DE102016206243A1 - Method and forecasting system for the computer-aided prognosis of measured values - Google Patents

Abstract

For the computer-aided prognosis of measured values of a plurality of measuring devices (S1,..., SN), a sequence (TTS1,..., TTSN) of measured values is read in for a respective measuring device of a training set of measuring devices. The read-in sequences (TTS1,..., TTSN) are subjected to cluster analysis, whereby groups (C1,..., CM) are determined by sequences which are similar to one another with regard to a similarity measure (PX), and where for a respective group (C1, ..., CM) a group-typical sequence (P1, ..., PM) is determined. The determined group-typical sequences (P1,..., PM) are each extrapolated, resulting extrapolation values (EP1,..., EPM) being stored group-specifically. Furthermore, the following method steps are carried out for a respective measuring device (S1, ..., SN) of the plurality of measuring devices: A sequence (TS1,..., TSN) of measured values to be extrapolated is read in, for the sequence to be extrapolated a combination of the determined group-typical sequences (P1, ..., PM), which approximates the sequence to be extrapolated, the stored extrapolation values (EP1, ..., EPM) of the group-typical sequences (P1, ..., PM) are linked in accordance with the determined combination and a result (ETS1, ..., ETSN) of this link is output as a forecast.

Description

In the control of complex technical systems, e.g. Power grids, manufacturing plants, gas turbines, wind turbines, robots or vehicles, it is often desirable to predict system behavior, at least in the short term, e.g. to match resource requirements to a resource supply, to recognize imminent resource bottlenecks in time and / or to take advantageous planning measures.

The system behavior of sufficiently complex technical systems is usually determined by means of a large number of measuring devices. Thus, e.g. Consumer loads in a power grid usually measured by a large variety of electricity meters. So-called smart meters, often referred to as smart meters, can automatically transmit their readings to utility companies to provide a foundation for predictive grid and resource control.

A prognosis of measured values of a large number of measuring devices of a technical system generally allows a good prognosis of the system behavior.

It is known to measure smart meter readings by machine learning techniques, e.g. using regression models and / or artificial neural networks. A separate prediction of the measurements of e.g. However, all smart meters in a power supply network requires a hardly calculable computational effort due to the often very large number of smart meters. On the other hand, if aggregated total consumption is taken into account, information about individual consumption patterns is lost, which often results in a deterioration in forecasting accuracy.

It is an object of the present invention to provide a method and a prognosis system for the computer-aided prognosis of measured values of a multiplicity of measuring devices, which permit more accurate prognoses and / or require less computational effort.

This object is achieved by a method having the features of patent claim 1, by a prognosis system having the features of patent claim 11, by a computer program product having the features of patent claim 12 and by a computer-readable storage medium having the features of patent claim 13.

For the computer-aided prognosis of measured values of a plurality of measuring devices, a sequence of measured values is read in for a respective measuring device of a training set of measuring devices. The measuring devices may be, for example, measuring devices for measuring mechanical, optical, acoustic, magnetic, thermoelectric, capacitive, inductive, chemical, physical and / or other measured variables; e.g. Sensors, soft sensors or electricity meters, especially smart meters. The read-in sequences are subjected to cluster analysis, whereby groups of sequences which are similar to one another with respect to a similarity measure are determined, and a group-typical sequence is determined for a respective group. Such groups are often referred to as clusters. The determined group-typical sequences are each extrapolated, with extrapolation values obtained being stored group-specifically. Furthermore, the following method steps are carried out for a respective measuring device of the plurality of measuring devices: A sequence of measured values to be extrapolated is read in, for the sequence to be extrapolated a combination of the group-typical sequences is determined, which approximates the sequence to be extrapolated, the stored extrapolation values of the group-typical sequences become linked according to the determined combination and a result of this link is output as a forecast.

To carry out the method according to the invention, a prognosis system, a computer program product and a computer-readable storage medium are provided.

The inventive method or the prediction system according to the invention can be implemented or implemented, for example, by means of one or more processors, application-specific integrated circuits (ASIC), digital signal processors (DSP) and / or so-called field programmable gate arrays (FPGA).

