Method and system for monitoring a voltage grid, method for training an artificial intelligence to predict a future state of a voltage grid, computer program, and computer-readable data carrier

Abstract

The invention relates to a method for monitoring a voltage grid (12). According to the method, at least one first characteristic variable is detected at a first point in time, and the first characteristic variable (u) of the voltage grid (12) is detected at a second point in time. The first characteristic variable detected at the first point in time and the first characteristic variable detected at the second point in time are fed to a processor unit (16) which processes the first characteristic variable detected at the two points in time together such that a future state of the voltage grid (12) is predicted on the basis of the first characteristic variable detected at the at least two different points in time, said predicted future state of the voltage grid (12) being output. The invention additionally relates to a method for training an artificial intelligence (24), to a system (10), to a computer program (18), and to a computer-readable data carrier (20).

Description

[0001] Method and system for monitoring a voltage network, method for training an artificial intelligence to predict a future state of a voltage network, computer program and computer-readable data carrier

[0002] The invention relates to a method for monitoring a power grid. Furthermore, the invention relates to a method for training an artificial intelligence system to predict the future state of a power grid. Furthermore, the invention relates to a system for monitoring a power grid, a computer program, and a computer-readable data carrier.

[0003] Today, voltage grids are already significantly more volatile than was the case in the past. This is due, among other things, to the constantly rising but fluctuating energy demand and the increasing, even more irregular, feed-in of renewable energies into the supply grid. Furthermore, it can be assumed that voltage grids will become even more volatile in the future, as the proportion of renewable energy fed into the grid will continue to rise and, at the same time, the voltage grids will be subject to greater short-term loads, for example, due to the increasing use of electromobility, particularly in the automotive sector. However, the increasing volatility of voltage grids is resulting in a decline in grid quality, which can lead to increased disruptions and even outages, which must be avoided.

[0004] It is therefore already of interest to grid operators, electricity users, and customers to record the current status of the voltage grid, analyzing past grid-disturbing events to identify potential faults. Smart devices such as the DEHN record are available for this purpose. These devices record and analyze a characteristic parameter of the voltage grid to determine the current power quality of the voltage grid. Such smart devices are used, for example, by grid operators, electricity users, or customers.

[0005] From EP 3 336 995 A1, for example, a method and a system are known by which a current or instantaneous operating state of a sub-network of an energy supply network is determined.

[0006] In this respect, the voltage grid being analyzed can be the local voltage grid, also known as the building grid, i.e., the voltage grid at the customer's location, such as a private household or an industrial building. In other words, a low-voltage grid is being monitored.

[0007] However, due to the increasing volatility of the voltage grids in the future, the analysis of the current grid quality will probably no longer be sufficient, so that improved analysis and monitoring of the voltage grid will be required.

[0008] The object of the invention is to provide a possibility with which the voltage network can be monitored cost-effectively and with high quality.

[0009] The object is achieved according to the invention by a method for monitoring a voltage network, which comprises the steps:

[0010] Recording at least one first characteristic parameter of the voltage network at a first point in time,

[0011] Recording the first characteristic parameter of the voltage network at a second time which is different from the first time,

[0012] Feeding the first characteristic parameter detected at the first time and the first characteristic parameter detected at the second time into a processor unit which jointly processes the first characteristic parameter detected at the first time and the first characteristic parameter detected at the second time in such a way that a future state of the voltage network is predicted based on the first characteristic parameter detected at the at least two different times, and outputting the predicted future state of the voltage network.

[0013] The basic idea of ​​the invention is that future grid quality is estimated or forecast based on historical and / or current data from the voltage grid, enabling the grid operator and / or the electricity customer / user to detect future disruptions at an early stage. In other words, the grid operator and / or the electricity customer / user can predict the time of an impending grid disruption and its intensity. This makes it possible to initiate appropriate countermeasures to avert the impending grid disruption, at least reduce its intensity, or generally minimize the impact of the impending grid disruption on certain areas.

