Patent Application
20190094286
This application claims priority to European application No. 17193136.3 having a filing date of Sep. 26, 2017 the entire contents both of which are hereby incorporated by reference.
The following relates to a method and an in-field apparatus for automatic localization of a fault having occurred at power transmission lines of a power supply system.
A power supply system of a power supply grid can comprise several power transmission lines to transport electric currents between nodes of a power supply system. Most power supply systems comprise 3-phased transmission lines L1, L2, L3. During the distribution of electrical power via the 3-phased power transmission lines of the power supply grid unexpected failures can occur due to various causes. For example, one or more electrical loops formed by the power transmission lines can comprise a short circuit caused by a falling tree or branch during a storm. This event might cause a short circuit between at least two power transmission lines or between two power transmission lines and earth. Further, a power transmission line might be interrupted by a falling tree or other events. Such failures impact a reliable and safe operation of the power supply system. Accordingly, it is necessary to detect these kind of failures or faults by protective devices within the power supply system. Further, measures have to be taken in the power supply system to prevent a propagation of faults within the remaining power supply grid.
In a conventional power supply system protective devices are normally installed at both ends of power transmission lines connecting two nodes of the power supply system. This protective device can constantly monitor an electric current voltage of every loop formed by the power transmission lines. In case a fault does occur, the normal power supply signal forming a phased sine wave is disturbed by the anomaly or fault. In this case the protective device does trigger the generation of a fault record and can shut down the affected power transmission lines. Given the fault record that does contain the raw signals i.e. the current voltage signals of every power transmission line within a predetermined time interval of e.g. 200 ms as the error occurs the protective device can execute a series of steps to determine the fault location for instance by performing impedance calculations based on the parameters optimal measurement point and optimal window size. The accuracy of the calculated fault location does strongly depend on the choice of these two parameters that are in turn determined based on the raw signals.
An aspect relates to a method and an apparatus for automatic localization of a fault having occurred at power transmission lines of a power supply system which is robust and provides an accurate localization of the fault.
The following provides according to a first aspect an in-field apparatus for automatic localization of a fault having occurred at power transmission lines of a power supply system, said in-field apparatus comprising:
a preprocessing unit adapted to process measured voltage and/or current raw time series data of the power transmission lines to provide a normalized raw data and/or feature representation of the measured raw time series data; and
an artificial intelligence module configured to predict an optimal evaluation time used for evaluation of the measured voltage and/or current raw time series data to localize the fault based on the normalized raw data and/or feature representation.
In a possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the preprocessing unit is adapted to perform a z-score normalization of measured voltage and/or current raw data of the power transmission lines.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module comprises a window detection unit adapted to determine an optimal window size of a window used for feature extraction from the measured voltage and/or current raw time series data.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module further comprises a time point detection unit adapted to determine an optimal evaluation time point for the evaluation of features extracted from the measured voltage and/or current raw time series data.
In a further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the window detection unit is adapted to generate learned regression curves for different predefined window sizes.
In a further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module comprises a false positive detection unit adapted to determine a probability whether the predicted optimal evaluation time does lead to a valid fault localization.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module is adapted to perform a feature-based fault classification based on the feature representation of the preprocessed measured raw time series data provided by the preprocessing unit.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module is configured to perform the feature-based fault classification using a random forest, a decision tree, a support vector machine, and/or a fully connected neural network.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module is adapted to perform a not feature-based direct fault classification based on the normalized measured raw time series data provided by the preprocessing unit.
In a further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the artificial intelligence module is configured to perform the not feature-based direct fault classification using a feed-forward neural network, in particular a convolutional neural network or a recurrent neural network.
In a further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the preprocessing unit is adapted to normalize raw data of a measured fault record.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the measured fault record applied to the preprocessing unit comprises voltage raw data of different voltage phases transported through corresponding power transmission lines, current raw data of associated electrical currents transported through the respective power transmission lines, and current raw data of an electrical current flowing through earth.
In a further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the voltage and/or current raw data of the measured fault record processed by said preprocessing unit comprises data samples measured by corresponding sensors within a time window.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention a database or an online stream of fault records with known fault types is provided and used to train the artificial intelligence module of the in-field detection apparatus.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention the database of fault records comprises simulated fault records and/or fault records of historical faults having occurred at power transmission lines of the power supply system.
