WO2023168471A1

WO2023168471A1 - Method of controlling a thermal energy supply network

Info

Publication number: WO2023168471A1
Application number: PCT/AT2023/060045
Authority: WO
Inventors: Günther Beck
Original assignee: Beck & Partner Kg
Priority date: 2022-03-08
Filing date: 2023-02-14
Publication date: 2023-09-14
Also published as: AT525571A4; AT525571B1

Abstract

The invention relates to a computer-implemented method for detecting and rectifying anomalies, in particular undersupply states, for example during the use of load limitations on the consumer side in a thermal energy supply network having at least one energy source (1 ) which is connected via a primary distribution network (2) to at least one transfer station (10, 10', 10") of at least one consumer (3, 3', 3").

Description

Method for controlling a thermal energy supply network

The invention relates to a computer-implemented method for detecting and eliminating anomalies in a thermal energy supply network, in particular undersupply conditions on the consumer side, which are caused, for example, by the use of load limitations.

Methods and devices for controlling thermal energy supply networks, in particular district heating networks and district cooling networks, which are based on the long-distance transmission of a thermal medium, are known from the prior art. Such energy supply networks generally include at least one energy source, which is connected via a distribution network to at least one transfer station of at least one customer. At the transfer station, the energy is transferred from the primary distribution network to a secondary distribution network (consumption network).

Load control methods are known for such energy supply networks, in which the energy supply to individual consumers is temporarily reduced or interrupted without resulting in a noticeable reduction in comfort for the user (so-called Demand Side Management, DSM). Load peaks can be reduced by deliberately shifting the time of the consumers (load shifting) or by limiting the possible consumption capacity of individual consumers. The sum of the loads in the primary network is limited by partially limiting the output of selected consumers. The technical implementation of the DSM can be done by using a temperature shift of the set point calculated by the controller. Another method is to directly specify the set point (usually secondary to the flow temperature) by overwriting the controller result. The performance of the controller can also be temporarily limited by setting a limit from the maximum performance (set on the controller). There are also other setting parameters, such as triggering a buffer or hot water tank charging process or blocking one. This triggers additional load or prevents load due to hot water demand. Such procedures are usually based on monitoring individual indicators of the customers and apply statistical relationships to shed or shift burdens. For example, the provision of hot water can be reduced at certain times of the day. This is well suited for large customers such as factories, hotels or office buildings, but has its limits for small units, especially private households, as such small units have a much higher statistical fluctuation range.

When using conventional customer-side load control methods for private households, undersupply conditions often arise, which in turn means that load management can only be used with less intensity. Such undersupply conditions are perceived as anomalies by the buyer. The object of the invention is to provide a method and a device in order to detect and eliminate such anomalies, in particular the undersupply conditions described, at an early stage and to increase the intensity and efficiency of the use of load management.

These and other problems are solved according to the invention with a computer-implemented method and a device according to the independent patent claims.

A method according to the invention is designed to detect and eliminate anomalies, in particular undersupply conditions, in a thermal energy supply network, for example during the use of load limitations on the consumer side. However, the method can also be used in other situations, for example to identify bottlenecks or when loads for a specific customer need to be reduced in favor of another customer. Any deviations from the normal state of the energy supply network can be described as an anomaly; The invention is therefore not limited to the use of DSM. A thermal energy supply network for using the method has at least one thermal energy source, which is connected to at least one transfer station of at least one customer via a primary distribution network for distributing a medium, for example water. A number i = 1, 2, ..., N consumers can be provided in the energy supply network. The transfer station is used to transfer energy from the primary distribution network to the secondary distribution network, for example into a house distribution network for heat or cold. As a rule, the transfer station includes a heat exchanger, at least one valve and at least one regulator that regulates the temperature and/or the flow of the medium.