[0014] The invention is based on the realization that, due to the volatile nature of the voltage grid, the future behavior of the voltage grid in terms of its network quality (network status) is often more important for the grid operator and / or the electricity customer / user than the current network status, which cannot be responded to anyway. Looking ahead makes it possible to initiate appropriate measures in a timely manner to counteract the predicted grid disruption, i.e., the deterioration of the voltage grid's condition.

[0015] In other words, the grid operator and / or the electricity customer / user will be able to make the voltage grid more resilient to disruptions expected in the future.

[0016] The future state of the voltage grid can be predicted for the next few hours, days or weeks, giving the grid operator and / or the electricity customer / user sufficient time to initiate any countermeasures in a timely manner to counteract any forecast disruption or deterioration.

[0017] The countermeasures to be taken could, for example, consist of avoiding further loads during a predicted weak phase of the voltage grid, thereby ensuring the fail-safe operation of devices and / or machines whose operation should not be disrupted. In this respect, the countermeasure could consist of disconnecting non-essential machines or devices from the grid during the predicted weak phase of the voltage grid or not connecting them to the voltage grid in order to reduce the load or prevent it from increasing further.

[0018] In principle, the voltage grid can be a local network, also known as a building network, assigned to a private household or an industrial building. The voltage grid can therefore be a low-voltage network, for example, a building network connected to a utility grid.

[0019] For forecasting, the first characteristic parameter can be recorded multiple times, i.e., at least at two different points in time. The values ​​of the first characteristic parameter recorded at the different points in time are fed into the processor unit for evaluation. Thus, the future time profile can be forecast from a historically recorded time profile of the first characteristic parameter of the voltage grid, which can extend up to the present. The future time profile can relate to the first characteristic parameter, so that its future time profile is forecast, which corresponds to the forecast future state of the voltage grid.

[0020] The first characteristic parameter alone allows for characterizing the current state of the voltage grid. Since the processor unit jointly evaluates at least two values ​​of the first characteristic parameter, recorded at different times, it is now possible to predict future states of the voltage grid accordingly, which would not be possible by evaluating a single characteristic parameter alone.

[0021] In particular, correlations between the at least two different points in time of the first characteristic parameter are exploited, on the basis of which a future state of the voltage network is reliably predicted.

[0022] In this respect, correlations can be inferred based on the joint evaluation of the same measured variable at different points in time. A second variable that differs from the first can also be used. The second variable can be a separate variable that is not necessarily a parameter of the voltage network.

[0023] The second quantity may be obtained from a database to include, for example, environmental data, in particular weather data, time data (working day, weekend, holiday, day and / or night) or usage data of systems in the environment, for example a train timetable or usage data of electrical charging systems in the environment.

[0024] In this respect, the second variable can be used to obtain information that encompasses a difference between the two different points in time. In particular, this makes it possible to determine a relationship between two different values ​​of the first characteristic parameter recorded at the two different points in time.

[0025] However, the second quantity can also be a second characteristic parameter of the voltage network that is different from the first characteristic parameter of the voltage network.

[0026] In particular, the temporal progression of the first characteristic parameter, which was recorded at least at the first time and at the second time, can be evaluated together with the second parameter, which is different from the first characteristic parameter of the voltage grid. In this respect, a time series of the first characteristic parameter can be evaluated together with other data, for example, data obtained from a database such as weather data, time data, and / or usage data from systems in the surrounding area.

[0027] This can result in a data fusion of two different variables, which allows a glimpse into the future state of the voltage network. In other words, the joint processing of the two different variables exploits a correlation between these variables, which allows conclusions to be drawn about the future behavior of the voltage network. Therefore, the future state of the voltage network can be predicted accordingly.

