Methods and systems for managing the behaviour of electrical transformers and / or components when an electricity distribution grid is turned back on
By applying regression analysis and machine learning to predict load behavior, the method addresses CLPU challenges, reducing inrush currents and preventing equipment failure through strategic power restoration management.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- HYDRO QUEBEC CORP
- Filing Date
- 2025-11-20
- Publication Date
- 2026-06-04
Smart Images
Figure CA2025051561_04062026_PF_FP_ABST
Abstract
Description
METHODS AND SYSTEMS FOR MANAGING THE BEHAVIOR OF TRANSFORMERS AND / OR ELECTRICAL COMPONENTS DURING THE RE-ENGINEERING OF AN ELECTRICITY DISTRIBUTION NETWORK CORRESPONDING REQUEST
[0001] This application claims priority from provisional patent application US 63 / 725,206, the contents of which are incorporated by reference into this application. TECHNICAL FIELD
[0002] The present invention relates to cold load pickup (CLPU) management, including in particular the prediction of the load recovery of electrical equipment, such as power transformers, after a power outage. The invention may also relate to the identification of electrical equipment at risk of failure and the protection of this at-risk electrical equipment and / or other equipment in the distribution network. STATE OF THE ART
[0003] Cold load recovery (CLPU) poses a major challenge in electricity distribution networks, particularly after prolonged outages. The sudden restoration of power often triggers high inrush currents, which puts significant stress on electrical equipment, such as transformers and other electrical components, including protective devices and conductors.
[0004] The increasing demand for residential electricity, along with the growing adoption of electric vehicles and high-power appliances, has increased the risk of overloading distribution networks when restoring power.
[0005] Therefore, there is a need for new methods and systems to better understand the phenomenon of recovery after load, in order to protect the distribution network and the electrical devices and equipment of the network. SUMMARY OF THE INVENTION
[0006] According to one aspect, a method is proposed for predicting the load behavior of electrical devices after a power outage during the restoration of power to an electrical distribution network following a power outage. The method may include the following steps: accessing the energy consumption data and outdoor temperature data associated with each of the electrical devices among a plurality of electrical devices; applying a regression analysis to the energy consumption data and outdoor temperature data for each electrical device to define their load profiles; retrieving the regression coefficient(s) and the regression analysis parameter(s) for each electrical device to define its load profile;predict the post-break load behavior of each electrical device based on their respective load profiles by entering all or part of the regression coefficients and parameters and data associated with power outages and electrical devices into a trained machine learning model, the post-break load behavior being indicative or corresponding to the inrush current.
[0007] In one aspect, a method is proposed. This method protects transformers or electrical components associated with transformers in an electrical distribution network against overload conditions during network restoration after a power outage. The method may include the following steps: accessing energy consumption profiles and outdoor temperature data associated with each transformer within a plurality of transformers, with the energy consumption profiles extracted from smart meters associated with electrical installations served by one of these transformers. The method may also include a step of applying a regression analysis to the energy consumption profiles and outdoor temperature data for each transformer.The method may also include a step to retrieve, from linear regression analysis, at least one coefficient and at least one parameter for each of the transformers, in order to define a profile of. The method may include a step to predict the load recovery behavior after a power outage for each transformer based on the load profiles. This is achieved by providing the following inputs to a trained machine learning model: at least one coefficient and at least one parameter associated with each transformer; data related to the power outage; and data related to the transformers. The method may also include a step to identify transformers at risk when their predicted load behavior after the outage exceeds a given threshold. The method may further include a step to defer or delay the re-energization of transformers at risk and / or modify the behavior of electrical components associated with transformers at risk when the power outage is restored.
[0008] According to another aspect, the invention may also relate to a method for identifying electrical equipment (transformers or electrical components associated with transformers) that are at risk of failure due to overload conditions during the restoration of power to an electricity distribution network after a power outage. This method may include steps similar to those defined above and may also include the following step: identifying electrical equipment at risk when its expected load behavior after the outage exceeds a given threshold.
[0009] In another aspect, the invention may also relate to a method for protecting electrical equipment against overloads during the restoration of power to a distribution network after a power outage. This method may include steps similar to those defined above and may also include a step for implementing a protection strategy, such as postponing the restoration of power to at-risk electrical equipment and / or replacing the protective devices associated with at-risk equipment.
[0010] In one possible embodiment, the electrical devices may be power transformers. The electrical devices may have a nominal phase-to-phase voltage greater than 240 V, and possibly greater than 500 V or 750 V and less than 44,000 V.
[0011] In one possible embodiment, the application of linear regression analysis generates a regression curve, with at least one coefficient comprising: the slope of the regression curve, the y-intercept of the regression curve, and the coefficient of determination R² of the regression curve. Data relating to power outages may include at least one of the following: an estimated duration of the power outage, an outside temperature at the time of the outage, and the day of the week of the outage. Data relating to transformers may include at least one of the following: a geographical indication of the transformers, a rated capacity of the transformers in kVA, and the number of electrical installations served by each transformer according to the capacity of the installations in amperes.
[0012] According to one possible embodiment, regression analysis is performed for each electrical device, using daily load peaks (in kW / kVA) and / or voltage troughs (in V).
[0013] In one possible embodiment, the electrical components may include protective devices such as fuses and / or circuit breakers. These protective devices may be checked and / or replaced as part of a protection strategy. The curves governing the behavior of the protective devices may be modified during the restart process.
[0014] According to one possible embodiment, for each transformer, the outside temperature includes a measured or inferred daily average temperature in the vicinity of the transformer.
