Method and arrangement for estimating a grid state of a power distribution grid

Abstract

A method estimates a grid state of an electrical power distribution grid having a multiplicity of network sections, in which a central computer arrangement is used to receive measured values from measuring devices. A state estimation device is used to make a prediction of a future grid state, wherein a voltage and a phase angle are respectively ascertained for each network section, and in that a naive Bayes method is used for the prediction.

Claims

1. A method for estimating a grid state of an electrical power distribution grid having a plurality of network sections, where a grid state contains a statement about the voltage magnitude or a voltage and the voltage angle or phase angle on each network section, which comprises the steps of: measuring, with measuring devices, a voltage and a phase angle for each of the plurality of network sections of the electrical power distribution grid and generating corresponding measured values; receiving, via a central computer configuration, the measured values from the measuring devices; using a state estimation device to make a prediction of a future grid state based on the measured values, wherein a voltage and a phase angle are respectively ascertained for each of the network sections; and using a machine learning method, being a naive Bayes method, for formulating the prediction; taking into consideration, for the prediction, specific predictions for individual ones of the measuring devices that each have a measurement-location-specific degree of error; and determining a measurement-location-specific weighting factor, using a respective degree of error, such that the specific predictions are weighted all more highly a lower their degree of error; and presenting the predictions to a central controller, and initiating countermeasures to the grid based on the predictions when necessary.

2. The method according to claim 1, wherein a low-voltage grid is used for at least part of the electrical power distribution grid.

3. The method according to claim 1, which further comprising disposing at least some of the measuring devices in substations.

4. The method according to claim 1, which further comprises disposing at least some of the measuring devices at feed points of the network sections.

5. The method according to claim 1, wherein the prediction estimates a grid state at least 5 hours in advance.

6. The method according to claim 1, wherein the prediction has a mean average percentage error of no more than 10%.

7. A central computer configuration for estimating a grid state of an electrical power distribution grid having a plurality of network sections, where a grid state contains a statement about the voltage magnitude or a voltage and the voltage angle or phase angle on each network section, the central computer configuration comprising: measuring devices configured to measure a voltage and a phase angle for each of the plurality of network sections of the electrical power distribution grid and generating corresponding measured values a receiving device for receiving the measured values from said measuring devices disposed in the electrical power distribution grid; and a state estimation device configured to make a prediction of a future grid state based on the measured values, wherein a voltage and a phase angle are respectively ascertained for each of the network sections, said state estimation device is configured to use a machine learning method, being a naive Bayes method, for formulating the prediction, and specific predictions for individual ones of the measuring devices that each have a measurement-location-specific degree of error are taken into consideration for the prediction, and a respective degree of error is used to determine a measurement-location-specific weighting factor such that the specific predictions are weighted all more highly a lower their degree of error; and a central controller configured for presenting the predictions, and initiating countermeasures to the grid based on the predictions when necessary.

8. The central computer configuration according to claim 7, wherein at least part of the electrical power distribution grid has a low-voltage grid.

9. The central computer configuration according to claim 7, wherein said state estimation device is configured to take into consideration for the prediction specific predictions for individual ones of the measuring devices that each have a measurement-location-specific degree of error, wherein a respective degree of error is used to determine a measurement-location-specific weighting factor such that the specific predictions are weighted all more highly a lower their degree of error.

10. The central computer configuration according to claim 7, wherein at least some of the measuring devices are disposed in substations.

11. The central computer configuration according to claim 7, wherein at least some of the measuring devices are disposed at feed points of the network sections.

12. The central computer configuration according to claim 7, wherein said state estimation device is configured to use the prediction to estimate the grid state at least 5 hours in advance.

13. The central computer configuration according to claim 7, wherein said state estimation device is configured with a mean average percentage error of no more than 10%.

14. A non-transitory computer readable medium having computer executable instructions for performing a method for estimating a grid state of an electrical power distribution grid having a plurality of network sections, where a grid state contains a statement about the voltage magnitude or a voltage and the voltage angle or phase angle on each network section, which comprises the steps of: measuring, with measuring devices, a voltage and a phase angle for each of the plurality of network sections of the electrical power distribution grid and generating corresponding measured values; receiving, via a central computer configuration, the measured values from the measuring devices; using a state estimation device to make a prediction of a future grid state based on the measured values, wherein a voltage and a phase angle are respectively ascertained for each of the network sections; and using a machine learning method, being a naive Bayes method, for generating the prediction; and taking into consideration, for the prediction, specific predictions for individual ones of the measuring devices that each have a measurement-location-specific degree of error; and determining a measurement-location-specific weighting factor, using a respective degree of error, such that the specific predictions are weighted all more highly a lower their degree of error; and presenting the predictions via a central controller, and initiating countermeasures to the grid based on the predictions when necessary.