[0028] In principle, the procedure for monitoring a voltage network can also include the following steps:

[0029] Recording at least one first characteristic parameter of the voltage network, in particular at a first time and a second time,

[0030] Recording at least one second characteristic parameter of the voltage network which is different from the first characteristic parameter of the voltage network,

[0031] Feeding the first characteristic parameter, in particular the values ​​of the first characteristic parameter recorded at the first time and at the second time, and the second characteristic parameter into a processor unit which processes the two different characteristic parameters together in such a way that a future state of the voltage network is forecast based on the two different characteristic parameters, and

[0032] - Output of the predicted future state of the voltage network.

[0033] One aspect provides that the first characteristic parameter and / or the second characteristic parameter is a voltage, a current, a power, a frequency, a distortion ("Total Harmonic Distortion" - THD), a harmonic (up to the 50th harmonic), a reactive power, and / or an energy value. The corresponding characteristic parameter can be recorded for each phase of a multi-phase voltage network, for example, for phases L1 to L3.

[0034] A further aspect provides that the first characteristic parameter is recorded multiple times, so that a time sequence of the first characteristic parameter comprising more than two points in time is available, which is processed. Alternatively or additionally, the second parameter, in particular the second characteristic parameter, can also be recorded multiple times, so that a time sequence of the second parameter is available, which is processed. In this respect, the respective parameter, i.e. the first characteristic parameter or the second parameter, is recorded multiple times in succession, in particular periodically, in order to obtain a time series or time sequence of the corresponding parameter. Based on the corresponding time series or time sequence, the future state of the voltage network is forecast. In other words, at least the historical behavior of the first characteristic parameter is recorded and taken into account when forecasting the future state.This can also be done accordingly for the second variable, for example the second characteristic parameter. For example, a change in the first characteristic parameter can be determined which, together with the temporal change in the second variable, is clearly specific to a specific future state of the voltage grid. The temporal change can be a gradual deterioration, but also general characteristic oscillations or trends that, considered on their own, would initially have no significance. The change in the second variable can be, for example, a different day, a changing weather situation, a change due to the timetable, a change in charging system usage, or something similar.By combining the two variables, it is possible to obtain information that is relevant for the future state of the voltage network, provided that the variables influence each other.

[0035] According to one embodiment, the processor unit comprises an artificial intelligence that receives at least the first characteristic parameter recorded at at least two different points in time as an input and outputs the future state of the voltage grid as an output. The artificial intelligence can be trained to recognize the corresponding correlation between the values ​​of the first characteristic parameter recorded at the different points in time and to learn corresponding technical relationships in the voltage grid, thereby enabling reliable or very accurate forecasts regarding the future state of the voltage grid. The invention therefore also relates to a method for training an artificial intelligence to forecast a future state of a voltage grid. The training method comprises the steps:

[0036] Providing a training data set for the artificial intelligence, which comprises at least a first characteristic parameter of the voltage network at a first point in time, the first characteristic parameter of the voltage network at a second point in time and an actual state of the voltage network at a third point in time, which is later in time than the first point in time and the second point in time,

[0037] Feeding the first characteristic parameter at the first time and the first characteristic parameter at the second time into a processor unit having the artificial intelligence to be trained, wherein the processor unit having the artificial intelligence jointly processes the first characteristic parameter recorded at the different times and outputs a predicted future state of the voltage network at the third time,

[0038] Comparing the predicted future state of the voltage network at the third time point with the actual state of the voltage network at the third time point, which is part of the training data set, in order to determine a deviation between the predicted future state of the voltage network at the third time point and the actual state of the voltage network at the third time point, and

[0039] Feeding back the deviation between the predicted future state of the voltage grid at the third time point and the actual state of the voltage grid at the third time point in order to adjust weighting factors of the artificial intelligence to be trained, provided that the deviation lies outside a tolerance range.

[0040] The artificial intelligence can therefore be trained using at least one training data set, which includes the first characteristic parameter both at a first point in time and at a second point in time, in particular a corresponding time sequence or time series of the first characteristic parameter, as well as an actual future state of the voltage network that exists at the third point in time, which is chronologically subsequent to the points in time at which the first characteristic parameter was determined or recorded, i.e., corresponding values ​​of the first characteristic parameter. In particular, the first point in time and the second point in time do not coincide, so that the first characteristic parameter was recorded at two different points in time, in particular the respective values ​​of the first characteristic parameter.