[0015] According to one possible embodiment, the method includes the retrieval of electrical profiles generated by electrical meters deployed within the electrical distribution network, each electrical meter being associated with an electrical installation served by one of the electrical devices, such as transformers.
[0016] According to one possible embodiment, for each electrical device, the outside temperature used in the regression analysis includes a daily average temperature measured near the electrical device or indicative of the temperature near the transformer.
[0017] According to one possible embodiment, the regression analysis is a linear regression analysis.
[0018] According to one possible embodiment, the expected behavior of electrical devices after a failure is expressed as a value per unit (PU) based on the respective theoretical load capacities of the electrical devices in the electricity distribution network.
[0019] According to one possible embodiment, the threshold for identifying risky electrical devices is between 60% and 70% of the theoretical load capacity of a given electrical device.
[0020] According to one possible embodiment, a protection strategy may consist of reactivating risky electrical devices after a random or predetermined delay, while reactivating other devices without delay.
[0021] According to one possible embodiment, the machine learning model is trained using a training dataset comprising data associated with past power outages, the past power outage data comprising at least one of the following: the duration of the power outage and the outside temperature during the power outage.
[0022] According to one possible embodiment, the method may include a step of verifying whether the data relating to past power outages correspond to an actual power outage by comparing the electrical consumption of the electrical installations with the periods during which a power outage is presumed to have occurred.
[0023] In one possible embodiment, the training dataset includes one or more of the following: the location of each of the electrical devices; the rated capacity of the electrical device; the ordinates of the regression analyses in addition to the rates (or slopes) and coefficients of determination; whether the power outage occurs during the weekend or on a weekday; and the number and type of electrical installations served by each electrical device.
[0024] According to one possible embodiment, the method may include steps consisting of validating and testing the post-failure load behavior predictions of the machine learning model trained. A validation step and a testing step of the predictions of the load recovery behavior after a power failure of transformers or electrical components associated with transformers, can be carried out using historical datasets.
[0025] According to one possible embodiment, the trained machine learning model is a regressive machine learning model, such as a light gradient boosting machine learning model.
[0026] In another aspect, a system is provided to protect electrical appliances against overcurrents when restoring power to a distribution network after a power outage. The system may include one or more databases containing electrical consumption data and outdoor temperature data associated with each electrical appliance from a plurality of appliances. The system may also include a regression analysis tool to apply regression analysis to the electrical consumption and outdoor temperature data for each appliance, in order to obtain at least one rate (or slope) and a coefficient of determination from the regression analysis for each appliance, thus defining its load profile.
[0027] According to one aspect, a non-transient memory medium readable by one or more processors is provided, comprising instructions executable by the processor(s) to carry out the steps of the methods defined previously, according to any one of the embodiments.
[0028] In another aspect, a protection system is used to protect transformers or electrical components associated with transformers in an electrical distribution network against overload conditions during network restoration after a power outage. The system may include databases containing energy consumption profiles and outdoor temperature data associated with each transformer in a plurality of transformers. The energy consumption profiles are extracted from smart meters associated with electrical installations served by one of these transformers. The system may also include a configured regression analysis tool. The system applies a regression analysis to the energy consumption profiles and outside temperature data for each transformer, and generates, from the linear regression analysis, at least one coefficient and at least one parameter for each transformer, in order to define a load profile for each transformer, with load profiles associated with each transformer. The system may include a machine learning model trained to predict the load recovery behavior after a power outage for each transformer based on the load profiles, using: at least one coefficient and at least one parameter associated with each transformer; power outage data derived from the energy consumption profiles; and transformer data.The system may include an analysis tool to identify transformers at risk when their predicted load behavior after a failure exceeds a given threshold. The system may also include a control system adapted to manage the at-risk transformers and / or the electrical components associated with these transformers in order to delay their re-energization once power is restored.
[0029] The parameter set may include one or more of the following: rates and coefficients of determination from the regression analysis; and data related to the power outage and electrical appliances. The system may also include an analytical tool to identify electrical appliances at risk when their expected load resumption exceeds a given threshold.
[0030] Finally, the system may include a suitable control system to implement a protection strategy, such as the control of electrical equipment in order to delay the re-energizing of hazardous electrical equipment and / or the replacement of other protective equipment and devices. BRIEF DESCRIPTION OF THE FIGURES
[0031] Possible embodiments of the invention will be described below with reference to the following figures: FIGURE 1 shows a schematic representation of a distribution network with an electricity meter infrastructure. FIGURE 2 shows a flow diagram of a method for protecting electrical devices against overcurrents when restoring power to an electrical distribution network after a power outage. FIGURE 3 shows a graph representing a linear regression applied to electrical profiles retrieved from electricity meters powered by a given transformer, as a function of outside temperature. FIGURE 4A shows a graph representing the expected load of an electrical device, such as a transformer, before, during and after a power outage, using a trained AI model. FIGURE 4B shows a graph representing the load after a power outage, after implementing a protection strategy. DETAILED DESCRIPTION
[0032] In one aspect, systems and methods are provided for predicting the behavior of electrical equipment, including transformers and transformer-related electrical components such as protective devices and conductors, upon re-energization following a power outage. In possible embodiments, the load pickup behavior after a power outage is predicted. This behavior may correspond to, or be indicative of, the load pickup magnitude (CLPU) or inrush current of electrical equipment, such as transformers, in a distribution network. The prediction can be made using power profile data obtained from electricity meters associated with transformers in a distribution network and using local temperature data in the vicinity of the transformers during the outage.According to one possible embodiment of the invention, a regression between power peaks obtained from electrical profile data and local transformer temperatures is applied. For all transformers in the electrical network, parameters characterizing the respective regressions are used, along with other parameters related to the power outage, to train a model. Artificial intelligence (AI) (such as a machine learning (ML) model) can be used. Once trained, the model can accurately predict load recovery behavior after a failure and estimate the inrush current of transformers or other electrical equipment in the network. Based on this information, transformers (or associated components such as circuit breakers and fuses) that are likely to fail during power restoration can be identified, and strategies to protect them against sudden power restorations can be implemented.