For cluster analysis, it is possible to use standard numerical methods, which are generally well scalable and can also be carried out efficiently for very large numbers of measured value sequences. In many cases, relatively few group-typical sequences can already be sufficient to approximate a multiplicity of sequences to be extrapolated. Thus, often only relatively few explicit extrapolations and thus a relatively small amount of computation are required to determine a plurality of relatively accurate meter individual forecasts.

Advantageous embodiments and further developments of the invention are specified in the dependent claims.

According to an advantageous embodiment of the invention, the extrapolation of the group-typical sequences can be carried out by means of a data-driven trainable regressor, an artificial neural network, a support vector machine, a decision tree and / or a deep learning method. The above learning-based systems and methods generally permit good recognition of patterns in the measurement sequences and good extrapolation of these patterns.

Preferably, the training amount of measuring devices may include the plurality of measuring devices, a plurality of measuring devices similar thereto, and / or a selection of the plurality of measuring devices.

According to an advantageous embodiment of the invention, a respective sequence of measured values can be coded as a vector in a high-dimensional vector space. In this case, as the similarity measure, a distance measure of the vector space or a correlation of the sequences may be used. Such a measure of similarity is often referred to as the measure of proximity. A smaller distance or a larger correlation can represent a greater similarity here.

Furthermore, as a group-typical sequence of a group, a sequence located in this group can be selected. The selected sequence then forms a representative of this group.

Alternatively or additionally, the group-typical sequence of a respective group can be determined on the basis of a distribution characteristic of the sequences in this group. In this way, e.g. an average, median, maximum or minimum of sequences in the group can be determined as a group-typical sequence. The mean value, the median, the maximum or the minimum can be determined here in particular by specific weighting of the group members.

Furthermore, a linear combination and / or a non-linear combination of group-typical sequences can be determined in each case as a combination. While a linear combination can usually be determined with less computational effort, a nonlinear combination often allows a more accurate approximation.

Furthermore, the cluster analysis may include a method for hierarchical, partitioning and / or density-based cluster analysis. As a partitioning cluster analysis, a so-called "fuzzy C-means" method can be used which allows a very efficient determination of the groups of similar sequences.

According to an advantageous development of the invention, the determined group-typical sequences can be correlated with each other. Preferably, a group-typical sequence whose correlation with another group-typical sequence fulfills a predetermined criterion can not be taken into account in the combination. In this way, group-typical sequences that have high similarity or proximity to one or more other group-typical sequences can be excluded from the combination. Insofar as group-typical sequences which are highly similar to each other lead to unstable forecasts, a prognosis quality can be significantly improved in this way.

According to an advantageous embodiment of the invention, preferably current external influence data can be detected. This may be, for example, weather data such as temperature, cloudiness and / or wind speed and / or calendar data such as date, time and / or an indication of whether a working day, holiday or holiday is present. The extrapolation can then be done depending on the external influence data. In this way, external influencing factors can also be taken into account when forecasting the measured values, as a result of which the quality of prognosis can often be considerably improved.

An embodiment of the invention will be explained in more detail with reference to the drawing. In each case show in a schematic representation:

1 a prognosis system according to the invention in a training mode and

2 the forecasting system according to the invention in a forecasting mode.

Unless otherwise indicated, like reference characters designate like objects throughout the figures.

1 shows a schematic representation of an inventive prognosis system PS when operating in a training mode. The prediction system PS has one or more processors PROC for executing all method steps of the prognosis system PS as well as a memory MEM.

The forecasting system PS is coupled to a multiplicity of measuring devices S ₁ ,..., S _N. The measuring devices S ₁ , ..., S _N are used in training mode for training the forecasting system PS and thus form a training amount of measuring devices. A respective measuring device S ₁ ,... Or S _N can be a measuring device for measuring mechanical, optical, acoustic, magnetic, thermoelectric, capacitive, inductive, chemical and / or physical quantities or combinations thereof, for example, a sensor, a counter or an electricity meter.

In the present exemplary embodiment, a prognosis of individual consumer loads in a power grid is considered. Such consumer loads are often measured by smart meters, so-called smart meters. Accordingly, the measuring devices S ₁ , ..., S _N are implemented in the present embodiment as a smart meter. In a power grid, a number N of smart meters can be very large and quite several millions. Nevertheless, by means of the invention, a prognosis for the individual smart meters S ₁ ,..., S _{N can} also be determined in an efficient manner for large power grids.