[0041] In principle, when training the artificial intelligence, it can also be provided that the artificial intelligence does not output the state of the voltage network at a discrete point in time, but at an interval that includes the third point in time, i.e. predicts the future state of the voltage network for a future period.

[0042] The corresponding tolerance range for the deviation can be predefined and / or set by a user, e.g. as a percentage or variance.

[0043] In principle, artificial intelligence trained using the method described above can be used to predict the future state of the voltage grid.

[0044] In other words, the processor unit comprises a trained artificial intelligence, in particular an artificial intelligence trained according to the method described above.

[0045] Another aspect provides for the artificial intelligence to have at least one artificial neural network, e.g., a recurrent neural network (RNN) or a convolutional neural network (CNN). The artificial intelligence can have a long short-term memory (LSTM) network or a gated recurrent unit (GRU). Such neural networks allow a prediction of a future state based on time series. The artificial intelligence thus learns from past data (historical data), predicting the future state of the voltage network.

[0046] In particular, a multidimensional vector is generated which comprises the first characteristic parameter at the different points in time, i.e. the first characteristic parameter recorded at at least two different points in time, wherein the multidimensional vector is processed by the processor unit. The artificial neural network is accordingly configured to process a multidimensional vector which, for example, comprises several points in time of the first characteristic parameter in order to thus predict the future state of the voltage network. In addition to the first characteristic parameter, the multidimensional vector can also comprise further variables, for example at least one second variable, in particular at several points in time. The second variable can be a variable obtained from a database or a second characteristic variable of the voltage network.For example, the multidimensional vector has a dimension of up to eight, so the vector can contain data from up to eight different characteristic quantities, such as information regarding voltage, current, power, frequency, distortion, harmonics, reactive power, and / or energy. The additional quantities besides the first characteristic quantity can generally be singular values ​​or (also) corresponding time series.

[0047] Furthermore, at least a future temporal course of a characteristic parameter can be forecast as the future state of the voltage grid. The artificial intelligence can therefore, for example, predict a future voltage curve for a specific time interval. Artificial intelligence can also predict a future current curve for the corresponding time interval. The time interval can be defined by a user, so that forecasts for a targeted time interval are possible. This is particularly interesting when certain machines or devices connected to the voltage grid to be monitored must operate without disruption within a defined period of time. The operator or grid operator therefore has an interest in ensuring that no or only very minor disruptions to the voltage grid occur, at least during this period.

[0048] In principle, the future temporal course of a characteristic parameter can be predicted, for example, the future temporal course of the previously recorded characteristic parameter. However, the future temporal course of a different characteristic parameter can also be predicted than the one recorded and fed into the processor unit.

[0049] Furthermore, the invention achieves the object by providing a system for monitoring a voltage network, wherein the system comprises a processor unit configured to perform one of the aforementioned methods. The system can be a measuring and analysis device integrated into the voltage network to be monitored, e.g., on a DIN rail. The aforementioned advantages thus apply analogously to the system.

[0050] In particular, the system may comprise a device, for example a measuring and / or evaluation device or a protective device such as a surge protection device (SPD), wherein the device comprises the processor unit.

[0051] The system, in particular the device, can also comprise at least one data transmission device, whereby the system, in particular the device, can communicate with other devices, for example, and in particular receive data that is additionally used for the evaluation or prediction of the future state of the power grid. Furthermore, the predicted future state of the power grid can be output via the data transmission device, which is designed, for example, as a communication interface.

[0052] Furthermore, the invention provides a computer program comprising program code means for performing the steps of one of the aforementioned methods when the computer program is executed on a processor unit, e.g., the processor unit of the aforementioned system. Furthermore, the invention provides a computer-readable data carrier on which the computer program of the aforementioned type is stored.