The method according to the invention comprises the following steps:

First, a measurement is carried out by a large number of sensors of current measurement data M _i and a determination, by the controller, of current operating states B _i of the customer i, as well as a transmission of the measurement data M _i and operating states B _i to a data processing unit. The data can be transmitted via a wired or wireless interface, in particular via an Internet router. In particular, it can be an Internet router that is integrated into the controller.

The data processing unit can be designed as a microcontroller or microcomputer and a central processing unit (CPU), a volatile semiconductor memory (RAM), a non-volatile semiconductor memory (ROM, SSD hard drive), a magnetic memory (hard drive) and / or an optical memory (CD ROM) as well as interface units (Ethernet, USB) and the like.

The data processing unit then queries a database and extracts at least one classification model K _i assigned to the customer. The database can be provided as a software module in the data processing unit, in a computer separate from the data processing unit or in an external server. The components of the data processing unit and database according to the invention are fundamentally known to those skilled in the art. The classification model K _i can be any method from the field of machine learning, for example a decision tree generated by special algorithms. These can be machine learning models that allow a prediction (classification) of whether a particular data set is likely to be perceived as an anomaly (particularly as an undersupply). In the simplest case, this is a classification into “OK” and “not OK”, but it is also possible to calculate a numerical probability. In the first case, a flag that is set triggers the process that is intended to eliminate the anomaly. In the second case, a probability value is used as a threshold. The basic idea is that a customer-specific learning process is used to arrive at models that provide a forecast as to whether the measurement data M _i and operating status data B _i at the current point in time for customer i indicate an anomaly, in particular an undersupply.

In particular, to create a decision tree, the J48 algorithm can be used, an open source implementation of a C4.5 algorithm (see e.g. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann Publishers, 1993), which can be used to to create a decision tree from a variety of training data that can be used to classify data sets. Of course, other supervised or unsupervised machine learning algorithms can also be used to create the classification models. What is essential to the invention is that the classification model K _i is customer-specific, so that an individual classification model K _i is stored in the database for each customer. The selection of the relevant measurement data M _i and operating states B _i can also be specific for each customer.

The data processing unit then applies the selected classification model K _i to the current measurement data M _i and the current operating states B _i of the customer, and determines whether there is an undersupply condition or another anomaly. If an undersupply condition or another anomaly has been detected, a control takes place by sending a data signal F' from the data processing unit, a controller or a control valve in the customer's transfer station to eliminate the undersupply condition or the anomaly.

The assessment of whether there is an undersupply condition or another anomaly is not based on a threshold value, but rather by applying a classification model K _i created individually for the customer to a data set comprising measurement data M _i and operating states B _i of the customer or the customer's transfer station . The classification model K _i stored individually for each consumer in the database represents the consumer-typical combination of measurement data M _i and operating states B _i that occur in the event of an anomaly or undersupply for the respective consumer.

This allows automated detection and elimination of anomalies during operation without having to set global threshold values. If an undersupply is detected, then existing load limitations are overridden in order to eliminate the undersupply by lifting or increasing the permitted limit. If no undersupply is detected, any load limitation set is adopted without change. So a changed data signal is not always sent out.

The undersupply conditions are eliminated, for example, by weakening or canceling an existing load limitation at the customer by sending a changed data signal F' to the transfer station. However, load management as such, especially for other customers, remains fundamentally unaffected. If the customer is no longer undersupplied, the load limitations that provide for load management in general or for the specific customer automatically become valid again. According to the invention, it can be provided that the classification model K _i continuously, preferably at intervals of less than 30 minutes, particularly preferably at intervals of less than 15 minutes, preferably at intervals of 1 minute to 10 minutes, responds to the current measurement data M _i and operating states B _i des of the customer is applied.