[0033] In possible embodiments, the proposed systems and methods leverage data collected from smart energy meters to analyze load patterns before, during, and after power outages or failures, while simultaneously examining historical outage data to identify common characteristics of cold load pick-ups (CLPUs) and peak demand periods. AI models, such as machine learning models, are used to predict load behavior after an outage. These models can be trained to estimate inrush current based on variables such as outage duration, ambient temperature, and load type (transformer type, number of electrical installations served by the transformer, etc.).In possible embodiments, to mitigate the overload on the distribution network, a gradual restoration strategy can be implemented to protect loads at risk of failure and / or critical loads, and to gradually restore the power supply. This method can be partially implemented within the distribution management system (DMS) of the distribution network for automated control purposes.
[0034] The proposed cold load pick-up (CLPU) systems and methods can reduce inrush current spikes and minimize the risk of secondary failures, thereby increasing the stability and reliability of the distribution network.
[0035] In the detailed description that follows, an electrical profile is defined as a time series of electrical inputs. This input can be active power consumption, apparent power consumption, reactive power consumption, voltage, current, or any other variable of an electrical nature. Electrical profiles for single-phase and two-phase installations generally include at least voltage (in V) and active power (in W) profiles. kWh), while for multiphase electrical installations, the recovered profiles include the total active energy (in kWh) and per electrical phase (EA, EB, EC), a measure of the total apparent energy (in kVAh) and per electrical phase, a measure of the total reactive energy (kVARh) and per electrical phase, a measure of the voltages (in V) per phase and a measure of the currents (in A) per phase.
[0036] In the following description, the profiles used are generated by an electricity meter, also known as a smart meter. By electricity meter, we mean an electrical measurement component integrated into an advanced metering infrastructure that produces, among other things, electrical profiles from an electrical installation connected to a low-voltage network (for example, a network where the nominal phase-to-phase voltage does not exceed 750 V) or a medium-voltage network (for example, where the nominal phase-to-phase voltage is greater than 750 V and less than 44,000 V). These electricity meters, whose main function is to measure energy for billing purposes, are sometimes called electricity meters, smart meters, communicating meters, or next-generation meters.
[0037] The following description refers to electrical devices or equipment. In one possible implementation, these electrical devices include distribution network transformers. These transformers are typically medium-voltage / low-voltage (MV / LV) transformers, designed to transform the medium voltage of the distribution lines to the low voltage required for electrical service entrances / installations. In other implementations, it may be possible to apply the methods and systems described below to other types of electrical equipment, such as high-voltage transformers, or to electricity meters associated with distribution network customers. In an electrical network, electricity flows through conductors suspended from pylons, running from power stations to substations—which reduce the voltage—and then to satellite substations—which further reduce the voltage.Electricity can leave satellite substations via underground lines. A certain distance from the substations, the network becomes overhead, and low- or medium-power transformers attached to poles further reduce the voltage. Some of these transformers can be prone to failure due to overcurrent, which often occurs when power is restored after an outage or fault. In addition, the transformers are also subject to high inrush currents during power restoration, which can place additional stress on the equipment. equipment. Other electrical devices and equipment may also be exposed to risks, including protective devices such as circuit breakers and fuses, as well as electrical conductors (cables), which may melt and / or cause a fire in the event of overcurrent.
[0038] For the purposes of the following description, a customer is defined as a user associated with an electrical installation connected to the low-voltage or medium-voltage electricity network. This connection is made via the electrical installation (including, for example, an electrical panel or distribution board). By electrical installation, we mean the electrical components necessary to supply a customer's electrical loads. A customer's load may include household appliances, a lighting system, or a heating and air conditioning system. Most installations include, but are not limited to, an electricity meter suitable for the nature and size of the loads, as well as one or more distribution boards also suitable for the nature and size of the loads. The electrical panels distribute electricity to the customer's various electrical equipment.
[0039] In general, this document describes a method for predicting the recovery behavior of electrical equipment in a distribution network, particularly transformers and their associated components. The proposed method is especially suited to predicting the recovery behavior of low- and medium-power transformers in a distribution network, as mentioned above. The method aims to predict the recovery behavior of distribution network transformers when they are re-energized after a fault or power outage. The process may involve accessing energy consumption data, notably from power profiles generated by meters, and ambient temperatures in the vicinity of transformers serving the electrical installations to which the power profiles are linked.Energy consumption data can therefore be extracted from the electrical profiles generated by energy meters for all installations served by the transformers. The outside temperature can be obtained from meteorological information sources, based on the geographical location of the transformers. Depending on the size of the distribution network, there may be more than 100,000 distribution transformers (e.g., MV / LV type), and in some networks, more than 500,000 distribution transformers. In some implementations, the method includes a step involving the application of a... Regression analysis is performed on energy consumption and outdoor temperature data for each electrical device (typically each transformer). From these regression analyses, parameters characterizing the regression can be extracted. In the case of linear regression, these include the rate (or slope) and a coefficient of determination (R²). 2These parameters can be obtained for each electrical device. They define the load profile of each transformer. Using these parameters, along with other parameters related to the power outage and / or the electrical installations served by the transformers, the cold start-up of each transformer can be predicted using a trained machine learning model. By applying thresholds to the load start-up predictions, at-risk electrical devices can be identified. Various protection strategies can be implemented to safeguard these at-risk transformers and / or associated electrical devices, for example, by delaying their re-energization after the power outage has been restored.