[0053] The advantages mentioned above thus arise in an analogous manner for the computer program and the computer-readable data carrier.

[0054] Further advantages and features of the invention will become apparent from the following description and the drawings, to which reference is made. In the drawings:

[0055] Figure 1 is a schematic representation of a system according to the invention,

[0056] Figure 2 is an overview illustrating the method according to the invention for monitoring a voltage network, and

[0057] Figure 3 is an overview illustrating the inventive method for training an artificial intelligence.

[0058] Figure 1 shows a system 10 that monitors a voltage network 12, which in the illustrated embodiment is a multi-phase building network connected to a supply network 13. In this respect, the voltage network 12 is, for example, a low-voltage network.

[0059] The system 10 comprises a plurality of measuring devices 14, which are connected to a device 15 for signal transmission, for example, a protection device or a measuring and / or evaluation device. In the illustrated embodiment, both the measuring devices 14 and the device 15 are connected or integrated with the voltage network 12 to be monitored. Alternatively, the device 15 can also be designed separately from the voltage network 12, so that it is only connected to the measuring devices 14 for signal transmission.

[0060] The device 15 comprises a processor unit 16 which is configured to execute a computer program 18 with program code means in order to monitor the voltage network 12.

[0061] The computer program 18 can be installed on the processor unit 16 from a computer-readable data carrier 20 on which the computer program 18 is stored, so that the system 10, in particular the device 15 having the processor unit 16, is able to carry out a corresponding method for monitoring the voltage network 12.

[0062] The device 15 therefore has a corresponding interface via which the data carrier 20 can be coupled to the device 15 in order to install the computer program 18.

[0063] The computer program 18 can also be transmitted via a data transmission device 22, which is designed, for example, as a communication interface, for example for communication, in particular wireless communication, with the measuring devices 14 and / or a server. It can also be a LAN interface, via which communication of the device 15 is possible.

[0064] The method which is carried out by the computer program 18 is explained below with reference to Figure 2.

[0065] In a first step S1, at least a first characteristic parameter of the voltage network 12 is recorded at a first point in time. This can be done using a first of the measuring devices 14 or via the device 15 itself. In any case, it is ensured that the device 15, in particular its processor unit 16, records the first characteristic parameter of the voltage network 12.

[0066] In a second step S2, the first characteristic parameter of the voltage network 12 is recorded at a second time which is different from the first time.

[0067] In principle, the first characteristic parameter of the voltage network 12 can be recorded at multiple points in time. The first characteristic parameter is thus recorded several times consecutively, in particular periodically, so that a time series / sequence of the first characteristic parameter is obtained.

[0068] In addition, it can be provided that at least one second variable, for example a second characteristic parameter of the voltage network 12, is recorded, which is different from the first characteristic parameter of the voltage network 12. The second variable can be recorded independently of the voltage network 12, for example read from a database to which the device 15 has access, in particular by means of the communication interface 22. Data can be obtained from the database that is included in the analysis, for example environmental data such as weather data, time data (working day, weekend, holiday, day and / or night) or usage data of systems in the surrounding area, for example a timetable for trains or usage data of electrical charging systems in the surrounding area.This allows additional information to be obtained that explains any difference between the two values ​​of the first characteristic parameter of the voltage network 12 that were measured at the two points in time.

[0069] The second variable obtained from the database is transmitted to the processor unit 16 of the device 15 so that the processor unit 16 detects the second variable of the voltage network 12.

[0070] However, the second variable can also be a second characteristic parameter of the voltage network 12, which is detected by one of the measuring devices 14 or by the device 15 itself, wherein the second characteristic parameter is transmitted to the processor unit 16 of the device 15 so that the processor unit 16 detects the second characteristic parameter of the voltage network 12. If two characteristic parameters are determined, the two characteristic parameters differ from one another.