Furthermore, it can be provided that several classification models assigned to the respective customer are stored in the database. In this case, the method steps can be carried out for several classification models K _i assigned to the customer. In particular, structurally different classification models for the respective customer can be stored in the database, for example a neural network, a decision tree and a linear regression model. In this case, the determination of whether an anomaly exists can be made based on a combination of the results of several different classification models K _i . For example, an anomaly can be detected if at least two of three classification models used detect an anomaly. In principle, however, any statistical or combinatorial method is conceivable. A majority vote can be used, as can a statistical method in which the results of the models are combined based on their probability to determine the overall probability of error. For example, a first model may predict an anomaly with 95% probability and a second model with 80% probability. If the models were combined, the error probability would be reduced to a value of 0.05*0.20 = 1%. The probability of error can be displayed to the user as a decision criterion.

In addition to or instead of a structural difference in the classification models used, it can also be provided that the classification models use different measurement data as input parameters. For example, two neural networks - which are mathematically identical - can be used as classification models, but they differ in the selection of input parameters and are therefore different models for the specific application. The measurement data Mj can be, for example, the flow temperature, the return temperature and the flow of the primary distribution network, the flow temperature, the return temperature and the flow of the secondary distribution network, and / or an outside temperature or an inside temperature. Preferably, the measurement data M _i are transmitted from sensors in the primary and/or secondary distribution network to the data processing unit, for example via an Internet router.

The operating states B _i can, for example, be the status of a pump or a valve in the customer's transfer station, a current hot water requirement, a flow temperature setpoint, the temperature spread between the primary and secondary distribution network, and/or the grade of the heat exchanger . The operating states B _i are preferably transmitted from the controller or another unit in the transfer station to the data processing unit, for example via an Internet router.

According to the invention, it can be provided that the database is queried in order to determine those measurement data M _i and operating states B _i that are relevant for the classification model K _i , and subsequently the classification model K _i only to the measurement data M _i that has been determined to be relevant for the provider and operating states B _i is applied. The connection between the classification models K _i and the relevant measurement data M _i and operating states B _i can be stored in the database for this purpose.

The invention further relates to a computer-implemented classification model K _i for use in a method according to the invention, in particular in the form of a decision tree constructed specifically for this purpose. According to the invention, the classification model is individually adapted to each customer and is stored in a database.

The invention further relates to a computer-implemented method for creating an individual classification model K _i for use in a method according to the invention. A method according to the invention comprises the following steps: In a first step, receiving a training data set comprising historical measurement data M _i ' and operating states B _i ' of the customer as well as reference values R, which characterize whether an anomaly, in particular an undersupply state, is present. The measurement data M _i ' and operating states B _i ' are those data that were collected over time, typically a few months back, but possibly also several years. In order to create a classification model for the first time, data from at least 4 to 8 weeks ago is used. Training data from comparison periods in previous years can also be used, for example from the respective month or season in previous years. Combinations are also possible, for example for March 2022 the model based on the historical data from February 2022 can be combined with the historical data from March 2021.

The reference values can be a yes/no classification of the training data with regard to the presence of an anomaly, in particular an undersupply. However, it can also be a probability that there is an anomaly, in particular an undersupply, in the relevant training data.

In a further step, a classification model is determined, applied to the training data set, and the results of the classification model K _i are received and compared with the reference value. The classification model K _i can be any method from the field of machine learning, for example an algorithm for creating a decision tree.

In a further step, the model parameters of the classification model K _i are adjusted in order to reduce the difference between the result of applying the classification model K _i and the reference value. The steps listed are repeated with the customer's individual training data and reference values until the desired prediction quality is achieved. For example, the training can be repeated until the differences between the results and the reference values remain below a predetermined threshold. When comparing several classification models, the result can also be that the model with the best prediction quality for this customer is selected. Any metric can be used to select the classification model, for example the classification model that delivers results for the customer the fastest can be selected.

Finally, the individual classification model K _i that is most suitable for the customer is stored in a database or in another electronic way, for example in the main memory of the data processing unit, in order to be able to use it later to detect and correct anomalies.

In this step, several suitable classification models can also be saved in the database for each customer. Furthermore, if necessary, the measurement data and operating states relevant to the customer and the selected classification model are stored in the database.