By the prediction system PS, respectively, a sequence of measured values TTS _1, ... TTS and _N is read in the form of a time series for training the prediction system PS of a respective measuring device S _1, ..., and S _N. A time series here is a chronological sequence of measured values, whereby the time resolution and time horizon for an application of the invention can be almost arbitrary. The training time series TTS ₁ ,..., TTS _N are preferably each represented as a vector in a high-dimensional vector space. As training time series TTS ₁ , ..., TTS _N historical time series of the measuring devices S ₁ , ..., S _N or similar measuring devices can be used.

The training time series TTS ₁ ,..., TTS _N are transmitted to a cluster analyzer CA of the prediction system PS, which serves for the numerical cluster analysis of the training time series TTS ₁ ,..., TTS _N.

A cluster analysis is a numerical analysis method for the detection of similarity structures in preferably large amounts of data, here the training time series TTS ₁ , ..., TTS _N. Here, groups of similar objects are to be recognized, identified and / or differentiated from other similarity groups. Such groups of similar objects are often referred to as clusters, grouping as clustering. The cluster analysis can be performed hierarchically, partitioning, graph theoretic, density based and / or optimizing. Advantageously, a so-called fuzzy C-methods method can be used for partitioning cluster analysis.

The cluster analyzer CA uses a proximity measure PX as a predetermined measure of similarity for the training time series TTS ₁ ,..., TTS _N to quantify similarities. The proximity measure PX can be, for example, a distance in the high-dimensional vector space of the time series TTS ₁ ,..., TTS _N.

In the present embodiment, _M C are the Clusteranalysator CA from the training time series TTS _1, ..., _N TTS cluster C _1, ..., determined. A respective cluster C ₁ ,... Or C _M is a group of time series TTS ₁ ,..., TTS _N similar to the proximity measure PX. The clusters C ₁ ,..., C _M are transmitted by the cluster analyzer CA to a selection module SEL. By means of the selection module SEL, a group-typical time series P ₁ ,... Or P _M for each of the clusters C ₁ ,..., C _{M is} determined as a prototype for the relevant cluster. Such a prototype is often referred to as a cluster center or representative of the cluster. Preferably, as a prototype P ₁ ,... Or P _M, a time series located in the relevant cluster C ₁ ,... Or C _M can be selected, for example that time series which is closest to an average of the cluster. Alternatively or additionally, the prototype P ₁ ,... Or P _{M of} a respective cluster C ₁ ,... Or C _M can be determined on the basis of a distribution characteristic of the time series located in this cluster.

Preferably, the prototypes P ₁ ,..., P _{M determined in this way} can be correlated with one another, ie a respective correlation function can be calculated in order to determine linear dependencies between the prototypes P ₁ ,..., P _M. Subsequently, those of the prototypes P ₁ ,..., P _{M can be} selected which have only a relatively low correlation or linear dependence to one another. In the following, it is assumed that out of the prototypes P ₁ ,..., P _M , the non-selected prototypes have already been rejected by the selection module SEL.

Studies have shown that with a number of 1000 smart meters, i. N = 1000, already three prototypes, i. M = 3 can suffice to map a large range of time series patterns with sufficient accuracy. In this way, the computational resources required for forecasting can be significantly reduced.

The selected prototypes P ₁ ,..., P _M are stored in the memory MEM in association with the respective cluster C ₁ ,... Or C _M. In addition, the prototypes P ₁ , ..., P _{M are} transmitted to a data-driven trainable regressor NN.

The regressor NN serves for the numerical extrapolation, ie here for the prognosis of a further temporal course of the prototypes P ₁ ,..., P _M. The regressor NN is implemented in the present embodiment as a neural network NN. Alternatively or additionally, a support vector machine, a decision tree and / or a deep learning method can be used for extrapolation.

In addition, external influence data EXT are detected by the prognosis system PS and transmitted to the neural network NN. The external influencing data EXT can be, for example, weather data such as temperature, cloudiness and / or wind speed and / or calendar data such as date, time and / or an indication of whether a working day, holiday or holiday is present. The extrapolation of the prototypes P ₁ ,..., P _M then takes place as a function of the external influencing data EXT.