[0071] In principle, the first characteristic parameter of the voltage network 12 can be a voltage, a current, a power, a frequency, a distortion, a harmonic, a reactive power and / or an energy value of the voltage network 12, in particular of a phase of the multi-phase voltage network 12. This applies analogously to the second characteristic parameter, provided that this is additionally recorded.

[0072] As already mentioned above, a time series / time sequence of the first characteristic parameter of the voltage network 12 is recorded. This can also apply to the optionally recorded second parameter. For example, the second parameter is always recorded when the first characteristic parameter of the voltage network 12 is recorded, so that the values ​​of the respective parameters are recorded in parallel, in particular simultaneously.

[0073] This means in particular that when the first characteristic parameter of the voltage network 12 is recorded at different times, a corresponding data set is also read out from the database, which represents the second parameter, for example the weather data available at the respective times.

[0074] In a third step S3, at least the first characteristic parameter detected at the first time and the first characteristic parameter detected at the second time are fed into the processor unit 16, which jointly processes the first characteristic parameter detected at the first time and the first characteristic parameter detected at the second time, so that a future state of the voltage network 12 is predicted by the processor unit 16 based on the first characteristic parameter detected at the at least two different times.

[0075] In addition, it can be provided that the second variable, which is recorded at least once and is different from the first characteristic parameter of the voltage network 12, is also processed together in order to predict the future state of the voltage network 12 by the processor unit 16.

[0076] For the forecast, the processor unit 16 has an artificial intelligence 24 which receives at least the first characteristic parameter recorded at at least two different points in time as an input variable and outputs the future state of the voltage network 12 as an output variable.

[0077] The artificial intelligence 24 may comprise at least one artificial neural network, for example a convolutional neural network (CNN) or an artificial recurrent neural network (RNN), such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU).

[0078] Accordingly, the artificial intelligence 24 is capable of predicting a future state of the voltage network 12 based at least on the obtained time series, i.e., the time sequence of the first characteristic parameter. To this end, the artificial intelligence 24 processes the multiple values ​​of the first characteristic parameter, which were recorded at different times, by generating them as a multidimensional vector and processing them accordingly by the artificial intelligence 24.

[0079] The multidimensional vector, for example, has up to eight dimensions, e.g., eight dimensions per phase of the multiphase voltage network 12. The eight dimensions can be represented by voltage, current, power, frequency, distortion, harmonics, reactive power, and energy. However, the dimension of the vector can also be higher, depending on the application.

[0080] The artificial intelligence 24 may have been previously trained using a method in which the artificial intelligence 24 was trained to predict the future state of the voltage network 12 based on values ​​of the first characteristic parameter that were recorded at at least two different points in time. Consequently, the artificial intelligence 24 is a trained artificial intelligence 24.

[0081] The corresponding method for training the artificial intelligence 24 is shown in Figure 3, to which reference is made below.

[0082] In a first training step T1, a training data set is provided for the artificial intelligence 24, which includes at least the first characteristic parameter of the voltage network 12 at a first point in time, the first characteristic parameter of the voltage network 12 at a second point in time, and an actual state of the voltage network 12 at a third point in time. The third point in time is later than the first point in time and the second point in time, so that it is a future point in time to be forecast based on the first point in time and the second point in time.

[0083] Additionally, the training data set may contain a second variable that is different from the first characteristic variable, so that the training data set comprises at least two different variables. The second variable may be a second characteristic variable of the voltage network 12 that is different from the first characteristic variable.

[0084] In particular, the training data set may comprise a time series or time sequence of the first characteristic parameter, so that several values, preferably more than two values, of the first characteristic parameter are present, which have been measured or recorded at different times.

[0085] The optionally provided second variable can also be included in the training data set as a time series or time sequence. The training data set can thus comprise data from at least two different variables, in particular two different characteristic parameters of the voltage network 12, for a specific period of time, as well as information on the state of the voltage network 12 obtained at a time later than the specific period of time.