According to the invention, it can be provided that the method for creating an individual classification model K _i is repeated at certain intervals, for example every 2, 4, or 6 weeks or at the beginning of each season, in order to reflect changes in user behavior. The repetition can be done automatically, for example by running a script on the data processing unit. What is important here is that the classification models K _i used and their parameters generally not only differ from building to building, but that the classification models K _i can also change over time for one and the same building. This can be explained by changes in user behavior, by changes to the building structure such as thermal insulation, changes in settings and especially by changes over the season (shading, passive solar input, wind...). The invention further relates to the use of methods according to the invention for detecting and eliminating anomalies, in particular undersupply states on the customer side, in a district heating network, a district cooling network or a combined district heating and district cooling network.

According to the invention, it can be provided that the data processing unit and preferably also the database are integrated into the customer's transfer station or are locally connected to it. This allows the process to operate completely independently, provided the individual classification models have been defined.

According to the invention, it can also be provided that the data processing unit and preferably also the database are integrated into a control center of the primary distribution network that is connected to the customer's transfer station via the Internet or another data network or is locally connected to it. This means that other customers can also access the control center’s data processing unit.

According to the invention, it can also be provided that the database is implemented on a server on the Internet to which the data processing unit has access. This can in particular be a cloud server that provides the various classification models for the data processing unit.

The invention further relates to a data processing unit comprising means for carrying out a method according to the invention.

The invention further relates to a computer-readable storage medium, comprising instructions which cause a data processing unit to carry out a method according to the invention.

Further features according to the invention result from the patent claims, the description of the exemplary embodiments and the figures.

The invention is explained in more detail below using exemplary embodiments. Show it: Fig. 1a: a schematic representation of an energy supply network;

Fig. 1 b: a schematic representation of an embodiment of a device according to the invention;

Fig. 2 is a schematic representation of program modules provided in the data processing unit;

3 shows a schematic flowchart of a method for creating customer-specific classification models K _i ;

4 shows a schematic flowchart of a method for detecting and eliminating undersupply states of an energy supply network on the consumer side;

Figs. 5a - 5c show different network topologies for using a method according to the invention.

Fig. 1a shows a schematic representation of a thermal energy supply network with an energy source 1, a primary distribution network 2 and a number of consumers 3, 3 ', 3".

The consumers 3, 3', 3" each have a transfer station 10, 10', 10", which serves as a connection between the primary distribution network 2 and the secondary distribution networks 12, 12', 12" in the consumers 3, 3', 3 “serves. The primary distribution network 2 can be, for example, a district heating network or a district cooling network.

Fig. 1b shows a schematic representation of an embodiment of a device according to the invention. The primary distribution network 2 is shown, which leads to a transfer station 10 of a customer 3. A heat exchanger 11 is provided in the transfer station 10, which acts as a connection between the primary distribution network 2 and the secondary distribution network 12 (consumption network) in the customer 3. A large number of primary sensors 6, 6 'are provided in the primary distribution network 2. A large number of secondary sensors 7, 7 'are also provided in the secondary distribution network 12. The sensors 6, 6', 7, 7' deliver measurement data M _{i from} the transfer station 10 to a data processing unit 4. A regulator 9 and a control valve 8 are also provided in the transfer station 10, which regulate the flow and thus the energy output of the medium flowing in the primary energy distribution network 2 to the secondary distribution network 12. The controller 9 continuously determines the current operating states B _i of the customer 3 and transmits these to the data processing unit 4. The data processing unit 4 is connected to the controller 9 and, in this embodiment, is also connected to a database 5.

Fig. 2 shows a schematic representation of program modules E, G and H provided in the data processing unit 4 and their interaction with the database 5 and a customer 3. The program E represents a load management method according to the prior art. In this exemplary embodiment Invention, the program E accesses the database 5 to determine whether a load limit has been stored for the relevant customer.

