EP-4742141-A2 - METHOD AND SYSTEM FOR DETERMINING NON-TECHNICAL ENERGY LOSSES IN AN ELECTRICAL ENERY DISTRIBUTION NETWORK

EP4742141A2EP 4742141 A2EP4742141 A2EP 4742141A2EP-4742141-A2

Abstract

Method for determining losses due to non-physical and non-systematic effects of electrical energy in an electrical distribution network, which network comprises a plurality of stations for measuring time series of electrical parameters which characterize the consumption of electrical energy at said measuring stations and wherein data relating to one or more of said electrical parameters are analyzed to determine whether any loss of energy is due to a systematic physical factor or to an anomaly resulting from incidental malfunctions of the network or from abusive withdrawals of electricity, said analysis being performed by means of a machine learning algorithm trained to classify said data as relating to losses due to systematic physical effects or losses due to malfunctions of the network or to actions of abusive energy withdrawals, and wherein said algorithm consists of a so-called autoencoder, which is trained on a dataset comprising at least said one or more electrical parameters.

Inventors

TALUCCI, Fabio
VERONESE, Fabio
TARANTINO, Sergio
GUGLIETTI, Fabrizio
Vaccari, Stefano

Assignees

Enel Grids S.r.l.

Dates

Publication Date: 20260513
Application Date: 20241022

Claims (12)

Method for determining losses due to non-physical and non-systematic effects of electrical energy in an electrical distribution network, which network comprises a plurality of stations for measuring time series of electrical parameters which characterize the consumption of electrical energy at said measuring stations and wherein data relating to one or more of said electrical parameters are analyzed to determine whether any loss of energy is due to a systematic physical factor or to an anomaly resulting from incidental malfunctions of the network or from abusive withdrawals of electricity, said analysis being performed by means of a machine learning algorithm trained to classify said data as relating to losses due to systematic physical effects or losses due to malfunctions of the network or to actions of abusive energy withdrawals, said method being characterized in that said algorithm consists of a so-called autoencoder, which is trained on a dataset comprising at least said one or more electrical parameters.
The method according to claim 1, wherein the training (430) of the autoencoder is performed using records comprising data from different datasets and which consist at least of data relating to the electrical parameters of the supply and to time trend of the energy withdrawn relative to all historical cases that have been subject to inspection and for which the output is known in relation to whether they relate to the presence or absence of non-technical energy losses; random samples within the cases for which the output regarding being related to technical losses and non-technical losses is not known ex-ante but is assumed to have a pre-established state; records for which the quality of whether or not they correspond to non-technical losses is not known ex ante and for which said probability of assuming a pre-established condition is the lowest, said probability being evaluated using a classifier.
Method according to claim 2, wherein training records relating to samples are used in which it is known or estimated even without this being known ex ante that the sample represented by the data record corresponds to a condition of absence of non-technical loss.
Method according to one or more of the previous claims, wherein a step is provided for evaluating the correspondence between the input record to the autoencoder and the reconstructed output record which provides the steps of determining a scale of correspondence values and a threshold value on the basis of which to discriminate the reconstructed records that are to be considered corresponding to the input records and the reconstructed records that are to be considered not corresponding to the input records, the outcome of correspondence being associated with an evaluation of absence of technical loss and the outcome of non-correspondence being associated with a condition of non-technical loss and said outcome being used to generate one or more of the following activities: request of information from the user; generation of a suggestion or order for the execution of a technical inspection to verify in the field the condition of the network and its operating units, such as concentrators and/or meters in particular; opening of administrative activities aimed at carrying out legal actions and/or recovering the cost of the energy relating to the non-technical loss if the outcome of the inspection were to confirm the condition detected by the algorithm.
Method according to one or more of the previous claims, wherein the information relating to the training records and to the input records for the qualitative determination of whether the sample is to be considered a case relating to a non-technical loss condition or to a technical loss condition can comprise, in addition to the aforementioned electrical parameters and to the measurements relating to the time trend of the energy withdrawn, also further demographic data.
Method according to one or more of the previous claims, wherein the lower probability of corresponding to a condition of absence of non-technical loss is determined through a preliminary step of selecting these records by processing a classifier, such as for example state of the art classifiers for determining the presence or absence of non-technical losses.
Method according to one or more of the previous claims, wherein the type of data categories is selected by means of a machine learning algorithm which selects, from the individual categories of data that form the training records of the autoencoder, one or more categories of data whose variation has a greater effect on determining the condition of presence or absence of non-technical losses.
System configured to implement a method according to one or more of the previous claims, which system comprises at least one processing unit (100) with at least one memory (170) in which a software is stored, in which software the instructions for the execution of at least one autoencoder algorithm are encoded and which software makes said processing unit (100) capable of executing the steps provided by the algorithm on one or more input data records; at least one interface (130) for loading one or more records which comprise data relating at least to electrical parameters which data constitute the input data to said autoencoder algorithm.
System according to claim 8, further comprising at least one comparator (191) of an input record with a reconstructed output record, which comparator (191) executes software for comparing and evaluating the outcome of said comparison which comprises the instructions for setting one or more measurement metrics of one or more variables representing the differences between input record and reconstructed output record and calculating the value of said one or more variables and the instructions for setting a pre-established common and/or specific threshold value for said one or more variables, based on which threshold value it is defined if the value of said one or more variables corresponds to a condition of presence or absence of non-technical losses relating to the sample represented by the input record data, said system comprises a memory (140, 190) in which a program is loaded for activating communication activities of the outcome of the processing and/or activating one or more technical verification and/or administrative intervention actions.
System according to claims 8 or 9, wherein the interface (130) for loading one or more records, which comprise data relating at least to electrical parameters and/or measurements of the time trend of the electricity withdrawal in one or more measuring stations and which data are the input data to said autoencoder algorithm, is a communication interface (130) with a system for monitoring the operating conditions of an electrical distribution network and/or an administrative management system of the administrative activities relating to the supply of electricity through said electricity distribution network.
System according to any of claims 8 to 10, wherein the processing unit (100) can comprise at least one memory (150) in which a software is stored which makes said processing unit (100) capable of executing a training step of said autoencoder algorithm.
System according to claim 11, wherein the processing unit (100) can comprise at least one memory (160) in which a software is stored which comprises the instructions for said processing unit (100) relating to a step of preventive processing of the data provided for the input record, said software comprising the instructions for the execution by said processing unit (100) of processing steps of said records provided by one or by a combination of one or more of the following algorithms: a machine learning algorithm or a combination of machine learning algorithms for optimizing and/or selecting the descriptive variables of data provided for the input records, a machine learning algorithm such as a classifier or a clustering algorithm for determining the probabilities that a pre-established record is associated with the condition of presence or absence of non-technical losses, an algorithm for projecting the records from an initial multidimensional space to a multidimensional space with fewer dimensions than the initial one, a domain adaptation algorithm.