[0086] However, it is also possible that the training data set includes corresponding information from more than just two different variables, in particular from more than two different characteristic parameters of the voltage network 12, whereby overall more information or data is made available, whereby the training is more extensive and the informative value of the correspondingly trained artificial intelligence 24 is higher.

[0087] In a second training step T2, at least the first characteristic parameter acquired at the first time and the first characteristic parameter acquired at the second time, in particular the time series or time sequence of the first characteristic parameter, are fed into the processor unit 16, which has the artificial intelligence 24 to be trained. The processor unit 16, which has the artificial intelligence 24, jointly processes the first characteristic parameter acquired at the different times, in particular the time series or time sequence, and outputs a predicted future state of the voltage network 12 at the third time, at which the training data set comprises the actual state of the voltage network 12.

[0088] Accordingly, during training, the artificial intelligence 24 learns corresponding relationships between the first characteristic parameter recorded at the two different points in time and their effect(s) on the later state of the voltage network 12, so that the artificial intelligence 24 is trained to predict the future state of the voltage network 12 based on the past or current data.

[0089] In addition, the second variable can optionally be included, whereby correlations between the different variables are detected during training, i.e. correlations between the values ​​of the first characteristic variable, which were recorded at different times, and the second variable, which was recorded once or also at different times.

[0090] In a third training step T3, the predicted future state of the voltage network 12 at the third time point is compared with the actual state of the voltage network 12 at the third time point, the latter being included in the training data set. This comparison determines any deviation between the predicted future state of the voltage network 12 and the actual state of the voltage network 12. Thus, the training determines how accurate the forecast of the artificial intelligence 24 already is, i.e., how well the forecast matches the actual state.

[0091] In a fourth training step T4, the determined deviation between the predicted future state of the voltage network 12 and the actual state of the voltage network 12 is fed back into the artificial intelligence 24 to be trained in order to adjust the weighting factors of the artificial intelligence 24 to be trained if the deviation lies outside a tolerance range. This is indicated in Figure 3 by the corresponding arrow. The tolerance range can be specified and / or set by a user.

[0092] Subsequently, at least training step T3 is repeated, with the deviation determined during the comparison in training step T3 being progressively reduced. After a certain number of repetitions (iterations), the deviation is so small that it lies within the tolerance range, so that the deviation is no longer fed back. At this point, the artificial intelligence 24 has at least reached a (pre-)trained state for the training data set, so that it can be used.

[0093] The artificial intelligence 24 can then be further trained with the same training steps T1-T4 with corresponding iterations, so that the artificial intelligence 24 is trained, for example, on further characteristic parameters of the voltage network 12, in particular different pairings of characteristic parameters of the voltage network 12.

[0094] In particular, the training of the artificial intelligence 24 can also include the input of more than two different variables, in particular characteristic parameters of the voltage network 12, for example, up to eight different variables or more. The corresponding training set used for this purpose therefore has more data that is provided and input.

[0095] Typically, training steps T1 to T4 are repeated for several different actual states of the voltage network 12 and / or several different data sets of the characteristic parameters, in particular for several different characteristic parameters, in order to train the artificial intelligence 24. In the final training step T4, as already described, the weighting factors of the artificial intelligence 24 to be trained are adjusted such that the respective predicted future state of the voltage network 12 always lies within the tolerance range.

[0096] The artificial intelligence 24 used in the method for monitoring the voltage network 12 has been trained according to the aforementioned training method, so that it is a trained artificial intelligence 24 which predicts the future state of the voltage network 12 based on at least two values ​​of the first characteristic parameter which have been recorded at different times.

[0097] The processor unit 16, which comprises the trained artificial intelligence 24, accordingly outputs the predicted future state of the voltage grid 12 in a fourth step S4. This can be a temporal profile of a characteristic parameter of the voltage grid 12, in particular a characteristic parameter that is different from the first characteristic parameter (and the second characteristic parameter). For example, the artificial intelligence 24 is designed, among other things, to predict the future profile of a total harmonic distortion (TH D) of the voltage grid 12, whereby this is done based on the historical data (time series) of the current and / or the voltage. In principle, this makes it possible for the grid operator and / or the electricity customer or the electricity user to predict future disruptions orTo detect weak phases in the voltage network 12 in a timely manner and to initiate appropriate countermeasures in order to ensure stable operation of the voltage network 12.