In the event of a load limitation for the relevant consumer 3, a data signal F is output, which is passed on to the consumer 3. In other exemplary embodiments, however, the program E can also work completely autonomously and without access to the database 5.

Access to the database 5 can serve to filter out those consumers who are suitable for load management, since a data signal F is only sent to these consumers. However, manual control of the load management can also be provided by a boiler house control or a heating supervisor.

The program G includes the creation of the classification models KL ES obtains historical measurement data M _i , operating status data B _i and reference values R from the database 5 and uses them to determine an individual classification model K _i for each customer 3 to detect undersupply. The classification model K _i is stored in the database 5. The program H represents the ongoing application of the classification models K _i to the current measurement data M _i and operating status data B _i . The program H obtains the customer-specific classification model K _i from the database 5 and uses it to determine a classification that indicates an undersupply or not (yes /no- classification). In other embodiments of the invention, a probability value for the undersupply of the customer can also be determined. If an undersupply is detected, program H ensures that any existing load limitation is reduced or eliminated. The modified data signal F' generated from this is then transmitted to the customer 3.

Fig. 3 shows a schematic flow diagram of a method for creating customer-specific classification models K _i by the data processing unit 4. This is the subprogram G described in Fig. 2. It is carried out in such a way that at least one valid classification model K _i is created for each customer 3 is created. A periodic update of the classification model K _i at intervals of several weeks or once per season can be provided.

The database 5 contains historical measurement data M _i ' and operating state data B _i ' as training data, as well as the corresponding reference values R.

The reference values can be yes-no statements as to whether an undersupply was determined for the relevant sets of measurement data M _i ' and operating state data B _i '. But it can also be a probability value.

The following steps are carried out for each customer 3. Step S301 extracts the historical data from the database 5. In step S302, incomplete or incorrect data from step S301 is eliminated. Subsequently, in step S303, irrelevant or redundant features are removed from the remaining data. For example, certain measurement data M _i or operating state data B _i are known to be irrelevant for the classification. Step S304 provides an identification of those data sets from step S303 for which an undersupply is obvious, is to be feared, or could be determined. This process can be carried out by an expert who carries out the labeling, or through automatic comparison with threshold values and whether they are exceeded or fallen below as part of an assistance program. In order to obtain a valid model, a sufficient number of data sets for each of the two classifications (undersupply YES or NO) must be provided in the training data.

In step S305, a machine learning model is trained. To do this, a reference model for classification is first created. A multinomial logistic regression is used as the algorithm and the result is evaluated. The Quinlan C4.5 algorithm (also called “J48”) is then used for comparison and evaluated in comparison to the reference model. Furthermore, models can be created and evaluated using any other algorithms. Finally, the algorithm that delivers the most suitable results in the evaluation is selected as the classification model K _i .

Finally, in step S306, the results are determined for all customers 3 using the classification model K _i determined for the respective customers in step S305.

The classification models K _i from step S306 are stored in the database in step S307 for later use. It is possible to select one classification model for all customers as well as different classification models. Not only do the model parameters differ from customer to customer, but also the basic model structure.

Fig. 4 shows a schematic flow diagram of the subprogram H described in Fig. 2. The steps in program H are carried out for all customers 3 at short intervals of approximately 1 - 10 minutes. In step S401, those consumers i for which a load limitation is active are selected by querying the database 5. For these customers, the measurement data M _i , operating state data B _i and the individual classification models K _i are read out from the database 5 in step S402. In step S403, irrelevant or redundant features of the data are eliminated. In step S404, the classification models K _i are applied to the data to obtain a classification of underprovision (yes/no).

In step S405 it is checked whether an undersupply has been detected. If no undersupply was detected, there is no change to the existing load limitation (step S406). If an undersupply has been detected, it is checked in step S407 whether an instruction to increase the load limitation is stored in the database 5. If this is the case, then in step S409 a new signal F' is sent to the recipient on this basis. If no instruction to increase the load limit is found in the database, the load limit is completely lifted for the relevant customer and a corresponding signal F" is sent to the customer in step S408. Within the consumer 3, the heating controller 9 processes the signal from step S408 or step S409. In the last step, the control valve 8 reacts to the changed setting of the controller 9.