[0040] Figure 1 shows the various components that may be necessary to implement methods for predicting and identifying transformers at risk, and for protecting transformers and associated components, during a cold restart following a power outage affecting part of the distribution network, including upstream components. A simplified, exemplary electrical distribution network (100) is shown, comprising a plurality of single-phase electrical installations (110). In some other embodiments, multiphase electrical installations may be involved. Although only a few electrical installations are shown in Figure 1, it should be noted that an electrical distribution network can comprise several thousand, or even several million, electrical installations.Electrical installations are connected to electrical devices, such as transformers (116), which are in turn connected to electrical conductors of the distribution network (100), which converge at distribution substations (or stations), not shown in FIG. 1.
[0041] Each electrical installation (110, 120) is connected to a distribution transformer (116). Each meter (120) includes measuring means and data transmission means. The measurements taken by the energy meters (data and profiles) can be routed, once processed, to a Metering Data Management System (MDMS). The Data can be collected by a front-end acquisition system and transferred to various systems, including a distribution management system (170), commonly known as a DMS (Distribution Management System). Each meter also includes control means for interrupting the power supply to the electrical installation to which it is connected. These means can be activated by sending a signal from the central monitoring and management system (a request to open a control element located in the meter) to the meter. Distribution substations can also be equipped with control systems (114) including, for example, circuit breakers for switching specific transformers in a region or zone on or off.
[0042] The front-end acquisition (FEA) system (160) includes a database (172) for storing the raw measurements transmitted by the electricity meters, from which energy consumption profiles (i.e., energy consumption over time) are derived. The FEA system (160) and the database (172) can be located on one or more servers in the same building, or distributed across several servers located in different locations, for example, in a cloud data infrastructure. As shown in FIG. 1, the meters do not communicate directly with the DMS. The meters can exchange information with each other or send it directly to a router (115). The router communicates with collectors (130), which in turn transmit the information to the FEA system, and possibly to other systems, including, for example, the DMS (170) via a wide area network (WAN) (140).Thus, the measurements taken by the meters are routed to the front-end data acquisition system (160), and then to the DMS (170). A firewall security system (150) is used to protect the meter data. Of course, other network configurations are also possible.
[0043] The method for predicting the magnitude of the recovery of electrical equipment (including, but not limited to, transformers) and / or low-voltage networks after a power outage can be implemented using data processing software tools, analytical tools, databases, machine learning models (190), and graphical user interfaces, which can be implemented in a dedicated software application or provided as modules within the DMS. This application (10) can be deployed in a computer system (180) that may include an algorithmic processing unit, comprising one or more processors and central or distributed storage memory. The memory of this server or these servers may include instructions executable by one or more processors to perform the steps of the methods detailed below. The system (180) may also include one or more servers and a database (182). The latter is used to store, among other things, the electrical profiles of the DMS, the topological data of the distribution network (from the geographic information system, GIS, of the electricity utility), personal data related to a meter and an electrical installation, as well as meteorological data indicating local weather conditions. Local weather conditions, including outside temperature and wind speeds, may be obtained from local weather stations.Local weather conditions, including outside temperature and wind speed, can be obtained from local weather stations (118) associated with the nearest transformer (116). The correlation between the GPS coordinates of the transformers and the positions of the weather stations can be used to determine which weather stations should be used as a source of representative weather conditions for a given transformer. It is also possible to interpolate the temperature at the transformer based on data from weather stations near a given transformer.
[0044] Figure 2 illustrates the overall process (200) of protecting electrical equipment in the distribution network against overloads (or overcurrents) during the restoration of power to the distribution network after a power outage. This process is divided into several steps. A first step (210) involves accessing the outside temperature data associated with each electrical device (such as power transformers) for a plurality of electrical devices. In one possible implementation, meteorological data, which may include temperature data as well as wind speed, are extracted and processed so that they can be used to apply regression analyses and to train the machine learning model.In one possible implementation, a data table (array, grid, database, or spreadsheet) can be created to associate each targeted transformer with its corresponding weather data. For example, each entry in the table includes a transformer identifier, the temperature in degrees Celsius, and the wind speed for a given period, such as each day and for all hours. This data is available for all transformers in the network. distribution networks typically consist of several hundred, or even several thousand, transformers. This data can be used to calculate the average daily temperature per transformer. In one possible embodiment, the average wind speed (wind chill) can also be taken into account. The data can then be exported to a table whose entries include transformer IDs, dates, average temperatures per transformer, and optionally the average wind speed per day. Other temperature indicators are possible, including, for example, minimum and / or maximum temperatures, humidity levels, etc.