Claims

Patent claims 1. Method for monitoring a voltage network (12), comprising the steps: Recording at least one first characteristic parameter of the voltage network (12) at a first time point in time, Recording the first characteristic parameter of the voltage network (12) at a second time point in time that differs from the first time point in time, Feeding the first characteristic parameter recorded at the first time and the first characteristic parameter recorded at the second time into a processor unit (16), which processes the first characteristic parameter recorded at the first time and the first characteristic parameter recorded at the second time together in such a way that a future state of the voltage network (12) is predicted based on the first characteristic parameter recorded at the at least two different times, and - Output of the predicted future state of the voltage network (12).

2. Method according to claim 1, characterized in that the first characteristic parameter is a voltage, a current, a power, a frequency, a distortion, a harmonic, a reactive power and / or an energy value, in particular for a phase of a multiphase voltage network (12).

3. Method according to claim 1 or 2, characterized in that the first characteristic parameter is recorded multiple times, such that a time sequence of the first characteristic parameter comprising more than two time points is available, which is processed.

4. Method according to one of the preceding claims, characterized in that the processor unit (16) comprises an artificial intelligence (24) which at least performs the following tasks at the at least two different times The first characteristic parameter is received as an input and the future state of the voltage network (12) is output as an output parameter.

5. Method according to claim 4, characterized in that the artificial intelligence (24) comprises at least one artificial neural network, for example an artificial convolutional neural network or an artificial recurrent neural network, in particular wherein the artificial intelligence comprises a Long Short-Term Memory (LSTM) network or a Gated recurrent unit (GRU).

6. Method according to one of the preceding claims, characterized in that a multidimensional vector is generated which comprises the first characteristic parameter recorded at the at least two different times, wherein the multidimensional vector is processed by the processor unit (16).

7. Method according to one of the preceding claims, characterized in that at least one future temporal progression of a characteristic parameter is predicted as the future state of the voltage network (12).

8. Method for training an artificial intelligence (24) to predict a future state of a power grid (12), comprising the following steps: Providing a training dataset for the artificial intelligence (24) which includes at least a first characteristic parameter of the voltage network (12) at a first time point, the first characteristic parameter of the voltage network (12) at a second time point, and an actual state of the voltage network (12) at a third time point, which is later in time than the first time point and the second time point, Feeding the first characteristic parameter at the first time point and the first characteristic parameter at the second time point into a processor unit (16) that has the artificial intelligence (24) to be trained, wherein the processor unit (16) having the artificial intelligence (24) recorded the parameters at the different time points, first characteristic parameter is processed together and outputs a predicted future state of the voltage network (12) at the third time point, Comparing the predicted future state of the voltage network (12) at the third time point with the actual state of the voltage network (12) at the third time point, which is part of the training data set, in order to determine a deviation between the predicted future state of the voltage network (12) at the third time point and the actual state of the voltage network (12) at the third time point, and Feeding back the deviation between the predicted future state of the voltage network (12) at the third time point and the actual state of the voltage network (12) at the third time point in order to adjust weighting factors of the artificial intelligence (24) to be trained, provided that the deviation is outside a tolerance range.

9. System (10) for monitoring a voltage network (12), comprising at least one processor unit (16) configured to perform a method according to one of the preceding claims.

10. Computer program (18) comprising program code means for performing the steps of a method according to any one of claims 1 to 8 when the computer program (18) is executed on a processor unit (16), in particular the processor unit (16) of the system (10) according to claim 9.

11. Computer-readable data carrier (20) on which the computer program (18) according to claim 10 is stored.

Images