5a, the data processing unit 4 and the database 5, like the local controller 9, are integrated into the transfer station 10 of a customer 3. This enables the process to operate completely independently. The data processing unit 4 and the database 5 can, for example, be provided in the software of the controller 9 of a customer product.

5b, the data processing unit 4 and the database 5 are integrated in a control center 13 of the primary distribution network 2, which is connected to the transfer station 10 of a customer 3 or the controller 9 arranged therein via the Internet. This has the advantage that different customers can use the central data processing unit 5 and database 5 of the control center 13 at the same time. For this purpose, the control center can provide an http server to which the transfer stations 10 of the customers 3 gain access. By sending current measurement data, a transfer station 10 can thus immediately receive an answer by querying the assigned classification model K _i , in particular an evaluation or a forecast calculation of the current state.

According to Fig. 5c, the database 5 with the individual classification models K _i is implemented on an external cloud server 14 on the Internet, to which the data processing unit 4 has access. This allows the classification models K _i to be maintained separately via the Internet without having to intervene in the control center 13 of the primary distribution network 2.

The scope of protection of the patent is determined by the following patent claims and is not limited to the application for detecting and eliminating undersupply conditions, but also includes applications for detecting and eliminating other undesirable situations (anomalies) in thermal energy supply networks.

Reference symbol list

1 energy source

2 Primary distribution network

3, 3', 3" takers

4 data processing unit

5 database

6, 6' sensor

7, 7' sensor

8 control valve

9 regulators

10 transfer station

11 heat exchangers

12, 12', 12" secondary distribution network

13 control center

14 cloud servers

Claims

Patent claims

1 . Computer-implemented method for detecting and eliminating anomalies, in particular undersupply conditions, for example during the use of consumer-side load limitations in a thermal energy supply network with at least one energy source (1) which is connected via a primary distribution network (2) with at least one transfer station (10, 10 ', 10 “) of at least one customer (3, 3', 3”), characterized in that the method comprises the following steps: a. Measurement, by a large number of sensors (6, 6', 7, 7'), of current measurement data M _i and determination, by a controller (9) in the transfer station (10), of current operating states B _i of the customer (3, 3' , 3”) and transmission of the measurement data M _i and operating states B _i to a data processing unit (4), b. Query, by the data processing unit (4), of a database (5) and extraction of a classification model K _i assigned to the customer (3, 3 ', 3"), c. Application, by the data processing unit (4), of the classification model K _i to the current measurement data M _i and the current operating states B _i of the customer (3, 3', 3"), and determining whether an anomaly exists, i.e. If an anomaly has been detected, control by sending a data signal F 'from the data processing unit (4), the controller (9) or a control valve (8) in the transfer station (10) of the customer (3, 3 ', 3") to Fix the anomaly.

2. The method according to claim 1, characterized in that steps c. - d. of the process can be carried out continuously, preferably at intervals of less than 30 minutes, particularly preferably at intervals of less than 15 minutes, preferably at intervals of 1 minute to 10 minutes.

3. Method according to one of claims 1 or 2, characterized in that steps b. - c. for several, preferably structurally different, classification models K _i assigned to the customer (3, 3', 3"), the determination of whether an anomaly exists being made on the basis of a combination of the results of the classification models K _i .

4. The method according to one of claims 1 to 3, characterized in that the sensors (6, 6 ', 7, 7') as measurement data Mi the flow temperature, the return temperature and the flow of the primary distribution network (2), the flow temperature, the Return temperature and the flow of the secondary distribution network (12, 12', 12"), and/or an outside temperature or an inside temperature.