In a still further possible embodiment of the in-field apparatus according to the first aspect of embodiments of the present invention countermeasures to remove a fault are initiated automatically by a controller depending on a detected fault type of the localized fault.
The following further provides according to a second aspect a method for automatic localization of a fault comprising the features of claim 15.
The following provides according to the second aspect a method for automatic localization of a fault having occurred at power transmission lines of a power supply system, the method comprising the steps of:
preprocessing measured voltage and/or current raw time series data of the power transmission lines to provide a normalized raw data and/or feature representation of the raw time series data, calculating an optimal evaluation time point and/or evaluation time window based on the normalized raw data and/or feature representation using a trained artificial intelligence module and evaluating by the trained artificial intelligence module the measured voltage and/or current raw time series data using the calculated optimal evaluation time point and/or evaluation time window to localize the fault.
Some of the embodiments will be described in detail, with references to the following Figures, wherein like designations denote like members, wherein:
As can be seen in the block diagram of
The in-field apparatus 1 comprises an artificial intelligence module 3. This artificial intelligence module 3 is configured to predict an optimal evaluation time used for evaluation of the measured voltage and/or current raw time series data to localize the fault based on the normalized raw data and/or features representation provided by the preprocessing unit 2. As can be seen in the embodiment illustrated in
The artificial intelligence module 3 of the in-field apparatus 1 can be adapted to perform a feature-based fault classification based on the feature representation of the preprocessed measured raw time series data provided by the preprocessing unit 2. In a possible embodiment the artificial intelligence module 3 can be configured to perform the feature-based fault classification using a random forest, a decision tree, a support vector machine and/or a fully connected neural network.
In an alternative embodiment the in-field apparatus 1 comprises an artificial intelligence module 3 which is adapted to perform a not feature-based direct fault classification based on the normalized measured raw time series data provided by the preprocessing unit 2 of the in-field apparatus 1. The artificial intelligence module 3 can be configured to perform the not feature-based direct fault classification using in a possible embodiment a feed-forward neural network, in particular a convolutional neural network CNN or a recurrent neural network RNN.
In the embodiment illustrated in
A database or online stream of fault records with known fault types can be used to train the artificial intelligence module 3 of the in-field detection apparatus 1 as illustrated in the embodiments of
The in-field apparatus 1 according to embodiments of the present invention comprises the artificial intelligence module 3 which allows to automatically determine specific parameters, in particular an optimal evaluation time which can be used for evaluation of the measured voltage and/or current raw time series data. The parameters can comprise an optimal window size of a window used for feature extraction from the measured voltage and/or current raw time series data. Further, the parameters can comprise an optimal evaluation time point for the evaluation of features extracted from the measured voltage and/or current raw time series data. The artificial intelligence module 3 can execute an artificial neural network ANN stored in a local memory of the in-field apparatus 1. The memory footprint of the artificial neural network stored in the local memory is comparatively low (i.e. the memory space required by the artificial intelligence module 3 is low). In a possible embodiment the in-field apparatus 1 according to the first aspect of embodiments of the present invention can be implemented on a SIPROTEC device of a power supply system.
In the illustrated embodiment the method comprises three main steps S1, S2, S3.
In a first step S1 measures voltage and/or current raw time series data of the power transmission lines are preprocessed to provide a normalized raw data and/or feature representation of the raw time series data. This preprocessing can comprise a Z-score normalization of the measured voltage and/or current raw data of the power transmission lines.
In a further step S2 an optimal evaluation time point and/or evaluation time window is calculated based on the normalized raw data and/or feature representation using a trained artificial intelligence module.
In a third step S3 the trained artificial intelligence module is evaluating the measured voltage and/or current raw time series data using the calculated optimal evaluation time point and/or evaluation time window to localize the fault.