As a result of the extrapolation, the neural network NN supplies an extrapolated time series EP ₁ ,... Or EP _M for a respective prototype P ₁ ,... Or P _M. The extrapolated values of the time series EP ₁ ,..., EP _M are assigned to the respective cluster C ₁ ,... Or C _M and / or assigned to the respective prototype P ₁ ,... Or P _M in the memory MEM saved.

After training the forecasting system PS by cluster analysis, selection and extrapolation, the forecasting system PS can be used for the meter-specific forecast for all smart meters S ₁ ,..., S _N.

2 shows a schematic representation of the prognosis system PS according to the invention in a forecasting mode. The forecasting system PS is coupled to a technical system T, eg a power grid. The technical system T has a control device CTL in order to control the technical system T on the basis of the measuring device-specific forecasts.

In addition, a large number of measuring devices is coupled to the forecasting system PS whose measured value time series are to be extrapolated. In the present embodiment are coupled as measuring devices, the smart meter S ₁ , ..., S _N , for the forecasting system PS - as related to 1 described - was trained. Alternatively or additionally, measuring devices of the same type or similar to the smart meters S ₁ ,..., S _N can also be coupled to the prognosis system PS for prognosis.

In addition to the related 1 The prediction system PS has an approximation module APX, a multiplier module MUL and an output interface OUT.

The smart meters S ₁ , ..., S _N transmit current time series TS ₁ ,..., TS _N , ie sequences of current measured values whose further course is to be predicted, to the forecasting module PS. The time series to be extrapolated TS ₁ , ..., TS _N are preferably each represented as a vector in a high-dimensional vector space.

The time series TS ₁ ,..., TS _N are read in by the approximation module APX and approximated by a combination of the prototypes P ₁ ,..., P _M. The prototypes P ₁ ,..., P _M are read out of the memory MEM by the approximation module APX.

In the present exemplary embodiment, the time series TS ₁ ,..., TS _{N are} respectively approximated by a linear combination of the prototypes P ₁ ,..., P _M. In this case, coefficients K _{ij of} the linear combinations with i = 1,..., N and j = 1,..., M are numerically determined so that a respective linear combination approximates a respective time series TS _i with the lowest possible numerical error. This can be done, for example, by means of a standard minimization method. For a respective time series, TS _i = Σ _j K _ij * P _j , where i = 1,..., N and the sum runs over j = 1,..., M. The time series TS ₁ , ..., TS _N are thus developed numerically according to prototypes P ₁ , ..., P _M with K _ij as development _coefficients . Thus, the time series TS ₁ ,..., TS _N can be reconstructed, as it were, with the prototypes P ₁ ,..., P _M , with K _ij as the reconstruction matrix. The time series TS ₁ ,..., TS _N are encoded to a certain extent by the coefficients K _ij .

In an at least approximate representation of the time series TS ₁ ,..., TS _N by a non-linear combination of the prototypes P ₁ ,..., P _M , coefficients or parameters of this nonlinear combination can be regarded as a non-linear reconstruction rule, on the basis of which the time series TS ₁ , ..., TS _{N to} some extent reconstructed.

The coefficients K _ij thus determined are transmitted from the approximation module APX to the multiplier module MUL. Furthermore, the extrapolated time series EP ₁ ,..., EP _{M are} read out of the memory MEM by the multiplier module MUL. The multiplier module MUL now links the extrapolated time series EP ₁ ,..., EP _M with the coefficients K _ij in a linear combination corresponding to the above linear combination in order to produce as the result of the combination an extrapolation ETS _{i of} the time series TS _i , i = 1. N, according to ETS _i = Σ _j K _ij * EP _j , where the sum in each case runs over j = 1,... N. Since each extrapolated time series EP ₁ ,..., EP _{M is represented} by a vector, this combination corresponds to a matrix multiplication.

In the case of a representation of the time series TS ₁ ,..., TS _N by a non-linear combination of the prototypes P ₁ ,..., P _M , a recombination module for non-linear recombination or reconstruction of the extrapolations ETS _i can be used instead of the multiplier module MUL.

The invention requires no complex regressorgestützte extrapolation of each of the N time series TS ₁ , ..., TS _N. Instead, a regressor-supported extrapolation of the generally small number of prototypes P ₁ ,..., P _M together with a simple linear combination is sufficient.