Description

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING

(1) FIG. 1 is an illustration showing an example of a power distribution grid;

(2) FIG. 2 is an illustration showing an approach for a naive Bayes method; and

(3) FIG. 3 is a block diagram showing an example of a method according to the invention.

DETAILED DESCRIPTION OF THE INVENTION

(4) Referring now to the figures of the drawings in detail and first, particularly to FIG. 1 thereof, there is shown a simple example of a power distribution grid 1. The power distribution grid 1 has a medium-voltage level 2 connected via a substation to a measurement location 8 and via a transformer 11 to a low-voltage level 3 of the distribution grid. The distribution grid 3 has two sections 6, 7 or branches. Each branch has multiple generators 12 of electric power, which are each depicted as a circuit. These can be wind turbines or photovoltaic installations, for example. Electrical loads 5 are depicted as arrows, the electrical loads 5 being able to be single-family homes or the like, for example. In section 7, there is provision for a measurement location 9 in the connecting line to the distribution grid 3. In the section 6, there is provision for a measurement location 10 in the connecting line to the distribution grid 3.

(5) The increasing feed from generators of renewable energy at distribution grid level means that it is becoming increasingly important to be able to make reliable predictions of the grid state. Only on the basis of a resilient prediction is it possible for an operator of a control center to detect problems and initiate countermeasures in good time. To date, this has been accomplished for the most part by making predictions for individual pieces of equipment such as e.g. the generators 12 and loads 5 cited at the outset. This information is used to calculate power flows in the power grid. This approach has several disadvantages:

(6) a) the individual predictions for the equipment 5, 12 frequently contain comparatively large errors, because individual pieces of equipment 5, 12 are by their nature subject to greater fluctuations than aggregated considerations—that is to say e.g. the measurement data obtained at the measurement locations 8, 9, 10.
b) Every feed needs to be predicted.
c) All predictions are used regardless of the prediction quality, in order to make a summarized prediction.
d) Weather forecasts are usually available for larger areas and not selectively for regions in which individual installations are installed.

(7) By contrast, the invention proposes making predictions for the measurement locations 8, 9, 10, at which there is already a high level of aggregation of the individual influences of the loads 5 and generators 12. his has the particular advantage that there are comparatively small errors for the prediction. Since not all feeds from the generators 12 into the grid are known, a power flow calculation can no longer be performed with sufficient accuracy. Instead, the invention requires a state estimation to be performed that relates to the whole power distribution grid 1.

(8) The accuracy of the predictions can be increased further if specific predictions for individual measuring devices that each have a measurement-location-specific degree of error are taken into consideration for the prediction, wherein the respective degree of error is used to determine a measurement-location-specific weighting factor such that specific predictions are weighted all the more highly the lower their degree of error. More reliable estimates are therefore weighted more highly for the final prediction.

(9) The invention proposes using what is known as a naive Bayes approach. The input data required for this are for example:

(10) a) historical measurement data pertaining to the real power, which preferably have a comparable time resolution to the prediction to be provided as the end result; e.g. 15 minutes or one hour. In this context, “historical” denotes longer measurement periods for capturing the measurement data provided with timestamps, e.g. over days, weeks or months, an identification of the type of day (weekday, workday, weekend, public holiday, etc.) also being advantageous, in particular. Furthermore, the historical data pertaining to the real power relate to grids without correction measures such as e.g. downward regulation. Should the historical data pertaining to the real power already relate to “corrected” grids, it must be known what has been corrected in order to perform appropriate conversion (backward correction) of the historical data.
b) historical and predicted weather data: ambient temperature, wet-bulb temperature (the lowest temperature that can be reached as a result of direct evaporative cooling), dew point, relative humidity, wind speed, light intensity.
c) where available, historical and predicted market data such as e.g. the price of electricity etc. can also be used.

(11) In the naive Bayes approach, a classification is performed such that the probability of the predicted value (“forecast”) is provided by the measurement data (“predictor”, e.g. ambient temperature, wet-bulb temperature, etc.):
P(Forecasted.sub.F|Predictor)=
(P(Predictor|Forecasted.sub.F)*P(Forecasted.sub.F))/P(Predictor)

(12) The designations here are as follows:

(13) Forecasted.sub.F variable to be predicted: real power at the respective measurement location.

(14) Predictor prediction variable correlated with the value to be predicted, real power. It can be e.g. ambient temperature and/or wet-bulb temperature.

(15) P(Predictor|Forecasted.sub.F) dependent probability of prediction variable assuming the value v if variable to be predicted has the value w.

(16) P(Forecasted.sub.F) probability of the variable to be predicted, real power, assuming the value f.

(17) P(Predictor) probability of the prediction variable assuming the value v.