In a preferred embodiment the steps S1, S2, S3 of the method as illustrated in
The method and apparatus according to embodiments of the present invention requires very limited hardware resources, in particular storage resources and/or calculation resources. The method according to embodiments of the present invention leads to an improved prediction of fault localization estimations compared to a conventional expert-driven rule-based prediction system. In a preferred embodiment the in-field apparatus 1 even provides information whether a fault record is valid or invalid. It can automatically recognize whether it is able to make a reliable prediction of a fault localization given a specific error report. The artificial intelligence module 3 can be trained on a combination of synthetic fault records which can be generated efficiently by existing simulation systems. Further, the artificial intelligence module 3 can also be trained with historical fault records of faults having occurred in the past. The artificial intelligence module 3 is trained and can learn to interpret data patterns in data which can in turn be used to process new previously unseen data samples. The in-field apparatus 1 according to embodiments of the present invention does not use a rule-based module that requires an extensive manual effort to be extended. The in-field apparatus 1 comprises an artificial intelligence module 3 which can be automatically extending its knowledge about faults by training on a growing set of real or simulated fault records. In a preferred embodiment the artificial intelligence module 3 executes an artificial neural network ANN stored in a local memory of the in-field apparatus 1. This artificial neural network ANN can comprise an artificial deep neural network. The artificial deep neural network can take as an input numerical values, e.g. raw or preprocessed electrical currents and/or electrical voltages of the different power line loops of the power supply system. This data can be passed through a network of weighted non-linear functions and results in several output values. In order to obtain an optimal measurement point, first a continuous output is learnt that can correspond to the deviation between the impedance value that would be computed given a certain measurement point and the true value. This normalized impedance value can be generated for various window sizes. In a possible embodiment the artificial intelligence module 3 can generate a binary output or flag indicating whether a fault record is invalid or valid. To obtain a system that can reliably perform automatic calculations of fault locations for unseen data, the artificial neural network ANN can be trained to detect patterns by sequentially passing data samples with a known output value (e.g. a function output for regression, class label for classification) through the artificial neural ANN network and by comparing the generated output value with the expected value.
The preprocessing unit 2 of the in-field apparatus 1 can take as input in a possible embodiment seven raw input values from a fault record and applies a series of preprocessing operations. In general, these preprocessing operations can include a mandatory normalization step (z-score normalization) and an optional step to compute a set of features on the received raw data. The resulting values can then be fed into the artificial intelligence module 3.
The artificial intelligence module 3 comprises in a possible embodiment as illustrated in
The time point detection unit 3B takes as an input the predicted normalized impedance values for all measurements points for a single fault record. In a possible embodiment for every window size a moving mean value is applied to smooth the resulting regression curve. Afterwards the optimal window size and measurement point can be identified by minimizing the smoothed normalized impedance values across the window sizes and measurement points.
In a possible embodiment the artificial intelligence module 3 can further comprise a false positive detection module 3C. This false positive detection module 3C can take as input all predicted and smoothed impedance values. The false positive detection module can output a score in the range of [0,1], i.e. a class probability whether the predicted optimal measurement window will lead to a valid fault location computation or not.
The in-field apparatus 1 according to embodiments of the present invention can be embedded in an in-field protection device such as SIPROTEC having limited computational and storage resources.
The artificial intelligence module 3 of the in-field apparatus 1 comprises an artificial neural network ANN which can be trained in a training procedure. In a possible embodiment first a database (or online stream) of fault records with known fault locations can be generated by a simulation algorithm. The capabilities of the model depend on the number and variation of fault records used for training the artificial neural network of the artificial intelligence module 3. Further, the performance of the artificial intelligence module 3 can be increased by adding as many real fault records to the dataset as possible. Second, impedances can be computed for a number of defined window sizes and can be normalized by known fault location. The normalized impedance can serve as an output. Lastly, the dataset can be split into a training and in a validation dataset. The artificial intelligence module 3 can then be optimized on the training dataset and then tested on the validation dataset which is not used during training. Once the module training of the artificial intelligence module 3 has been completed (i.e. error is minimized) the trained artificial neural network ANN can be saved with a predefined architecture with learned weights.
During inference a new data sample (e.g. a fault record) can be passed through the artificial intelligence module 3, yet without knowing the intended output value. Thus, in a possible embodiment only a forward pass is needed without back-propagating any error.