The method steps described above are preferably carried out in parallel by the approximation module APX and the multiplier module MUL for a plurality of the time series TS ₁ ,..., TS _N. This natural parallelizability of the process also contributes to high efficiency.

The extrapolations ETS _{i of} the time series TS _i are respectively output by the multiplier module MUL as a device-specific prognosis via the output interface OUT.

In the present exemplary embodiment, the extrapolations ETS ₁ ,..., ETS _N are transmitted to the control device CTL of the technical system T in order to control the latter on the basis of the forecasts. In this way, in the technical system T, for example, a predicted resource requirement can be matched to a predicted supply of resources, imminent resource bottlenecks can be detected in good time, and / or advantageous planning measures can be taken.

Claims (13)

Method for the computer-aided prognosis of measured values of a plurality of measuring devices (S ₁ ,..., S _N ) wherein a) for a respective measuring device (S ₁ , ..., S _N ) of a training set of measuring devices a sequence (TTS ₁ ,. .., TTS _N ) is read in from measured values, b) the read-in sequences (TTS ₁ ,..., TTS _N ) are subjected to a cluster analysis, groups (C ₁ ,..., C _M ) of having a similarity measure ( PX) similar sequences are determined, and wherein for a respective group (C ₁ , ..., C _M ) a group-typical sequence (P ₁ , ..., P _M ) is determined, c) the determined group-typical sequences (P ₁ , ..., P _M ) are extrapolated in each case, resulting extrapolation values (EP ₁ ,..., EP _M ) being stored in a group-specific manner, and d) for a respective measuring device (S ₁ ,..., S _N ) Variety of measuring devices - a sequence to be extrapolated (TS ₁ , ..., TS _N ) of measured values is read in, - f for the sequence to be extrapolated (TS ₁ , ..., TS _N ) a combination of the group-typical sequences (P ₁ , ..., P _M ) is determined, which approximates the sequence to be extrapolated, - the stored extrapolation values (EP ₁ , ..., EP _M ) of the group-typical sequences (P ₁ ,..., P _M ) are linked according to the determined combination, and - a result (ETS ₁ ,..., ETS _N ) of this combination is output as a forecast. Method according to Claim 1, characterized in that the extrapolation of the group-typical sequences (P ₁ ,..., P _M ) is carried out by means of a data-driven trainable regressor, an artificial neural network (NN), a support vector machine, a decision tree and / or a deep learning procedure. Method according to one of the preceding claims, characterized in that the training amount of measuring devices (S ₁ , ..., S _N ) comprises the plurality of measuring devices, a plurality of similar to these measuring devices and / or a selection of the plurality of measuring devices. Method according to one of the preceding claims, characterized in that a respective sequence of measured values (TTS ₁ , ..., TTS _N , TS ₁ , ..., TS _N ) is coded as a vector in a high-dimensional vector space, and that as a similarity measure (PX) a distance measure of the vector space or a correlation of the sequences is used. Method according to one of the preceding claims, characterized in that as a group-typical sequence (P ₁ , ..., P _M ) of a group (C ₁ , ..., C _M ) a sequence located in this group is selected. Method according to one of the preceding claims, characterized in that the group-typical sequence (P ₁ , ..., P _M ) of a respective group (C ₁ , ..., C _M ) is determined on the basis of a distribution characteristic of the sequences in this group , Method according to one of the preceding claims, characterized in that in each case a linear combination and / or a non-linear combination of group-typical sequences (P ₁ , ..., P _M ) is determined as a combination. Method according to one of the preceding claims, characterized in that the cluster analysis comprises a method for hierarchical, partitioning and / or density-based cluster analysis. Method according to one of the preceding claims, characterized in that the determined group-typical sequences (P ₁ , ..., P _M ) are correlated with one another, and that a group-typical sequence (P ₁ ,..., P _M ) whose correlation with another group-typical sequence fulfills a predetermined criterion is not taken into account in the combination. Method according to one of the preceding claims, characterized in that external influence data (EXT) are detected, and that the extrapolation is performed depending on the external influence data. Prediction system (PS) for the computer-aided prognosis of measured values of a plurality of measuring devices (S ₁ , ..., S _N ), arranged for carrying out a method according to one of the preceding claims. Computer program product adapted to carry out a method according to one of claims 1 to 10. Computer-readable storage medium with a computer program product according to claim 12.

Images