(18) Furthermore, the naive Bayes classification allows the prediction model to be substantially simplified. In this regard, FIG. 2 shows that the real power on a load 20—on the left in FIG. 2—is calculated in previous approaches from a multiplicity of individually calculated prediction variables for e.g. previous real power on the load, hour of the day, evaporative cooling, dew point, temperature, relative humidity or type of day (weekday, weekend, public holiday, etc.). This is indicated by the arrows pointing from the individual prediction variables to the variable to be predicted, real power 20. This is very computation intensive, however.

(19) By contrast, a reversal of the arrow direction—in the right-hand part of FIG. 2—is obtained, since the prediction variables are trained for the real power in the naive Bayes approach. This “turnaround of the arrows” causes separation of the conditional probability for each prediction variable.

(20) To use the naive Bayes approach, there can be provision for the following steps:

(21) a) iterative calculation of a vector for the envisaged classes,

(22) b) in each class, calculation of:

(23) $\begin{array}{cc}P\left({\mathrm{Class}}_i.Math.\mathrm{Predictors}\right)=& i)\\ \underset{j=1}{\overset{M}{.Math.}}\frac{\left(P\left({\mathrm{Predictor}}_j.Math.{\mathrm{Class}}_i\right)*P\left({\mathrm{Class}}_i\right)\right)}{P\left(\mathrm{Predictor\_j}\right)}& \mathrm{ii})\end{array}$

(24) identification of that class for which the highest value is achieved from the preceding step:
Arg max
k∈{1,k}(P(Class.sub.i|Predictors)

(25) In order now to predict the real power, two steps are performed in succession:

(26) a) training

(27) b) work step.

(28) During the training (A), the following substeps can be performed:

(29) A1) collecting historical data

(30) A2) conditioning the historical data in order e.g. to remove outliers and obvious errors. It is also possible for missing data points to be added by means of estimations in order to obtain a complete time series.

(31) A3) if necessary converting the historical data into discrete values, e.g. by means of an “equal bin approach”.

(32) A4) providing a training data record and an evaluation data record, each having the conditioned historical data.

(33) A5) calculating conditional probability tables for each pair comprising variable to be predicted (e.g. real power at the respective measurement location) and prediction variable, based on the training data record. Preferably, the tables are determined for each type of day (workday, weekend, public holiday, etc.) and in accordance with the time resolution for each time of day, that is to say e.g. for desired prediction of one hour in the future with hourly resolution.

(34) A6) performing naive Bayes with the training data record or the probability tables.

(35) A7) checking the prediction accuracy from A7 using the evaluation data record.

(36) A8) ascertaining the respective prediction error for each prediction at a measurement location. A measurement-location-specific degree of error is calculated as “mean average percentage error (MAPE)”.

(37) In the work step (B), e.g. the following substeps can be performed:

(38) B1) if necessary converting the measurement data from each measurement location into discrete values, e.g. by means of an “equal bin approach”. Preferably, measurement data from the same point in time are used, the available time resolution and the desired prediction accuracy being able to be used to stipulate the size of the time window of measurement data that is supposed to be used.

(39) B2) providing a measurement data record containing the measurement data from step (B1).

(40) B3) performing naive Bayes with the measurement data record and with the probability tables.

(41) B4) ascertaining and storing the most probable grid state from step (B3) for the respective point in time.

(42) When used e.g. in a state estimation device such as for example the DSSE described at the outset, the following steps can be added:

(43) B5) stipulating a point in time in the future at which the prediction is supposed to indicate the grid state.

(44) B6) consulting prediction data and optionally forecast timetables for generators and/or loads from a prediction database that each relate to the measurement locations. Additionally, prediction data pertaining to individual feed points into the grid can also be used.

(45) B7) stipulating a prediction error for the prediction data from step (B6), a percentage of between 0% and 100% being stipulated. The following holds for the respective prediction error w:
W=1/prediction error with w=100 for prediction error=0.

(46) B8) performing the state estimation for the point in time in the future.

(47) B9) evaluating the prediction for the grid state at the future point in time such that it is stipulated whether permissible threshold values for electrical quantities are infringed. These can involve a prescribed voltage band and/or a maximum current, for example. They can also involve a phase angle. This check is performed for all network sections, lines and transformers.

(48) FIG. 3 shows an example of a method according to the invention. Predictions 34, which, in particular, have weighting factors corresponding to the prediction errors from step B(7), are taken from a prediction database 31 for the measurement locations. Furthermore, network data 32 pertaining to the topology and electrical properties of the grid and of the equipment are retrieved. Scheduled switching actions 33 are also included too.

(49) These input data are used in a state estimation device 30 to perform the estimation and provide a prediction of a future grid state 38.

(50) Optionally, predictions for timetables 36 and load predictions 37 can also be taken into consideration.