The system can be parameterized by the window sizes provided by the method and apparatus according to embodiments of the present invention. Normalized impedances can be computed for every window size and can serve as samples for a target function. For every defined window size a single neural network can be trained until a mean logarithmic error of the predicted function samples and the real ones is minimized. Similarly, an artificial neural network can be defined to identify invalid fault records. A goal in this case can be that the resulting fault location of the post-processing module deviates e.g. not more than 5 km from the true value. Optionally the neural networks can be coupled by sharing the weights of the first layers.
In a possible embodiment the method and apparatus according to embodiments of the present invention do not require any feature computation on the raw input signals as the artificial neural network ANN is capable of learning the important or relevant characteristics directly in an end-to-end fashion. A feature-based machine learning approach does require more domain understanding and manual efforts to design good features. This also bears the risk of missing important patterns that cannot be recognized by the predefined feature set. The performance relies on the raw input data with a mild normalization step and thus can easily be extended or improved by feeding newly simulated or real fault records.
In a possible embodiment the size of the artificial intelligence module 3 can be significantly reduced by using domain knowledge to process the raw input data, thereby generating features that are known to be beneficial for the task at hand. In this embodiment a moving window (e.g. having a window size of 32) and e.g. seven input time series data (i.e. electrical current and electrical voltage of every transmission loop) can then be used to compute a set of e.g. seven statistical features such as sum, mean value, median value, min value, max value, variance or standard deviation. The statistical features can be used to describe the properties over time. In a possible embodiment it is possible to compute a moving Fourier transform (window size=32) to expose the range of frequencies in the signals. In a possible embodiment instead of using a convolutional neural network CNN, a fully connected neural network can be implemented with a single hidden layer and a relatively small number of hidden nodes (e.g. number of hidden nodes n=50).
In a possible embodiment the method and apparatus according to embodiments of the present invention can also be implemented in a standard computer or PC as part of an analytics framework such as the SIGRA system that allows an offline analysis of fault records and a computation of fault locations.
The method and apparatus according to embodiments of the present invention is datadriven. The method and apparatus have a performance which does improve with more data or received data records. The method and apparatus according to embodiments of the present invention is not rule-based and consequently does require less expert domain knowledge. Improvements of the artificial intelligence module 3 can be provided without adding new rules and thresholds. The artificial intelligence module 3 of the in-field apparatus 1 can be trained both on real and/or simulated fault data. It is not necessary to compute domain-specific features as the features are implicitly learned by the deep neural network implemented in the artificial intelligence module 3. Further, new and improved versions of the artificial intelligence module 3 can be generated autonomously i.e. the neural network of the artificial intelligence module 3 can be retrained on new or additional data without any human interaction.
In a possible embodiment the in-field apparatus 1 according to embodiments of the present invention can be implemented on a SIPROTEC protection device of a power supply system. In a further embodiment the artificial neural network 3 can provide a deep learning neural network which requires a computation time of less than 200 ms. The number of layers and the artificial neural network can vary depending on the use case. The artificial intelligence module 3 allows for more accurate results with a small memory footprint and requires few computational resources. The in-field apparatus 1 according to embodiments of the present invention can be integrated in a protection device such as a digital protection relay of a power supply system. The in-field apparatus 1 can comprise in a possible embodiment a preprocessing unit 2, an artificial intelligence module 3 and a controller. The controller can perform automatically countermeasures depending on a detected fault type of the localized fault. These countermeasures can for instance comprise switching operations performed by switches connected to the controller of the in-field apparatus 1. The calculated fault location can be output to a user via user interface. Further, the calculated fault location can be used for further processing by other entities of the system.
Although the invention has been illustrated and described in greater detail with reference to the preferred exemplary embodiment, the invention is not limited to the examples disclosed, and further variations can be inferred by a person skilled in the art, without departing from the scope of protection of the invention.
For the sake of clarity, it is to be understood that the use of “a” or “an” throughout this application does not exclude a plurality, and “comprising” does not exclude other steps or elements.
| Number | Date | Country | Kind |
|---|---|---|---|
| 17193136.3 | Sep 2017 | EP | regional |
