frequency domain (FD) approach may predict 303 the future samples of the original load current and voltage frequency components.

The machine learning component 260 after the pre-processing 302 step may use a time domain machine learning based method and/or a frequency domain machine learning based method to estimate and predict samples. In time domain ML based approach, an ML algorithm may predict 303 the future samples of a load 101 such as for instance an electric arc furnaces current waveform and then, MSRF + ES may be used as a part of the pre-processing to detect the harmonics and interharmonics. In a frequency domain machine learning based approach, samples from power quality events and variations may be extracted using sliding DFT and further are followed by a machine learning algo rithm to predict 303 power quality events and variations such as harmonics and inter harmonics components. A benefit of accurately predicting 303 future samples is that it may significantly reduce response time of any estimation method. Furthermore, the re sponse time delay problem for the active power filter device 100 may be eliminated so that real-time operation may be possible. This is specifically powerful for highly time- and frequency-varying spectrum cases such as nonlinear-load currents.

The machine learning based method 300 utilized in the disclosure may provide a method that essentially reduces compensation delay time of active power filter devices 100 to zero, by modelling and prediction 303 of the frequency components of the load current signal. Response and reaction times of the active power filter device 100 may be re duced to zero, by estimating the future samples of power quality variations and events such as the harmonics and interharmonics of the non-linear load currents, the algorithms may be applied in parallel for each frequency component using graphical processing units 201. The method 300 may extract demanded frequency domain component, whether it is a harmonic or an interharmonic at any desired resolution. The method may perform prediction on all of the frequency components of signal simultaneously.

The pre-processing 302 step may further comprise the steps of;

- Extracting and detecting features by using a MSRF matrix method

- Processing said features using low-pass filtering and exponential smoothing.

The MSRF+ES method is utilized to decompose the signal into the phases and amplitudes of its frequency components. The MSRF+ES algorithm may have a minimum possible response time of at least 1ms with a sampling rate of at least 25KHz, which corresponds to 512 samples/cycle for the fundamental frequency 50.0Hz, 1ms approximately equals to 25 samples

The extracted and detected features from the pre-processing step 302 may be at least one of individual harmonics, inter-harmonics, fundamental frequency, flicker, and/or any other disturbances in the signals. The term “other disturbances” refers to other power quality events such as voltage fluctuations, voltage and current unbalance, high frequency voltage noise or any other events.

The machine learning component 260 may predict 303 future voltage and current by computing the amplitude and/or phases of at least one of harmonics, inter-harmonics or fundamental frequency at different frequencies for at least one future cycle. Furthermore, the machine learning component 260 may predict 303 future voltage and current for less than one future cycle. Further, the machine learning component 260 may adjust the number of future cycles calculated when predicting 303 said future voltage and current.

In other words, said machine learning component may predict 303 the future voltage and future current levels by computing the amplitude and/or phases of at least one of harmonics, inter-harmonics or fundamental frequency at different frequencies for an adjustable amount of cycles. Accordingly, the predict-ahead time is a parameter, which may introduce a trade-off between desired prediction 303 and compensation accuracy and response time, which may be adjusted by the user of the active power filter device 100, in other words the disclosure provides an adjustable prediction time horizon. The method allows for mitigation of active power filter reaction time delay.

The machine learning component 260 may comprise of a structure capable of containing a learning algorithm, said learning algorithms being Artificial Neural Networks, Support Vector Machines, Decision Tree, Bayes Classifier or any combination thereof. The structure may be a computational structure.

A type artificial neural networks that may be used is convolutional Neural Networks (CNN) or Long Short-Term Memory (LSTM) or a type of recurrent neural networks (RNN) or any combination thereof.

LSTM units are composed of cells, which can model and remember non-linear and time- varing signal and automatically learn their pattern. In the disclosure, LSTM may be used by the machine learning component 260 for prediction 303 and classification of time se ries data such as harmonics, interharmonics and raw current signal itself. An LSTM cell is a stateful operator that, based on the input at some time, computes an output 107 and updates its internal state. The input may be a current waveform segment or frequency components extracted from such a segment. A current from a load 101 may be seg mented and paired to a corresponding ground truth prediction. At least a part of the power supply system may supply at least one non-linear load 101. At least a part of the power supply system may supply electric arc furnaces or variable frequency drives. The disclosure may be specifically beneficial for electric arc furnaces. The power supply system may be a power grid supplying any non-linear load 101. Elec tric arc furnaces may result in highly nonlinear and stochastic load characteristics in the power supply system due to their operation principles.