[0045] Another step (220) may involve accessing all the electricity consumption profiles (load and voltage profiles) from the electricity meters in a region. Since the electricity meter data may be incomplete or contain inconsistencies, data processing is applied, such as interval resampling, duplicate handling, and validation of the number of profiles. The electricity meter data may also be filtered to remove data corresponding to power outages, which could bias the regression curves. In one possible embodiment, before performing the regression analysis, it is possible to filter the data to retain only temperatures below a threshold temperature, for example, +15 degrees Celsius.Depending on the geographical location of the distribution network (a network further north, with harsh winters, or further south, with hot summers and significant air conditioning needs), other temperature thresholds may be used, for example, +30°C. Depending on the geographical region, the thresholds can be adjusted (+20°C, +25°C, etc.). Once processed, the electrical profiles are used to calculate the daily peak demand (such as peak power in kVA, peak voltage in V, and / or peak current in A) for each transformer in the region. Another data table (table, grid, data tables, spreadsheet) can be created to combine the daily peak demands and the average daily temperature (or other equivalent meteorological data) into a single dataset for all transformers, for every day within a given period. This period can extend over several months or several years.The daily peak power consumption of a transformer can be derived from the energy consumption profiles of the meters that are served by that transformer.
[0046] Another step (230) involves applying regression analyses to the data in this table using a regression analysis tool. This includes data on electricity consumption and outdoor temperature for all or some electrical appliances (such as all transformers). In one possible embodiment, the regression analysis models used can be adapted or trained for distinct seasonal regimes, where one model is trained to process datasets for temperatures below a certain threshold (e.g., winter regime), and another model is trained to process datasets for temperatures above a certain threshold (e.g., summer regime).Data relating to daily load and / or peak current demand and average temperature over a given period, such as the previous year, can be entered into a regression analysis, for example, a linear regression. In one possible embodiment, for each transformer, the slope of a line, its ordinate, and the coefficient of determination R² are generated. Thus, as indicated in step (240), for each transformer, the rate (or slope), the ordinate, and the coefficient of determination (R²) can be obtained in a context where the chosen regression analysis is a linear regression. These coefficients and / or regression parameters provide an indication of the transformer's load profile and allow it to be characterized.This type of regression analysis, applied to daily consumption peaks and average temperatures to characterize the load profile of each transformer in a distribution network, is unprecedented.
[0047] Referring to FIG. 3, an example of linear regression is shown for a transformer in a distribution network, where the X-axis represents the average daily temperature (over 24 hours), the Y-axis represents the daily peak demand, for example in kW / kVA, and each point on the graph represents one day during a given period. From the regression graph, an equation of the form a = mx + b can be derived, where "m" corresponds to the slope of the graph, b to the y-coordinate, and R 2The coefficient of determination (or distribution) of the regression is a key factor. A steeper slope indicates a greater load amplitude at the transformer when power is restored to the distribution network. Regression analysis can be applied to all or some of the targeted transformers in the distribution network. Regression analysis can also provide information on the nature of the loads, the quality of building insulation, and can help normalize winter peak loads.
[0048] Referring again to FIG. 2, the data relating to power outages or interruptions are accessed and processed (step 250). Data relating to power outages can be retrieved, for example, from the DMS. The DMS includes databases containing data relating to power outages accumulated over several years, for example, 10 years. A limitation of this data is that some data ranges may appear to indicate outages when they do not. The proposed method may include a step of identifying actual power outages or interruptions from the DMS data by comparing them with the profiles of the electricity meters and / or the periods during which service calls were logged. For example, a data range that appears to indicate a power outage can be excluded if the associated electricity profiles show electricity consumption during that period.Once the actual power outages or failures have been validated, their respective durations can be determined using the start and end times of the outages. This step 250 therefore makes it possible to identify the start and end times of the power outages, and their geographical extent (based on the affected meters, and therefore the associated transformers).
[0049] In one possible implementation, a peak load dataset can be extracted from the DMS and then processed to train a machine learning model. Raw outage data can be extracted from available databases over a given period. In another implementation, this raw outage data can be correlated with service calls. The resulting outage data will represent most network outages. An outage table can be created, listing all outages with their respective start and end times, as well as the location or area where they occurred. The geographic area of the outage can be delimited by the identifiers of the transformers affected by a given outage or determined using GPS coordinates.Consumption profiles do not include measurements taken during a power outage. Since each meter is connected to a specific transformer, and the location of each transformer is known, it is possible to delimit the geographical area of the outages based on the meter consumption profiles.
[0050] Advantageously, the entries in this table can be correlated with power or voltage measurements from electricity meters (based on power profiles) once the outages have ended: as mentioned above, genuine power outages can be distinguished from false outages using the power profiles of the electricity meters. Consumption data should be zero for meters affected by the supposed outage. Using an algorithm based on the gap and island problems algorithm, it is possible to obtain a list of genuine power outages detected by the electricity meters.According to one possible embodiment, missing data in a given profile (for example, due to a meter failure) can be distinguished from measurements that could not be taken due to a failure, for example by validating the number of consecutive periods (e.g. 15 minutes) with measurements (normal operation), without measurements (failure or interruption) or with partial measurements (meter problem, but not necessarily a failure).
[0051] The power outage data obtained can be filtered to ensure that no electricity meter data is missing and that power restorations actually correspond to the end of an outage. The resulting table includes a list of confirmed power outages over a given period, for example, several years, with the date and duration of each outage. In one possible implementation, for each power outage and restoration found in the previous step, the affected transformers are identified, along with the outdoor temperature data for those transformers corresponding to the power restoration. This combination of power outage data and the corresponding average temperature data for the affected transformer can be exported to another table, which then includes the transformers affected by the outages and the temperatures during the outages.