Description

FIELD OF THE INVENTION The present invention concerns a method for determining losses, due to non-physical and non-systematic effects, of electrical energy in an electrical distribution network, which network comprises a plurality of stations measuring time series of different electrical parameters which characterize the conditions of withdrawal of the electrical energy in the various stations, and the time trend of the quantity of electrical energy withdrawn measured at the various stations, i.e. the consumption of electrical energy at the measuring stations, and wherein at least the data relating to one or more of the electrical parameters and/or the time trend of the quantity of electrical energy withdrawn from the network in correspondence with the one or more measuring stations, together with the context data of the meter (meter demographic data, distribution contract details, etc.), are analyzed to determine whether any loss of energy is due to a systematic physical factor or to an anomaly resulting from incidental malfunctions of the network and/or of the measuring systems, or is due to abusive withdrawals of electrical energy, the analysis being performed by means of a machine learning algorithm trained to classify the one or more measured data as relating to losses due to systematic physical effects, or losses due to malfunctions of the network components (measuring apparatuses), or due to actions of abusive energy withdrawals. The present invention addresses the technical problem known as the "Energy Recovery Problem", i.e. the problem of identifying and evaluating anomalies and potential frauds by analyzing the behavior relating to energy consumption and other data provided by measuring devices, and also relating to customer domain, with the ultimate aim of billing utilities for the amount of unrecorded energy. BACKGROUND OF THE INVENTION Document WO2014043287 discloses the use of predictive artificial intelligence or machine learning algorithms to determine the probability that a loss of energy is due to non-technical causes, such as those defined above, in order to determine which utilities are the most likely cause of such non-technical losses. Document WO2014043287 addresses, in greater detail, the use of statistical tools to calculate the probabilities that there are non-technical losses in the network, and only superficially mentions analysis tools that use machine learning or artificial intelligence algorithms, indicating them as potential alternative technologies to statistical tools. In the state of the art there are systems and methods that provide to use a supervised classifier that is trained on the basis of training databases which comprise, as input data, a combination of data including at least user profiling data, data on the operating conditions of the network and the operating units provided therein, and data on measurements of the energy withdrawn from the network, while as output data they provide the indication of an evaluation made with alternative tools regarding the condition of whether the input data configure non-technical losses or normal technical losses resulting from systematic physical effects. The solutions proposed in relation to using systems that operate using artificial intelligence algorithms, and in particular machine learning algorithms, are theoretically possible solutions, but they have drawbacks with regard to the specific selection of the algorithm used since it does not appear suitable, due to its nature, for the practical conditions that define the problem. In fact, the systems that provide to use machine learning algorithms with supervised training have a number of important limitations. A first drawback lies in the fact that supervised training, such as a classifier or suchlike, requires a large number of training records for which the evaluation regarding the type of loss, i.e. the presence or absence of a non-technical loss as detected by an on-site inspection, is known in relation to the input data. Since the correctness of this training data is essential for the correct training of the algorithm, it is necessary that the expected outputs for the various inputs are correct. In this specific case, the reliable determination of the outputs, as an evaluation of the presence or absence of a non-technical loss, requires the intervention of human personnel, and in the worst case even technical inspections to verify the conditions of the sampling points, i.e. the meters or other measuring stations. Therefore, generating a training database of size and quality suitable to solve the problem described above through the supervised training of a classifier and/or regressor is a long and expensive process. Even acknowledging that this supervised training process has to only be done once, and that after this the use of the algorithm with new data admittedly allows to keep the output provided increasingly efficient and precise, so that it can be worth investing