The method 300 may optionally further comprise a step of storing and processing data related to controlling and monitoring the active power filter device 100 in a network cloud-based repository 102. Accordingly, all the current and previous states of current and voltage signals across a power supply system may be stored and processed in the cloud-based repository 102. Further, all data related to pre-processing 302 such as the extraction and detection of features may also be stored and processed in the cloud- based repository 102. Further, the output 107 of the active power filter device 100 may be stored and processed in a cloud-based repository 102. In the cloud-based repository 102 the data may be analyzed and stored in order to use such data and analysis in other events/applications/devices, allowing for different power supply systems to make use of the data for a more efficient and proactive handling of power quality events with the help of machine learning components. This is disclosed in figure 1 where data relating to all steps in the method may be stored in a cloud-based repository 102.

The machine learning component 260 may be updated by downloading data stored in the cloud-based repository 102. This may help the machine learning component 260 to recognize patterns and previous data, and adjust the computing of the amplitude and/or phases of at least one of harmonics, inter-harmonics or fundamental frequency at differ ent frequencies for future cycles. Figure 2 discloses this where there is seen at least one machine learning component/module 260 in a process having an input interface 205 and an output interface 206. The output signal may be computed and may be trained/updated regularly to minimise the error of the output 205. The processing and control unit 201 may be using learning algorithms such as Artificial Neural Networks, Support Vector Machines, Decision Tree, Bayes Classifier or any combination thereof. The error of the output 206 may be computed by an oracle unit, which may depend on the input 205. The method 300 may be performed by further comprising a switch control strategy which is shown in figure 4. Figure 4 discloses a method 400 for controlling and monitoring an active power filter device 100, comprising the steps of:

- measuring 401 current and previous states of current and voltage signals across at least a part of a power supply system and store these in a memory 202;

- pre-processing 402 said current and voltage signals by using a machine learning component 260, allowing for extraction and detection of features 103;

- wherein the machine learning component 260 further predicts 303 future current and voltage levels 104;

- wherein the predicted future current and voltage levels are provided to a second machine learning component which computes a switch control strategy for the active power filter

- actuating 404 the switch control strategy in the active power filter 100, al- lowing for the desired output 107 to be obtained.

The generated switch control strategy is based on the predictions from the machine learning component 260. Accordingly, the processing and control unit 201 may further comprise a second machine learning component, provided for computing 404 a switch control strategy for the active power filter 100. An exemplary machine learning component 500 is outlined in Figure 5. The machine learning component 500 is provided an input 503 which is used by a model 501 to compute an output 504. The model 501 which can be an algorithm built on a mathematical model may be trained/updated regularly to minimise the error of the output 504. The error may be computed by an oracle unit 502, which may depend also on some external input 503 (e.g. observation of the environment, additional measurements within or outside the self-programmable power electronic device, list of known correct outputs for all inputs).

The pre-processing steps, and the machine learning training (performed by a machine learning component) and inference may be run on dedicated processing hardware, such as a central processing unit, graphics processing unit, field programmable gate array etc for accelerated parallel computation and any other suitable calculation.

In Figure 6 an exemplary application of a machine learning component 500 is shown in relation to a self-programmable power electronic device 600 which may be an active power filter 100. We propose a modification of the switch control component in a wide range of power electronic devices 600, which in part comprise replacing the control component with a machine learning component 601 as seen in figure 5, which can automatically learn the optimal control strategy for the particular power electronic device 600. A power electronic device 600 may be a rectifier or an inverter, or an active power filter or any other device. The switching strategy is empirically altered automatically over time, and adapts to changes in the environment such as placement, nearby devices, and wear of device components. By monitoring and analysing how the switching behaviour changes over time it can be possible to predict lifetime of components and report potential future component failures. As an alternative, the power electronic device 600 can be equipped with a dedicated software component for modelling and predicting wear of the device hardware. Accordingly, the method 300 may be directed to controlling and monitoring any power electronic device 600. As seen in Figure 6, the power electronic device 600 may comprise a power electronic component 606 having an input 602 and an output 603, wherein the input 602, the output 605, and the power electronic component 606 may be controlled and monitored by a machine learning component 601.