5. The method according to one of claims 1 to 4, characterized in that the controller (9) as operating states B _i the status of a pump or a valve in the transfer station (10), the current hot water requirement, the temperature spread between the primary distribution network ( 2) and secondary distribution network (12, 12', 12"), program parameters set on the customer (3), the grade of the heat exchanger and/or other specific status data of the customer (3).

6. The method according to any one of claims 1 to 5, characterized in that after step b. the database (5) is queried in order to determine those measurement data Mi and operating states B _i that are relevant for the classification model K _i , and application, in step c., of the classification model K _i only to those for the provider (3). relevantly determined measurement data Mi and operating states B _i .

7. Computer-implemented classification model K _i for detecting and eliminating anomalies, in particular undersupply conditions, for example during the use of consumer-side load limitations in an energy supply network with at least one energy source (1) which has at least one energy source (1) via a primary distribution network (2) with at least one transfer station (10). a customer (3, 3', 3"), characterized in that the classification model K _i is designed for use in a method according to one of claims 1 to 6, in particular in the form of a correspondingly trained neural network, a linear regression model or in the form of a decision tree.

8. Computer-implemented method for creating an individual classification model K _i for detecting and eliminating anomalies, in particular undersupply conditions, for example during the use of consumer-side load limitations in an energy supply network with at least one energy source (1) which is connected via a primary distribution network (2) with at least one transfer station (10) of at least one customer (3, 3', 3"), in particular a method for creating the classification model K _i from claim 6, characterized in that the method comprises the following steps: a. Receiving a training data set comprising historical measurement data Mi' and operating states B _i ʻ of the customer (3, 3', 3") as well as reference values R which indicate whether an anomaly exists; b. Determining a classification model K _i and applying the classification model K _i to the training data set, receiving the result of the classification model K _i and comparing the result with the reference value; c. Adjusting the model parameters of the classification model K _i to reduce the difference between the result and the reference value; d. Repeat steps a. - c. with the user's individual training data and reference values (3, 3', 3") until the differences between the results and the reference values remain below a predetermined threshold, e. Storage of the individual classification model K _i and, if necessary, the measurement data Mi' and operating states B _i ʻ found to be relevant for the classification model K _i in a database (5).

9. The method according to claim 8, characterized in that to determine the classification model in step b. First, a reference model is created that applies a multinomial logistic regression to the training data set, and then for comparison, a C4.5 algorithm to create a decision tree is applied to the training data set and is evaluated in comparison to the reference model.

10. The method according to claim 9, characterized in that to determine the classification model, further reference models are created and evaluated with other algorithms from the field of machine learning, and finally that algorithm is selected as the classification model K _i which delivers the most suitable results in the evaluation .

11. The method according to any one of claims 8 to 10, characterized in that steps a. - e. be repeated at an interval of less than 2 weeks, less than 4 weeks, or less than 8 weeks, preferably once per season, preferably automatically, in particular taking into account and using at least one comparison period from previous years.

12. Use of a computer-implemented method according to one of claims 1 to 11 for detecting and eliminating anomalies, in particular customer-side undersupply states of a district heating network, a district cooling network or a combined district heating and district cooling network.

13. Use according to claim 12, characterized in that the data processing unit (4) and preferably also the database (5) are integrated into or with the transfer station (10, 10 ', 10") of the customer (3, 3', 3") this is locally connected.

14. Use according to claim 13, characterized in that the data processing unit (4) and preferably also the database (5) are in one with the transfer station (10, 10 ', 10") of the customer (3, 3', 3"). Control center (13) of the primary distribution network (2) connected via the Internet or another data network or is locally connected to it.

15. Use according to claim 13 or 14, characterized in that the database (5) is implemented on a server on the Internet to which the data processing unit (4) has access.

16. Data processing unit (4), comprising means for carrying out a method according to one of claims 1 to 11.

17. Computer-readable storage medium, comprising instructions that cause a data processing unit (4), in particular a data processing unit according to claim 16, to carry out a method according to one of claims 1 to 11.