[0052] At this stage, the method may include a step of merging the entries in this table listing the transformers affected by the outages and the average outside temperatures during the outages with the regression parameters identified for each transformer. A second validation of the power outage data can be performed at this stage by further filtering the data using a table listing the power profiles that includes partial data for the period corresponding to the power outages. One possible filtering method is This involves using load profile data for all meters at regular intervals, for example, every 15 minutes. This data indicates whether the interval is complete, partial, or missing, with each state potentially associated with a corresponding indicator (e.g., complete = 0, partial = 2, or missing = 4). Since meters do not have a load profile during a fault, this information is crucial for distinguishing a genuine fault from missing data. Furthermore, because the first interval (e.g., 15 minutes) during a fault is a partial interval, this partial interval is generally omitted from databases. In one possible embodiment, the prediction algorithm can be configured to reconstruct the first interval in order to estimate the actual peak consumption of a given electrical installation.Furthermore, using two separate data sources to verify that a power outage has occurred helps ensure the integrity of the dataset that will be used to train the machine learning model.
[0053] Referring to step 260 of the diagram in FIG. 2, to supplement the collected data (outage duration, affected transformers, parameters derived from regressions to establish transformer load profiles, average temperature during outages), additional databases can be used to include nominal data associated with the power outages, the transformers, and / or the electrical components connected to the transformers. For example, databases that correct the associations between transformers and customers and their electrical installations can be used. Impedance matrices can be used to verify the accuracy of the association in the database. This step ensures that the electrical profiles associated with the transformers are indeed derived from the electricity meters and the electrical installations they serve.Additional nominal data may include one or more of the following parameters: the location of each transformer, such as GPS coordinates; the transformer's nominal capacity (e.g., 10, 15, 25, 50, 75, 100, or 167 kVA); the ordinates of regression analyses in addition to the rates (or slopes) and coefficients of determination; whether the power outage occurs on a weekend or weekday; and the number and type of electrical installations (60A, 100A, 200A, etc.) served by each transformer. The number of electrical installations may include, as an example only, the number of buildings or apartments supplied by a given transformer and the electrical inlets of these installations (60A, 100A, 200A, 400A, or 600A).
[0054] The next step is to train the machine learning model (step 270). Although different sets of parameters can be used, it has been shown that including linear regression parameters, including at least the slope and coefficient of determination for each transformer, combined with data on the date and duration of failures as well as the temperature at the time of failures, allows for rapid and accurate convergence in predicting the post-failure behavior of transformers in an electrical distribution network.In one possible embodiment, all or some of the following parameters can be used to train the model: duration of outages; geographical location of the transformers where the outages occurred; rated capacity of the transformers (in kVA); outside temperatures during the power outages; slope (m); coefficient of determination (R²); and intercept of linear regressions for each transformer; day of the week or weekend; number of customers associated with each transformer; and the current (in A) of their electrical input. During the machine learning model training, the value to be predicted is also provided, corresponding in this case to the starting load (power in kW) or the inrush current (in A).
[0055] The dataset containing these parameters can be divided into three sets: one for training, one for validation, and one for testing. In possible implementations, the machine learning model is a regression machine learning model, including, for example, a Light Gradient Boosting Machine (or Extreme Gradient Boosting Machine), an Extra Trees Regressor (or Random Forest Regressor), a Bayesian Ridge model, an Elastic Net model, and so on. In one possible implementation, a Light Gradient Boosting Machine is preferred as the trained model, based on the accuracy of its prediction performance, but other types of models can offer adequate performance.
[0056] Once the machine learning model is trained, it can be used to predict load behavior after a failure, for example the inrush current in transformers (step 280), using parameters similar to those used for its training, i.e. including at least: all or part of the set of coefficients and parameters from the regression analysis, such as rates / slopes and the coefficients of determination R 2and the data associated with the power outage and transformers. In one possible implementation, to simplify the analysis, theoretical outage data based on utility company standards or regional standards can be used to generate theoretical post-outage power demand data for each selected transformer in the network. This theoretical post-outage power demand data can then be used to normalize forecasts. For example, normalization factors can be used to generate a list of predicted load resumptions for all single-phase overhead transformers in a region (such as a state or province), which could include several thousand transformers.In one possible implementation, the expected load take-up values can be evaluated against a theoretical load capacity (or theoretical capacity limit) of the low-voltage distribution network. The theoretical capacity limit of transformers can be established for a primary voltage, such as 25 kV, and for different transformer types, such as from 10 kVA to 167 kVA.
[0057] In possible implementations, the load pickup expected for transformers and / or associated electrical components (circuit breakers, fuses, conductors, etc.) can be expressed as a value per unit (PU) based on the respective theoretical load capacities of the electrical equipment in the power distribution network. Predicted load pickup values can be converted into PU values. If the load pickup of one of the transformers, expressed in PU, exceeds a predetermined threshold, resumption of operation after a power outage will most likely result in the failure of the affected power transformer. The threshold may vary depending on the conditions used to determine the theoretical failure data. For example, the theoretical failure data may be based on several factors such as the duration of the outage, the temperature during the outage, and the temporal and / or climatic environment.The value chosen for these different factors will be used to determine an appropriate threshold based on the degree of severity of the cut chosen as a theoretical reference.