As seen in Figure 6, the solution may be reported to an external device located in an external network, e.g. a cloud based solution. While online machine learning is performed in small scale on individual power electronic devices 600, large-scale optimisation can be done in centralized servers to solve general high-computational problems that potentially benefit multiple power electronic devices 600. The result from the centralised large-scale optimisation is broadcasted to individual power electronic devices 600 that may accept the new strategy or partially merge with the local model in order to continuously increase the performance.

Continuous modelling of the grid for active power quality compensation is possible. For instance, for AC-DC inverters, the control strategy is learned automatically by observing the signal on the grid and optimising the switching strategy with respect to the corresponding output of the inverter, which implies an implicit modelling of the grid. In particular, this apples to disturbances in the grid, such as harmonics and power quality events. By allowing post deployment online learning, the inverter may as such also learn to compensate such disturbances.

It may be possible to adopt methodology from the emerging field of cooperative/ federal/distributed machine learning and utilize the computational power of many power electronic devices 600 solve machine learning problems. Accordingly, each individual power electronic device 600 continuously solves the switching and compensation problem based on local grid observation, and shares potentially generally useful strategies/experiences with other power electronic devices 600, which can choose to incorporate the received knowledge. The size of a predict-ahead time is a parameter, which introduces the trade-off between desired prediction and compensation accuracy and response time, the solution of which depends on the application. The end-used can hence tune the response time of a power electronic device 600. It should be noted that the word “comprising” does not exclude the presence of other elements or steps than those listed and the words “a” or “an” preceding an element do not exclude the presence of a plurality of such elements. It should further be noted that any reference signs do not limit the scope of the claims, that the invention may be at least in part implemented by means of both hardware and software, and that several “means” or “units” may be represented by the same item of hardware.

The above mentioned and described embodiments are only given as examples and should not be limiting to the present invention. Other solutions, uses, objectives, and functions within the scope of the invention as claimed in the below described patent embodiments should be apparent for the person skilled in the art.

Claims

1. A method (300) for controlling and monitoring an active power filter device (100), comprising the steps of:

- measuring (301) current and previous states of current and voltage sig nals across at least a part of a power supply system and store the current and voltage signals in a memory (202);

- pre-processing (302) said current and voltage signals, allowing for extrac tion and detection of features;

- wherein a machine learning component (260) further predicts (303) future current and voltage levels;

- wherein the predicted future current and voltage levels are provided to said active power filter device (100), allowing for a desired output (107) to be obtained.

2. The method (300) according to claim 1 , wherein said pre-processing (302) comprise the steps of;

- extracting and detecting features by using a kernel-based modified synchronous reference frame matrix method

- processing said features using low-pass filtering and exponential smoothing.

3. The method (300) according to claim 1 or 2, wherein said pre-processing (302) further comprise performing at least one of a discrete Fourier transform, fast Fourier transform, recursive discrete Fourier transform or any combination thereof.

4. The method (300) according to any of claims 1-3, wherein said extracted and detected features are at least one of individual harmonics, inter-harmonics, fundamental frequency, flicker, and/or any other disturbances in the signals.

5. The method (300) according to any of claims 1-4, wherein said machine learning component (260) predicts future voltage and current by predicting the amplitude and/or phases of at least one of harmonics, inter-harmonics or fundamental frequency at different frequencies for at least one future cycle.

6. The method (300) according to any of claims 1-5, wherein said machine learning component (260) comprises of a computational structure capable of containing a learning algorithm, said learning algorithm being Artificial Neural Networks, Support Vector Machines, Decision Tree, Bayes Classifier or any combination thereof.

7. The method (300) according to any of claims 1-6, wherein at least a part of said power supply system supplies electric arc furnaces.

8. The method (300) according to any of claims 1-7, further comprising the step of storing and processing data related to controlling and monitoring the active power filter device (100) in a cloud-based repository (102), wherein the machine learning component (260) can be updated by downloading data stored in said cloud-based repository (102).

9. The method (300) according to any of claims 1-8, further comprising the step of compensating (305) any power quality events and variations based on the predicted voltage and current levels.

10. An active power filter device (100), comprising at least one processing and control unit (201), at least one memory (202), at least one input interface (106) and at least one output interface (107), wherein the processing and control unit

(201) is arranged to execute instruction sets for operating the method according to any of above claims.

11. A computer-readable medium storing instruction sets for controlling and monitoring the output (107) of an active power filter device (100), the instruction sets arranged to be executed in a processing device and arranged to perform the method according to claims 1-10.