[0058] In possible implementations, transformers and / or electrical devices at risk can be identified by an analysis tool, by applying a threshold to the predicted load take-up values (step 300). For example, referring to FIG. 4A, a graph representing the predicted load behavior of power transformers is shown, to which a threshold has been applied. The graph shows the load (in kW) before the outage (April 4 and 5), the load during the outage (April 6 and 7), and the load after the outage (April 8 and beyond), corresponding to what is predicted by the AI model. During the recovery period following the outage (between April 7 and 8), a peak consumption (at 290 kW) can be observed. Depending on the distribution network, the threshold can be adjusted and chosen between 60% and 70% of the theoretical load capacity of a given electrical device. For example, in the case of power transformers, transformers at risk of failure during the restoration of power to the distribution network after a power outage can be identified when their predicted load recovery is greater than 0.65 PU.Load resumption forecasts can be performed periodically, in advance, to identify at-risk transformers and / or at-risk sections of the distribution network, and to implement a protection strategy for these transformers, either at the DMS level or in the distribution substations, before the next power outage. In other embodiments, forecasts can be performed in real time, during power outages, and protection strategies can be dynamically adjusted based on the real-time forecasts.
[0059] In possible implementations, a protection strategy applied to at-risk transformers could involve delaying their re-energization when power is restored. The re-energization of at-risk electrical equipment can be delayed by re-energizing it after a random or predetermined period, while other equipment on the distribution network is re-energized without delay. In one possible embodiment, for each identified at-risk transformer, a specific number of kVA to be shed can be defined. This value reduces the number of electrical installations that will be disconnected to prevent overloading a section of the network. This value can be established using the predicted value – (safety factor * network capacity) = kVA to be shed.The protection strategy can be implemented via the DMS, with control commands sent to substations, for example, by remotely controlling the re-energization of at-risk transformers via controllable circuit breakers (or other control devices). Other protection strategies may include replacing protective devices, such as circuit breakers, associated with zones or network segments linked to the identified transformers. In one possible embodiment, the behavior of protective devices (e.g., circuit breakers, switchgear) associated with at-risk transformers can be modified. For example, the curves. The rules governing the behavior of protective devices can be modified during the restart. This may involve pushing or transmitting specific protection curves to be followed by certain circuit breakers on the distribution line, associated with the transformers at risk. These instructions can be transmitted by servers that are part of a network protection system. These curves can be called i2t curves (current squared per time) and can govern the behavior of protective devices. To prevent nuisance tripping of circuit breakers, once the risk zones of a network section have been identified, specific protection curves (more or less permissive, depending on the desired outcome) can be transmitted to these circuit breakers, so that they are applied by the circuit breakers for a given period. These curves can be transmitted via a communication network (e.g.,cellular network) in the form of instructions executable by the processor controlling the circuit breaker's behavior, including the breaker opening mechanism. The graph in FIG. 4B shows the load (energy consumption) for a set of meters supplied by a given transformer, where a protection strategy has been implemented. It can be seen that the load is more consistent after the fault and exhibits no load spikes, with the maximum load being approximately 210 kW.
[0060] In one possible embodiment, the proposed method can be executed at least partially by a smart meter (i.e., equipped with processing capabilities (processor(s) and memory)). The smart meter can detect a power outage or failure, and if this exceeds a given duration threshold, for example, 8 hours, the meter can be configured / programmed to locally open an internal switch and perform a phased and / or random re-closing in the absence of a command from the central control system of the distribution network. In one possible option, the meter can prioritize a power recovery management command from an energy management system.
[0061] The proposed methods and systems described above advantageously leverage electrical profiles retrieved from electricity meters, beyond their conventional role in billing and monitoring, highlighting their potential to enhance grid resilience. By integrating machine learning for load resumption forecasting, predictive maintenance and fault management can be improved. Furthermore, the implementation of a restoration progressive power supply offers a practical and scalable solution to the challenges posed by CLPII, which can be adopted by utilities worldwide.
[0062] The proposed CLPII protection strategy, based on energy meter data, addresses one of the most pressing problems in the field of electricity distribution. By leveraging advanced metering infrastructure and predictive analytics, electricity utilities can significantly improve power outage management, thereby ensuring a more reliable and resilient power supply.
[0063] Although the concepts, data flows and methods associated with the invention and results have been illustrated in the attached drawings and described above, it will be obvious to those versed in the art that modifications may be made to these embodiments without departing from the invention.
Claims
DEMANDS 1. Method for protecting transformers or electrical components associated with transformers in an electrical distribution network against overload conditions during network re-energization after a power outage, the method comprising: accessing energy consumption profiles and outside temperature data associated with each transformer among a plurality of transformers, the energy consumption profiles being extracted from smart meters associated with electrical installations served by one of said transformers; applying a regression analysis on the energy consumption profiles and outside temperature data for each of the transformers;to retrieve, from linear regression analysis, at least one coefficient and at least one parameter for each transformer, in order to define a load profile per transformer, with load profiles respectively associated with said transformers; to predict the load recovery behavior after the power failure for each transformer based on the load profiles, by providing as input to a trained machine learning model: at least one coefficient and at least one parameter associated with each transformer; data relating to the power failure; and data relating to the transformers; to identify the transformers at risk when their predicted load behavior after the failure exceeds a given threshold; and; delay the re-energizing of at-risk transformers and / or modify the behavior of electrical components associated with at-risk transformers when the power outage is restored.
2. A method for identifying transformers or electrical components associated with transformers in an electrical distribution network at risk of failure under overload conditions when the network is re-energized following a power outage, the method comprising: accessing energy consumption profiles and outdoor temperature data associated with each transformer among a plurality of transformers, the energy consumption profiles being extracted from smart meters associated with electrical installations served by one of said transformers; applying a regression analysis on the energy consumption profiles and outdoor temperature data for each of the transformers;to retrieve, from linear regression analysis, at least one coefficient and at least one parameter for each of the transformers, in order to define a load profile per transformer, with load profiles being respectively associated with said transformers; to predict the load recovery behavior after the power failure for each of the transformers based on the load profiles, by providing as input to a trained machine learning model: at least one coefficient and at least one parameter associated with each of the transformers; data relating to the power failure; and data relating to the transformers; to identify the transformers or electrical components associated with the transformers at risk when their predicted load behavior after the failure exceeds a given threshold.
3. A method for predicting the behavior of transformers or electrical components associated with transformers in an electrical distribution network during network re-energization following a power outage, the method comprising: accessing energy consumption profiles and outdoor temperature data associated with each transformer among a plurality of transformers, the energy consumption profiles being generated by smart meters associated with electrical installations served by one of said transformers; applying a regression analysis on the energy consumption profiles and outdoor temperature data for each of the transformers; retrieving, from the regression analysis, at least one coefficient for each of the transformers, in order to define a load profile per transformer, load profiles being respectively associated with said transformers;predict the load recovery behavior after a power failure for each of the transformers or electrical components associated with the transformers based on load profiles, by providing as input to a trained machine learning model: at least one coefficient associated with each of the transformers; data relating to the power failure; and data relating to the transformers; the load recovery behavior after the power failure being indicative of or corresponding to the inrush current of each transformer.
4. The method according to any one of claims 1 to 3, wherein the transformers have a nominal phase-to-phase voltage equal to or greater than 240 V and less than 44,000 V.
5. The method according to any one of claims 1 to 4, wherein the application of linear regression analysis generates a regression curve, the at least one coefficient comprising: the slope of the regression curve, the y-intercept of the regression curve, the coefficient of determination R² of the regression curve 6. The method according to any one of claims 1 to 5, wherein the data relating to power outages include at least one element of: an estimated duration of the power outage, an outside temperature at the time of the outage and the day of the week of the outage.
7. The method according to any one of claims 1 to 6, wherein the data relating to the transformers includes at least one element among: a geographical indication of the transformers, a nominal capacity of the transformers in kVA, a number of electrical installations served by each transformer according to the capacity in amperes of the installations.
8. The method according to any one of claims 1 to 7, wherein the energy consumption profiles include load peaks (in kWh) and / or voltage peaks (in V), per day.
9. The method according to any one of claims 1 to 8, wherein the electrical components include protective devices including fuses and / or circuit breakers.
10. The method according to any one of claims 1 to 9, for each transformer, the outside temperature includes a daily average temperature measured in the vicinity of the transformer.
11. The method according to any one of claims 1 to 10, wherein the regression analysis is a linear regression analysis.
12. The method according to any one of claims 1 to 11, wherein the load recovery behavior of transformers or electrical components associated with transformers after failure is expressed as a ratio or value per unit (PU) based on the respective theoretical load capacities of the transformers or electrical components associated with the transformers.
13. The method according to any one of claims 1 to 12, wherein the threshold for identifying transformers and / or electrical components at risk is between 60% and 70% of the theoretical load capacity of said transformers or electrical components.
14. The method according to any one of claims 1 to 13, wherein delaying the re-energizing of the transformers at risk and / or modifying the behavior of the electrical components, comprises: delaying the re-energizing after a random or predetermined delay, while the other transformers and associated electrical components are re-energized without delay, and / or applying specific protection curves.
15. The method according to any one of claims 1 to 13, comprising a step of training the machine learning model with a training dataset including data relating to the duration of past power outages and the outside temperature during these past power outages.
16. The method according to claim 15, comprising a step of verifying whether the data relating to past power outages correspond to actual power outages by correlating the data from the energy consumption profiles for the periods during which the past power outages took place.
17. The method according to claim 15 or 16, wherein the training step comprises at least one of: a geographical indication of the transformers; a nominal capacity of the transformers in kVA, a number of electrical installations served by each transformer and the capacity in amperes of the installations; coefficients and / or coordinates from regression analyses applied to historical energy consumption profiles and to outside temperature data; whether past power outages occur during the weekend or on weekdays.
18. The method according to any one of claims 15 to 17, comprising a validation step and a test step of predictions of load recovery behavior after a power failure of transformers or electrical components associated with transformers, using historical datasets.
19. The method according to any one of claims 15 to 18, wherein the trained machine learning model is of the regression model type.
20. The method according to claim 19, wherein the trained machine learning model is a light up-gradient machine learning model.
21. A non-transient memory medium readable by one or more processors, comprising instructions executable by the processor(s) to carry out the steps of the methods defined according to any one of claims 1 to 20.
22. A system for protecting transformers or electrical components associated with transformers in an electrical distribution network against overload conditions when the network is re-energized after a power outage, the system comprising: databases including energy consumption profiles and outside temperature data associated with each transformer of a plurality of transformers, the energy consumption profiles being extracted from smart meters associated with electrical installations served by one of said transformers;a regression analysis tool configured to apply a regression analysis to the energy consumption profiles and to the outside temperature data for each of the transformers, and to generate, from the linear regression analysis, at least one coefficient and at least one parameter for each of the transformers, in order to define a load profile; per transformer, load profiles being respectively associated with said transformers; a machine learning model trained to predict the load recovery behavior after the power failure for each of the transformers based on the load profiles, based on: at least one coefficient and at least one parameter associated with each of the transformers; power failure data taken from the power consumption profiles; and data relating to the transformers; an analysis tool to identify transformers at risk when their predicted load behavior after the failure exceeds a given threshold; and a suitable control system to control the transformers at risk and / or the electrical components associated with these transformers in order to delay their re-energization when the power supply is restored.